2012-03-02 03:38:13 +00:00
|
|
|
/*******************************************************************
|
2005-04-17 21:05:31 +00:00
|
|
|
* This file is part of the Emulex Linux Device Driver for *
|
2005-06-25 14:34:39 +00:00
|
|
|
* Fibre Channel Host Bus Adapters. *
|
2023-01-09 23:33:17 +00:00
|
|
|
* Copyright (C) 2017-2023 Broadcom. All Rights Reserved. The term *
|
2021-03-01 17:18:21 +00:00
|
|
|
* “Broadcom” refers to Broadcom Inc. and/or its subsidiaries. *
|
2016-07-06 19:36:13 +00:00
|
|
|
* Copyright (C) 2004-2016 Emulex. All rights reserved. *
|
2005-06-25 14:34:39 +00:00
|
|
|
* EMULEX and SLI are trademarks of Emulex. *
|
2017-02-12 21:52:39 +00:00
|
|
|
* www.broadcom.com *
|
2005-06-25 14:34:39 +00:00
|
|
|
* Portions Copyright (C) 2004-2005 Christoph Hellwig *
|
2005-04-17 21:05:31 +00:00
|
|
|
* *
|
|
|
|
|
* This program is free software; you can redistribute it and/or *
|
2005-06-25 14:34:39 +00:00
|
|
|
* modify it under the terms of version 2 of the GNU General *
|
|
|
|
|
* Public License as published by the Free Software Foundation. *
|
|
|
|
|
* This program is distributed in the hope that it will be useful. *
|
|
|
|
|
* ALL EXPRESS OR IMPLIED CONDITIONS, REPRESENTATIONS AND *
|
|
|
|
|
* WARRANTIES, INCLUDING ANY IMPLIED WARRANTY OF MERCHANTABILITY, *
|
|
|
|
|
* FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT, ARE *
|
|
|
|
|
* DISCLAIMED, EXCEPT TO THE EXTENT THAT SUCH DISCLAIMERS ARE HELD *
|
|
|
|
|
* TO BE LEGALLY INVALID. See the GNU General Public License for *
|
|
|
|
|
* more details, a copy of which can be found in the file COPYING *
|
|
|
|
|
* included with this package. *
|
2005-04-17 21:05:31 +00:00
|
|
|
*******************************************************************/
|
|
|
|
|
#include <linux/pci.h>
|
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
|
|
|
#include <linux/slab.h>
|
2005-04-17 21:05:31 +00:00
|
|
|
#include <linux/interrupt.h>
|
2011-05-27 13:37:25 +00:00
|
|
|
#include <linux/export.h>
|
2006-08-17 15:58:04 +00:00
|
|
|
#include <linux/delay.h>
|
2008-12-05 03:40:02 +00:00
|
|
|
#include <asm/unaligned.h>
|
2017-06-29 18:31:12 +00:00
|
|
|
#include <linux/t10-pi.h>
|
2013-04-18 00:14:49 +00:00
|
|
|
#include <linux/crc-t10dif.h>
|
2021-06-08 04:35:51 +00:00
|
|
|
#include <linux/blk-cgroup.h>
|
2013-04-18 00:14:49 +00:00
|
|
|
#include <net/checksum.h>
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
#include <scsi/scsi.h>
|
|
|
|
|
#include <scsi/scsi_device.h>
|
2008-12-05 03:40:02 +00:00
|
|
|
#include <scsi/scsi_eh.h>
|
2005-04-17 21:05:31 +00:00
|
|
|
#include <scsi/scsi_host.h>
|
|
|
|
|
#include <scsi/scsi_tcq.h>
|
|
|
|
|
#include <scsi/scsi_transport_fc.h>
|
|
|
|
|
|
|
|
|
|
#include "lpfc_version.h"
|
2009-05-22 18:51:39 +00:00
|
|
|
#include "lpfc_hw4.h"
|
2005-04-17 21:05:31 +00:00
|
|
|
#include "lpfc_hw.h"
|
|
|
|
|
#include "lpfc_sli.h"
|
2009-05-22 18:51:39 +00:00
|
|
|
#include "lpfc_sli4.h"
|
2008-09-07 15:52:10 +00:00
|
|
|
#include "lpfc_nl.h"
|
2005-04-17 21:05:31 +00:00
|
|
|
#include "lpfc_disc.h"
|
|
|
|
|
#include "lpfc.h"
|
2012-03-02 03:37:42 +00:00
|
|
|
#include "lpfc_scsi.h"
|
2005-04-17 21:05:31 +00:00
|
|
|
#include "lpfc_logmsg.h"
|
|
|
|
|
#include "lpfc_crtn.h"
|
2007-06-18 00:56:38 +00:00
|
|
|
#include "lpfc_vport.h"
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
#define LPFC_RESET_WAIT 2
|
|
|
|
|
#define LPFC_ABORT_WAIT 2
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
static char *dif_op_str[] = {
|
2012-03-02 03:37:42 +00:00
|
|
|
"PROT_NORMAL",
|
|
|
|
|
"PROT_READ_INSERT",
|
|
|
|
|
"PROT_WRITE_STRIP",
|
|
|
|
|
"PROT_READ_STRIP",
|
|
|
|
|
"PROT_WRITE_INSERT",
|
|
|
|
|
"PROT_READ_PASS",
|
|
|
|
|
"PROT_WRITE_PASS",
|
|
|
|
|
};
|
|
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
struct scsi_dif_tuple {
|
|
|
|
|
__be16 guard_tag; /* Checksum */
|
|
|
|
|
__be16 app_tag; /* Opaque storage */
|
|
|
|
|
__be32 ref_tag; /* Target LBA or indirect LBA */
|
|
|
|
|
};
|
|
|
|
|
|
2014-02-20 14:56:45 +00:00
|
|
|
static struct lpfc_rport_data *
|
|
|
|
|
lpfc_rport_data_from_scsi_device(struct scsi_device *sdev)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *)sdev->host->hostdata;
|
|
|
|
|
|
2014-04-04 17:52:21 +00:00
|
|
|
if (vport->phba->cfg_fof)
|
2014-02-20 14:56:45 +00:00
|
|
|
return ((struct lpfc_device_data *)sdev->hostdata)->rport_data;
|
|
|
|
|
else
|
|
|
|
|
return (struct lpfc_rport_data *)sdev->hostdata;
|
|
|
|
|
}
|
|
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_release_scsi_buf_s4(struct lpfc_hba *phba, struct lpfc_io_buf *psb);
|
2009-11-18 20:40:49 +00:00
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_release_scsi_buf_s3(struct lpfc_hba *phba, struct lpfc_io_buf *psb);
|
2013-05-31 21:03:39 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_prot_group_type(struct lpfc_hba *phba, struct scsi_cmnd *sc);
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2009-06-10 21:22:44 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_sli4_set_rsp_sgl_last - Set the last bit in the response sge.
|
|
|
|
|
* @phba: Pointer to HBA object.
|
|
|
|
|
* @lpfc_cmd: lpfc scsi command object pointer.
|
|
|
|
|
*
|
|
|
|
|
* This function is called from the lpfc_prep_task_mgmt_cmd function to
|
|
|
|
|
* set the last bit in the response sge entry.
|
|
|
|
|
**/
|
|
|
|
|
static void
|
|
|
|
|
lpfc_sli4_set_rsp_sgl_last(struct lpfc_hba *phba,
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd)
|
2009-06-10 21:22:44 +00:00
|
|
|
{
|
2019-01-28 19:14:19 +00:00
|
|
|
struct sli4_sge *sgl = (struct sli4_sge *)lpfc_cmd->dma_sgl;
|
2009-06-10 21:22:44 +00:00
|
|
|
if (sgl) {
|
|
|
|
|
sgl += 1;
|
|
|
|
|
sgl->word2 = le32_to_cpu(sgl->word2);
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_rampdown_queue_depth - Post RAMP_DOWN_QUEUE event to worker thread
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
*
|
|
|
|
|
* This routine is called when there is resource error in driver or firmware.
|
|
|
|
|
* This routine posts WORKER_RAMP_DOWN_QUEUE event for @phba. This routine
|
|
|
|
|
* posts at most 1 event each second. This routine wakes up worker thread of
|
|
|
|
|
* @phba to process WORKER_RAM_DOWN_EVENT event.
|
|
|
|
|
*
|
|
|
|
|
* This routine should be called with no lock held.
|
|
|
|
|
**/
|
2007-06-18 00:56:38 +00:00
|
|
|
void
|
2008-12-05 03:39:29 +00:00
|
|
|
lpfc_rampdown_queue_depth(struct lpfc_hba *phba)
|
2007-06-18 00:56:38 +00:00
|
|
|
{
|
|
|
|
|
unsigned long flags;
|
2008-06-15 02:52:53 +00:00
|
|
|
uint32_t evt_posted;
|
2014-09-03 16:55:58 +00:00
|
|
|
unsigned long expires;
|
2007-06-18 00:56:38 +00:00
|
|
|
|
|
|
|
|
spin_lock_irqsave(&phba->hbalock, flags);
|
|
|
|
|
atomic_inc(&phba->num_rsrc_err);
|
|
|
|
|
phba->last_rsrc_error_time = jiffies;
|
|
|
|
|
|
2014-09-03 16:55:58 +00:00
|
|
|
expires = phba->last_ramp_down_time + QUEUE_RAMP_DOWN_INTERVAL;
|
|
|
|
|
if (time_after(expires, jiffies)) {
|
2007-06-18 00:56:38 +00:00
|
|
|
spin_unlock_irqrestore(&phba->hbalock, flags);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
phba->last_ramp_down_time = jiffies;
|
|
|
|
|
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, flags);
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&phba->pport->work_port_lock, flags);
|
2008-06-15 02:52:53 +00:00
|
|
|
evt_posted = phba->pport->work_port_events & WORKER_RAMP_DOWN_QUEUE;
|
|
|
|
|
if (!evt_posted)
|
2007-06-18 00:56:38 +00:00
|
|
|
phba->pport->work_port_events |= WORKER_RAMP_DOWN_QUEUE;
|
|
|
|
|
spin_unlock_irqrestore(&phba->pport->work_port_lock, flags);
|
|
|
|
|
|
2008-06-15 02:52:53 +00:00
|
|
|
if (!evt_posted)
|
|
|
|
|
lpfc_worker_wake_up(phba);
|
2007-06-18 00:56:38 +00:00
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_ramp_down_queue_handler - WORKER_RAMP_DOWN_QUEUE event handler
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
*
|
|
|
|
|
* This routine is called to process WORKER_RAMP_DOWN_QUEUE event for worker
|
|
|
|
|
* thread.This routine reduces queue depth for all scsi device on each vport
|
|
|
|
|
* associated with @phba.
|
|
|
|
|
**/
|
2007-06-18 00:56:38 +00:00
|
|
|
void
|
|
|
|
|
lpfc_ramp_down_queue_handler(struct lpfc_hba *phba)
|
|
|
|
|
{
|
2007-08-02 15:09:51 +00:00
|
|
|
struct lpfc_vport **vports;
|
|
|
|
|
struct Scsi_Host *shost;
|
2007-06-18 00:56:38 +00:00
|
|
|
struct scsi_device *sdev;
|
2009-11-18 20:39:44 +00:00
|
|
|
unsigned long new_queue_depth;
|
2007-06-18 00:56:38 +00:00
|
|
|
unsigned long num_rsrc_err, num_cmd_success;
|
2007-08-02 15:09:51 +00:00
|
|
|
int i;
|
2007-06-18 00:56:38 +00:00
|
|
|
|
|
|
|
|
num_rsrc_err = atomic_read(&phba->num_rsrc_err);
|
|
|
|
|
num_cmd_success = atomic_read(&phba->num_cmd_success);
|
|
|
|
|
|
2012-05-10 01:18:40 +00:00
|
|
|
/*
|
|
|
|
|
* The error and success command counters are global per
|
|
|
|
|
* driver instance. If another handler has already
|
|
|
|
|
* operated on this error event, just exit.
|
|
|
|
|
*/
|
|
|
|
|
if (num_rsrc_err == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
2007-08-02 15:09:51 +00:00
|
|
|
vports = lpfc_create_vport_work_array(phba);
|
|
|
|
|
if (vports != NULL)
|
2009-05-22 18:53:21 +00:00
|
|
|
for (i = 0; i <= phba->max_vports && vports[i] != NULL; i++) {
|
2007-08-02 15:09:51 +00:00
|
|
|
shost = lpfc_shost_from_vport(vports[i]);
|
|
|
|
|
shost_for_each_device(sdev, shost) {
|
2007-06-18 00:56:38 +00:00
|
|
|
new_queue_depth =
|
2007-08-02 15:09:51 +00:00
|
|
|
sdev->queue_depth * num_rsrc_err /
|
|
|
|
|
(num_rsrc_err + num_cmd_success);
|
|
|
|
|
if (!new_queue_depth)
|
|
|
|
|
new_queue_depth = sdev->queue_depth - 1;
|
|
|
|
|
else
|
|
|
|
|
new_queue_depth = sdev->queue_depth -
|
|
|
|
|
new_queue_depth;
|
2014-11-13 14:08:42 +00:00
|
|
|
scsi_change_queue_depth(sdev, new_queue_depth);
|
2007-08-02 15:09:51 +00:00
|
|
|
}
|
2007-06-18 00:56:38 +00:00
|
|
|
}
|
2008-01-11 06:52:54 +00:00
|
|
|
lpfc_destroy_vport_work_array(phba, vports);
|
2007-06-18 00:56:38 +00:00
|
|
|
atomic_set(&phba->num_rsrc_err, 0);
|
|
|
|
|
atomic_set(&phba->num_cmd_success, 0);
|
|
|
|
|
}
|
|
|
|
|
|
2008-08-25 01:50:11 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_scsi_dev_block - set all scsi hosts to block state
|
2008-08-25 01:50:11 +00:00
|
|
|
* @phba: Pointer to HBA context object.
|
|
|
|
|
*
|
|
|
|
|
* This function walks vport list and set each SCSI host to block state
|
|
|
|
|
* by invoking fc_remote_port_delete() routine. This function is invoked
|
|
|
|
|
* with EEH when device's PCI slot has been permanently disabled.
|
|
|
|
|
**/
|
|
|
|
|
void
|
|
|
|
|
lpfc_scsi_dev_block(struct lpfc_hba *phba)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_vport **vports;
|
|
|
|
|
struct Scsi_Host *shost;
|
|
|
|
|
struct scsi_device *sdev;
|
|
|
|
|
struct fc_rport *rport;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
vports = lpfc_create_vport_work_array(phba);
|
|
|
|
|
if (vports != NULL)
|
2009-05-22 18:53:21 +00:00
|
|
|
for (i = 0; i <= phba->max_vports && vports[i] != NULL; i++) {
|
2008-08-25 01:50:11 +00:00
|
|
|
shost = lpfc_shost_from_vport(vports[i]);
|
|
|
|
|
shost_for_each_device(sdev, shost) {
|
|
|
|
|
rport = starget_to_rport(scsi_target(sdev));
|
|
|
|
|
fc_remote_port_delete(rport);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
lpfc_destroy_vport_work_array(phba, vports);
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* lpfc_new_scsi_buf_s3 - Scsi buffer allocator for HBA with SLI3 IF spec
|
2008-12-05 03:39:02 +00:00
|
|
|
* @vport: The virtual port for which this call being executed.
|
2020-11-02 14:23:43 +00:00
|
|
|
* @num_to_alloc: The requested number of buffers to allocate.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* This routine allocates a scsi buffer for device with SLI-3 interface spec,
|
|
|
|
|
* the scsi buffer contains all the necessary information needed to initiate
|
|
|
|
|
* a SCSI I/O. The non-DMAable buffer region contains information to build
|
|
|
|
|
* the IOCB. The DMAable region contains memory for the FCP CMND, FCP RSP,
|
|
|
|
|
* and the initial BPL. In addition to allocating memory, the FCP CMND and
|
|
|
|
|
* FCP RSP BDEs are setup in the BPL and the BPL BDE is setup in the IOCB.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* Return codes:
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* int - number of scsi buffers that were allocated.
|
|
|
|
|
* 0 = failure, less than num_to_alloc is a partial failure.
|
2008-12-05 03:39:02 +00:00
|
|
|
**/
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_new_scsi_buf_s3(struct lpfc_vport *vport, int num_to_alloc)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_hba *phba = vport->phba;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *psb;
|
2005-04-17 21:05:31 +00:00
|
|
|
struct ulp_bde64 *bpl;
|
|
|
|
|
IOCB_t *iocb;
|
2008-08-25 01:49:55 +00:00
|
|
|
dma_addr_t pdma_phys_fcp_cmd;
|
|
|
|
|
dma_addr_t pdma_phys_fcp_rsp;
|
2019-01-28 19:14:19 +00:00
|
|
|
dma_addr_t pdma_phys_sgl;
|
2005-10-29 15:28:33 +00:00
|
|
|
uint16_t iotag;
|
2013-04-18 00:16:15 +00:00
|
|
|
int bcnt, bpl_size;
|
|
|
|
|
|
|
|
|
|
bpl_size = phba->cfg_sg_dma_buf_size -
|
|
|
|
|
(sizeof(struct fcp_cmnd) + sizeof(struct fcp_rsp));
|
|
|
|
|
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
|
|
|
|
"9067 ALLOC %d scsi_bufs: %d (%d + %d + %d)\n",
|
|
|
|
|
num_to_alloc, phba->cfg_sg_dma_buf_size,
|
|
|
|
|
(int)sizeof(struct fcp_cmnd),
|
|
|
|
|
(int)sizeof(struct fcp_rsp), bpl_size);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
for (bcnt = 0; bcnt < num_to_alloc; bcnt++) {
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
psb = kzalloc(sizeof(struct lpfc_io_buf), GFP_KERNEL);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
if (!psb)
|
|
|
|
|
break;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
/*
|
|
|
|
|
* Get memory from the pci pool to map the virt space to pci
|
|
|
|
|
* bus space for an I/O. The DMA buffer includes space for the
|
|
|
|
|
* struct fcp_cmnd, struct fcp_rsp and the number of bde's
|
|
|
|
|
* necessary to support the sg_tablesize.
|
|
|
|
|
*/
|
2017-07-06 08:13:05 +00:00
|
|
|
psb->data = dma_pool_zalloc(phba->lpfc_sg_dma_buf_pool,
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
GFP_KERNEL, &psb->dma_handle);
|
|
|
|
|
if (!psb->data) {
|
|
|
|
|
kfree(psb);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* Allocate iotag for psb->cur_iocbq. */
|
|
|
|
|
iotag = lpfc_sli_next_iotag(phba, &psb->cur_iocbq);
|
|
|
|
|
if (iotag == 0) {
|
2017-07-06 08:13:05 +00:00
|
|
|
dma_pool_free(phba->lpfc_sg_dma_buf_pool,
|
2017-02-12 21:52:30 +00:00
|
|
|
psb->data, psb->dma_handle);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
kfree(psb);
|
|
|
|
|
break;
|
|
|
|
|
}
|
2022-02-25 02:22:52 +00:00
|
|
|
psb->cur_iocbq.cmd_flag |= LPFC_IO_FCP;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
|
|
|
|
|
psb->fcp_cmnd = psb->data;
|
|
|
|
|
psb->fcp_rsp = psb->data + sizeof(struct fcp_cmnd);
|
2019-01-28 19:14:19 +00:00
|
|
|
psb->dma_sgl = psb->data + sizeof(struct fcp_cmnd) +
|
2008-08-25 01:49:55 +00:00
|
|
|
sizeof(struct fcp_rsp);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
/* Initialize local short-hand pointers. */
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
bpl = (struct ulp_bde64 *)psb->dma_sgl;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
pdma_phys_fcp_cmd = psb->dma_handle;
|
|
|
|
|
pdma_phys_fcp_rsp = psb->dma_handle + sizeof(struct fcp_cmnd);
|
2019-01-28 19:14:19 +00:00
|
|
|
pdma_phys_sgl = psb->dma_handle + sizeof(struct fcp_cmnd) +
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
sizeof(struct fcp_rsp);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The first two bdes are the FCP_CMD and FCP_RSP. The balance
|
|
|
|
|
* are sg list bdes. Initialize the first two and leave the
|
|
|
|
|
* rest for queuecommand.
|
|
|
|
|
*/
|
|
|
|
|
bpl[0].addrHigh = le32_to_cpu(putPaddrHigh(pdma_phys_fcp_cmd));
|
|
|
|
|
bpl[0].addrLow = le32_to_cpu(putPaddrLow(pdma_phys_fcp_cmd));
|
|
|
|
|
bpl[0].tus.f.bdeSize = sizeof(struct fcp_cmnd);
|
|
|
|
|
bpl[0].tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
bpl[0].tus.w = le32_to_cpu(bpl[0].tus.w);
|
|
|
|
|
|
|
|
|
|
/* Setup the physical region for the FCP RSP */
|
|
|
|
|
bpl[1].addrHigh = le32_to_cpu(putPaddrHigh(pdma_phys_fcp_rsp));
|
|
|
|
|
bpl[1].addrLow = le32_to_cpu(putPaddrLow(pdma_phys_fcp_rsp));
|
|
|
|
|
bpl[1].tus.f.bdeSize = sizeof(struct fcp_rsp);
|
|
|
|
|
bpl[1].tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
bpl[1].tus.w = le32_to_cpu(bpl[1].tus.w);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Since the IOCB for the FCP I/O is built into this
|
|
|
|
|
* lpfc_scsi_buf, initialize it with all known data now.
|
|
|
|
|
*/
|
|
|
|
|
iocb = &psb->cur_iocbq.iocb;
|
|
|
|
|
iocb->un.fcpi64.bdl.ulpIoTag32 = 0;
|
|
|
|
|
if ((phba->sli_rev == 3) &&
|
|
|
|
|
!(phba->sli3_options & LPFC_SLI3_BG_ENABLED)) {
|
|
|
|
|
/* fill in immediate fcp command BDE */
|
|
|
|
|
iocb->un.fcpi64.bdl.bdeFlags = BUFF_TYPE_BDE_IMMED;
|
|
|
|
|
iocb->un.fcpi64.bdl.bdeSize = sizeof(struct fcp_cmnd);
|
|
|
|
|
iocb->un.fcpi64.bdl.addrLow = offsetof(IOCB_t,
|
|
|
|
|
unsli3.fcp_ext.icd);
|
|
|
|
|
iocb->un.fcpi64.bdl.addrHigh = 0;
|
|
|
|
|
iocb->ulpBdeCount = 0;
|
|
|
|
|
iocb->ulpLe = 0;
|
2011-03-31 01:57:33 +00:00
|
|
|
/* fill in response BDE */
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
iocb->unsli3.fcp_ext.rbde.tus.f.bdeFlags =
|
|
|
|
|
BUFF_TYPE_BDE_64;
|
|
|
|
|
iocb->unsli3.fcp_ext.rbde.tus.f.bdeSize =
|
|
|
|
|
sizeof(struct fcp_rsp);
|
|
|
|
|
iocb->unsli3.fcp_ext.rbde.addrLow =
|
|
|
|
|
putPaddrLow(pdma_phys_fcp_rsp);
|
|
|
|
|
iocb->unsli3.fcp_ext.rbde.addrHigh =
|
|
|
|
|
putPaddrHigh(pdma_phys_fcp_rsp);
|
|
|
|
|
} else {
|
|
|
|
|
iocb->un.fcpi64.bdl.bdeFlags = BUFF_TYPE_BLP_64;
|
|
|
|
|
iocb->un.fcpi64.bdl.bdeSize =
|
|
|
|
|
(2 * sizeof(struct ulp_bde64));
|
|
|
|
|
iocb->un.fcpi64.bdl.addrLow =
|
2019-01-28 19:14:19 +00:00
|
|
|
putPaddrLow(pdma_phys_sgl);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
iocb->un.fcpi64.bdl.addrHigh =
|
2019-01-28 19:14:19 +00:00
|
|
|
putPaddrHigh(pdma_phys_sgl);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
iocb->ulpBdeCount = 1;
|
|
|
|
|
iocb->ulpLe = 1;
|
|
|
|
|
}
|
|
|
|
|
iocb->ulpClass = CLASS3;
|
|
|
|
|
psb->status = IOSTAT_SUCCESS;
|
2009-05-22 18:51:39 +00:00
|
|
|
/* Put it back into the SCSI buffer list */
|
2022-04-12 22:20:02 +00:00
|
|
|
psb->cur_iocbq.io_buf = psb;
|
2019-01-28 19:14:37 +00:00
|
|
|
spin_lock_init(&psb->buf_lock);
|
2009-11-18 20:40:49 +00:00
|
|
|
lpfc_release_scsi_buf_s3(phba, psb);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2008-08-25 01:49:55 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
return bcnt;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2011-02-16 17:39:35 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_sli4_vport_delete_fcp_xri_aborted -Remove all ndlp references for vport
|
|
|
|
|
* @vport: pointer to lpfc vport data structure.
|
|
|
|
|
*
|
|
|
|
|
* This routine is invoked by the vport cleanup for deletions and the cleanup
|
|
|
|
|
* for an ndlp on removal.
|
|
|
|
|
**/
|
|
|
|
|
void
|
|
|
|
|
lpfc_sli4_vport_delete_fcp_xri_aborted(struct lpfc_vport *vport)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *psb, *next_psb;
|
2019-01-28 19:14:22 +00:00
|
|
|
struct lpfc_sli4_hdw_queue *qp;
|
2011-02-16 17:39:35 +00:00
|
|
|
unsigned long iflag = 0;
|
2019-01-28 19:14:22 +00:00
|
|
|
int idx;
|
2011-02-16 17:39:35 +00:00
|
|
|
|
2019-01-28 19:14:38 +00:00
|
|
|
if (!(vport->cfg_enable_fc4_type & LPFC_ENABLE_FCP))
|
2017-02-12 21:52:30 +00:00
|
|
|
return;
|
2019-01-28 19:14:22 +00:00
|
|
|
|
2011-02-16 17:39:35 +00:00
|
|
|
spin_lock_irqsave(&phba->hbalock, iflag);
|
2019-01-28 19:14:22 +00:00
|
|
|
for (idx = 0; idx < phba->cfg_hdw_queue; idx++) {
|
|
|
|
|
qp = &phba->sli4_hba.hdwq[idx];
|
|
|
|
|
|
2019-08-14 23:57:11 +00:00
|
|
|
spin_lock(&qp->abts_io_buf_list_lock);
|
2019-01-28 19:14:22 +00:00
|
|
|
list_for_each_entry_safe(psb, next_psb,
|
2019-08-14 23:57:11 +00:00
|
|
|
&qp->lpfc_abts_io_buf_list, list) {
|
2022-02-25 02:22:52 +00:00
|
|
|
if (psb->cur_iocbq.cmd_flag & LPFC_IO_NVME)
|
2019-08-14 23:57:11 +00:00
|
|
|
continue;
|
|
|
|
|
|
2019-01-28 19:14:22 +00:00
|
|
|
if (psb->rdata && psb->rdata->pnode &&
|
|
|
|
|
psb->rdata->pnode->vport == vport)
|
|
|
|
|
psb->rdata = NULL;
|
|
|
|
|
}
|
2019-08-14 23:57:11 +00:00
|
|
|
spin_unlock(&qp->abts_io_buf_list_lock);
|
2011-02-16 17:39:35 +00:00
|
|
|
}
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, iflag);
|
|
|
|
|
}
|
|
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
/**
|
2019-08-14 23:57:11 +00:00
|
|
|
* lpfc_sli4_io_xri_aborted - Fast-path process of fcp xri abort
|
2009-05-22 18:51:39 +00:00
|
|
|
* @phba: pointer to lpfc hba data structure.
|
|
|
|
|
* @axri: pointer to the fcp xri abort wcqe structure.
|
2020-11-02 14:23:43 +00:00
|
|
|
* @idx: index into hdwq
|
2009-05-22 18:51:39 +00:00
|
|
|
*
|
|
|
|
|
* This routine is invoked by the worker thread to process a SLI4 fast-path
|
2019-08-14 23:57:11 +00:00
|
|
|
* FCP or NVME aborted xri.
|
2009-05-22 18:51:39 +00:00
|
|
|
**/
|
|
|
|
|
void
|
2019-08-14 23:57:11 +00:00
|
|
|
lpfc_sli4_io_xri_aborted(struct lpfc_hba *phba,
|
|
|
|
|
struct sli4_wcqe_xri_aborted *axri, int idx)
|
2009-05-22 18:51:39 +00:00
|
|
|
{
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
u16 xri = 0;
|
|
|
|
|
u16 rxid = 0;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *psb, *next_psb;
|
2019-01-28 19:14:22 +00:00
|
|
|
struct lpfc_sli4_hdw_queue *qp;
|
2009-05-22 18:51:39 +00:00
|
|
|
unsigned long iflag = 0;
|
2010-02-26 19:14:23 +00:00
|
|
|
struct lpfc_iocbq *iocbq;
|
|
|
|
|
int i;
|
2010-11-21 04:11:55 +00:00
|
|
|
struct lpfc_nodelist *ndlp;
|
|
|
|
|
int rrq_empty = 0;
|
2017-02-12 21:52:30 +00:00
|
|
|
struct lpfc_sli_ring *pring = phba->sli4_hba.els_wq->pring;
|
2021-05-14 19:55:57 +00:00
|
|
|
struct scsi_cmnd *cmd;
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
int offline = 0;
|
2009-05-22 18:51:39 +00:00
|
|
|
|
2017-02-12 21:52:30 +00:00
|
|
|
if (!(phba->cfg_enable_fc4_type & LPFC_ENABLE_FCP))
|
|
|
|
|
return;
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
offline = pci_channel_offline(phba->pcidev);
|
|
|
|
|
if (!offline) {
|
|
|
|
|
xri = bf_get(lpfc_wcqe_xa_xri, axri);
|
|
|
|
|
rxid = bf_get(lpfc_wcqe_xa_remote_xid, axri);
|
|
|
|
|
}
|
2019-01-28 19:14:22 +00:00
|
|
|
qp = &phba->sli4_hba.hdwq[idx];
|
2010-02-26 19:14:23 +00:00
|
|
|
spin_lock_irqsave(&phba->hbalock, iflag);
|
2019-08-14 23:57:11 +00:00
|
|
|
spin_lock(&qp->abts_io_buf_list_lock);
|
2009-05-22 18:51:39 +00:00
|
|
|
list_for_each_entry_safe(psb, next_psb,
|
2019-08-14 23:57:11 +00:00
|
|
|
&qp->lpfc_abts_io_buf_list, list) {
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
if (offline)
|
|
|
|
|
xri = psb->cur_iocbq.sli4_xritag;
|
2009-05-22 18:51:39 +00:00
|
|
|
if (psb->cur_iocbq.sli4_xritag == xri) {
|
2019-08-14 23:57:11 +00:00
|
|
|
list_del_init(&psb->list);
|
2019-10-18 21:18:21 +00:00
|
|
|
psb->flags &= ~LPFC_SBUF_XBUSY;
|
2009-05-22 18:51:39 +00:00
|
|
|
psb->status = IOSTAT_SUCCESS;
|
2022-02-25 02:22:52 +00:00
|
|
|
if (psb->cur_iocbq.cmd_flag & LPFC_IO_NVME) {
|
2019-08-14 23:57:11 +00:00
|
|
|
qp->abts_nvme_io_bufs--;
|
|
|
|
|
spin_unlock(&qp->abts_io_buf_list_lock);
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, iflag);
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
if (!offline) {
|
|
|
|
|
lpfc_sli4_nvme_xri_aborted(phba, axri,
|
|
|
|
|
psb);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
lpfc_sli4_nvme_pci_offline_aborted(phba, psb);
|
|
|
|
|
spin_lock_irqsave(&phba->hbalock, iflag);
|
|
|
|
|
spin_lock(&qp->abts_io_buf_list_lock);
|
|
|
|
|
continue;
|
2019-08-14 23:57:11 +00:00
|
|
|
}
|
|
|
|
|
qp->abts_scsi_io_bufs--;
|
|
|
|
|
spin_unlock(&qp->abts_io_buf_list_lock);
|
|
|
|
|
|
2011-02-16 17:39:35 +00:00
|
|
|
if (psb->rdata && psb->rdata->pnode)
|
|
|
|
|
ndlp = psb->rdata->pnode;
|
|
|
|
|
else
|
|
|
|
|
ndlp = NULL;
|
|
|
|
|
|
2010-11-21 04:11:55 +00:00
|
|
|
rrq_empty = list_empty(&phba->active_rrq_list);
|
2010-02-26 19:14:23 +00:00
|
|
|
spin_unlock_irqrestore(&phba->hbalock, iflag);
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
if (ndlp && !offline) {
|
2012-05-10 01:19:14 +00:00
|
|
|
lpfc_set_rrq_active(phba, ndlp,
|
|
|
|
|
psb->cur_iocbq.sli4_lxritag, rxid, 1);
|
2011-12-13 18:21:57 +00:00
|
|
|
lpfc_sli4_abts_err_handler(phba, ndlp, axri);
|
|
|
|
|
}
|
2021-05-14 19:55:57 +00:00
|
|
|
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
if (phba->cfg_fcp_wait_abts_rsp || offline) {
|
2021-05-14 19:55:57 +00:00
|
|
|
spin_lock_irqsave(&psb->buf_lock, iflag);
|
|
|
|
|
cmd = psb->pCmd;
|
|
|
|
|
psb->pCmd = NULL;
|
|
|
|
|
spin_unlock_irqrestore(&psb->buf_lock, iflag);
|
|
|
|
|
|
|
|
|
|
/* The sdev is not guaranteed to be valid post
|
|
|
|
|
* scsi_done upcall.
|
|
|
|
|
*/
|
|
|
|
|
if (cmd)
|
2021-10-07 20:28:41 +00:00
|
|
|
scsi_done(cmd);
|
2021-05-14 19:55:57 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We expect there is an abort thread waiting
|
|
|
|
|
* for command completion wake up the thread.
|
|
|
|
|
*/
|
|
|
|
|
spin_lock_irqsave(&psb->buf_lock, iflag);
|
2022-02-25 02:22:52 +00:00
|
|
|
psb->cur_iocbq.cmd_flag &=
|
2021-05-14 19:55:57 +00:00
|
|
|
~LPFC_DRIVER_ABORTED;
|
|
|
|
|
if (psb->waitq)
|
|
|
|
|
wake_up(psb->waitq);
|
|
|
|
|
spin_unlock_irqrestore(&psb->buf_lock, iflag);
|
|
|
|
|
}
|
|
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
lpfc_release_scsi_buf_s4(phba, psb);
|
2010-11-21 04:11:55 +00:00
|
|
|
if (rrq_empty)
|
|
|
|
|
lpfc_worker_wake_up(phba);
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
if (!offline)
|
|
|
|
|
return;
|
|
|
|
|
spin_lock_irqsave(&phba->hbalock, iflag);
|
|
|
|
|
spin_lock(&qp->abts_io_buf_list_lock);
|
|
|
|
|
continue;
|
2009-05-22 18:51:39 +00:00
|
|
|
}
|
|
|
|
|
}
|
2019-08-14 23:57:11 +00:00
|
|
|
spin_unlock(&qp->abts_io_buf_list_lock);
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
if (!offline) {
|
|
|
|
|
for (i = 1; i <= phba->sli.last_iotag; i++) {
|
|
|
|
|
iocbq = phba->sli.iocbq_lookup[i];
|
2010-02-26 19:14:23 +00:00
|
|
|
|
2022-02-25 02:22:52 +00:00
|
|
|
if (!(iocbq->cmd_flag & LPFC_IO_FCP) ||
|
|
|
|
|
(iocbq->cmd_flag & LPFC_IO_LIBDFC))
|
scsi: lpfc: Fix EEH support for NVMe I/O
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-09-10 23:31:54 +00:00
|
|
|
continue;
|
|
|
|
|
if (iocbq->sli4_xritag != xri)
|
|
|
|
|
continue;
|
|
|
|
|
psb = container_of(iocbq, struct lpfc_io_buf, cur_iocbq);
|
|
|
|
|
psb->flags &= ~LPFC_SBUF_XBUSY;
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, iflag);
|
|
|
|
|
if (!list_empty(&pring->txq))
|
|
|
|
|
lpfc_worker_wake_up(phba);
|
|
|
|
|
return;
|
|
|
|
|
}
|
2010-02-26 19:14:23 +00:00
|
|
|
}
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, iflag);
|
2009-05-22 18:51:39 +00:00
|
|
|
}
|
|
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
/**
|
2010-11-21 04:11:55 +00:00
|
|
|
* lpfc_get_scsi_buf_s3 - Get a scsi buffer from lpfc_scsi_buf_list of the HBA
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* @phba: The HBA for which this call is being executed.
|
2020-11-02 14:23:43 +00:00
|
|
|
* @ndlp: pointer to a node-list data structure.
|
|
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* This routine removes a scsi buffer from head of @phba lpfc_scsi_buf_list list
|
|
|
|
|
* and returns to caller.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* NULL - Error
|
|
|
|
|
* Pointer to lpfc_scsi_buf - Success
|
|
|
|
|
**/
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
static struct lpfc_io_buf *
|
2019-01-28 19:14:27 +00:00
|
|
|
lpfc_get_scsi_buf_s3(struct lpfc_hba *phba, struct lpfc_nodelist *ndlp,
|
|
|
|
|
struct scsi_cmnd *cmnd)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd = NULL;
|
2013-04-18 00:17:40 +00:00
|
|
|
struct list_head *scsi_buf_list_get = &phba->lpfc_scsi_buf_list_get;
|
2013-09-06 16:22:38 +00:00
|
|
|
unsigned long iflag = 0;
|
2013-04-18 00:17:40 +00:00
|
|
|
|
2013-09-06 16:22:38 +00:00
|
|
|
spin_lock_irqsave(&phba->scsi_buf_list_get_lock, iflag);
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
list_remove_head(scsi_buf_list_get, lpfc_cmd, struct lpfc_io_buf,
|
2013-04-18 00:17:40 +00:00
|
|
|
list);
|
|
|
|
|
if (!lpfc_cmd) {
|
2013-09-06 16:22:38 +00:00
|
|
|
spin_lock(&phba->scsi_buf_list_put_lock);
|
2013-04-18 00:17:40 +00:00
|
|
|
list_splice(&phba->lpfc_scsi_buf_list_put,
|
|
|
|
|
&phba->lpfc_scsi_buf_list_get);
|
|
|
|
|
INIT_LIST_HEAD(&phba->lpfc_scsi_buf_list_put);
|
|
|
|
|
list_remove_head(scsi_buf_list_get, lpfc_cmd,
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf, list);
|
2013-09-06 16:22:38 +00:00
|
|
|
spin_unlock(&phba->scsi_buf_list_put_lock);
|
2007-04-25 13:51:30 +00:00
|
|
|
}
|
2013-09-06 16:22:38 +00:00
|
|
|
spin_unlock_irqrestore(&phba->scsi_buf_list_get_lock, iflag);
|
2018-08-01 00:23:22 +00:00
|
|
|
|
|
|
|
|
if (lpfc_ndlp_check_qdepth(phba, ndlp) && lpfc_cmd) {
|
|
|
|
|
atomic_inc(&ndlp->cmd_pending);
|
|
|
|
|
lpfc_cmd->flags |= LPFC_SBUF_BUMP_QDEPTH;
|
|
|
|
|
}
|
2005-10-29 00:30:02 +00:00
|
|
|
return lpfc_cmd;
|
|
|
|
|
}
|
2010-11-21 04:11:55 +00:00
|
|
|
/**
|
2019-01-28 19:14:22 +00:00
|
|
|
* lpfc_get_scsi_buf_s4 - Get a scsi buffer from io_buf_list of the HBA
|
2010-11-21 04:11:55 +00:00
|
|
|
* @phba: The HBA for which this call is being executed.
|
2020-11-02 14:23:43 +00:00
|
|
|
* @ndlp: pointer to a node-list data structure.
|
|
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
2010-11-21 04:11:55 +00:00
|
|
|
*
|
2019-01-28 19:14:22 +00:00
|
|
|
* This routine removes a scsi buffer from head of @hdwq io_buf_list
|
2010-11-21 04:11:55 +00:00
|
|
|
* and returns to caller.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* NULL - Error
|
|
|
|
|
* Pointer to lpfc_scsi_buf - Success
|
|
|
|
|
**/
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
static struct lpfc_io_buf *
|
2019-01-28 19:14:27 +00:00
|
|
|
lpfc_get_scsi_buf_s4(struct lpfc_hba *phba, struct lpfc_nodelist *ndlp,
|
|
|
|
|
struct scsi_cmnd *cmnd)
|
2010-11-21 04:11:55 +00:00
|
|
|
{
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd;
|
2019-01-28 19:14:22 +00:00
|
|
|
struct lpfc_sli4_hdw_queue *qp;
|
2019-01-28 19:14:19 +00:00
|
|
|
struct sli4_sge *sgl;
|
|
|
|
|
dma_addr_t pdma_phys_fcp_rsp;
|
|
|
|
|
dma_addr_t pdma_phys_fcp_cmd;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
uint32_t cpu, idx;
|
2019-01-28 19:14:27 +00:00
|
|
|
int tag;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
struct fcp_cmd_rsp_buf *tmp = NULL;
|
2010-11-21 04:11:55 +00:00
|
|
|
|
2019-03-28 18:06:22 +00:00
|
|
|
cpu = raw_smp_processor_id();
|
2019-01-28 19:14:29 +00:00
|
|
|
if (cmnd && phba->cfg_fcp_io_sched == LPFC_FCP_SCHED_BY_HDWQ) {
|
2021-08-09 23:03:30 +00:00
|
|
|
tag = blk_mq_unique_tag(scsi_cmd_to_rq(cmnd));
|
2019-01-28 19:14:27 +00:00
|
|
|
idx = blk_mq_unique_tag_to_hwq(tag);
|
|
|
|
|
} else {
|
2019-01-28 19:14:31 +00:00
|
|
|
idx = phba->sli4_hba.cpu_map[cpu].hdwq;
|
2019-01-28 19:14:27 +00:00
|
|
|
}
|
2019-01-28 19:14:22 +00:00
|
|
|
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_cmd = lpfc_get_io_buf(phba, ndlp, idx,
|
|
|
|
|
!phba->cfg_xri_rebalancing);
|
|
|
|
|
if (!lpfc_cmd) {
|
|
|
|
|
qp = &phba->sli4_hba.hdwq[idx];
|
2019-01-28 19:14:22 +00:00
|
|
|
qp->empty_io_bufs++;
|
2011-02-16 17:39:35 +00:00
|
|
|
return NULL;
|
2019-01-28 19:14:22 +00:00
|
|
|
}
|
2018-08-01 00:23:22 +00:00
|
|
|
|
2019-01-28 19:14:19 +00:00
|
|
|
/* Setup key fields in buffer that may have been changed
|
|
|
|
|
* if other protocols used this buffer.
|
|
|
|
|
*/
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag = LPFC_IO_FCP;
|
2019-01-28 19:14:19 +00:00
|
|
|
lpfc_cmd->prot_seg_cnt = 0;
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
lpfc_cmd->timeout = 0;
|
|
|
|
|
lpfc_cmd->flags = 0;
|
|
|
|
|
lpfc_cmd->start_time = jiffies;
|
|
|
|
|
lpfc_cmd->waitq = NULL;
|
2019-01-28 19:14:22 +00:00
|
|
|
lpfc_cmd->cpu = cpu;
|
2019-01-28 19:14:19 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
lpfc_cmd->prot_data_type = 0;
|
|
|
|
|
#endif
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
tmp = lpfc_get_cmd_rsp_buf_per_hdwq(phba, lpfc_cmd);
|
2020-01-28 00:23:02 +00:00
|
|
|
if (!tmp) {
|
|
|
|
|
lpfc_release_io_buf(phba, lpfc_cmd, lpfc_cmd->hdwq);
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
return NULL;
|
2020-01-28 00:23:02 +00:00
|
|
|
}
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
|
|
|
|
|
lpfc_cmd->fcp_cmnd = tmp->fcp_cmnd;
|
|
|
|
|
lpfc_cmd->fcp_rsp = tmp->fcp_rsp;
|
2019-01-28 19:14:19 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The first two SGEs are the FCP_CMD and FCP_RSP.
|
|
|
|
|
* The balance are sg list bdes. Initialize the
|
|
|
|
|
* first two and leave the rest for queuecommand.
|
|
|
|
|
*/
|
|
|
|
|
sgl = (struct sli4_sge *)lpfc_cmd->dma_sgl;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
pdma_phys_fcp_cmd = tmp->fcp_cmd_rsp_dma_handle;
|
2019-01-28 19:14:19 +00:00
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(pdma_phys_fcp_cmd));
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(pdma_phys_fcp_cmd));
|
|
|
|
|
sgl->word2 = le32_to_cpu(sgl->word2);
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 0);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(sizeof(struct fcp_cmnd));
|
|
|
|
|
sgl++;
|
|
|
|
|
|
|
|
|
|
/* Setup the physical region for the FCP RSP */
|
|
|
|
|
pdma_phys_fcp_rsp = pdma_phys_fcp_cmd + sizeof(struct fcp_cmnd);
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(pdma_phys_fcp_rsp));
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(pdma_phys_fcp_rsp));
|
|
|
|
|
sgl->word2 = le32_to_cpu(sgl->word2);
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(sizeof(struct fcp_rsp));
|
|
|
|
|
|
2018-08-23 13:56:33 +00:00
|
|
|
if (lpfc_ndlp_check_qdepth(phba, ndlp)) {
|
2018-08-01 00:23:22 +00:00
|
|
|
atomic_inc(&ndlp->cmd_pending);
|
|
|
|
|
lpfc_cmd->flags |= LPFC_SBUF_BUMP_QDEPTH;
|
|
|
|
|
}
|
2013-04-18 00:17:40 +00:00
|
|
|
return lpfc_cmd;
|
2010-11-21 04:11:55 +00:00
|
|
|
}
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_get_scsi_buf - Get a scsi buffer from lpfc_scsi_buf_list of the HBA
|
|
|
|
|
* @phba: The HBA for which this call is being executed.
|
2020-11-02 14:23:43 +00:00
|
|
|
* @ndlp: pointer to a node-list data structure.
|
|
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
2010-11-21 04:11:55 +00:00
|
|
|
*
|
|
|
|
|
* This routine removes a scsi buffer from head of @phba lpfc_scsi_buf_list list
|
|
|
|
|
* and returns to caller.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* NULL - Error
|
|
|
|
|
* Pointer to lpfc_scsi_buf - Success
|
|
|
|
|
**/
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
static struct lpfc_io_buf*
|
2019-01-28 19:14:27 +00:00
|
|
|
lpfc_get_scsi_buf(struct lpfc_hba *phba, struct lpfc_nodelist *ndlp,
|
|
|
|
|
struct scsi_cmnd *cmnd)
|
2010-11-21 04:11:55 +00:00
|
|
|
{
|
2019-01-28 19:14:27 +00:00
|
|
|
return phba->lpfc_get_scsi_buf(phba, ndlp, cmnd);
|
2010-11-21 04:11:55 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2021-03-12 09:47:11 +00:00
|
|
|
* lpfc_release_scsi_buf_s3 - Return a scsi buffer back to hba scsi buf list
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @psb: The scsi buffer which is being released.
|
|
|
|
|
*
|
|
|
|
|
* This routine releases @psb scsi buffer by adding it to tail of @phba
|
|
|
|
|
* lpfc_scsi_buf_list list.
|
|
|
|
|
**/
|
2005-10-29 00:30:02 +00:00
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_release_scsi_buf_s3(struct lpfc_hba *phba, struct lpfc_io_buf *psb)
|
2005-10-29 00:30:02 +00:00
|
|
|
{
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
unsigned long iflag = 0;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2013-04-18 00:17:40 +00:00
|
|
|
psb->seg_cnt = 0;
|
|
|
|
|
psb->prot_seg_cnt = 0;
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&phba->scsi_buf_list_put_lock, iflag);
|
2005-10-29 00:30:02 +00:00
|
|
|
psb->pCmd = NULL;
|
2022-02-25 02:22:52 +00:00
|
|
|
psb->cur_iocbq.cmd_flag = LPFC_IO_FCP;
|
2013-04-18 00:17:40 +00:00
|
|
|
list_add_tail(&psb->list, &phba->lpfc_scsi_buf_list_put);
|
|
|
|
|
spin_unlock_irqrestore(&phba->scsi_buf_list_put_lock, iflag);
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_release_scsi_buf_s4: Return a scsi buffer back to hba scsi buf list.
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @psb: The scsi buffer which is being released.
|
|
|
|
|
*
|
2019-01-28 19:14:22 +00:00
|
|
|
* This routine releases @psb scsi buffer by adding it to tail of @hdwq
|
|
|
|
|
* io_buf_list list. For SLI4 XRI's are tied to the scsi buffer
|
2009-05-22 18:51:39 +00:00
|
|
|
* and cannot be reused for at least RA_TOV amount of time if it was
|
|
|
|
|
* aborted.
|
|
|
|
|
**/
|
|
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_release_scsi_buf_s4(struct lpfc_hba *phba, struct lpfc_io_buf *psb)
|
2009-05-22 18:51:39 +00:00
|
|
|
{
|
2019-01-28 19:14:22 +00:00
|
|
|
struct lpfc_sli4_hdw_queue *qp;
|
2009-05-22 18:51:39 +00:00
|
|
|
unsigned long iflag = 0;
|
|
|
|
|
|
2013-04-18 00:17:40 +00:00
|
|
|
psb->seg_cnt = 0;
|
|
|
|
|
psb->prot_seg_cnt = 0;
|
|
|
|
|
|
2019-01-28 19:14:26 +00:00
|
|
|
qp = psb->hdwq;
|
2019-10-18 21:18:21 +00:00
|
|
|
if (psb->flags & LPFC_SBUF_XBUSY) {
|
2019-08-14 23:57:11 +00:00
|
|
|
spin_lock_irqsave(&qp->abts_io_buf_list_lock, iflag);
|
2021-05-14 19:55:57 +00:00
|
|
|
if (!phba->cfg_fcp_wait_abts_rsp)
|
|
|
|
|
psb->pCmd = NULL;
|
2019-08-14 23:57:11 +00:00
|
|
|
list_add_tail(&psb->list, &qp->lpfc_abts_io_buf_list);
|
2019-01-28 19:14:22 +00:00
|
|
|
qp->abts_scsi_io_bufs++;
|
2019-08-14 23:57:11 +00:00
|
|
|
spin_unlock_irqrestore(&qp->abts_io_buf_list_lock, iflag);
|
2009-05-22 18:51:39 +00:00
|
|
|
} else {
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_release_io_buf(phba, (struct lpfc_io_buf *)psb, qp);
|
2009-05-22 18:51:39 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* lpfc_release_scsi_buf: Return a scsi buffer back to hba scsi buf list.
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @psb: The scsi buffer which is being released.
|
|
|
|
|
*
|
|
|
|
|
* This routine releases @psb scsi buffer by adding it to tail of @phba
|
|
|
|
|
* lpfc_scsi_buf_list list.
|
|
|
|
|
**/
|
|
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_release_scsi_buf(struct lpfc_hba *phba, struct lpfc_io_buf *psb)
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
{
|
2018-08-01 00:23:22 +00:00
|
|
|
if ((psb->flags & LPFC_SBUF_BUMP_QDEPTH) && psb->ndlp)
|
|
|
|
|
atomic_dec(&psb->ndlp->cmd_pending);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
|
2018-08-01 00:23:22 +00:00
|
|
|
psb->flags &= ~LPFC_SBUF_BUMP_QDEPTH;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
phba->lpfc_release_scsi_buf(phba, psb);
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-15 19:26:42 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_fcpcmd_to_iocb - copy the fcp_cmd data into the IOCB
|
|
|
|
|
* @data: A pointer to the immediate command data portion of the IOCB.
|
|
|
|
|
* @fcp_cmnd: The FCP Command that is provided by the SCSI layer.
|
|
|
|
|
*
|
|
|
|
|
* The routine copies the entire FCP command from @fcp_cmnd to @data while
|
|
|
|
|
* byte swapping the data to big endian format for transmission on the wire.
|
|
|
|
|
**/
|
|
|
|
|
static void
|
|
|
|
|
lpfc_fcpcmd_to_iocb(u8 *data, struct fcp_cmnd *fcp_cmnd)
|
|
|
|
|
{
|
|
|
|
|
int i, j;
|
|
|
|
|
|
|
|
|
|
for (i = 0, j = 0; i < sizeof(struct fcp_cmnd);
|
|
|
|
|
i += sizeof(uint32_t), j++) {
|
|
|
|
|
((uint32_t *)data)[j] = cpu_to_be32(((uint32_t *)fcp_cmnd)[j]);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_dma_buf_s3 - DMA mapping for scsi buffer to SLI3 IF spec
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be mapped.
|
|
|
|
|
*
|
|
|
|
|
* This routine does the pci dma mapping for scatter-gather list of scsi cmnd
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* field of @lpfc_cmd for device with SLI-3 interface spec. This routine scans
|
2016-10-13 22:06:02 +00:00
|
|
|
* through sg elements and format the bde. This routine also initializes all
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* IOCB fields which are dependent on scsi command request buffer.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* 1 - Error
|
|
|
|
|
* 0 - Success
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_scsi_prep_dma_buf_s3(struct lpfc_hba *phba, struct lpfc_io_buf *lpfc_cmd)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct scatterlist *sgel = NULL;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct ulp_bde64 *bpl = (struct ulp_bde64 *)lpfc_cmd->dma_sgl;
|
2010-02-26 19:14:23 +00:00
|
|
|
struct lpfc_iocbq *iocbq = &lpfc_cmd->cur_iocbq;
|
2005-04-17 21:05:31 +00:00
|
|
|
IOCB_t *iocb_cmd = &lpfc_cmd->cur_iocbq.iocb;
|
2008-08-25 01:49:55 +00:00
|
|
|
struct ulp_bde64 *data_bde = iocb_cmd->unsli3.fcp_ext.dbde;
|
2005-04-17 21:05:31 +00:00
|
|
|
dma_addr_t physaddr;
|
2008-08-25 01:49:55 +00:00
|
|
|
uint32_t num_bde = 0;
|
2007-06-18 00:56:39 +00:00
|
|
|
int nseg, datadir = scsi_cmnd->sc_data_direction;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* There are three possibilities here - use scatter-gather segment, use
|
|
|
|
|
* the single mapping, or neither. Start the lpfc command prep by
|
|
|
|
|
* bumping the bpl beyond the fcp_cmnd and fcp_rsp regions to the first
|
|
|
|
|
* data bde entry.
|
|
|
|
|
*/
|
|
|
|
|
bpl += 2;
|
2007-07-04 13:03:11 +00:00
|
|
|
if (scsi_sg_count(scsi_cmnd)) {
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
2021-09-25 12:53:23 +00:00
|
|
|
* The driver stores the segment count returned from dma_map_sg
|
2005-04-17 21:05:31 +00:00
|
|
|
* because this a count of dma-mappings used to map the use_sg
|
|
|
|
|
* pages. They are not guaranteed to be the same for those
|
|
|
|
|
* architectures that implement an IOMMU.
|
|
|
|
|
*/
|
|
|
|
|
|
2007-07-04 13:03:11 +00:00
|
|
|
nseg = dma_map_sg(&phba->pcidev->dev, scsi_sglist(scsi_cmnd),
|
|
|
|
|
scsi_sg_count(scsi_cmnd), datadir);
|
|
|
|
|
if (unlikely(!nseg))
|
|
|
|
|
return 1;
|
|
|
|
|
|
2007-06-18 00:56:39 +00:00
|
|
|
lpfc_cmd->seg_cnt = nseg;
|
2005-04-17 21:05:31 +00:00
|
|
|
if (lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"9064 BLKGRD: %s: Too many sg segments"
|
|
|
|
|
" from dma_map_sg. Config %d, seg_cnt"
|
|
|
|
|
" %d\n", __func__, phba->cfg_sg_seg_cnt,
|
|
|
|
|
lpfc_cmd->seg_cnt);
|
2019-08-14 23:56:54 +00:00
|
|
|
WARN_ON_ONCE(lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt);
|
2013-04-18 00:16:15 +00:00
|
|
|
lpfc_cmd->seg_cnt = 0;
|
2007-06-18 00:56:39 +00:00
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
2019-08-14 23:56:54 +00:00
|
|
|
return 2;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The driver established a maximum scatter-gather segment count
|
|
|
|
|
* during probe that limits the number of sg elements in any
|
|
|
|
|
* single scsi command. Just run through the seg_cnt and format
|
|
|
|
|
* the bde's.
|
2008-08-25 01:49:55 +00:00
|
|
|
* When using SLI-3 the driver will try to fit all the BDEs into
|
|
|
|
|
* the IOCB. If it can't then the BDEs get added to a BPL as it
|
|
|
|
|
* does for SLI-2 mode.
|
2005-04-17 21:05:31 +00:00
|
|
|
*/
|
2008-08-25 01:49:55 +00:00
|
|
|
scsi_for_each_sg(scsi_cmnd, sgel, nseg, num_bde) {
|
2005-04-17 21:05:31 +00:00
|
|
|
physaddr = sg_dma_address(sgel);
|
2008-08-25 01:49:55 +00:00
|
|
|
if (phba->sli_rev == 3 &&
|
2008-12-05 03:40:02 +00:00
|
|
|
!(phba->sli3_options & LPFC_SLI3_BG_ENABLED) &&
|
2022-02-25 02:22:52 +00:00
|
|
|
!(iocbq->cmd_flag & DSS_SECURITY_OP) &&
|
2008-08-25 01:49:55 +00:00
|
|
|
nseg <= LPFC_EXT_DATA_BDE_COUNT) {
|
|
|
|
|
data_bde->tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
data_bde->tus.f.bdeSize = sg_dma_len(sgel);
|
|
|
|
|
data_bde->addrLow = putPaddrLow(physaddr);
|
|
|
|
|
data_bde->addrHigh = putPaddrHigh(physaddr);
|
|
|
|
|
data_bde++;
|
|
|
|
|
} else {
|
|
|
|
|
bpl->tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
bpl->tus.f.bdeSize = sg_dma_len(sgel);
|
|
|
|
|
bpl->tus.w = le32_to_cpu(bpl->tus.w);
|
|
|
|
|
bpl->addrLow =
|
|
|
|
|
le32_to_cpu(putPaddrLow(physaddr));
|
|
|
|
|
bpl->addrHigh =
|
|
|
|
|
le32_to_cpu(putPaddrHigh(physaddr));
|
|
|
|
|
bpl++;
|
|
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
2007-07-04 13:03:11 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Finish initializing those IOCB fields that are dependent on the
|
2008-08-25 01:49:55 +00:00
|
|
|
* scsi_cmnd request_buffer. Note that for SLI-2 the bdeSize is
|
|
|
|
|
* explicitly reinitialized and for SLI-3 the extended bde count is
|
|
|
|
|
* explicitly reinitialized since all iocb memory resources are reused.
|
2005-04-17 21:05:31 +00:00
|
|
|
*/
|
2008-12-05 03:40:02 +00:00
|
|
|
if (phba->sli_rev == 3 &&
|
2010-02-26 19:14:23 +00:00
|
|
|
!(phba->sli3_options & LPFC_SLI3_BG_ENABLED) &&
|
2022-02-25 02:22:52 +00:00
|
|
|
!(iocbq->cmd_flag & DSS_SECURITY_OP)) {
|
2008-08-25 01:49:55 +00:00
|
|
|
if (num_bde > LPFC_EXT_DATA_BDE_COUNT) {
|
|
|
|
|
/*
|
|
|
|
|
* The extended IOCB format can only fit 3 BDE or a BPL.
|
|
|
|
|
* This I/O has more than 3 BDE so the 1st data bde will
|
|
|
|
|
* be a BPL that is filled in here.
|
|
|
|
|
*/
|
|
|
|
|
physaddr = lpfc_cmd->dma_handle;
|
|
|
|
|
data_bde->tus.f.bdeFlags = BUFF_TYPE_BLP_64;
|
|
|
|
|
data_bde->tus.f.bdeSize = (num_bde *
|
|
|
|
|
sizeof(struct ulp_bde64));
|
|
|
|
|
physaddr += (sizeof(struct fcp_cmnd) +
|
|
|
|
|
sizeof(struct fcp_rsp) +
|
|
|
|
|
(2 * sizeof(struct ulp_bde64)));
|
|
|
|
|
data_bde->addrHigh = putPaddrHigh(physaddr);
|
|
|
|
|
data_bde->addrLow = putPaddrLow(physaddr);
|
2011-03-31 01:57:33 +00:00
|
|
|
/* ebde count includes the response bde and data bpl */
|
2008-08-25 01:49:55 +00:00
|
|
|
iocb_cmd->unsli3.fcp_ext.ebde_count = 2;
|
|
|
|
|
} else {
|
2011-03-31 01:57:33 +00:00
|
|
|
/* ebde count includes the response bde and data bdes */
|
2008-08-25 01:49:55 +00:00
|
|
|
iocb_cmd->unsli3.fcp_ext.ebde_count = (num_bde + 1);
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
iocb_cmd->un.fcpi64.bdl.bdeSize =
|
|
|
|
|
((num_bde + 2) * sizeof(struct ulp_bde64));
|
2010-02-26 19:14:23 +00:00
|
|
|
iocb_cmd->unsli3.fcp_ext.ebde_count = (num_bde + 1);
|
2008-08-25 01:49:55 +00:00
|
|
|
}
|
2008-01-11 06:52:54 +00:00
|
|
|
fcp_cmnd->fcpDl = cpu_to_be32(scsi_bufflen(scsi_cmnd));
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Due to difference in data length between DIF/non-DIF paths,
|
|
|
|
|
* we need to set word 4 of IOCB here
|
|
|
|
|
*/
|
2009-04-06 22:48:10 +00:00
|
|
|
iocb_cmd->un.fcpi.fcpi_parm = scsi_bufflen(scsi_cmnd);
|
2020-11-15 19:26:42 +00:00
|
|
|
lpfc_fcpcmd_to_iocb(iocb_cmd->unsli3.fcp_ext.icd, fcp_cmnd);
|
2008-12-05 03:40:02 +00:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2016-10-13 22:06:02 +00:00
|
|
|
/* Return BG_ERR_INIT if error injection is detected by Initiator */
|
2012-03-02 03:37:42 +00:00
|
|
|
#define BG_ERR_INIT 0x1
|
2016-10-13 22:06:02 +00:00
|
|
|
/* Return BG_ERR_TGT if error injection is detected by Target */
|
2012-03-02 03:37:42 +00:00
|
|
|
#define BG_ERR_TGT 0x2
|
2016-10-13 22:06:02 +00:00
|
|
|
/* Return BG_ERR_SWAP if swapping CSUM<-->CRC is required for error injection */
|
2012-03-02 03:37:42 +00:00
|
|
|
#define BG_ERR_SWAP 0x10
|
2021-03-12 09:47:11 +00:00
|
|
|
/*
|
2016-10-13 22:06:02 +00:00
|
|
|
* Return BG_ERR_CHECK if disabling Guard/Ref/App checking is required for
|
|
|
|
|
* error injection
|
2021-03-12 09:47:11 +00:00
|
|
|
*/
|
2012-03-02 03:37:42 +00:00
|
|
|
#define BG_ERR_CHECK 0x20
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_bg_err_inject - Determine if we should inject an error
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
2011-10-11 01:34:11 +00:00
|
|
|
* @sc: The SCSI command to examine
|
|
|
|
|
* @reftag: (out) BlockGuard reference tag for transmitted data
|
|
|
|
|
* @apptag: (out) BlockGuard application tag for transmitted data
|
2020-11-02 14:23:43 +00:00
|
|
|
* @new_guard: (in) Value to replace CRC with if needed
|
2011-10-11 01:34:11 +00:00
|
|
|
*
|
2012-03-02 03:37:42 +00:00
|
|
|
* Returns BG_ERR_* bit mask or 0 if request ignored
|
2012-01-18 21:25:09 +00:00
|
|
|
**/
|
2011-10-11 01:34:11 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_bg_err_inject(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
|
|
|
|
uint32_t *reftag, uint16_t *apptag, uint32_t new_guard)
|
|
|
|
|
{
|
|
|
|
|
struct scatterlist *sgpe; /* s/g prot entry */
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd = NULL;
|
2012-01-18 21:25:09 +00:00
|
|
|
struct scsi_dif_tuple *src = NULL;
|
2012-03-02 03:38:29 +00:00
|
|
|
struct lpfc_nodelist *ndlp;
|
|
|
|
|
struct lpfc_rport_data *rdata;
|
2011-10-11 01:34:11 +00:00
|
|
|
uint32_t op = scsi_get_prot_op(sc);
|
|
|
|
|
uint32_t blksize;
|
|
|
|
|
uint32_t numblks;
|
2021-03-01 17:18:02 +00:00
|
|
|
u32 lba;
|
2011-10-11 01:34:11 +00:00
|
|
|
int rc = 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
int blockoff = 0;
|
2011-10-11 01:34:11 +00:00
|
|
|
|
|
|
|
|
if (op == SCSI_PROT_NORMAL)
|
|
|
|
|
return 0;
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
sgpe = scsi_prot_sglist(sc);
|
2021-08-17 02:50:14 +00:00
|
|
|
lba = scsi_prot_ref_tag(sc);
|
2012-03-02 03:38:29 +00:00
|
|
|
|
|
|
|
|
/* First check if we need to match the LBA */
|
2011-10-11 01:34:11 +00:00
|
|
|
if (phba->lpfc_injerr_lba != LPFC_INJERR_LBA_OFF) {
|
2021-08-17 02:50:14 +00:00
|
|
|
blksize = scsi_prot_interval(sc);
|
2011-10-11 01:34:11 +00:00
|
|
|
numblks = (scsi_bufflen(sc) + blksize - 1) / blksize;
|
|
|
|
|
|
|
|
|
|
/* Make sure we have the right LBA if one is specified */
|
2021-03-01 17:18:02 +00:00
|
|
|
if (phba->lpfc_injerr_lba < (u64)lba ||
|
|
|
|
|
(phba->lpfc_injerr_lba >= (u64)(lba + numblks)))
|
2011-10-11 01:34:11 +00:00
|
|
|
return 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
if (sgpe) {
|
2021-03-01 17:18:02 +00:00
|
|
|
blockoff = phba->lpfc_injerr_lba - (u64)lba;
|
2012-01-18 21:25:09 +00:00
|
|
|
numblks = sg_dma_len(sgpe) /
|
|
|
|
|
sizeof(struct scsi_dif_tuple);
|
|
|
|
|
if (numblks < blockoff)
|
|
|
|
|
blockoff = numblks;
|
|
|
|
|
}
|
2011-10-11 01:34:11 +00:00
|
|
|
}
|
|
|
|
|
|
2012-03-02 03:38:29 +00:00
|
|
|
/* Next check if we need to match the remote NPortID or WWPN */
|
2014-02-20 14:56:45 +00:00
|
|
|
rdata = lpfc_rport_data_from_scsi_device(sc->device);
|
2012-03-02 03:38:29 +00:00
|
|
|
if (rdata && rdata->pnode) {
|
|
|
|
|
ndlp = rdata->pnode;
|
|
|
|
|
|
|
|
|
|
/* Make sure we have the right NPortID if one is specified */
|
|
|
|
|
if (phba->lpfc_injerr_nportid &&
|
|
|
|
|
(phba->lpfc_injerr_nportid != ndlp->nlp_DID))
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Make sure we have the right WWPN if one is specified.
|
|
|
|
|
* wwn[0] should be a non-zero NAA in a good WWPN.
|
|
|
|
|
*/
|
|
|
|
|
if (phba->lpfc_injerr_wwpn.u.wwn[0] &&
|
|
|
|
|
(memcmp(&ndlp->nlp_portname, &phba->lpfc_injerr_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name)) != 0))
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Setup a ptr to the protection data if the SCSI host provides it */
|
|
|
|
|
if (sgpe) {
|
|
|
|
|
src = (struct scsi_dif_tuple *)sg_virt(sgpe);
|
|
|
|
|
src += blockoff;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_cmd = (struct lpfc_io_buf *)sc->host_scribble;
|
2012-03-02 03:38:29 +00:00
|
|
|
}
|
|
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
/* Should we change the Reference Tag */
|
|
|
|
|
if (reftag) {
|
2012-01-18 21:25:09 +00:00
|
|
|
if (phba->lpfc_injerr_wref_cnt) {
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2012-03-02 03:37:42 +00:00
|
|
|
if (src) {
|
|
|
|
|
/*
|
|
|
|
|
* For WRITE_PASS, force the error
|
|
|
|
|
* to be sent on the wire. It should
|
|
|
|
|
* be detected by the Target.
|
|
|
|
|
* If blockoff != 0 error will be
|
|
|
|
|
* inserted in middle of the IO.
|
|
|
|
|
*/
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR,
|
|
|
|
|
LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"9076 BLKGRD: Injecting reftag error: "
|
|
|
|
|
"write lba x%lx + x%x oldrefTag x%x\n",
|
|
|
|
|
(unsigned long)lba, blockoff,
|
2012-03-02 03:37:42 +00:00
|
|
|
be32_to_cpu(src->ref_tag));
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* Save the old ref_tag so we can
|
|
|
|
|
* restore it on completion.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
2012-03-02 03:37:42 +00:00
|
|
|
if (lpfc_cmd) {
|
|
|
|
|
lpfc_cmd->prot_data_type =
|
|
|
|
|
LPFC_INJERR_REFTAG;
|
|
|
|
|
lpfc_cmd->prot_data_segment =
|
|
|
|
|
src;
|
|
|
|
|
lpfc_cmd->prot_data =
|
|
|
|
|
src->ref_tag;
|
|
|
|
|
}
|
|
|
|
|
src->ref_tag = cpu_to_be32(0xDEADBEEF);
|
2012-01-18 21:25:09 +00:00
|
|
|
phba->lpfc_injerr_wref_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wref_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_TGT | BG_ERR_CHECK;
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
}
|
2020-08-23 22:36:59 +00:00
|
|
|
fallthrough;
|
2012-03-02 03:37:42 +00:00
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* For WRITE_INSERT, force the error
|
|
|
|
|
* to be sent on the wire. It should be
|
|
|
|
|
* detected by the Target.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
2012-03-02 03:37:42 +00:00
|
|
|
/* DEADBEEF will be the reftag on the wire */
|
2012-01-18 21:25:09 +00:00
|
|
|
*reftag = 0xDEADBEEF;
|
|
|
|
|
phba->lpfc_injerr_wref_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wref_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_TGT | BG_ERR_CHECK;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-03-02 03:37:42 +00:00
|
|
|
"9078 BLKGRD: Injecting reftag error: "
|
2012-01-18 21:25:09 +00:00
|
|
|
"write lba x%lx\n", (unsigned long)lba);
|
|
|
|
|
break;
|
2012-03-02 03:37:42 +00:00
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* For WRITE_STRIP and WRITE_PASS,
|
|
|
|
|
* force the error on data
|
|
|
|
|
* being copied from SLI-Host to SLI-Port.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
2011-10-11 01:34:11 +00:00
|
|
|
*reftag = 0xDEADBEEF;
|
|
|
|
|
phba->lpfc_injerr_wref_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wref_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_INIT;
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-03-02 03:37:42 +00:00
|
|
|
"9077 BLKGRD: Injecting reftag error: "
|
2011-10-11 01:34:11 +00:00
|
|
|
"write lba x%lx\n", (unsigned long)lba);
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
2011-10-11 01:34:11 +00:00
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
}
|
|
|
|
|
if (phba->lpfc_injerr_rref_cnt) {
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
/*
|
|
|
|
|
* For READ_STRIP and READ_PASS, force the
|
|
|
|
|
* error on data being read off the wire. It
|
|
|
|
|
* should force an IO error to the driver.
|
|
|
|
|
*/
|
2011-10-11 01:34:11 +00:00
|
|
|
*reftag = 0xDEADBEEF;
|
|
|
|
|
phba->lpfc_injerr_rref_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_rref_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
rc = BG_ERR_INIT;
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"9079 BLKGRD: Injecting reftag error: "
|
2011-10-11 01:34:11 +00:00
|
|
|
"read lba x%lx\n", (unsigned long)lba);
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
2011-10-11 01:34:11 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Should we change the Application Tag */
|
|
|
|
|
if (apptag) {
|
2012-01-18 21:25:09 +00:00
|
|
|
if (phba->lpfc_injerr_wapp_cnt) {
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2012-03-02 03:38:29 +00:00
|
|
|
if (src) {
|
2012-03-02 03:37:42 +00:00
|
|
|
/*
|
|
|
|
|
* For WRITE_PASS, force the error
|
|
|
|
|
* to be sent on the wire. It should
|
|
|
|
|
* be detected by the Target.
|
|
|
|
|
* If blockoff != 0 error will be
|
|
|
|
|
* inserted in middle of the IO.
|
|
|
|
|
*/
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR,
|
|
|
|
|
LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"9080 BLKGRD: Injecting apptag error: "
|
|
|
|
|
"write lba x%lx + x%x oldappTag x%x\n",
|
|
|
|
|
(unsigned long)lba, blockoff,
|
2012-03-02 03:37:42 +00:00
|
|
|
be16_to_cpu(src->app_tag));
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* Save the old app_tag so we can
|
|
|
|
|
* restore it on completion.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
2012-03-02 03:37:42 +00:00
|
|
|
if (lpfc_cmd) {
|
|
|
|
|
lpfc_cmd->prot_data_type =
|
|
|
|
|
LPFC_INJERR_APPTAG;
|
|
|
|
|
lpfc_cmd->prot_data_segment =
|
|
|
|
|
src;
|
|
|
|
|
lpfc_cmd->prot_data =
|
|
|
|
|
src->app_tag;
|
|
|
|
|
}
|
|
|
|
|
src->app_tag = cpu_to_be16(0xDEAD);
|
2012-01-18 21:25:09 +00:00
|
|
|
phba->lpfc_injerr_wapp_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wapp_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_TGT | BG_ERR_CHECK;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
}
|
2020-08-23 22:36:59 +00:00
|
|
|
fallthrough;
|
2012-03-02 03:37:42 +00:00
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* For WRITE_INSERT, force the
|
|
|
|
|
* error to be sent on the wire. It should be
|
|
|
|
|
* detected by the Target.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
2012-03-02 03:37:42 +00:00
|
|
|
/* DEAD will be the apptag on the wire */
|
2012-01-18 21:25:09 +00:00
|
|
|
*apptag = 0xDEAD;
|
|
|
|
|
phba->lpfc_injerr_wapp_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wapp_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_TGT | BG_ERR_CHECK;
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-03-02 03:37:42 +00:00
|
|
|
"0813 BLKGRD: Injecting apptag error: "
|
2012-01-18 21:25:09 +00:00
|
|
|
"write lba x%lx\n", (unsigned long)lba);
|
|
|
|
|
break;
|
2012-03-02 03:37:42 +00:00
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* For WRITE_STRIP and WRITE_PASS,
|
|
|
|
|
* force the error on data
|
|
|
|
|
* being copied from SLI-Host to SLI-Port.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
2011-10-11 01:34:11 +00:00
|
|
|
*apptag = 0xDEAD;
|
|
|
|
|
phba->lpfc_injerr_wapp_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wapp_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_INIT;
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-03-02 03:37:42 +00:00
|
|
|
"0812 BLKGRD: Injecting apptag error: "
|
2011-10-11 01:34:11 +00:00
|
|
|
"write lba x%lx\n", (unsigned long)lba);
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
2011-10-11 01:34:11 +00:00
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
}
|
|
|
|
|
if (phba->lpfc_injerr_rapp_cnt) {
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
/*
|
|
|
|
|
* For READ_STRIP and READ_PASS, force the
|
|
|
|
|
* error on data being read off the wire. It
|
|
|
|
|
* should force an IO error to the driver.
|
|
|
|
|
*/
|
2011-10-11 01:34:11 +00:00
|
|
|
*apptag = 0xDEAD;
|
|
|
|
|
phba->lpfc_injerr_rapp_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_rapp_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
rc = BG_ERR_INIT;
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"0814 BLKGRD: Injecting apptag error: "
|
2011-10-11 01:34:11 +00:00
|
|
|
"read lba x%lx\n", (unsigned long)lba);
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
2011-10-11 01:34:11 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
/* Should we change the Guard Tag */
|
2012-01-18 21:25:09 +00:00
|
|
|
if (new_guard) {
|
|
|
|
|
if (phba->lpfc_injerr_wgrd_cnt) {
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_CHECK;
|
2020-08-23 22:36:59 +00:00
|
|
|
fallthrough;
|
2012-03-02 03:37:42 +00:00
|
|
|
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* For WRITE_INSERT, force the
|
|
|
|
|
* error to be sent on the wire. It should be
|
|
|
|
|
* detected by the Target.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
|
|
|
|
phba->lpfc_injerr_wgrd_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wgrd_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2012-03-02 03:37:42 +00:00
|
|
|
rc |= BG_ERR_TGT | BG_ERR_SWAP;
|
2012-01-18 21:25:09 +00:00
|
|
|
/* Signals the caller to swap CRC->CSUM */
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-03-02 03:37:42 +00:00
|
|
|
"0817 BLKGRD: Injecting guard error: "
|
2012-01-18 21:25:09 +00:00
|
|
|
"write lba x%lx\n", (unsigned long)lba);
|
|
|
|
|
break;
|
2012-03-02 03:37:42 +00:00
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
2012-03-02 03:37:42 +00:00
|
|
|
* For WRITE_STRIP and WRITE_PASS,
|
|
|
|
|
* force the error on data
|
|
|
|
|
* being copied from SLI-Host to SLI-Port.
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
|
|
|
|
phba->lpfc_injerr_wgrd_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_wgrd_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2011-10-11 01:34:11 +00:00
|
|
|
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_INIT | BG_ERR_SWAP;
|
2012-01-18 21:25:09 +00:00
|
|
|
/* Signals the caller to swap CRC->CSUM */
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-03-02 03:37:42 +00:00
|
|
|
"0816 BLKGRD: Injecting guard error: "
|
2012-01-18 21:25:09 +00:00
|
|
|
"write lba x%lx\n", (unsigned long)lba);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (phba->lpfc_injerr_rgrd_cnt) {
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
/*
|
|
|
|
|
* For READ_STRIP and READ_PASS, force the
|
|
|
|
|
* error on data being read off the wire. It
|
|
|
|
|
* should force an IO error to the driver.
|
|
|
|
|
*/
|
|
|
|
|
phba->lpfc_injerr_rgrd_cnt--;
|
2012-03-02 03:38:29 +00:00
|
|
|
if (phba->lpfc_injerr_rgrd_cnt == 0) {
|
|
|
|
|
phba->lpfc_injerr_nportid = 0;
|
|
|
|
|
phba->lpfc_injerr_lba =
|
|
|
|
|
LPFC_INJERR_LBA_OFF;
|
|
|
|
|
memset(&phba->lpfc_injerr_wwpn,
|
|
|
|
|
0, sizeof(struct lpfc_name));
|
|
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2012-03-02 03:37:42 +00:00
|
|
|
rc = BG_ERR_INIT | BG_ERR_SWAP;
|
2012-01-18 21:25:09 +00:00
|
|
|
/* Signals the caller to swap CRC->CSUM */
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"0818 BLKGRD: Injecting guard error: "
|
|
|
|
|
"read lba x%lx\n", (unsigned long)lba);
|
|
|
|
|
}
|
2011-10-11 01:34:11 +00:00
|
|
|
}
|
|
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_sc_to_bg_opcodes - Determine the BlockGuard opcodes to be used with
|
|
|
|
|
* the specified SCSI command.
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
2010-04-06 18:49:53 +00:00
|
|
|
* @sc: The SCSI command to examine
|
2020-11-02 14:23:43 +00:00
|
|
|
* @txop: (out) BlockGuard operation for transmitted data
|
|
|
|
|
* @rxop: (out) BlockGuard operation for received data
|
2010-04-06 18:49:53 +00:00
|
|
|
*
|
|
|
|
|
* Returns: zero on success; non-zero if tx and/or rx op cannot be determined
|
|
|
|
|
*
|
2012-01-18 21:25:09 +00:00
|
|
|
**/
|
2008-12-05 03:40:02 +00:00
|
|
|
static int
|
2010-04-06 18:49:53 +00:00
|
|
|
lpfc_sc_to_bg_opcodes(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
|
|
|
|
uint8_t *txop, uint8_t *rxop)
|
2008-12-05 03:40:02 +00:00
|
|
|
{
|
2010-04-06 18:49:53 +00:00
|
|
|
uint8_t ret = 0;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_IP_CHECKSUM) {
|
2008-12-05 03:40:02 +00:00
|
|
|
switch (scsi_get_prot_op(sc)) {
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
2010-04-06 18:49:53 +00:00
|
|
|
*rxop = BG_OP_IN_NODIF_OUT_CSUM;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_CSUM_OUT_NODIF;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
2010-04-06 18:49:53 +00:00
|
|
|
*rxop = BG_OP_IN_CRC_OUT_NODIF;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_NODIF_OUT_CRC;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
|
2009-09-18 21:32:59 +00:00
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2010-04-06 18:49:53 +00:00
|
|
|
*rxop = BG_OP_IN_CRC_OUT_CSUM;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_CSUM_OUT_CRC;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_NORMAL:
|
|
|
|
|
default:
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2011-07-22 22:36:25 +00:00
|
|
|
"9063 BLKGRD: Bad op/guard:%d/IP combination\n",
|
|
|
|
|
scsi_get_prot_op(sc));
|
2010-04-06 18:49:53 +00:00
|
|
|
ret = 1;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
}
|
2011-07-22 22:36:25 +00:00
|
|
|
} else {
|
2008-12-05 03:40:02 +00:00
|
|
|
switch (scsi_get_prot_op(sc)) {
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
2010-04-06 18:49:53 +00:00
|
|
|
*rxop = BG_OP_IN_CRC_OUT_NODIF;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_NODIF_OUT_CRC;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2010-04-06 18:49:53 +00:00
|
|
|
*rxop = BG_OP_IN_CRC_OUT_CRC;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_CRC_OUT_CRC;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
2011-07-22 22:36:25 +00:00
|
|
|
*rxop = BG_OP_IN_NODIF_OUT_CRC;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_CRC_OUT_NODIF;
|
2011-07-22 22:36:25 +00:00
|
|
|
break;
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
case SCSI_PROT_NORMAL:
|
|
|
|
|
default:
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2011-07-22 22:36:25 +00:00
|
|
|
"9075 BLKGRD: Bad op/guard:%d/CRC combination\n",
|
|
|
|
|
scsi_get_prot_op(sc));
|
2010-04-06 18:49:53 +00:00
|
|
|
ret = 1;
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
return ret;
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_bg_err_opcodes - reDetermine the BlockGuard opcodes to be used with
|
|
|
|
|
* the specified SCSI command in order to force a guard tag error.
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @sc: The SCSI command to examine
|
2020-11-02 14:23:43 +00:00
|
|
|
* @txop: (out) BlockGuard operation for transmitted data
|
|
|
|
|
* @rxop: (out) BlockGuard operation for received data
|
2012-01-18 21:25:09 +00:00
|
|
|
*
|
|
|
|
|
* Returns: zero on success; non-zero if tx and/or rx op cannot be determined
|
|
|
|
|
*
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_bg_err_opcodes(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
|
|
|
|
uint8_t *txop, uint8_t *rxop)
|
|
|
|
|
{
|
|
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_IP_CHECKSUM) {
|
2012-01-18 21:25:09 +00:00
|
|
|
switch (scsi_get_prot_op(sc)) {
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
|
|
|
|
*rxop = BG_OP_IN_NODIF_OUT_CRC;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_CRC_OUT_NODIF;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
|
|
|
|
*rxop = BG_OP_IN_CSUM_OUT_NODIF;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_NODIF_OUT_CSUM;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2012-03-02 03:38:29 +00:00
|
|
|
*rxop = BG_OP_IN_CSUM_OUT_CRC;
|
2012-03-02 03:37:42 +00:00
|
|
|
*txop = BG_OP_IN_CRC_OUT_CSUM;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_NORMAL:
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
switch (scsi_get_prot_op(sc)) {
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
|
|
|
|
*rxop = BG_OP_IN_CSUM_OUT_NODIF;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_NODIF_OUT_CSUM;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
2012-03-02 03:38:29 +00:00
|
|
|
*rxop = BG_OP_IN_CSUM_OUT_CSUM;
|
2012-03-02 03:37:42 +00:00
|
|
|
*txop = BG_OP_IN_CSUM_OUT_CSUM;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
|
|
|
|
*rxop = BG_OP_IN_NODIF_OUT_CSUM;
|
2012-03-02 03:38:29 +00:00
|
|
|
*txop = BG_OP_IN_CSUM_OUT_NODIF;
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case SCSI_PROT_NORMAL:
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2021-08-31 11:40:58 +00:00
|
|
|
return 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_bg_setup_bpl - Setup BlockGuard BPL with no protection data
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @sc: pointer to scsi command we're working on
|
|
|
|
|
* @bpl: pointer to buffer list for protection groups
|
2020-11-02 14:23:43 +00:00
|
|
|
* @datasegcnt: number of segments of data that have been dma mapped
|
2012-01-18 21:25:09 +00:00
|
|
|
*
|
|
|
|
|
* This function sets up BPL buffer list for protection groups of
|
2008-12-05 03:40:02 +00:00
|
|
|
* type LPFC_PG_TYPE_NO_DIF
|
|
|
|
|
*
|
|
|
|
|
* This is usually used when the HBA is instructed to generate
|
|
|
|
|
* DIFs and insert them into data stream (or strip DIF from
|
|
|
|
|
* incoming data stream)
|
|
|
|
|
*
|
|
|
|
|
* The buffer list consists of just one protection group described
|
|
|
|
|
* below:
|
|
|
|
|
* +-------------------------+
|
2010-04-06 18:49:53 +00:00
|
|
|
* start of prot group --> | PDE_5 |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* | PDE_6 |
|
2008-12-05 03:40:02 +00:00
|
|
|
* +-------------------------+
|
|
|
|
|
* | Data BDE |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* |more Data BDE's ... (opt)|
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
*
|
|
|
|
|
*
|
|
|
|
|
* Note: Data s/g buffers have been dma mapped
|
2012-01-18 21:25:09 +00:00
|
|
|
*
|
|
|
|
|
* Returns the number of BDEs added to the BPL.
|
|
|
|
|
**/
|
2008-12-05 03:40:02 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_bg_setup_bpl(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
|
|
|
|
struct ulp_bde64 *bpl, int datasegcnt)
|
|
|
|
|
{
|
|
|
|
|
struct scatterlist *sgde = NULL; /* s/g data entry */
|
2010-04-06 18:49:53 +00:00
|
|
|
struct lpfc_pde5 *pde5 = NULL;
|
|
|
|
|
struct lpfc_pde6 *pde6 = NULL;
|
2008-12-05 03:40:02 +00:00
|
|
|
dma_addr_t physaddr;
|
2010-04-06 18:49:53 +00:00
|
|
|
int i = 0, num_bde = 0, status;
|
2008-12-05 03:40:02 +00:00
|
|
|
int datadir = sc->sc_data_direction;
|
2012-05-10 01:18:12 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t rc;
|
2012-05-10 01:18:12 +00:00
|
|
|
#endif
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t checking = 1;
|
2008-12-05 03:40:02 +00:00
|
|
|
uint32_t reftag;
|
2010-04-06 18:49:53 +00:00
|
|
|
uint8_t txop, rxop;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
status = lpfc_sc_to_bg_opcodes(phba, sc, &txop, &rxop);
|
|
|
|
|
if (status)
|
2008-12-05 03:40:02 +00:00
|
|
|
goto out;
|
|
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
/* extract some info from the scsi command for pde*/
|
2021-08-17 02:50:14 +00:00
|
|
|
reftag = scsi_prot_ref_tag(sc);
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-03-02 03:38:29 +00:00
|
|
|
rc = lpfc_bg_err_inject(phba, sc, &reftag, NULL, 1);
|
2012-01-18 21:25:09 +00:00
|
|
|
if (rc) {
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_SWAP)
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_err_opcodes(phba, sc, &txop, &rxop);
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_CHECK)
|
2012-01-18 21:25:09 +00:00
|
|
|
checking = 0;
|
|
|
|
|
}
|
2011-10-11 01:34:11 +00:00
|
|
|
#endif
|
|
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
/* setup PDE5 with what we have */
|
|
|
|
|
pde5 = (struct lpfc_pde5 *) bpl;
|
|
|
|
|
memset(pde5, 0, sizeof(struct lpfc_pde5));
|
|
|
|
|
bf_set(pde5_type, pde5, LPFC_PDE5_DESCRIPTOR);
|
|
|
|
|
|
2010-08-04 20:11:18 +00:00
|
|
|
/* Endianness conversion if necessary for PDE5 */
|
2010-07-14 19:30:54 +00:00
|
|
|
pde5->word0 = cpu_to_le32(pde5->word0);
|
2011-07-22 22:36:25 +00:00
|
|
|
pde5->reftag = cpu_to_le32(reftag);
|
2010-07-14 19:30:54 +00:00
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
/* advance bpl and increment bde count */
|
|
|
|
|
num_bde++;
|
|
|
|
|
bpl++;
|
|
|
|
|
pde6 = (struct lpfc_pde6 *) bpl;
|
|
|
|
|
|
|
|
|
|
/* setup PDE6 with the rest of the info */
|
|
|
|
|
memset(pde6, 0, sizeof(struct lpfc_pde6));
|
|
|
|
|
bf_set(pde6_type, pde6, LPFC_PDE6_DESCRIPTOR);
|
|
|
|
|
bf_set(pde6_optx, pde6, txop);
|
|
|
|
|
bf_set(pde6_oprx, pde6, rxop);
|
2013-04-18 00:18:07 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We only need to check the data on READs, for WRITEs
|
|
|
|
|
* protection data is automatically generated, not checked.
|
|
|
|
|
*/
|
2010-04-06 18:49:53 +00:00
|
|
|
if (datadir == DMA_FROM_DEVICE) {
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_GUARD_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(pde6_ce, pde6, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(pde6_ce, pde6, 0);
|
|
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_REF_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(pde6_re, pde6, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(pde6_re, pde6, 0);
|
2010-04-06 18:49:53 +00:00
|
|
|
}
|
|
|
|
|
bf_set(pde6_ai, pde6, 1);
|
2011-07-22 22:36:25 +00:00
|
|
|
bf_set(pde6_ae, pde6, 0);
|
|
|
|
|
bf_set(pde6_apptagval, pde6, 0);
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2010-08-04 20:11:18 +00:00
|
|
|
/* Endianness conversion if necessary for PDE6 */
|
2010-07-14 19:30:54 +00:00
|
|
|
pde6->word0 = cpu_to_le32(pde6->word0);
|
|
|
|
|
pde6->word1 = cpu_to_le32(pde6->word1);
|
|
|
|
|
pde6->word2 = cpu_to_le32(pde6->word2);
|
|
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
/* advance bpl and increment bde count */
|
2008-12-05 03:40:02 +00:00
|
|
|
num_bde++;
|
|
|
|
|
bpl++;
|
|
|
|
|
|
|
|
|
|
/* assumption: caller has already run dma_map_sg on command data */
|
|
|
|
|
scsi_for_each_sg(sc, sgde, datasegcnt, i) {
|
|
|
|
|
physaddr = sg_dma_address(sgde);
|
|
|
|
|
bpl->addrLow = le32_to_cpu(putPaddrLow(physaddr));
|
|
|
|
|
bpl->addrHigh = le32_to_cpu(putPaddrHigh(physaddr));
|
|
|
|
|
bpl->tus.f.bdeSize = sg_dma_len(sgde);
|
|
|
|
|
if (datadir == DMA_TO_DEVICE)
|
|
|
|
|
bpl->tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
else
|
|
|
|
|
bpl->tus.f.bdeFlags = BUFF_TYPE_BDE_64I;
|
|
|
|
|
bpl->tus.w = le32_to_cpu(bpl->tus.w);
|
|
|
|
|
bpl++;
|
|
|
|
|
num_bde++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
out:
|
|
|
|
|
return num_bde;
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_bg_setup_bpl_prot - Setup BlockGuard BPL with protection data
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @sc: pointer to scsi command we're working on
|
|
|
|
|
* @bpl: pointer to buffer list for protection groups
|
|
|
|
|
* @datacnt: number of segments of data that have been dma mapped
|
|
|
|
|
* @protcnt: number of segment of protection data that have been dma mapped
|
|
|
|
|
*
|
|
|
|
|
* This function sets up BPL buffer list for protection groups of
|
|
|
|
|
* type LPFC_PG_TYPE_DIF
|
2008-12-05 03:40:02 +00:00
|
|
|
*
|
|
|
|
|
* This is usually used when DIFs are in their own buffers,
|
|
|
|
|
* separate from the data. The HBA can then by instructed
|
|
|
|
|
* to place the DIFs in the outgoing stream. For read operations,
|
|
|
|
|
* The HBA could extract the DIFs and place it in DIF buffers.
|
|
|
|
|
*
|
|
|
|
|
* The buffer list for this type consists of one or more of the
|
|
|
|
|
* protection groups described below:
|
|
|
|
|
* +-------------------------+
|
2010-04-06 18:49:53 +00:00
|
|
|
* start of first prot group --> | PDE_5 |
|
2008-12-05 03:40:02 +00:00
|
|
|
* +-------------------------+
|
2010-04-06 18:49:53 +00:00
|
|
|
* | PDE_6 |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* | PDE_7 (Prot BDE) |
|
2008-12-05 03:40:02 +00:00
|
|
|
* +-------------------------+
|
|
|
|
|
* | Data BDE |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* |more Data BDE's ... (opt)|
|
|
|
|
|
* +-------------------------+
|
2010-04-06 18:49:53 +00:00
|
|
|
* start of new prot group --> | PDE_5 |
|
2008-12-05 03:40:02 +00:00
|
|
|
* +-------------------------+
|
|
|
|
|
* | ... |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
*
|
|
|
|
|
* Note: It is assumed that both data and protection s/g buffers have been
|
|
|
|
|
* mapped for DMA
|
2012-01-18 21:25:09 +00:00
|
|
|
*
|
|
|
|
|
* Returns the number of BDEs added to the BPL.
|
|
|
|
|
**/
|
2008-12-05 03:40:02 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_bg_setup_bpl_prot(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
|
|
|
|
struct ulp_bde64 *bpl, int datacnt, int protcnt)
|
|
|
|
|
{
|
|
|
|
|
struct scatterlist *sgde = NULL; /* s/g data entry */
|
|
|
|
|
struct scatterlist *sgpe = NULL; /* s/g prot entry */
|
2010-04-06 18:49:53 +00:00
|
|
|
struct lpfc_pde5 *pde5 = NULL;
|
|
|
|
|
struct lpfc_pde6 *pde6 = NULL;
|
2011-03-11 21:05:52 +00:00
|
|
|
struct lpfc_pde7 *pde7 = NULL;
|
2008-12-05 03:40:02 +00:00
|
|
|
dma_addr_t dataphysaddr, protphysaddr;
|
2023-01-09 23:33:08 +00:00
|
|
|
unsigned short curr_prot = 0;
|
2011-03-11 21:05:52 +00:00
|
|
|
unsigned int split_offset;
|
|
|
|
|
unsigned int protgroup_len, protgroup_offset = 0, protgroup_remainder;
|
2008-12-05 03:40:02 +00:00
|
|
|
unsigned int protgrp_blks, protgrp_bytes;
|
|
|
|
|
unsigned int remainder, subtotal;
|
2010-04-06 18:49:53 +00:00
|
|
|
int status;
|
2008-12-05 03:40:02 +00:00
|
|
|
int datadir = sc->sc_data_direction;
|
|
|
|
|
unsigned char pgdone = 0, alldone = 0;
|
|
|
|
|
unsigned blksize;
|
2012-05-10 01:18:12 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t rc;
|
2012-05-10 01:18:12 +00:00
|
|
|
#endif
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t checking = 1;
|
2008-12-05 03:40:02 +00:00
|
|
|
uint32_t reftag;
|
2010-04-06 18:49:53 +00:00
|
|
|
uint8_t txop, rxop;
|
2008-12-05 03:40:02 +00:00
|
|
|
int num_bde = 0;
|
|
|
|
|
|
|
|
|
|
sgpe = scsi_prot_sglist(sc);
|
|
|
|
|
sgde = scsi_sglist(sc);
|
|
|
|
|
|
|
|
|
|
if (!sgpe || !sgde) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:57:06 +00:00
|
|
|
"9020 Invalid s/g entry: data=x%px prot=x%px\n",
|
2012-01-18 21:25:09 +00:00
|
|
|
sgpe, sgde);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
status = lpfc_sc_to_bg_opcodes(phba, sc, &txop, &rxop);
|
|
|
|
|
if (status)
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
/* extract some info from the scsi command */
|
2021-08-17 02:50:14 +00:00
|
|
|
blksize = scsi_prot_interval(sc);
|
|
|
|
|
reftag = scsi_prot_ref_tag(sc);
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-03-02 03:38:29 +00:00
|
|
|
rc = lpfc_bg_err_inject(phba, sc, &reftag, NULL, 1);
|
2012-01-18 21:25:09 +00:00
|
|
|
if (rc) {
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_SWAP)
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_err_opcodes(phba, sc, &txop, &rxop);
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_CHECK)
|
2012-01-18 21:25:09 +00:00
|
|
|
checking = 0;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
split_offset = 0;
|
|
|
|
|
do {
|
2013-04-18 00:16:15 +00:00
|
|
|
/* Check to see if we ran out of space */
|
|
|
|
|
if (num_bde >= (phba->cfg_total_seg_cnt - 2))
|
|
|
|
|
return num_bde + 3;
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* setup PDE5 with what we have */
|
|
|
|
|
pde5 = (struct lpfc_pde5 *) bpl;
|
|
|
|
|
memset(pde5, 0, sizeof(struct lpfc_pde5));
|
|
|
|
|
bf_set(pde5_type, pde5, LPFC_PDE5_DESCRIPTOR);
|
|
|
|
|
|
|
|
|
|
/* Endianness conversion if necessary for PDE5 */
|
|
|
|
|
pde5->word0 = cpu_to_le32(pde5->word0);
|
|
|
|
|
pde5->reftag = cpu_to_le32(reftag);
|
|
|
|
|
|
|
|
|
|
/* advance bpl and increment bde count */
|
|
|
|
|
num_bde++;
|
|
|
|
|
bpl++;
|
|
|
|
|
pde6 = (struct lpfc_pde6 *) bpl;
|
|
|
|
|
|
|
|
|
|
/* setup PDE6 with the rest of the info */
|
|
|
|
|
memset(pde6, 0, sizeof(struct lpfc_pde6));
|
|
|
|
|
bf_set(pde6_type, pde6, LPFC_PDE6_DESCRIPTOR);
|
|
|
|
|
bf_set(pde6_optx, pde6, txop);
|
|
|
|
|
bf_set(pde6_oprx, pde6, rxop);
|
2013-04-18 00:18:07 +00:00
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_GUARD_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(pde6_ce, pde6, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(pde6_ce, pde6, 0);
|
|
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_REF_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(pde6_re, pde6, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(pde6_re, pde6, 0);
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
bf_set(pde6_ai, pde6, 1);
|
|
|
|
|
bf_set(pde6_ae, pde6, 0);
|
|
|
|
|
bf_set(pde6_apptagval, pde6, 0);
|
|
|
|
|
|
|
|
|
|
/* Endianness conversion if necessary for PDE6 */
|
|
|
|
|
pde6->word0 = cpu_to_le32(pde6->word0);
|
|
|
|
|
pde6->word1 = cpu_to_le32(pde6->word1);
|
|
|
|
|
pde6->word2 = cpu_to_le32(pde6->word2);
|
|
|
|
|
|
|
|
|
|
/* advance bpl and increment bde count */
|
|
|
|
|
num_bde++;
|
|
|
|
|
bpl++;
|
|
|
|
|
|
|
|
|
|
/* setup the first BDE that points to protection buffer */
|
|
|
|
|
protphysaddr = sg_dma_address(sgpe) + protgroup_offset;
|
|
|
|
|
protgroup_len = sg_dma_len(sgpe) - protgroup_offset;
|
|
|
|
|
|
|
|
|
|
/* must be integer multiple of the DIF block length */
|
|
|
|
|
BUG_ON(protgroup_len % 8);
|
|
|
|
|
|
|
|
|
|
pde7 = (struct lpfc_pde7 *) bpl;
|
|
|
|
|
memset(pde7, 0, sizeof(struct lpfc_pde7));
|
|
|
|
|
bf_set(pde7_type, pde7, LPFC_PDE7_DESCRIPTOR);
|
|
|
|
|
|
|
|
|
|
pde7->addrHigh = le32_to_cpu(putPaddrHigh(protphysaddr));
|
|
|
|
|
pde7->addrLow = le32_to_cpu(putPaddrLow(protphysaddr));
|
|
|
|
|
|
|
|
|
|
protgrp_blks = protgroup_len / 8;
|
|
|
|
|
protgrp_bytes = protgrp_blks * blksize;
|
|
|
|
|
|
|
|
|
|
/* check if this pde is crossing the 4K boundary; if so split */
|
|
|
|
|
if ((pde7->addrLow & 0xfff) + protgroup_len > 0x1000) {
|
|
|
|
|
protgroup_remainder = 0x1000 - (pde7->addrLow & 0xfff);
|
|
|
|
|
protgroup_offset += protgroup_remainder;
|
|
|
|
|
protgrp_blks = protgroup_remainder / 8;
|
|
|
|
|
protgrp_bytes = protgrp_blks * blksize;
|
|
|
|
|
} else {
|
|
|
|
|
protgroup_offset = 0;
|
|
|
|
|
curr_prot++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
num_bde++;
|
|
|
|
|
|
|
|
|
|
/* setup BDE's for data blocks associated with DIF data */
|
|
|
|
|
pgdone = 0;
|
|
|
|
|
subtotal = 0; /* total bytes processed for current prot grp */
|
|
|
|
|
while (!pgdone) {
|
2013-04-18 00:16:15 +00:00
|
|
|
/* Check to see if we ran out of space */
|
|
|
|
|
if (num_bde >= phba->cfg_total_seg_cnt)
|
|
|
|
|
return num_bde + 1;
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
if (!sgde) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"9065 BLKGRD:%s Invalid data segment\n",
|
|
|
|
|
__func__);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
bpl++;
|
|
|
|
|
dataphysaddr = sg_dma_address(sgde) + split_offset;
|
|
|
|
|
bpl->addrLow = le32_to_cpu(putPaddrLow(dataphysaddr));
|
|
|
|
|
bpl->addrHigh = le32_to_cpu(putPaddrHigh(dataphysaddr));
|
|
|
|
|
|
|
|
|
|
remainder = sg_dma_len(sgde) - split_offset;
|
|
|
|
|
|
|
|
|
|
if ((subtotal + remainder) <= protgrp_bytes) {
|
|
|
|
|
/* we can use this whole buffer */
|
|
|
|
|
bpl->tus.f.bdeSize = remainder;
|
|
|
|
|
split_offset = 0;
|
|
|
|
|
|
|
|
|
|
if ((subtotal + remainder) == protgrp_bytes)
|
|
|
|
|
pgdone = 1;
|
|
|
|
|
} else {
|
|
|
|
|
/* must split this buffer with next prot grp */
|
|
|
|
|
bpl->tus.f.bdeSize = protgrp_bytes - subtotal;
|
|
|
|
|
split_offset += bpl->tus.f.bdeSize;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
subtotal += bpl->tus.f.bdeSize;
|
|
|
|
|
|
|
|
|
|
if (datadir == DMA_TO_DEVICE)
|
|
|
|
|
bpl->tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
else
|
|
|
|
|
bpl->tus.f.bdeFlags = BUFF_TYPE_BDE_64I;
|
|
|
|
|
bpl->tus.w = le32_to_cpu(bpl->tus.w);
|
|
|
|
|
|
|
|
|
|
num_bde++;
|
|
|
|
|
|
|
|
|
|
if (split_offset)
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
/* Move to the next s/g segment if possible */
|
|
|
|
|
sgde = sg_next(sgde);
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (protgroup_offset) {
|
|
|
|
|
/* update the reference tag */
|
|
|
|
|
reftag += protgrp_blks;
|
|
|
|
|
bpl++;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* are we done ? */
|
|
|
|
|
if (curr_prot == protcnt) {
|
|
|
|
|
alldone = 1;
|
|
|
|
|
} else if (curr_prot < protcnt) {
|
|
|
|
|
/* advance to next prot buffer */
|
|
|
|
|
sgpe = sg_next(sgpe);
|
|
|
|
|
bpl++;
|
|
|
|
|
|
|
|
|
|
/* update the reference tag */
|
|
|
|
|
reftag += protgrp_blks;
|
|
|
|
|
} else {
|
|
|
|
|
/* if we're here, we have a bug */
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"9054 BLKGRD: bug in %s\n", __func__);
|
2012-01-18 21:25:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
} while (!alldone);
|
|
|
|
|
out:
|
|
|
|
|
|
|
|
|
|
return num_bde;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_bg_setup_sgl - Setup BlockGuard SGL with no protection data
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @sc: pointer to scsi command we're working on
|
|
|
|
|
* @sgl: pointer to buffer list for protection groups
|
2020-11-02 14:23:43 +00:00
|
|
|
* @datasegcnt: number of segments of data that have been dma mapped
|
|
|
|
|
* @lpfc_cmd: lpfc scsi command object pointer.
|
2012-01-18 21:25:09 +00:00
|
|
|
*
|
|
|
|
|
* This function sets up SGL buffer list for protection groups of
|
|
|
|
|
* type LPFC_PG_TYPE_NO_DIF
|
|
|
|
|
*
|
|
|
|
|
* This is usually used when the HBA is instructed to generate
|
|
|
|
|
* DIFs and insert them into data stream (or strip DIF from
|
|
|
|
|
* incoming data stream)
|
|
|
|
|
*
|
|
|
|
|
* The buffer list consists of just one protection group described
|
|
|
|
|
* below:
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* start of prot group --> | DI_SEED |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* | Data SGE |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* |more Data SGE's ... (opt)|
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
*
|
|
|
|
|
*
|
|
|
|
|
* Note: Data s/g buffers have been dma mapped
|
|
|
|
|
*
|
|
|
|
|
* Returns the number of SGEs added to the SGL.
|
|
|
|
|
**/
|
2023-12-20 16:26:58 +00:00
|
|
|
static uint32_t
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_setup_sgl(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
struct sli4_sge *sgl, int datasegcnt,
|
|
|
|
|
struct lpfc_io_buf *lpfc_cmd)
|
2012-01-18 21:25:09 +00:00
|
|
|
{
|
|
|
|
|
struct scatterlist *sgde = NULL; /* s/g data entry */
|
|
|
|
|
struct sli4_sge_diseed *diseed = NULL;
|
|
|
|
|
dma_addr_t physaddr;
|
2023-12-20 16:26:58 +00:00
|
|
|
int i = 0, status;
|
|
|
|
|
uint32_t reftag, num_sge = 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
uint8_t txop, rxop;
|
2012-05-10 01:18:12 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t rc;
|
2012-05-10 01:18:12 +00:00
|
|
|
#endif
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t checking = 1;
|
|
|
|
|
uint32_t dma_len;
|
|
|
|
|
uint32_t dma_offset = 0;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
struct sli4_hybrid_sgl *sgl_xtra = NULL;
|
|
|
|
|
int j;
|
|
|
|
|
bool lsp_just_set = false;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
status = lpfc_sc_to_bg_opcodes(phba, sc, &txop, &rxop);
|
|
|
|
|
if (status)
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
/* extract some info from the scsi command for pde*/
|
2021-08-17 02:50:14 +00:00
|
|
|
reftag = scsi_prot_ref_tag(sc);
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-03-02 03:38:29 +00:00
|
|
|
rc = lpfc_bg_err_inject(phba, sc, &reftag, NULL, 1);
|
2012-01-18 21:25:09 +00:00
|
|
|
if (rc) {
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_SWAP)
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_err_opcodes(phba, sc, &txop, &rxop);
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_CHECK)
|
2012-01-18 21:25:09 +00:00
|
|
|
checking = 0;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
/* setup DISEED with what we have */
|
|
|
|
|
diseed = (struct sli4_sge_diseed *) sgl;
|
|
|
|
|
memset(diseed, 0, sizeof(struct sli4_sge_diseed));
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl, LPFC_SGE_TYPE_DISEED);
|
|
|
|
|
|
|
|
|
|
/* Endianness conversion if necessary */
|
|
|
|
|
diseed->ref_tag = cpu_to_le32(reftag);
|
|
|
|
|
diseed->ref_tag_tran = diseed->ref_tag;
|
|
|
|
|
|
2013-04-18 00:18:07 +00:00
|
|
|
/*
|
|
|
|
|
* We only need to check the data on READs, for WRITEs
|
|
|
|
|
* protection data is automatically generated, not checked.
|
|
|
|
|
*/
|
|
|
|
|
if (sc->sc_data_direction == DMA_FROM_DEVICE) {
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_GUARD_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(lpfc_sli4_sge_dif_ce, diseed, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_ce, diseed, 0);
|
|
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_REF_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(lpfc_sli4_sge_dif_re, diseed, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_re, diseed, 0);
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* setup DISEED with the rest of the info */
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_optx, diseed, txop);
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_oprx, diseed, rxop);
|
2013-04-18 00:18:07 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
bf_set(lpfc_sli4_sge_dif_ai, diseed, 1);
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_me, diseed, 0);
|
|
|
|
|
|
|
|
|
|
/* Endianness conversion if necessary for DISEED */
|
|
|
|
|
diseed->word2 = cpu_to_le32(diseed->word2);
|
|
|
|
|
diseed->word3 = cpu_to_le32(diseed->word3);
|
|
|
|
|
|
|
|
|
|
/* advance bpl and increment sge count */
|
|
|
|
|
num_sge++;
|
|
|
|
|
sgl++;
|
|
|
|
|
|
|
|
|
|
/* assumption: caller has already run dma_map_sg on command data */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
sgde = scsi_sglist(sc);
|
|
|
|
|
j = 3;
|
|
|
|
|
for (i = 0; i < datasegcnt; i++) {
|
|
|
|
|
/* clear it */
|
|
|
|
|
sgl->word2 = 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
/* do we need to expand the segment */
|
|
|
|
|
if (!lsp_just_set && !((j + 1) % phba->border_sge_num) &&
|
|
|
|
|
((datasegcnt - 1) != i)) {
|
|
|
|
|
/* set LSP type */
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl, LPFC_SGE_TYPE_LSP);
|
|
|
|
|
|
|
|
|
|
sgl_xtra = lpfc_get_sgl_per_hdwq(phba, lpfc_cmd);
|
|
|
|
|
|
|
|
|
|
if (unlikely(!sgl_xtra)) {
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(
|
|
|
|
|
sgl_xtra->dma_phys_sgl));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(
|
|
|
|
|
sgl_xtra->dma_phys_sgl));
|
|
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl, LPFC_SGE_TYPE_DATA);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!(bf_get(lpfc_sli4_sge_type, sgl) & LPFC_SGE_TYPE_LSP)) {
|
|
|
|
|
if ((datasegcnt - 1) == i)
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
|
|
|
|
physaddr = sg_dma_address(sgde);
|
|
|
|
|
dma_len = sg_dma_len(sgde);
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(physaddr));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(physaddr));
|
|
|
|
|
|
|
|
|
|
bf_set(lpfc_sli4_sge_offset, sgl, dma_offset);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(dma_len);
|
|
|
|
|
|
|
|
|
|
dma_offset += dma_len;
|
|
|
|
|
sgde = sg_next(sgde);
|
|
|
|
|
|
|
|
|
|
sgl++;
|
|
|
|
|
num_sge++;
|
|
|
|
|
lsp_just_set = false;
|
|
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(phba->cfg_sg_dma_buf_size);
|
|
|
|
|
|
|
|
|
|
sgl = (struct sli4_sge *)sgl_xtra->dma_sgl;
|
|
|
|
|
i = i - 1;
|
|
|
|
|
|
|
|
|
|
lsp_just_set = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
j++;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
out:
|
|
|
|
|
return num_sge;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_bg_setup_sgl_prot - Setup BlockGuard SGL with protection data
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @sc: pointer to scsi command we're working on
|
|
|
|
|
* @sgl: pointer to buffer list for protection groups
|
|
|
|
|
* @datacnt: number of segments of data that have been dma mapped
|
|
|
|
|
* @protcnt: number of segment of protection data that have been dma mapped
|
2020-11-02 14:23:43 +00:00
|
|
|
* @lpfc_cmd: lpfc scsi command object pointer.
|
2012-01-18 21:25:09 +00:00
|
|
|
*
|
|
|
|
|
* This function sets up SGL buffer list for protection groups of
|
|
|
|
|
* type LPFC_PG_TYPE_DIF
|
|
|
|
|
*
|
|
|
|
|
* This is usually used when DIFs are in their own buffers,
|
|
|
|
|
* separate from the data. The HBA can then by instructed
|
|
|
|
|
* to place the DIFs in the outgoing stream. For read operations,
|
|
|
|
|
* The HBA could extract the DIFs and place it in DIF buffers.
|
|
|
|
|
*
|
|
|
|
|
* The buffer list for this type consists of one or more of the
|
|
|
|
|
* protection groups described below:
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* start of first prot group --> | DISEED |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* | DIF (Prot SGE) |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* | Data SGE |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* |more Data SGE's ... (opt)|
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* start of new prot group --> | DISEED |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
* | ... |
|
|
|
|
|
* +-------------------------+
|
|
|
|
|
*
|
|
|
|
|
* Note: It is assumed that both data and protection s/g buffers have been
|
|
|
|
|
* mapped for DMA
|
|
|
|
|
*
|
|
|
|
|
* Returns the number of SGEs added to the SGL.
|
|
|
|
|
**/
|
2023-12-20 16:26:58 +00:00
|
|
|
static uint32_t
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_setup_sgl_prot(struct lpfc_hba *phba, struct scsi_cmnd *sc,
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
struct sli4_sge *sgl, int datacnt, int protcnt,
|
|
|
|
|
struct lpfc_io_buf *lpfc_cmd)
|
2012-01-18 21:25:09 +00:00
|
|
|
{
|
|
|
|
|
struct scatterlist *sgde = NULL; /* s/g data entry */
|
|
|
|
|
struct scatterlist *sgpe = NULL; /* s/g prot entry */
|
|
|
|
|
struct sli4_sge_diseed *diseed = NULL;
|
|
|
|
|
dma_addr_t dataphysaddr, protphysaddr;
|
2023-01-09 23:33:08 +00:00
|
|
|
unsigned short curr_prot = 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
unsigned int split_offset;
|
|
|
|
|
unsigned int protgroup_len, protgroup_offset = 0, protgroup_remainder;
|
|
|
|
|
unsigned int protgrp_blks, protgrp_bytes;
|
|
|
|
|
unsigned int remainder, subtotal;
|
|
|
|
|
int status;
|
|
|
|
|
unsigned char pgdone = 0, alldone = 0;
|
|
|
|
|
unsigned blksize;
|
|
|
|
|
uint32_t reftag;
|
|
|
|
|
uint8_t txop, rxop;
|
|
|
|
|
uint32_t dma_len;
|
2012-05-10 01:18:12 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t rc;
|
2012-05-10 01:18:12 +00:00
|
|
|
#endif
|
2012-01-18 21:25:09 +00:00
|
|
|
uint32_t checking = 1;
|
2023-12-20 16:26:58 +00:00
|
|
|
uint32_t dma_offset = 0, num_sge = 0;
|
|
|
|
|
int j = 2;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
struct sli4_hybrid_sgl *sgl_xtra = NULL;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
sgpe = scsi_prot_sglist(sc);
|
|
|
|
|
sgde = scsi_sglist(sc);
|
|
|
|
|
|
|
|
|
|
if (!sgpe || !sgde) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:57:06 +00:00
|
|
|
"9082 Invalid s/g entry: data=x%px prot=x%px\n",
|
2008-12-05 03:40:02 +00:00
|
|
|
sgpe, sgde);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
status = lpfc_sc_to_bg_opcodes(phba, sc, &txop, &rxop);
|
|
|
|
|
if (status)
|
2008-12-05 03:40:02 +00:00
|
|
|
goto out;
|
|
|
|
|
|
2010-04-06 18:49:53 +00:00
|
|
|
/* extract some info from the scsi command */
|
2021-08-17 02:50:14 +00:00
|
|
|
blksize = scsi_prot_interval(sc);
|
|
|
|
|
reftag = scsi_prot_ref_tag(sc);
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2011-10-11 01:34:11 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2012-03-02 03:38:29 +00:00
|
|
|
rc = lpfc_bg_err_inject(phba, sc, &reftag, NULL, 1);
|
2012-01-18 21:25:09 +00:00
|
|
|
if (rc) {
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_SWAP)
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_err_opcodes(phba, sc, &txop, &rxop);
|
2012-03-02 03:37:42 +00:00
|
|
|
if (rc & BG_ERR_CHECK)
|
2012-01-18 21:25:09 +00:00
|
|
|
checking = 0;
|
|
|
|
|
}
|
2011-10-11 01:34:11 +00:00
|
|
|
#endif
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
split_offset = 0;
|
|
|
|
|
do {
|
2013-04-18 00:16:15 +00:00
|
|
|
/* Check to see if we ran out of space */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if ((num_sge >= (phba->cfg_total_seg_cnt - 2)) &&
|
|
|
|
|
!(phba->cfg_xpsgl))
|
2013-04-18 00:16:15 +00:00
|
|
|
return num_sge + 3;
|
|
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
/* DISEED and DIF have to be together */
|
|
|
|
|
if (!((j + 1) % phba->border_sge_num) ||
|
|
|
|
|
!((j + 2) % phba->border_sge_num) ||
|
|
|
|
|
!((j + 3) % phba->border_sge_num)) {
|
|
|
|
|
sgl->word2 = 0;
|
|
|
|
|
|
|
|
|
|
/* set LSP type */
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl, LPFC_SGE_TYPE_LSP);
|
|
|
|
|
|
|
|
|
|
sgl_xtra = lpfc_get_sgl_per_hdwq(phba, lpfc_cmd);
|
|
|
|
|
|
|
|
|
|
if (unlikely(!sgl_xtra)) {
|
|
|
|
|
goto out;
|
|
|
|
|
} else {
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(
|
|
|
|
|
sgl_xtra->dma_phys_sgl));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(
|
|
|
|
|
sgl_xtra->dma_phys_sgl));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(phba->cfg_sg_dma_buf_size);
|
|
|
|
|
|
|
|
|
|
sgl = (struct sli4_sge *)sgl_xtra->dma_sgl;
|
|
|
|
|
j = 0;
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* setup DISEED with what we have */
|
|
|
|
|
diseed = (struct sli4_sge_diseed *) sgl;
|
|
|
|
|
memset(diseed, 0, sizeof(struct sli4_sge_diseed));
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl, LPFC_SGE_TYPE_DISEED);
|
|
|
|
|
|
|
|
|
|
/* Endianness conversion if necessary */
|
|
|
|
|
diseed->ref_tag = cpu_to_le32(reftag);
|
|
|
|
|
diseed->ref_tag_tran = diseed->ref_tag;
|
|
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_GUARD_CHECK) {
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(lpfc_sli4_sge_dif_ce, diseed, checking);
|
|
|
|
|
} else {
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_ce, diseed, 0);
|
|
|
|
|
/*
|
|
|
|
|
* When in this mode, the hardware will replace
|
|
|
|
|
* the guard tag from the host with a
|
|
|
|
|
* newly generated good CRC for the wire.
|
|
|
|
|
* Switch to raw mode here to avoid this
|
|
|
|
|
* behavior. What the host sends gets put on the wire.
|
|
|
|
|
*/
|
|
|
|
|
if (txop == BG_OP_IN_CRC_OUT_CRC) {
|
|
|
|
|
txop = BG_OP_RAW_MODE;
|
|
|
|
|
rxop = BG_OP_RAW_MODE;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
if (sc->prot_flags & SCSI_PROT_REF_CHECK)
|
2013-04-18 00:18:07 +00:00
|
|
|
bf_set(lpfc_sli4_sge_dif_re, diseed, checking);
|
|
|
|
|
else
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_re, diseed, 0);
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* setup DISEED with the rest of the info */
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_optx, diseed, txop);
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_oprx, diseed, rxop);
|
2013-04-18 00:18:07 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
bf_set(lpfc_sli4_sge_dif_ai, diseed, 1);
|
|
|
|
|
bf_set(lpfc_sli4_sge_dif_me, diseed, 0);
|
|
|
|
|
|
|
|
|
|
/* Endianness conversion if necessary for DISEED */
|
|
|
|
|
diseed->word2 = cpu_to_le32(diseed->word2);
|
|
|
|
|
diseed->word3 = cpu_to_le32(diseed->word3);
|
|
|
|
|
|
|
|
|
|
/* advance sgl and increment bde count */
|
|
|
|
|
num_sge++;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
sgl++;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
j++;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
/* setup the first BDE that points to protection buffer */
|
2011-03-11 21:05:52 +00:00
|
|
|
protphysaddr = sg_dma_address(sgpe) + protgroup_offset;
|
|
|
|
|
protgroup_len = sg_dma_len(sgpe) - protgroup_offset;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
/* must be integer multiple of the DIF block length */
|
|
|
|
|
BUG_ON(protgroup_len % 8);
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* Now setup DIF SGE */
|
|
|
|
|
sgl->word2 = 0;
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl, LPFC_SGE_TYPE_DIF);
|
|
|
|
|
sgl->addr_hi = le32_to_cpu(putPaddrHigh(protphysaddr));
|
|
|
|
|
sgl->addr_lo = le32_to_cpu(putPaddrLow(protphysaddr));
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
sgl->sge_len = 0;
|
2011-03-11 21:05:52 +00:00
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
protgrp_blks = protgroup_len / 8;
|
|
|
|
|
protgrp_bytes = protgrp_blks * blksize;
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* check if DIF SGE is crossing the 4K boundary; if so split */
|
|
|
|
|
if ((sgl->addr_lo & 0xfff) + protgroup_len > 0x1000) {
|
|
|
|
|
protgroup_remainder = 0x1000 - (sgl->addr_lo & 0xfff);
|
2011-03-11 21:05:52 +00:00
|
|
|
protgroup_offset += protgroup_remainder;
|
|
|
|
|
protgrp_blks = protgroup_remainder / 8;
|
2011-07-22 22:36:25 +00:00
|
|
|
protgrp_bytes = protgrp_blks * blksize;
|
2011-03-11 21:05:52 +00:00
|
|
|
} else {
|
|
|
|
|
protgroup_offset = 0;
|
|
|
|
|
curr_prot++;
|
|
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
num_sge++;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* setup SGE's for data blocks associated with DIF data */
|
2008-12-05 03:40:02 +00:00
|
|
|
pgdone = 0;
|
|
|
|
|
subtotal = 0; /* total bytes processed for current prot grp */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
|
|
|
|
|
sgl++;
|
|
|
|
|
j++;
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
while (!pgdone) {
|
2013-04-18 00:16:15 +00:00
|
|
|
/* Check to see if we ran out of space */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if ((num_sge >= phba->cfg_total_seg_cnt) &&
|
|
|
|
|
!phba->cfg_xpsgl)
|
2013-04-18 00:16:15 +00:00
|
|
|
return num_sge + 1;
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
if (!sgde) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"9086 BLKGRD:%s Invalid data segment\n",
|
2008-12-05 03:40:02 +00:00
|
|
|
__func__);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if (!((j + 1) % phba->border_sge_num)) {
|
|
|
|
|
sgl->word2 = 0;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
/* set LSP type */
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl,
|
|
|
|
|
LPFC_SGE_TYPE_LSP);
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
sgl_xtra = lpfc_get_sgl_per_hdwq(phba,
|
|
|
|
|
lpfc_cmd);
|
|
|
|
|
|
|
|
|
|
if (unlikely(!sgl_xtra)) {
|
|
|
|
|
goto out;
|
|
|
|
|
} else {
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(
|
|
|
|
|
putPaddrLow(sgl_xtra->dma_phys_sgl));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(
|
|
|
|
|
putPaddrHigh(sgl_xtra->dma_phys_sgl));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(
|
|
|
|
|
phba->cfg_sg_dma_buf_size);
|
|
|
|
|
|
|
|
|
|
sgl = (struct sli4_sge *)sgl_xtra->dma_sgl;
|
2008-12-05 03:40:02 +00:00
|
|
|
} else {
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
dataphysaddr = sg_dma_address(sgde) +
|
|
|
|
|
split_offset;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
remainder = sg_dma_len(sgde) - split_offset;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if ((subtotal + remainder) <= protgrp_bytes) {
|
|
|
|
|
/* we can use this whole buffer */
|
|
|
|
|
dma_len = remainder;
|
|
|
|
|
split_offset = 0;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if ((subtotal + remainder) ==
|
|
|
|
|
protgrp_bytes)
|
|
|
|
|
pgdone = 1;
|
|
|
|
|
} else {
|
|
|
|
|
/* must split this buffer with next
|
|
|
|
|
* prot grp
|
|
|
|
|
*/
|
|
|
|
|
dma_len = protgrp_bytes - subtotal;
|
|
|
|
|
split_offset += dma_len;
|
|
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
subtotal += dma_len;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
sgl->word2 = 0;
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(
|
|
|
|
|
dataphysaddr));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(
|
|
|
|
|
dataphysaddr));
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 0);
|
|
|
|
|
bf_set(lpfc_sli4_sge_offset, sgl, dma_offset);
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl,
|
|
|
|
|
LPFC_SGE_TYPE_DATA);
|
2008-12-05 03:40:02 +00:00
|
|
|
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
sgl->sge_len = cpu_to_le32(dma_len);
|
|
|
|
|
dma_offset += dma_len;
|
|
|
|
|
|
|
|
|
|
num_sge++;
|
|
|
|
|
|
|
|
|
|
if (split_offset) {
|
|
|
|
|
sgl++;
|
|
|
|
|
j++;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Move to the next s/g segment if possible */
|
|
|
|
|
sgde = sg_next(sgde);
|
|
|
|
|
|
|
|
|
|
sgl++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
j++;
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
2011-03-11 21:05:52 +00:00
|
|
|
if (protgroup_offset) {
|
|
|
|
|
/* update the reference tag */
|
|
|
|
|
reftag += protgrp_blks;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
/* are we done ? */
|
|
|
|
|
if (curr_prot == protcnt) {
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
/* mark the last SGL */
|
|
|
|
|
sgl--;
|
2012-01-18 21:25:09 +00:00
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
2008-12-05 03:40:02 +00:00
|
|
|
alldone = 1;
|
|
|
|
|
} else if (curr_prot < protcnt) {
|
|
|
|
|
/* advance to next prot buffer */
|
|
|
|
|
sgpe = sg_next(sgpe);
|
|
|
|
|
|
|
|
|
|
/* update the reference tag */
|
|
|
|
|
reftag += protgrp_blks;
|
|
|
|
|
} else {
|
|
|
|
|
/* if we're here, we have a bug */
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"9085 BLKGRD: bug in %s\n", __func__);
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
} while (!alldone);
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
out:
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
return num_sge;
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
2011-03-11 21:05:52 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_prot_group_type - Get prtotection group type of SCSI command
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @sc: pointer to scsi command we're working on
|
|
|
|
|
*
|
2008-12-05 03:40:02 +00:00
|
|
|
* Given a SCSI command that supports DIF, determine composition of protection
|
|
|
|
|
* groups involved in setting up buffer lists
|
|
|
|
|
*
|
2012-01-18 21:25:09 +00:00
|
|
|
* Returns: Protection group type (with or without DIF)
|
|
|
|
|
*
|
|
|
|
|
**/
|
2008-12-05 03:40:02 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_prot_group_type(struct lpfc_hba *phba, struct scsi_cmnd *sc)
|
|
|
|
|
{
|
|
|
|
|
int ret = LPFC_PG_TYPE_INVALID;
|
|
|
|
|
unsigned char op = scsi_get_prot_op(sc);
|
|
|
|
|
|
|
|
|
|
switch (op) {
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
|
|
|
|
ret = LPFC_PG_TYPE_NO_DIF;
|
|
|
|
|
break;
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
|
|
|
|
ret = LPFC_PG_TYPE_DIF_BUF;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
2013-05-31 21:03:39 +00:00
|
|
|
if (phba)
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2013-05-31 21:03:39 +00:00
|
|
|
"9021 Unsupported protection op:%d\n",
|
|
|
|
|
op);
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
2013-04-18 00:18:07 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_bg_scsi_adjust_dl - Adjust SCSI data length for BlockGuard
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be adjusted.
|
|
|
|
|
*
|
|
|
|
|
* Adjust the data length to account for how much data
|
|
|
|
|
* is actually on the wire.
|
|
|
|
|
*
|
|
|
|
|
* returns the adjusted data length
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_bg_scsi_adjust_dl(struct lpfc_hba *phba,
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd)
|
2013-04-18 00:18:07 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *sc = lpfc_cmd->pCmd;
|
|
|
|
|
int fcpdl;
|
|
|
|
|
|
|
|
|
|
fcpdl = scsi_bufflen(sc);
|
|
|
|
|
|
|
|
|
|
/* Check if there is protection data on the wire */
|
|
|
|
|
if (sc->sc_data_direction == DMA_FROM_DEVICE) {
|
2013-05-31 21:03:39 +00:00
|
|
|
/* Read check for protection data */
|
2013-04-18 00:18:07 +00:00
|
|
|
if (scsi_get_prot_op(sc) == SCSI_PROT_READ_INSERT)
|
|
|
|
|
return fcpdl;
|
|
|
|
|
|
|
|
|
|
} else {
|
2013-05-31 21:03:39 +00:00
|
|
|
/* Write check for protection data */
|
2013-04-18 00:18:07 +00:00
|
|
|
if (scsi_get_prot_op(sc) == SCSI_PROT_WRITE_STRIP)
|
|
|
|
|
return fcpdl;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If we are in DIF Type 1 mode every data block has a 8 byte
|
2013-05-31 21:03:39 +00:00
|
|
|
* DIF (trailer) attached to it. Must ajust FCP data length
|
|
|
|
|
* to account for the protection data.
|
2013-04-18 00:18:07 +00:00
|
|
|
*/
|
2021-08-17 02:50:14 +00:00
|
|
|
fcpdl += (fcpdl / scsi_prot_interval(sc)) * 8;
|
2013-04-18 00:18:07 +00:00
|
|
|
|
|
|
|
|
return fcpdl;
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_bg_scsi_prep_dma_buf_s3 - DMA mapping for scsi buffer to SLI3 IF spec
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be prep'ed.
|
|
|
|
|
*
|
2008-12-05 03:40:02 +00:00
|
|
|
* This is the protection/DIF aware version of
|
|
|
|
|
* lpfc_scsi_prep_dma_buf(). It may be a good idea to combine the
|
2019-08-14 23:56:54 +00:00
|
|
|
* two functions eventually, but for now, it's here.
|
|
|
|
|
* RETURNS 0 - SUCCESS,
|
|
|
|
|
* 1 - Failed DMA map, retry.
|
|
|
|
|
* 2 - Invalid scsi cmd or prot-type. Do not rety.
|
2012-01-18 21:25:09 +00:00
|
|
|
**/
|
2008-12-05 03:40:02 +00:00
|
|
|
static int
|
2012-01-18 21:25:09 +00:00
|
|
|
lpfc_bg_scsi_prep_dma_buf_s3(struct lpfc_hba *phba,
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd)
|
2008-12-05 03:40:02 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct ulp_bde64 *bpl = (struct ulp_bde64 *)lpfc_cmd->dma_sgl;
|
2008-12-05 03:40:02 +00:00
|
|
|
IOCB_t *iocb_cmd = &lpfc_cmd->cur_iocbq.iocb;
|
|
|
|
|
uint32_t num_bde = 0;
|
|
|
|
|
int datasegcnt, protsegcnt, datadir = scsi_cmnd->sc_data_direction;
|
|
|
|
|
int prot_group_type = 0;
|
2013-04-18 00:18:07 +00:00
|
|
|
int fcpdl;
|
2019-08-14 23:56:54 +00:00
|
|
|
int ret = 1;
|
2018-11-30 00:09:40 +00:00
|
|
|
struct lpfc_vport *vport = phba->pport;
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Start the lpfc command prep by bumping the bpl beyond fcp_cmnd
|
|
|
|
|
* fcp_rsp regions to the first data bde entry
|
|
|
|
|
*/
|
|
|
|
|
bpl += 2;
|
|
|
|
|
if (scsi_sg_count(scsi_cmnd)) {
|
|
|
|
|
/*
|
2021-09-25 12:53:23 +00:00
|
|
|
* The driver stores the segment count returned from dma_map_sg
|
2008-12-05 03:40:02 +00:00
|
|
|
* because this a count of dma-mappings used to map the use_sg
|
|
|
|
|
* pages. They are not guaranteed to be the same for those
|
|
|
|
|
* architectures that implement an IOMMU.
|
|
|
|
|
*/
|
|
|
|
|
datasegcnt = dma_map_sg(&phba->pcidev->dev,
|
|
|
|
|
scsi_sglist(scsi_cmnd),
|
|
|
|
|
scsi_sg_count(scsi_cmnd), datadir);
|
|
|
|
|
if (unlikely(!datasegcnt))
|
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->seg_cnt = datasegcnt;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
/* First check if data segment count from SCSI Layer is good */
|
2019-08-14 23:56:54 +00:00
|
|
|
if (lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt) {
|
|
|
|
|
WARN_ON_ONCE(lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt);
|
|
|
|
|
ret = 2;
|
2013-04-18 00:16:15 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
prot_group_type = lpfc_prot_group_type(phba, scsi_cmnd);
|
|
|
|
|
|
|
|
|
|
switch (prot_group_type) {
|
|
|
|
|
case LPFC_PG_TYPE_NO_DIF:
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
/* Here we need to add a PDE5 and PDE6 to the count */
|
2019-08-14 23:56:54 +00:00
|
|
|
if ((lpfc_cmd->seg_cnt + 2) > phba->cfg_total_seg_cnt) {
|
|
|
|
|
ret = 2;
|
2013-04-18 00:16:15 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2013-04-18 00:16:15 +00:00
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
num_bde = lpfc_bg_setup_bpl(phba, scsi_cmnd, bpl,
|
|
|
|
|
datasegcnt);
|
2009-12-18 20:40:42 +00:00
|
|
|
/* we should have 2 or more entries in buffer list */
|
2019-08-14 23:56:54 +00:00
|
|
|
if (num_bde < 2) {
|
|
|
|
|
ret = 2;
|
2008-12-05 03:40:02 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
case LPFC_PG_TYPE_DIF_BUF:
|
2008-12-05 03:40:02 +00:00
|
|
|
/*
|
|
|
|
|
* This type indicates that protection buffers are
|
|
|
|
|
* passed to the driver, so that needs to be prepared
|
|
|
|
|
* for DMA
|
|
|
|
|
*/
|
|
|
|
|
protsegcnt = dma_map_sg(&phba->pcidev->dev,
|
|
|
|
|
scsi_prot_sglist(scsi_cmnd),
|
|
|
|
|
scsi_prot_sg_count(scsi_cmnd), datadir);
|
|
|
|
|
if (unlikely(!protsegcnt)) {
|
|
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->prot_seg_cnt = protsegcnt;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* There is a minimun of 4 BPLs used for every
|
|
|
|
|
* protection data segment.
|
|
|
|
|
*/
|
|
|
|
|
if ((lpfc_cmd->prot_seg_cnt * 4) >
|
2019-08-14 23:56:54 +00:00
|
|
|
(phba->cfg_total_seg_cnt - 2)) {
|
|
|
|
|
ret = 2;
|
2013-04-18 00:16:15 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
num_bde = lpfc_bg_setup_bpl_prot(phba, scsi_cmnd, bpl,
|
|
|
|
|
datasegcnt, protsegcnt);
|
2009-12-18 20:40:42 +00:00
|
|
|
/* we should have 3 or more entries in buffer list */
|
2013-04-18 00:16:15 +00:00
|
|
|
if ((num_bde < 3) ||
|
2019-08-14 23:56:54 +00:00
|
|
|
(num_bde > phba->cfg_total_seg_cnt)) {
|
|
|
|
|
ret = 2;
|
2008-12-05 03:40:02 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
break;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
case LPFC_PG_TYPE_INVALID:
|
|
|
|
|
default:
|
2013-04-18 00:16:15 +00:00
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2008-12-05 03:40:02 +00:00
|
|
|
"9022 Unexpected protection group %i\n",
|
|
|
|
|
prot_group_type);
|
2019-08-14 23:56:54 +00:00
|
|
|
return 2;
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Finish initializing those IOCB fields that are dependent on the
|
|
|
|
|
* scsi_cmnd request_buffer. Note that the bdeSize is explicitly
|
|
|
|
|
* reinitialized since all iocb memory resources are used many times
|
|
|
|
|
* for transmit, receive, and continuation bpl's.
|
|
|
|
|
*/
|
|
|
|
|
iocb_cmd->un.fcpi64.bdl.bdeSize = (2 * sizeof(struct ulp_bde64));
|
|
|
|
|
iocb_cmd->un.fcpi64.bdl.bdeSize += (num_bde * sizeof(struct ulp_bde64));
|
|
|
|
|
iocb_cmd->ulpBdeCount = 1;
|
|
|
|
|
iocb_cmd->ulpLe = 1;
|
|
|
|
|
|
2013-04-18 00:18:07 +00:00
|
|
|
fcpdl = lpfc_bg_scsi_adjust_dl(phba, lpfc_cmd);
|
2008-12-05 03:40:02 +00:00
|
|
|
fcp_cmnd->fcpDl = be32_to_cpu(fcpdl);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Due to difference in data length between DIF/non-DIF paths,
|
|
|
|
|
* we need to set word 4 of IOCB here
|
|
|
|
|
*/
|
|
|
|
|
iocb_cmd->un.fcpi.fcpi_parm = fcpdl;
|
|
|
|
|
|
2018-11-30 00:09:40 +00:00
|
|
|
/*
|
|
|
|
|
* For First burst, we may need to adjust the initial transfer
|
|
|
|
|
* length for DIF
|
|
|
|
|
*/
|
|
|
|
|
if (iocb_cmd->un.fcpi.fcpi_XRdy &&
|
|
|
|
|
(fcpdl < vport->cfg_first_burst_size))
|
|
|
|
|
iocb_cmd->un.fcpi.fcpi_XRdy = fcpdl;
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
2008-12-05 03:40:02 +00:00
|
|
|
err:
|
2013-04-18 00:16:15 +00:00
|
|
|
if (lpfc_cmd->seg_cnt)
|
|
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
if (lpfc_cmd->prot_seg_cnt)
|
|
|
|
|
dma_unmap_sg(&phba->pcidev->dev, scsi_prot_sglist(scsi_cmnd),
|
|
|
|
|
scsi_prot_sg_count(scsi_cmnd),
|
|
|
|
|
scsi_cmnd->sc_data_direction);
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2013-04-18 00:16:15 +00:00
|
|
|
"9023 Cannot setup S/G List for HBA"
|
|
|
|
|
"IO segs %d/%d BPL %d SCSI %d: %d %d\n",
|
|
|
|
|
lpfc_cmd->seg_cnt, lpfc_cmd->prot_seg_cnt,
|
|
|
|
|
phba->cfg_total_seg_cnt, phba->cfg_sg_seg_cnt,
|
2008-12-05 03:40:02 +00:00
|
|
|
prot_group_type, num_bde);
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
lpfc_cmd->prot_seg_cnt = 0;
|
2019-08-14 23:56:54 +00:00
|
|
|
return ret;
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
2013-04-18 00:14:49 +00:00
|
|
|
/*
|
|
|
|
|
* This function calcuates the T10 DIF guard tag
|
|
|
|
|
* on the specified data using a CRC algorithmn
|
|
|
|
|
* using crc_t10dif.
|
|
|
|
|
*/
|
2014-09-03 16:55:36 +00:00
|
|
|
static uint16_t
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_bg_crc(uint8_t *data, int count)
|
|
|
|
|
{
|
|
|
|
|
uint16_t crc = 0;
|
|
|
|
|
uint16_t x;
|
|
|
|
|
|
|
|
|
|
crc = crc_t10dif(data, count);
|
|
|
|
|
x = cpu_to_be16(crc);
|
|
|
|
|
return x;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* This function calcuates the T10 DIF guard tag
|
|
|
|
|
* on the specified data using a CSUM algorithmn
|
|
|
|
|
* using ip_compute_csum.
|
|
|
|
|
*/
|
2014-09-03 16:55:36 +00:00
|
|
|
static uint16_t
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_bg_csum(uint8_t *data, int count)
|
|
|
|
|
{
|
|
|
|
|
uint16_t ret;
|
|
|
|
|
|
|
|
|
|
ret = ip_compute_csum(data, count);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* This function examines the protection data to try to determine
|
|
|
|
|
* what type of T10-DIF error occurred.
|
|
|
|
|
*/
|
2014-09-03 16:55:36 +00:00
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_calc_bg_err(struct lpfc_hba *phba, struct lpfc_io_buf *lpfc_cmd)
|
2013-04-18 00:14:49 +00:00
|
|
|
{
|
|
|
|
|
struct scatterlist *sgpe; /* s/g prot entry */
|
|
|
|
|
struct scatterlist *sgde; /* s/g data entry */
|
|
|
|
|
struct scsi_cmnd *cmd = lpfc_cmd->pCmd;
|
|
|
|
|
struct scsi_dif_tuple *src = NULL;
|
|
|
|
|
uint8_t *data_src = NULL;
|
2015-08-31 20:48:14 +00:00
|
|
|
uint16_t guard_tag;
|
2013-04-18 00:14:49 +00:00
|
|
|
uint16_t start_app_tag, app_tag;
|
|
|
|
|
uint32_t start_ref_tag, ref_tag;
|
|
|
|
|
int prot, protsegcnt;
|
|
|
|
|
int err_type, len, data_len;
|
|
|
|
|
int chk_ref, chk_app, chk_guard;
|
|
|
|
|
uint16_t sum;
|
|
|
|
|
unsigned blksize;
|
|
|
|
|
|
|
|
|
|
err_type = BGS_GUARD_ERR_MASK;
|
|
|
|
|
sum = 0;
|
|
|
|
|
guard_tag = 0;
|
|
|
|
|
|
|
|
|
|
/* First check to see if there is protection data to examine */
|
|
|
|
|
prot = scsi_get_prot_op(cmd);
|
|
|
|
|
if ((prot == SCSI_PROT_READ_STRIP) ||
|
|
|
|
|
(prot == SCSI_PROT_WRITE_INSERT) ||
|
|
|
|
|
(prot == SCSI_PROT_NORMAL))
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
/* Currently the driver just supports ref_tag and guard_tag checking */
|
|
|
|
|
chk_ref = 1;
|
|
|
|
|
chk_app = 0;
|
|
|
|
|
chk_guard = 0;
|
|
|
|
|
|
|
|
|
|
/* Setup a ptr to the protection data provided by the SCSI host */
|
|
|
|
|
sgpe = scsi_prot_sglist(cmd);
|
|
|
|
|
protsegcnt = lpfc_cmd->prot_seg_cnt;
|
|
|
|
|
|
|
|
|
|
if (sgpe && protsegcnt) {
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We will only try to verify guard tag if the segment
|
|
|
|
|
* data length is a multiple of the blksize.
|
|
|
|
|
*/
|
|
|
|
|
sgde = scsi_sglist(cmd);
|
2021-08-17 02:50:14 +00:00
|
|
|
blksize = scsi_prot_interval(cmd);
|
2013-04-18 00:14:49 +00:00
|
|
|
data_src = (uint8_t *)sg_virt(sgde);
|
|
|
|
|
data_len = sgde->length;
|
|
|
|
|
if ((data_len & (blksize - 1)) == 0)
|
|
|
|
|
chk_guard = 1;
|
|
|
|
|
|
2013-05-31 21:04:10 +00:00
|
|
|
src = (struct scsi_dif_tuple *)sg_virt(sgpe);
|
2021-08-17 02:50:14 +00:00
|
|
|
start_ref_tag = scsi_prot_ref_tag(cmd);
|
2013-04-18 00:14:49 +00:00
|
|
|
start_app_tag = src->app_tag;
|
|
|
|
|
len = sgpe->length;
|
|
|
|
|
while (src && protsegcnt) {
|
|
|
|
|
while (len) {
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* First check to see if a protection data
|
|
|
|
|
* check is valid
|
|
|
|
|
*/
|
2017-06-29 18:31:12 +00:00
|
|
|
if ((src->ref_tag == T10_PI_REF_ESCAPE) ||
|
|
|
|
|
(src->app_tag == T10_PI_APP_ESCAPE)) {
|
2013-04-18 00:14:49 +00:00
|
|
|
start_ref_tag++;
|
|
|
|
|
goto skipit;
|
|
|
|
|
}
|
|
|
|
|
|
2013-05-31 21:03:39 +00:00
|
|
|
/* First Guard Tag checking */
|
2013-04-18 00:14:49 +00:00
|
|
|
if (chk_guard) {
|
|
|
|
|
guard_tag = src->guard_tag;
|
2021-08-17 02:50:14 +00:00
|
|
|
if (cmd->prot_flags
|
|
|
|
|
& SCSI_PROT_IP_CHECKSUM)
|
2013-04-18 00:14:49 +00:00
|
|
|
sum = lpfc_bg_csum(data_src,
|
|
|
|
|
blksize);
|
|
|
|
|
else
|
|
|
|
|
sum = lpfc_bg_crc(data_src,
|
|
|
|
|
blksize);
|
|
|
|
|
if ((guard_tag != sum)) {
|
|
|
|
|
err_type = BGS_GUARD_ERR_MASK;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
}
|
2013-05-31 21:03:39 +00:00
|
|
|
|
|
|
|
|
/* Reference Tag checking */
|
|
|
|
|
ref_tag = be32_to_cpu(src->ref_tag);
|
|
|
|
|
if (chk_ref && (ref_tag != start_ref_tag)) {
|
|
|
|
|
err_type = BGS_REFTAG_ERR_MASK;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
start_ref_tag++;
|
|
|
|
|
|
|
|
|
|
/* App Tag checking */
|
|
|
|
|
app_tag = src->app_tag;
|
|
|
|
|
if (chk_app && (app_tag != start_app_tag)) {
|
|
|
|
|
err_type = BGS_APPTAG_ERR_MASK;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
2013-04-18 00:14:49 +00:00
|
|
|
skipit:
|
|
|
|
|
len -= sizeof(struct scsi_dif_tuple);
|
|
|
|
|
if (len < 0)
|
|
|
|
|
len = 0;
|
|
|
|
|
src++;
|
|
|
|
|
|
|
|
|
|
data_src += blksize;
|
|
|
|
|
data_len -= blksize;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Are we at the end of the Data segment?
|
|
|
|
|
* The data segment is only used for Guard
|
|
|
|
|
* tag checking.
|
|
|
|
|
*/
|
|
|
|
|
if (chk_guard && (data_len == 0)) {
|
|
|
|
|
chk_guard = 0;
|
|
|
|
|
sgde = sg_next(sgde);
|
|
|
|
|
if (!sgde)
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
data_src = (uint8_t *)sg_virt(sgde);
|
|
|
|
|
data_len = sgde->length;
|
|
|
|
|
if ((data_len & (blksize - 1)) == 0)
|
|
|
|
|
chk_guard = 1;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Goto the next Protection data segment */
|
|
|
|
|
sgpe = sg_next(sgpe);
|
|
|
|
|
if (sgpe) {
|
|
|
|
|
src = (struct scsi_dif_tuple *)sg_virt(sgpe);
|
|
|
|
|
len = sgpe->length;
|
|
|
|
|
} else {
|
|
|
|
|
src = NULL;
|
|
|
|
|
}
|
|
|
|
|
protsegcnt--;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
out:
|
|
|
|
|
if (err_type == BGS_GUARD_ERR_MASK) {
|
2021-04-27 08:30:13 +00:00
|
|
|
scsi_build_sense(cmd, 1, ILLEGAL_REQUEST, 0x10, 0x1);
|
|
|
|
|
set_host_byte(cmd, DID_ABORT);
|
2013-04-18 00:14:49 +00:00
|
|
|
phba->bg_guard_err_cnt++;
|
|
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9069 BLKGRD: reftag %x grd_tag err %x != %x\n",
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
2013-04-18 00:14:49 +00:00
|
|
|
sum, guard_tag);
|
|
|
|
|
|
|
|
|
|
} else if (err_type == BGS_REFTAG_ERR_MASK) {
|
2021-04-27 08:30:13 +00:00
|
|
|
scsi_build_sense(cmd, 1, ILLEGAL_REQUEST, 0x10, 0x3);
|
|
|
|
|
set_host_byte(cmd, DID_ABORT);
|
2013-04-18 00:14:49 +00:00
|
|
|
|
|
|
|
|
phba->bg_reftag_err_cnt++;
|
|
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9066 BLKGRD: reftag %x ref_tag err %x != %x\n",
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
2013-04-18 00:14:49 +00:00
|
|
|
ref_tag, start_ref_tag);
|
|
|
|
|
|
|
|
|
|
} else if (err_type == BGS_APPTAG_ERR_MASK) {
|
2021-04-27 08:30:13 +00:00
|
|
|
scsi_build_sense(cmd, 1, ILLEGAL_REQUEST, 0x10, 0x2);
|
|
|
|
|
set_host_byte(cmd, DID_ABORT);
|
2013-04-18 00:14:49 +00:00
|
|
|
|
|
|
|
|
phba->bg_apptag_err_cnt++;
|
|
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9041 BLKGRD: reftag %x app_tag err %x != %x\n",
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
2013-04-18 00:14:49 +00:00
|
|
|
app_tag, start_app_tag);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-15 19:26:43 +00:00
|
|
|
/*
|
|
|
|
|
* This function checks for BlockGuard errors detected by
|
|
|
|
|
* the HBA. In case of errors, the ASC/ASCQ fields in the
|
|
|
|
|
* sense buffer will be set accordingly, paired with
|
|
|
|
|
* ILLEGAL_REQUEST to signal to the kernel that the HBA
|
|
|
|
|
* detected corruption.
|
|
|
|
|
*
|
|
|
|
|
* Returns:
|
|
|
|
|
* 0 - No error found
|
|
|
|
|
* 1 - BlockGuard error found
|
|
|
|
|
* -1 - Internal error (bad profile, ...etc)
|
|
|
|
|
*/
|
|
|
|
|
static int
|
2022-02-25 02:23:04 +00:00
|
|
|
lpfc_parse_bg_err(struct lpfc_hba *phba, struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
struct lpfc_iocbq *pIocbOut)
|
2020-11-15 19:26:43 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *cmd = lpfc_cmd->pCmd;
|
2022-02-25 02:23:04 +00:00
|
|
|
struct sli3_bg_fields *bgf;
|
2020-11-15 19:26:43 +00:00
|
|
|
int ret = 0;
|
2022-02-25 02:23:04 +00:00
|
|
|
struct lpfc_wcqe_complete *wcqe;
|
|
|
|
|
u32 status;
|
2020-11-15 19:26:43 +00:00
|
|
|
u32 bghm = 0;
|
|
|
|
|
u32 bgstat = 0;
|
|
|
|
|
u64 failing_sector = 0;
|
|
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4) {
|
|
|
|
|
wcqe = &pIocbOut->wcqe_cmpl;
|
|
|
|
|
status = bf_get(lpfc_wcqe_c_status, wcqe);
|
2020-11-15 19:26:43 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
if (status == CQE_STATUS_DI_ERROR) {
|
|
|
|
|
/* Guard Check failed */
|
|
|
|
|
if (bf_get(lpfc_wcqe_c_bg_ge, wcqe))
|
|
|
|
|
bgstat |= BGS_GUARD_ERR_MASK;
|
2020-11-15 19:26:43 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
/* AppTag Check failed */
|
|
|
|
|
if (bf_get(lpfc_wcqe_c_bg_ae, wcqe))
|
|
|
|
|
bgstat |= BGS_APPTAG_ERR_MASK;
|
2020-11-15 19:26:43 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
/* RefTag Check failed */
|
|
|
|
|
if (bf_get(lpfc_wcqe_c_bg_re, wcqe))
|
|
|
|
|
bgstat |= BGS_REFTAG_ERR_MASK;
|
2020-11-15 19:26:43 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
/* Check to see if there was any good data before the
|
|
|
|
|
* error
|
|
|
|
|
*/
|
|
|
|
|
if (bf_get(lpfc_wcqe_c_bg_tdpv, wcqe)) {
|
|
|
|
|
bgstat |= BGS_HI_WATER_MARK_PRESENT_MASK;
|
|
|
|
|
bghm = wcqe->total_data_placed;
|
|
|
|
|
}
|
2020-11-15 19:26:43 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
/*
|
|
|
|
|
* Set ALL the error bits to indicate we don't know what
|
|
|
|
|
* type of error it is.
|
|
|
|
|
*/
|
|
|
|
|
if (!bgstat)
|
|
|
|
|
bgstat |= (BGS_REFTAG_ERR_MASK |
|
|
|
|
|
BGS_APPTAG_ERR_MASK |
|
|
|
|
|
BGS_GUARD_ERR_MASK);
|
2020-11-15 19:26:43 +00:00
|
|
|
}
|
|
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
} else {
|
|
|
|
|
bgf = &pIocbOut->iocb.unsli3.sli3_bg;
|
|
|
|
|
bghm = bgf->bghm;
|
|
|
|
|
bgstat = bgf->bgstat;
|
2020-11-15 19:26:43 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
if (lpfc_bgs_get_invalid_prof(bgstat)) {
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_ERROR << 16;
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9072 BLKGRD: Invalid BG Profile in cmd "
|
|
|
|
|
"0x%x reftag 0x%x blk cnt 0x%x "
|
2013-04-18 00:14:49 +00:00
|
|
|
"bgstat=x%x bghm=x%x\n", cmd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
|
|
|
|
scsi_logical_block_count(cmd), bgstat, bghm);
|
2008-12-05 03:40:02 +00:00
|
|
|
ret = (-1);
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (lpfc_bgs_get_uninit_dif_block(bgstat)) {
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_ERROR << 16;
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9073 BLKGRD: Invalid BG PDIF Block in cmd "
|
|
|
|
|
"0x%x reftag 0x%x blk cnt 0x%x "
|
2013-04-18 00:14:49 +00:00
|
|
|
"bgstat=x%x bghm=x%x\n", cmd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
|
|
|
|
scsi_logical_block_count(cmd), bgstat, bghm);
|
2008-12-05 03:40:02 +00:00
|
|
|
ret = (-1);
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (lpfc_bgs_get_guard_err(bgstat)) {
|
|
|
|
|
ret = 1;
|
2021-04-27 08:30:13 +00:00
|
|
|
scsi_build_sense(cmd, 1, ILLEGAL_REQUEST, 0x10, 0x1);
|
|
|
|
|
set_host_byte(cmd, DID_ABORT);
|
2008-12-05 03:40:02 +00:00
|
|
|
phba->bg_guard_err_cnt++;
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9055 BLKGRD: Guard Tag error in cmd "
|
|
|
|
|
"0x%x reftag 0x%x blk cnt 0x%x "
|
2013-04-18 00:14:49 +00:00
|
|
|
"bgstat=x%x bghm=x%x\n", cmd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
|
|
|
|
scsi_logical_block_count(cmd), bgstat, bghm);
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (lpfc_bgs_get_reftag_err(bgstat)) {
|
|
|
|
|
ret = 1;
|
2021-04-27 08:30:13 +00:00
|
|
|
scsi_build_sense(cmd, 1, ILLEGAL_REQUEST, 0x10, 0x3);
|
|
|
|
|
set_host_byte(cmd, DID_ABORT);
|
2008-12-05 03:40:02 +00:00
|
|
|
phba->bg_reftag_err_cnt++;
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9056 BLKGRD: Ref Tag error in cmd "
|
|
|
|
|
"0x%x reftag 0x%x blk cnt 0x%x "
|
2013-04-18 00:14:49 +00:00
|
|
|
"bgstat=x%x bghm=x%x\n", cmd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
|
|
|
|
scsi_logical_block_count(cmd), bgstat, bghm);
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (lpfc_bgs_get_apptag_err(bgstat)) {
|
|
|
|
|
ret = 1;
|
2021-04-27 08:30:13 +00:00
|
|
|
scsi_build_sense(cmd, 1, ILLEGAL_REQUEST, 0x10, 0x2);
|
|
|
|
|
set_host_byte(cmd, DID_ABORT);
|
2008-12-05 03:40:02 +00:00
|
|
|
phba->bg_apptag_err_cnt++;
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9061 BLKGRD: App Tag error in cmd "
|
|
|
|
|
"0x%x reftag 0x%x blk cnt 0x%x "
|
2013-04-18 00:14:49 +00:00
|
|
|
"bgstat=x%x bghm=x%x\n", cmd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
|
|
|
|
scsi_logical_block_count(cmd), bgstat, bghm);
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (lpfc_bgs_get_hi_water_mark_present(bgstat)) {
|
|
|
|
|
/*
|
|
|
|
|
* setup sense data descriptor 0 per SPC-4 as an information
|
2011-07-22 22:36:25 +00:00
|
|
|
* field, and put the failing LBA in it.
|
|
|
|
|
* This code assumes there was also a guard/app/ref tag error
|
|
|
|
|
* indication.
|
2008-12-05 03:40:02 +00:00
|
|
|
*/
|
2011-07-22 22:36:25 +00:00
|
|
|
cmd->sense_buffer[7] = 0xc; /* Additional sense length */
|
|
|
|
|
cmd->sense_buffer[8] = 0; /* Information descriptor type */
|
|
|
|
|
cmd->sense_buffer[9] = 0xa; /* Additional descriptor length */
|
|
|
|
|
cmd->sense_buffer[10] = 0x80; /* Validity bit */
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
/* bghm is a "on the wire" FC frame based count */
|
|
|
|
|
switch (scsi_get_prot_op(cmd)) {
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
|
|
|
|
bghm /= cmd->device->sector_size;
|
|
|
|
|
break;
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
|
|
|
|
bghm /= (cmd->device->sector_size +
|
|
|
|
|
sizeof(struct scsi_dif_tuple));
|
|
|
|
|
break;
|
|
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
|
|
|
|
|
failing_sector = scsi_get_lba(cmd);
|
|
|
|
|
failing_sector += bghm;
|
|
|
|
|
|
2011-07-22 22:36:25 +00:00
|
|
|
/* Descriptor Information */
|
|
|
|
|
put_unaligned_be64(failing_sector, &cmd->sense_buffer[12]);
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!ret) {
|
|
|
|
|
/* No error was reported - problem in FW? */
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_log(phba, KERN_WARNING, LOG_FCP | LOG_BG,
|
2021-03-01 17:18:02 +00:00
|
|
|
"9057 BLKGRD: Unknown error in cmd "
|
|
|
|
|
"0x%x reftag 0x%x blk cnt 0x%x "
|
2013-04-18 00:14:49 +00:00
|
|
|
"bgstat=x%x bghm=x%x\n", cmd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmd),
|
|
|
|
|
scsi_logical_block_count(cmd), bgstat, bghm);
|
2013-04-18 00:14:49 +00:00
|
|
|
|
2021-08-17 02:50:14 +00:00
|
|
|
/* Calculate what type of error it was */
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_calc_bg_err(phba, lpfc_cmd);
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
out:
|
|
|
|
|
return ret;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_dma_buf_s4 - DMA mapping for scsi buffer to SLI4 IF spec
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be mapped.
|
|
|
|
|
*
|
|
|
|
|
* This routine does the pci dma mapping for scatter-gather list of scsi cmnd
|
|
|
|
|
* field of @lpfc_cmd for device with SLI-4 interface spec.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
2019-08-14 23:56:54 +00:00
|
|
|
* 2 - Error - Do not retry
|
|
|
|
|
* 1 - Error - Retry
|
2010-04-06 18:49:53 +00:00
|
|
|
* 0 - Success
|
2009-05-22 18:51:39 +00:00
|
|
|
**/
|
|
|
|
|
static int
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_scsi_prep_dma_buf_s4(struct lpfc_hba *phba, struct lpfc_io_buf *lpfc_cmd)
|
2009-05-22 18:51:39 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct scatterlist *sgel = NULL;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
2019-01-28 19:14:19 +00:00
|
|
|
struct sli4_sge *sgl = (struct sli4_sge *)lpfc_cmd->dma_sgl;
|
2011-02-16 17:39:24 +00:00
|
|
|
struct sli4_sge *first_data_sgl;
|
2020-11-15 19:26:42 +00:00
|
|
|
struct lpfc_iocbq *pwqeq = &lpfc_cmd->cur_iocbq;
|
|
|
|
|
struct lpfc_vport *vport = phba->pport;
|
|
|
|
|
union lpfc_wqe128 *wqe = &pwqeq->wqe;
|
2009-05-22 18:51:39 +00:00
|
|
|
dma_addr_t physaddr;
|
|
|
|
|
uint32_t dma_len;
|
|
|
|
|
uint32_t dma_offset = 0;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
int nseg, i, j;
|
2011-02-16 17:39:24 +00:00
|
|
|
struct ulp_bde64 *bde;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
bool lsp_just_set = false;
|
|
|
|
|
struct sli4_hybrid_sgl *sgl_xtra = NULL;
|
2009-05-22 18:51:39 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* There are three possibilities here - use scatter-gather segment, use
|
|
|
|
|
* the single mapping, or neither. Start the lpfc command prep by
|
|
|
|
|
* bumping the bpl beyond the fcp_cmnd and fcp_rsp regions to the first
|
|
|
|
|
* data bde entry.
|
|
|
|
|
*/
|
|
|
|
|
if (scsi_sg_count(scsi_cmnd)) {
|
|
|
|
|
/*
|
2021-09-25 12:53:23 +00:00
|
|
|
* The driver stores the segment count returned from dma_map_sg
|
2009-05-22 18:51:39 +00:00
|
|
|
* because this a count of dma-mappings used to map the use_sg
|
|
|
|
|
* pages. They are not guaranteed to be the same for those
|
|
|
|
|
* architectures that implement an IOMMU.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
nseg = scsi_dma_map(scsi_cmnd);
|
2015-05-22 14:42:39 +00:00
|
|
|
if (unlikely(nseg <= 0))
|
2009-05-22 18:51:39 +00:00
|
|
|
return 1;
|
|
|
|
|
sgl += 1;
|
|
|
|
|
/* clear the last flag in the fcp_rsp map entry */
|
|
|
|
|
sgl->word2 = le32_to_cpu(sgl->word2);
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 0);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl += 1;
|
2011-02-16 17:39:24 +00:00
|
|
|
first_data_sgl = sgl;
|
2009-05-22 18:51:39 +00:00
|
|
|
lpfc_cmd->seg_cnt = nseg;
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if (!phba->cfg_xpsgl &&
|
|
|
|
|
lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"9074 BLKGRD:"
|
|
|
|
|
" %s: Too many sg segments from "
|
|
|
|
|
"dma_map_sg. Config %d, seg_cnt %d\n",
|
|
|
|
|
__func__, phba->cfg_sg_seg_cnt,
|
|
|
|
|
lpfc_cmd->seg_cnt);
|
2019-08-14 23:56:54 +00:00
|
|
|
WARN_ON_ONCE(lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt);
|
2013-04-18 00:16:15 +00:00
|
|
|
lpfc_cmd->seg_cnt = 0;
|
2009-05-22 18:51:39 +00:00
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
2019-08-14 23:56:54 +00:00
|
|
|
return 2;
|
2009-05-22 18:51:39 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The driver established a maximum scatter-gather segment count
|
|
|
|
|
* during probe that limits the number of sg elements in any
|
|
|
|
|
* single scsi command. Just run through the seg_cnt and format
|
|
|
|
|
* the sge's.
|
|
|
|
|
* When using SLI-3 the driver will try to fit all the BDEs into
|
|
|
|
|
* the IOCB. If it can't then the BDEs get added to a BPL as it
|
|
|
|
|
* does for SLI-2 mode.
|
|
|
|
|
*/
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
|
|
|
|
|
/* for tracking segment boundaries */
|
|
|
|
|
sgel = scsi_sglist(scsi_cmnd);
|
|
|
|
|
j = 2;
|
|
|
|
|
for (i = 0; i < nseg; i++) {
|
|
|
|
|
sgl->word2 = 0;
|
2021-09-10 23:31:58 +00:00
|
|
|
if (nseg == 1) {
|
2009-05-22 18:51:39 +00:00
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
bf_set(lpfc_sli4_sge_type, sgl,
|
|
|
|
|
LPFC_SGE_TYPE_DATA);
|
|
|
|
|
} else {
|
2009-05-22 18:51:39 +00:00
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 0);
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
|
|
|
|
|
/* do we need to expand the segment */
|
|
|
|
|
if (!lsp_just_set &&
|
|
|
|
|
!((j + 1) % phba->border_sge_num) &&
|
|
|
|
|
((nseg - 1) != i)) {
|
|
|
|
|
/* set LSP type */
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl,
|
|
|
|
|
LPFC_SGE_TYPE_LSP);
|
|
|
|
|
|
|
|
|
|
sgl_xtra = lpfc_get_sgl_per_hdwq(
|
|
|
|
|
phba, lpfc_cmd);
|
|
|
|
|
|
|
|
|
|
if (unlikely(!sgl_xtra)) {
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(
|
|
|
|
|
sgl_xtra->dma_phys_sgl));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(
|
|
|
|
|
sgl_xtra->dma_phys_sgl));
|
|
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
bf_set(lpfc_sli4_sge_type, sgl,
|
|
|
|
|
LPFC_SGE_TYPE_DATA);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!(bf_get(lpfc_sli4_sge_type, sgl) &
|
|
|
|
|
LPFC_SGE_TYPE_LSP)) {
|
|
|
|
|
if ((nseg - 1) == i)
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
|
|
|
|
|
|
|
|
|
physaddr = sg_dma_address(sgel);
|
|
|
|
|
dma_len = sg_dma_len(sgel);
|
|
|
|
|
sgl->addr_lo = cpu_to_le32(putPaddrLow(
|
|
|
|
|
physaddr));
|
|
|
|
|
sgl->addr_hi = cpu_to_le32(putPaddrHigh(
|
|
|
|
|
physaddr));
|
|
|
|
|
|
|
|
|
|
bf_set(lpfc_sli4_sge_offset, sgl, dma_offset);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(dma_len);
|
|
|
|
|
|
|
|
|
|
dma_offset += dma_len;
|
|
|
|
|
sgel = sg_next(sgel);
|
|
|
|
|
|
|
|
|
|
sgl++;
|
|
|
|
|
lsp_just_set = false;
|
|
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
sgl->sge_len = cpu_to_le32(
|
|
|
|
|
phba->cfg_sg_dma_buf_size);
|
|
|
|
|
|
|
|
|
|
sgl = (struct sli4_sge *)sgl_xtra->dma_sgl;
|
|
|
|
|
i = i - 1;
|
|
|
|
|
|
|
|
|
|
lsp_just_set = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
j++;
|
2009-05-22 18:51:39 +00:00
|
|
|
}
|
2021-09-10 23:31:58 +00:00
|
|
|
|
|
|
|
|
/* PBDE support for first data SGE only.
|
|
|
|
|
* For FCoE, we key off Performance Hints.
|
|
|
|
|
* For FC, we key off lpfc_enable_pbde.
|
2018-02-22 16:18:48 +00:00
|
|
|
*/
|
2021-09-10 23:31:58 +00:00
|
|
|
if (nseg == 1 &&
|
|
|
|
|
((phba->sli3_options & LPFC_SLI4_PERFH_ENABLED) ||
|
|
|
|
|
phba->cfg_enable_pbde)) {
|
|
|
|
|
/* Words 13-15 */
|
2011-02-16 17:39:24 +00:00
|
|
|
bde = (struct ulp_bde64 *)
|
2020-11-15 19:26:42 +00:00
|
|
|
&wqe->words[13];
|
2011-02-16 17:39:24 +00:00
|
|
|
bde->addrLow = first_data_sgl->addr_lo;
|
|
|
|
|
bde->addrHigh = first_data_sgl->addr_hi;
|
|
|
|
|
bde->tus.f.bdeSize =
|
|
|
|
|
le32_to_cpu(first_data_sgl->sge_len);
|
|
|
|
|
bde->tus.f.bdeFlags = BUFF_TYPE_BDE_64;
|
|
|
|
|
bde->tus.w = cpu_to_le32(bde->tus.w);
|
2020-11-15 19:26:42 +00:00
|
|
|
|
2021-09-10 23:31:58 +00:00
|
|
|
/* Word 11 - set PBDE bit */
|
|
|
|
|
bf_set(wqe_pbde, &wqe->generic.wqe_com, 1);
|
2020-11-15 19:26:42 +00:00
|
|
|
} else {
|
|
|
|
|
memset(&wqe->words[13], 0, (sizeof(uint32_t) * 3));
|
2021-09-10 23:31:58 +00:00
|
|
|
/* Word 11 - PBDE bit disabled by default template */
|
2011-02-16 17:39:24 +00:00
|
|
|
}
|
2009-05-22 18:51:39 +00:00
|
|
|
} else {
|
|
|
|
|
sgl += 1;
|
2021-09-10 23:31:58 +00:00
|
|
|
/* set the last flag in the fcp_rsp map entry */
|
2009-05-22 18:51:39 +00:00
|
|
|
sgl->word2 = le32_to_cpu(sgl->word2);
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 1);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
2018-06-26 15:24:26 +00:00
|
|
|
|
|
|
|
|
if ((phba->sli3_options & LPFC_SLI4_PERFH_ENABLED) ||
|
|
|
|
|
phba->cfg_enable_pbde) {
|
|
|
|
|
bde = (struct ulp_bde64 *)
|
2020-11-15 19:26:42 +00:00
|
|
|
&wqe->words[13];
|
2018-06-26 15:24:26 +00:00
|
|
|
memset(bde, 0, (sizeof(uint32_t) * 3));
|
|
|
|
|
}
|
2009-05-22 18:51:39 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Finish initializing those IOCB fields that are dependent on the
|
|
|
|
|
* scsi_cmnd request_buffer. Note that for SLI-2 the bdeSize is
|
|
|
|
|
* explicitly reinitialized.
|
|
|
|
|
* all iocb memory resources are reused.
|
|
|
|
|
*/
|
|
|
|
|
fcp_cmnd->fcpDl = cpu_to_be32(scsi_bufflen(scsi_cmnd));
|
2020-11-15 19:26:42 +00:00
|
|
|
/* Set first-burst provided it was successfully negotiated */
|
|
|
|
|
if (!(phba->hba_flag & HBA_FCOE_MODE) &&
|
|
|
|
|
vport->cfg_first_burst_size &&
|
|
|
|
|
scsi_cmnd->sc_data_direction == DMA_TO_DEVICE) {
|
|
|
|
|
u32 init_len, total_len;
|
|
|
|
|
|
|
|
|
|
total_len = be32_to_cpu(fcp_cmnd->fcpDl);
|
|
|
|
|
init_len = min(total_len, vport->cfg_first_burst_size);
|
|
|
|
|
|
|
|
|
|
/* Word 4 & 5 */
|
|
|
|
|
wqe->fcp_iwrite.initial_xfer_len = init_len;
|
|
|
|
|
wqe->fcp_iwrite.total_xfer_len = total_len;
|
|
|
|
|
} else {
|
|
|
|
|
/* Word 4 */
|
|
|
|
|
wqe->fcp_iwrite.total_xfer_len =
|
|
|
|
|
be32_to_cpu(fcp_cmnd->fcpDl);
|
|
|
|
|
}
|
2014-02-20 14:56:45 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If the OAS driver feature is enabled and the lun is enabled for
|
|
|
|
|
* OAS, set the oas iocb related flags.
|
|
|
|
|
*/
|
2014-04-04 17:52:21 +00:00
|
|
|
if ((phba->cfg_fof) && ((struct lpfc_device_data *)
|
2016-07-06 19:36:05 +00:00
|
|
|
scsi_cmnd->device->hostdata)->oas_enabled) {
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag |= (LPFC_IO_OAS | LPFC_IO_FOF);
|
2016-07-06 19:36:05 +00:00
|
|
|
lpfc_cmd->cur_iocbq.priority = ((struct lpfc_device_data *)
|
|
|
|
|
scsi_cmnd->device->hostdata)->priority;
|
2020-11-15 19:26:42 +00:00
|
|
|
|
|
|
|
|
/* Word 10 */
|
|
|
|
|
bf_set(wqe_oas, &wqe->generic.wqe_com, 1);
|
|
|
|
|
bf_set(wqe_ccpe, &wqe->generic.wqe_com, 1);
|
|
|
|
|
|
|
|
|
|
if (lpfc_cmd->cur_iocbq.priority)
|
|
|
|
|
bf_set(wqe_ccp, &wqe->generic.wqe_com,
|
|
|
|
|
(lpfc_cmd->cur_iocbq.priority << 1));
|
|
|
|
|
else
|
|
|
|
|
bf_set(wqe_ccp, &wqe->generic.wqe_com,
|
|
|
|
|
(phba->cfg_XLanePriority << 1));
|
2016-07-06 19:36:05 +00:00
|
|
|
}
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_bg_scsi_prep_dma_buf_s4 - DMA mapping for scsi buffer to SLI4 IF spec
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be mapped.
|
|
|
|
|
*
|
|
|
|
|
* This is the protection/DIF aware version of
|
|
|
|
|
* lpfc_scsi_prep_dma_buf(). It may be a good idea to combine the
|
|
|
|
|
* two functions eventually, but for now, it's here
|
2019-08-14 23:56:54 +00:00
|
|
|
* Return codes:
|
|
|
|
|
* 2 - Error - Do not retry
|
|
|
|
|
* 1 - Error - Retry
|
|
|
|
|
* 0 - Success
|
2012-01-18 21:25:09 +00:00
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_bg_scsi_prep_dma_buf_s4(struct lpfc_hba *phba,
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd)
|
2012-01-18 21:25:09 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
2019-01-28 19:14:19 +00:00
|
|
|
struct sli4_sge *sgl = (struct sli4_sge *)(lpfc_cmd->dma_sgl);
|
2020-11-15 19:26:42 +00:00
|
|
|
struct lpfc_iocbq *pwqeq = &lpfc_cmd->cur_iocbq;
|
|
|
|
|
union lpfc_wqe128 *wqe = &pwqeq->wqe;
|
2013-04-18 00:16:15 +00:00
|
|
|
uint32_t num_sge = 0;
|
2012-01-18 21:25:09 +00:00
|
|
|
int datasegcnt, protsegcnt, datadir = scsi_cmnd->sc_data_direction;
|
|
|
|
|
int prot_group_type = 0;
|
|
|
|
|
int fcpdl;
|
2019-08-14 23:56:54 +00:00
|
|
|
int ret = 1;
|
2018-11-30 00:09:40 +00:00
|
|
|
struct lpfc_vport *vport = phba->pport;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Start the lpfc command prep by bumping the sgl beyond fcp_cmnd
|
2013-04-18 00:16:15 +00:00
|
|
|
* fcp_rsp regions to the first data sge entry
|
2012-01-18 21:25:09 +00:00
|
|
|
*/
|
|
|
|
|
if (scsi_sg_count(scsi_cmnd)) {
|
|
|
|
|
/*
|
2021-09-25 12:53:23 +00:00
|
|
|
* The driver stores the segment count returned from dma_map_sg
|
2012-01-18 21:25:09 +00:00
|
|
|
* because this a count of dma-mappings used to map the use_sg
|
|
|
|
|
* pages. They are not guaranteed to be the same for those
|
|
|
|
|
* architectures that implement an IOMMU.
|
|
|
|
|
*/
|
|
|
|
|
datasegcnt = dma_map_sg(&phba->pcidev->dev,
|
|
|
|
|
scsi_sglist(scsi_cmnd),
|
|
|
|
|
scsi_sg_count(scsi_cmnd), datadir);
|
|
|
|
|
if (unlikely(!datasegcnt))
|
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
|
|
sgl += 1;
|
|
|
|
|
/* clear the last flag in the fcp_rsp map entry */
|
|
|
|
|
sgl->word2 = le32_to_cpu(sgl->word2);
|
|
|
|
|
bf_set(lpfc_sli4_sge_last, sgl, 0);
|
|
|
|
|
sgl->word2 = cpu_to_le32(sgl->word2);
|
|
|
|
|
|
|
|
|
|
sgl += 1;
|
|
|
|
|
lpfc_cmd->seg_cnt = datasegcnt;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
/* First check if data segment count from SCSI Layer is good */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if (lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt &&
|
|
|
|
|
!phba->cfg_xpsgl) {
|
2019-08-14 23:56:54 +00:00
|
|
|
WARN_ON_ONCE(lpfc_cmd->seg_cnt > phba->cfg_sg_seg_cnt);
|
|
|
|
|
ret = 2;
|
2013-04-18 00:16:15 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
|
|
|
|
|
prot_group_type = lpfc_prot_group_type(phba, scsi_cmnd);
|
|
|
|
|
|
|
|
|
|
switch (prot_group_type) {
|
|
|
|
|
case LPFC_PG_TYPE_NO_DIF:
|
2013-04-18 00:16:15 +00:00
|
|
|
/* Here we need to add a DISEED to the count */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if (((lpfc_cmd->seg_cnt + 1) >
|
|
|
|
|
phba->cfg_total_seg_cnt) &&
|
|
|
|
|
!phba->cfg_xpsgl) {
|
2019-08-14 23:56:54 +00:00
|
|
|
ret = 2;
|
2013-04-18 00:16:15 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
num_sge = lpfc_bg_setup_sgl(phba, scsi_cmnd, sgl,
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
datasegcnt, lpfc_cmd);
|
2013-04-18 00:16:15 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* we should have 2 or more entries in buffer list */
|
2019-08-14 23:56:54 +00:00
|
|
|
if (num_sge < 2) {
|
|
|
|
|
ret = 2;
|
2012-01-18 21:25:09 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
case LPFC_PG_TYPE_DIF_BUF:
|
2012-01-18 21:25:09 +00:00
|
|
|
/*
|
|
|
|
|
* This type indicates that protection buffers are
|
|
|
|
|
* passed to the driver, so that needs to be prepared
|
|
|
|
|
* for DMA
|
|
|
|
|
*/
|
|
|
|
|
protsegcnt = dma_map_sg(&phba->pcidev->dev,
|
|
|
|
|
scsi_prot_sglist(scsi_cmnd),
|
|
|
|
|
scsi_prot_sg_count(scsi_cmnd), datadir);
|
|
|
|
|
if (unlikely(!protsegcnt)) {
|
|
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->prot_seg_cnt = protsegcnt;
|
2013-04-18 00:16:15 +00:00
|
|
|
/*
|
|
|
|
|
* There is a minimun of 3 SGEs used for every
|
|
|
|
|
* protection data segment.
|
|
|
|
|
*/
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if (((lpfc_cmd->prot_seg_cnt * 3) >
|
|
|
|
|
(phba->cfg_total_seg_cnt - 2)) &&
|
|
|
|
|
!phba->cfg_xpsgl) {
|
2019-08-14 23:56:54 +00:00
|
|
|
ret = 2;
|
2013-04-18 00:16:15 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2013-04-18 00:16:15 +00:00
|
|
|
num_sge = lpfc_bg_setup_sgl_prot(phba, scsi_cmnd, sgl,
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
datasegcnt, protsegcnt, lpfc_cmd);
|
2013-04-18 00:16:15 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/* we should have 3 or more entries in buffer list */
|
scsi: lpfc: Support dynamic unbounded SGL lists on G7 hardware.
Typical SLI-4 hardware supports up to 2 4KB pages to be registered per XRI
to contain the exchanges Scatter/Gather List. This caps the number of SGL
elements that can be in the SGL. There are not extensions to extend the
list out of the 2 pages.
The G7 hardware adds a SGE type that allows the SGL to be vectored to a
different scatter/gather list segment. And that segment can contain a SGE
to go to another segment and so on. The initial segment must still be
pre-registered for the XRI, but it can be a much smaller amount (256Bytes)
as it can now be dynamically grown. This much smaller allocation can
handle the SG list for most normal I/O, and the dynamic aspect allows it to
support many MB's if needed.
The implementation creates a pool which contains "segments" and which is
initially sized to hold the initial small segment per xri. If an I/O
requires additional segments, they are allocated from the pool. If the
pool has no more segments, the pool is grown based on what is now
needed. After the I/O completes, the additional segments are returned to
the pool for use by other I/Os. Once allocated, the additional segments are
not released under the assumption of "if needed once, it will be needed
again". Pools are kept on a per-hardware queue basis, which is typically
1:1 per cpu, but may be shared by multiple cpus.
The switch to the smaller initial allocation significantly reduces the
memory footprint of the driver (which only grows if large ios are
issued). Based on the several K of XRIs for the adapter, the 8KB->256B
reduction can conserve 32MBs or more.
It has been observed with per-cpu resource pools that allocating a resource
on CPU A, may be put back on CPU B. While the get routines are distributed
evenly, only a limited subset of CPUs may be handling the put routines.
This can put a strain on the lpfc_put_cmd_rsp_buf_per_cpu routine because
all the resources are being put on a limited subset of CPUs.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-08-14 23:57:09 +00:00
|
|
|
if (num_sge < 3 ||
|
|
|
|
|
(num_sge > phba->cfg_total_seg_cnt &&
|
|
|
|
|
!phba->cfg_xpsgl)) {
|
2019-08-14 23:56:54 +00:00
|
|
|
ret = 2;
|
2012-01-18 21:25:09 +00:00
|
|
|
goto err;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2012-01-18 21:25:09 +00:00
|
|
|
break;
|
2013-04-18 00:16:15 +00:00
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
case LPFC_PG_TYPE_INVALID:
|
|
|
|
|
default:
|
2013-04-18 00:16:15 +00:00
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2012-01-18 21:25:09 +00:00
|
|
|
"9083 Unexpected protection group %i\n",
|
|
|
|
|
prot_group_type);
|
2019-08-14 23:56:54 +00:00
|
|
|
return 2;
|
2012-01-18 21:25:09 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2012-10-31 18:44:49 +00:00
|
|
|
switch (scsi_get_prot_op(scsi_cmnd)) {
|
|
|
|
|
case SCSI_PROT_WRITE_STRIP:
|
|
|
|
|
case SCSI_PROT_READ_STRIP:
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag |= LPFC_IO_DIF_STRIP;
|
2012-10-31 18:44:49 +00:00
|
|
|
break;
|
|
|
|
|
case SCSI_PROT_WRITE_INSERT:
|
|
|
|
|
case SCSI_PROT_READ_INSERT:
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag |= LPFC_IO_DIF_INSERT;
|
2012-10-31 18:44:49 +00:00
|
|
|
break;
|
|
|
|
|
case SCSI_PROT_WRITE_PASS:
|
|
|
|
|
case SCSI_PROT_READ_PASS:
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag |= LPFC_IO_DIF_PASS;
|
2012-10-31 18:44:49 +00:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
fcpdl = lpfc_bg_scsi_adjust_dl(phba, lpfc_cmd);
|
|
|
|
|
fcp_cmnd->fcpDl = be32_to_cpu(fcpdl);
|
|
|
|
|
|
2020-11-15 19:26:42 +00:00
|
|
|
/* Set first-burst provided it was successfully negotiated */
|
|
|
|
|
if (!(phba->hba_flag & HBA_FCOE_MODE) &&
|
|
|
|
|
vport->cfg_first_burst_size &&
|
|
|
|
|
scsi_cmnd->sc_data_direction == DMA_TO_DEVICE) {
|
|
|
|
|
u32 init_len, total_len;
|
2012-01-18 21:25:09 +00:00
|
|
|
|
2020-11-15 19:26:42 +00:00
|
|
|
total_len = be32_to_cpu(fcp_cmnd->fcpDl);
|
|
|
|
|
init_len = min(total_len, vport->cfg_first_burst_size);
|
|
|
|
|
|
|
|
|
|
/* Word 4 & 5 */
|
|
|
|
|
wqe->fcp_iwrite.initial_xfer_len = init_len;
|
|
|
|
|
wqe->fcp_iwrite.total_xfer_len = total_len;
|
|
|
|
|
} else {
|
|
|
|
|
/* Word 4 */
|
|
|
|
|
wqe->fcp_iwrite.total_xfer_len =
|
|
|
|
|
be32_to_cpu(fcp_cmnd->fcpDl);
|
|
|
|
|
}
|
2018-11-30 00:09:40 +00:00
|
|
|
|
2014-09-03 16:57:30 +00:00
|
|
|
/*
|
|
|
|
|
* If the OAS driver feature is enabled and the lun is enabled for
|
|
|
|
|
* OAS, set the oas iocb related flags.
|
|
|
|
|
*/
|
|
|
|
|
if ((phba->cfg_fof) && ((struct lpfc_device_data *)
|
2020-11-15 19:26:42 +00:00
|
|
|
scsi_cmnd->device->hostdata)->oas_enabled) {
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag |= (LPFC_IO_OAS | LPFC_IO_FOF);
|
2014-09-03 16:57:30 +00:00
|
|
|
|
2020-11-15 19:26:42 +00:00
|
|
|
/* Word 10 */
|
|
|
|
|
bf_set(wqe_oas, &wqe->generic.wqe_com, 1);
|
|
|
|
|
bf_set(wqe_ccpe, &wqe->generic.wqe_com, 1);
|
|
|
|
|
bf_set(wqe_ccp, &wqe->generic.wqe_com,
|
|
|
|
|
(phba->cfg_XLanePriority << 1));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Word 7. DIF Flags */
|
2022-02-25 02:22:52 +00:00
|
|
|
if (lpfc_cmd->cur_iocbq.cmd_flag & LPFC_IO_DIF_PASS)
|
2020-11-15 19:26:42 +00:00
|
|
|
bf_set(wqe_dif, &wqe->generic.wqe_com, LPFC_WQE_DIF_PASSTHRU);
|
2022-02-25 02:22:52 +00:00
|
|
|
else if (lpfc_cmd->cur_iocbq.cmd_flag & LPFC_IO_DIF_STRIP)
|
2020-11-15 19:26:42 +00:00
|
|
|
bf_set(wqe_dif, &wqe->generic.wqe_com, LPFC_WQE_DIF_STRIP);
|
2022-02-25 02:22:52 +00:00
|
|
|
else if (lpfc_cmd->cur_iocbq.cmd_flag & LPFC_IO_DIF_INSERT)
|
2020-11-15 19:26:42 +00:00
|
|
|
bf_set(wqe_dif, &wqe->generic.wqe_com, LPFC_WQE_DIF_INSERT);
|
|
|
|
|
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag &= ~(LPFC_IO_DIF_PASS |
|
2020-11-15 19:26:42 +00:00
|
|
|
LPFC_IO_DIF_STRIP | LPFC_IO_DIF_INSERT);
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
return 0;
|
|
|
|
|
err:
|
2013-04-18 00:16:15 +00:00
|
|
|
if (lpfc_cmd->seg_cnt)
|
|
|
|
|
scsi_dma_unmap(scsi_cmnd);
|
|
|
|
|
if (lpfc_cmd->prot_seg_cnt)
|
|
|
|
|
dma_unmap_sg(&phba->pcidev->dev, scsi_prot_sglist(scsi_cmnd),
|
|
|
|
|
scsi_prot_sg_count(scsi_cmnd),
|
|
|
|
|
scsi_cmnd->sc_data_direction);
|
|
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2023-08-03 21:19:32 +00:00
|
|
|
"9084 Cannot setup S/G List for HBA "
|
|
|
|
|
"IO segs %d/%d SGL %d SCSI %d: %d %d %d\n",
|
2013-04-18 00:16:15 +00:00
|
|
|
lpfc_cmd->seg_cnt, lpfc_cmd->prot_seg_cnt,
|
|
|
|
|
phba->cfg_total_seg_cnt, phba->cfg_sg_seg_cnt,
|
2023-08-03 21:19:32 +00:00
|
|
|
prot_group_type, num_sge, ret);
|
2013-04-18 00:16:15 +00:00
|
|
|
|
|
|
|
|
lpfc_cmd->seg_cnt = 0;
|
|
|
|
|
lpfc_cmd->prot_seg_cnt = 0;
|
2019-08-14 23:56:54 +00:00
|
|
|
return ret;
|
2012-01-18 21:25:09 +00:00
|
|
|
}
|
|
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_dma_buf - Wrapper function for DMA mapping of scsi buffer
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be mapped.
|
|
|
|
|
*
|
|
|
|
|
* This routine wraps the actual DMA mapping function pointer from the
|
|
|
|
|
* lpfc_hba struct.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
2010-04-06 18:49:53 +00:00
|
|
|
* 1 - Error
|
|
|
|
|
* 0 - Success
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
**/
|
|
|
|
|
static inline int
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_scsi_prep_dma_buf(struct lpfc_hba *phba, struct lpfc_io_buf *lpfc_cmd)
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
{
|
|
|
|
|
return phba->lpfc_scsi_prep_dma_buf(phba, lpfc_cmd);
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-18 21:25:09 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_bg_scsi_prep_dma_buf - Wrapper function for DMA mapping of scsi buffer
|
|
|
|
|
* using BlockGuard.
|
|
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be mapped.
|
|
|
|
|
*
|
|
|
|
|
* This routine wraps the actual DMA mapping function pointer from the
|
|
|
|
|
* lpfc_hba struct.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* 1 - Error
|
|
|
|
|
* 0 - Success
|
|
|
|
|
**/
|
|
|
|
|
static inline int
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_bg_scsi_prep_dma_buf(struct lpfc_hba *phba, struct lpfc_io_buf *lpfc_cmd)
|
2012-01-18 21:25:09 +00:00
|
|
|
{
|
|
|
|
|
return phba->lpfc_bg_scsi_prep_dma_buf(phba, lpfc_cmd);
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-15 19:26:42 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_cmnd_buf - Wrapper function for IOCB/WQE mapping of scsi
|
|
|
|
|
* buffer
|
2021-03-12 09:47:11 +00:00
|
|
|
* @vport: Pointer to vport object.
|
2020-11-15 19:26:42 +00:00
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be mapped.
|
|
|
|
|
* @tmo: Timeout value for IO
|
|
|
|
|
*
|
|
|
|
|
* This routine initializes IOCB/WQE data structure from scsi command
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* 1 - Error
|
|
|
|
|
* 0 - Success
|
|
|
|
|
**/
|
|
|
|
|
static inline int
|
|
|
|
|
lpfc_scsi_prep_cmnd_buf(struct lpfc_vport *vport, struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
uint8_t tmo)
|
|
|
|
|
{
|
|
|
|
|
return vport->phba->lpfc_scsi_prep_cmnd_buf(vport, lpfc_cmd, tmo);
|
|
|
|
|
}
|
|
|
|
|
|
2008-09-07 15:52:10 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_send_scsi_error_event - Posts an event when there is SCSI error
|
2008-09-07 15:52:10 +00:00
|
|
|
* @phba: Pointer to hba context object.
|
|
|
|
|
* @vport: Pointer to vport object.
|
|
|
|
|
* @lpfc_cmd: Pointer to lpfc scsi command which reported the error.
|
2021-03-12 09:47:11 +00:00
|
|
|
* @fcpi_parm: FCP Initiator parameter.
|
2008-09-07 15:52:10 +00:00
|
|
|
*
|
|
|
|
|
* This function posts an event when there is a SCSI command reporting
|
|
|
|
|
* error from the scsi device.
|
|
|
|
|
**/
|
|
|
|
|
static void
|
|
|
|
|
lpfc_send_scsi_error_event(struct lpfc_hba *phba, struct lpfc_vport *vport,
|
2020-11-15 19:26:43 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd, uint32_t fcpi_parm) {
|
2008-09-07 15:52:10 +00:00
|
|
|
struct scsi_cmnd *cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_rsp *fcprsp = lpfc_cmd->fcp_rsp;
|
|
|
|
|
uint32_t resp_info = fcprsp->rspStatus2;
|
|
|
|
|
uint32_t scsi_status = fcprsp->rspStatus3;
|
|
|
|
|
struct lpfc_fast_path_event *fast_path_evt = NULL;
|
|
|
|
|
struct lpfc_nodelist *pnode = lpfc_cmd->rdata->pnode;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!pnode)
|
2010-10-22 15:06:56 +00:00
|
|
|
return;
|
|
|
|
|
|
2008-09-07 15:52:10 +00:00
|
|
|
/* If there is queuefull or busy condition send a scsi event */
|
|
|
|
|
if ((cmnd->result == SAM_STAT_TASK_SET_FULL) ||
|
|
|
|
|
(cmnd->result == SAM_STAT_BUSY)) {
|
|
|
|
|
fast_path_evt = lpfc_alloc_fast_evt(phba);
|
|
|
|
|
if (!fast_path_evt)
|
|
|
|
|
return;
|
|
|
|
|
fast_path_evt->un.scsi_evt.event_type =
|
|
|
|
|
FC_REG_SCSI_EVENT;
|
|
|
|
|
fast_path_evt->un.scsi_evt.subcategory =
|
|
|
|
|
(cmnd->result == SAM_STAT_TASK_SET_FULL) ?
|
|
|
|
|
LPFC_EVENT_QFULL : LPFC_EVENT_DEVBSY;
|
|
|
|
|
fast_path_evt->un.scsi_evt.lun = cmnd->device->lun;
|
|
|
|
|
memcpy(&fast_path_evt->un.scsi_evt.wwpn,
|
|
|
|
|
&pnode->nlp_portname, sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(&fast_path_evt->un.scsi_evt.wwnn,
|
|
|
|
|
&pnode->nlp_nodename, sizeof(struct lpfc_name));
|
|
|
|
|
} else if ((resp_info & SNS_LEN_VALID) && fcprsp->rspSnsLen &&
|
|
|
|
|
((cmnd->cmnd[0] == READ_10) || (cmnd->cmnd[0] == WRITE_10))) {
|
|
|
|
|
fast_path_evt = lpfc_alloc_fast_evt(phba);
|
|
|
|
|
if (!fast_path_evt)
|
|
|
|
|
return;
|
|
|
|
|
fast_path_evt->un.check_cond_evt.scsi_event.event_type =
|
|
|
|
|
FC_REG_SCSI_EVENT;
|
|
|
|
|
fast_path_evt->un.check_cond_evt.scsi_event.subcategory =
|
|
|
|
|
LPFC_EVENT_CHECK_COND;
|
|
|
|
|
fast_path_evt->un.check_cond_evt.scsi_event.lun =
|
|
|
|
|
cmnd->device->lun;
|
|
|
|
|
memcpy(&fast_path_evt->un.check_cond_evt.scsi_event.wwpn,
|
|
|
|
|
&pnode->nlp_portname, sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(&fast_path_evt->un.check_cond_evt.scsi_event.wwnn,
|
|
|
|
|
&pnode->nlp_nodename, sizeof(struct lpfc_name));
|
|
|
|
|
fast_path_evt->un.check_cond_evt.sense_key =
|
|
|
|
|
cmnd->sense_buffer[2] & 0xf;
|
|
|
|
|
fast_path_evt->un.check_cond_evt.asc = cmnd->sense_buffer[12];
|
|
|
|
|
fast_path_evt->un.check_cond_evt.ascq = cmnd->sense_buffer[13];
|
|
|
|
|
} else if ((cmnd->sc_data_direction == DMA_FROM_DEVICE) &&
|
|
|
|
|
fcpi_parm &&
|
|
|
|
|
((be32_to_cpu(fcprsp->rspResId) != fcpi_parm) ||
|
|
|
|
|
((scsi_status == SAM_STAT_GOOD) &&
|
|
|
|
|
!(resp_info & (RESID_UNDER | RESID_OVER))))) {
|
|
|
|
|
/*
|
|
|
|
|
* If status is good or resid does not match with fcp_param and
|
|
|
|
|
* there is valid fcpi_parm, then there is a read_check error
|
|
|
|
|
*/
|
|
|
|
|
fast_path_evt = lpfc_alloc_fast_evt(phba);
|
|
|
|
|
if (!fast_path_evt)
|
|
|
|
|
return;
|
|
|
|
|
fast_path_evt->un.read_check_error.header.event_type =
|
|
|
|
|
FC_REG_FABRIC_EVENT;
|
|
|
|
|
fast_path_evt->un.read_check_error.header.subcategory =
|
|
|
|
|
LPFC_EVENT_FCPRDCHKERR;
|
|
|
|
|
memcpy(&fast_path_evt->un.read_check_error.header.wwpn,
|
|
|
|
|
&pnode->nlp_portname, sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(&fast_path_evt->un.read_check_error.header.wwnn,
|
|
|
|
|
&pnode->nlp_nodename, sizeof(struct lpfc_name));
|
|
|
|
|
fast_path_evt->un.read_check_error.lun = cmnd->device->lun;
|
|
|
|
|
fast_path_evt->un.read_check_error.opcode = cmnd->cmnd[0];
|
|
|
|
|
fast_path_evt->un.read_check_error.fcpiparam =
|
|
|
|
|
fcpi_parm;
|
|
|
|
|
} else
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
fast_path_evt->vport = vport;
|
|
|
|
|
spin_lock_irqsave(&phba->hbalock, flags);
|
|
|
|
|
list_add_tail(&fast_path_evt->work_evt.evt_listp, &phba->work_list);
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, flags);
|
|
|
|
|
lpfc_worker_wake_up(phba);
|
|
|
|
|
return;
|
|
|
|
|
}
|
2008-12-05 03:39:02 +00:00
|
|
|
|
|
|
|
|
/**
|
2009-06-10 21:22:44 +00:00
|
|
|
* lpfc_scsi_unprep_dma_buf - Un-map DMA mapping of SG-list for dev
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* @phba: The HBA for which this call is being executed.
|
2008-12-05 03:39:02 +00:00
|
|
|
* @psb: The scsi buffer which is going to be un-mapped.
|
|
|
|
|
*
|
|
|
|
|
* This routine does DMA un-mapping of scatter gather list of scsi command
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* field of @lpfc_cmd for device with SLI-3 interface spec.
|
2008-12-05 03:39:02 +00:00
|
|
|
**/
|
2006-07-06 19:50:08 +00:00
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_scsi_unprep_dma_buf(struct lpfc_hba *phba, struct lpfc_io_buf *psb)
|
2006-07-06 19:50:08 +00:00
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* There are only two special cases to consider. (1) the scsi command
|
|
|
|
|
* requested scatter-gather usage or (2) the scsi command allocated
|
|
|
|
|
* a request buffer, but did not request use_sg. There is a third
|
|
|
|
|
* case, but it does not require resource deallocation.
|
|
|
|
|
*/
|
2007-06-18 00:56:39 +00:00
|
|
|
if (psb->seg_cnt > 0)
|
|
|
|
|
scsi_dma_unmap(psb->pCmd);
|
2008-12-05 03:40:02 +00:00
|
|
|
if (psb->prot_seg_cnt > 0)
|
|
|
|
|
dma_unmap_sg(&phba->pcidev->dev, scsi_prot_sglist(psb->pCmd),
|
|
|
|
|
scsi_prot_sg_count(psb->pCmd),
|
|
|
|
|
psb->pCmd->sc_data_direction);
|
2006-07-06 19:50:08 +00:00
|
|
|
}
|
|
|
|
|
|
2021-08-16 16:28:54 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_unblock_requests - allow further commands to be queued.
|
|
|
|
|
* @phba: pointer to phba object
|
|
|
|
|
*
|
|
|
|
|
* For single vport, just call scsi_unblock_requests on physical port.
|
|
|
|
|
* For multiple vports, send scsi_unblock_requests for all the vports.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
lpfc_unblock_requests(struct lpfc_hba *phba)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_vport **vports;
|
|
|
|
|
struct Scsi_Host *shost;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4 &&
|
|
|
|
|
!phba->sli4_hba.max_cfg_param.vpi_used) {
|
|
|
|
|
shost = lpfc_shost_from_vport(phba->pport);
|
|
|
|
|
scsi_unblock_requests(shost);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
vports = lpfc_create_vport_work_array(phba);
|
|
|
|
|
if (vports != NULL)
|
|
|
|
|
for (i = 0; i <= phba->max_vports && vports[i] != NULL; i++) {
|
|
|
|
|
shost = lpfc_shost_from_vport(vports[i]);
|
|
|
|
|
scsi_unblock_requests(shost);
|
|
|
|
|
}
|
|
|
|
|
lpfc_destroy_vport_work_array(phba, vports);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_block_requests - prevent further commands from being queued.
|
|
|
|
|
* @phba: pointer to phba object
|
|
|
|
|
*
|
|
|
|
|
* For single vport, just call scsi_block_requests on physical port.
|
|
|
|
|
* For multiple vports, send scsi_block_requests for all the vports.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
lpfc_block_requests(struct lpfc_hba *phba)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_vport **vports;
|
|
|
|
|
struct Scsi_Host *shost;
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
if (atomic_read(&phba->cmf_stop_io))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4 &&
|
|
|
|
|
!phba->sli4_hba.max_cfg_param.vpi_used) {
|
|
|
|
|
shost = lpfc_shost_from_vport(phba->pport);
|
|
|
|
|
scsi_block_requests(shost);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
vports = lpfc_create_vport_work_array(phba);
|
|
|
|
|
if (vports != NULL)
|
|
|
|
|
for (i = 0; i <= phba->max_vports && vports[i] != NULL; i++) {
|
|
|
|
|
shost = lpfc_shost_from_vport(vports[i]);
|
|
|
|
|
scsi_block_requests(shost);
|
|
|
|
|
}
|
|
|
|
|
lpfc_destroy_vport_work_array(phba, vports);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_update_cmf_cmpl - Adjust CMF counters for IO completion
|
|
|
|
|
* @phba: The HBA for which this call is being executed.
|
|
|
|
|
* @time: The latency of the IO that completed (in ns)
|
|
|
|
|
* @size: The size of the IO that completed
|
|
|
|
|
* @shost: SCSI host the IO completed on (NULL for a NVME IO)
|
|
|
|
|
*
|
|
|
|
|
* The routine adjusts the various Burst and Bandwidth counters used in
|
|
|
|
|
* Congestion management and E2E. If time is set to LPFC_CGN_NOT_SENT,
|
|
|
|
|
* that means the IO was never issued to the HBA, so this routine is
|
|
|
|
|
* just being called to cleanup the counter from a previous
|
|
|
|
|
* lpfc_update_cmf_cmd call.
|
|
|
|
|
*/
|
|
|
|
|
int
|
|
|
|
|
lpfc_update_cmf_cmpl(struct lpfc_hba *phba,
|
|
|
|
|
uint64_t time, uint32_t size, struct Scsi_Host *shost)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_cgn_stat *cgs;
|
|
|
|
|
|
|
|
|
|
if (time != LPFC_CGN_NOT_SENT) {
|
|
|
|
|
/* lat is ns coming in, save latency in us */
|
|
|
|
|
if (time < 1000)
|
|
|
|
|
time = 1;
|
|
|
|
|
else
|
|
|
|
|
time = div_u64(time + 500, 1000); /* round it */
|
|
|
|
|
|
2022-04-12 22:19:57 +00:00
|
|
|
cgs = per_cpu_ptr(phba->cmf_stat, raw_smp_processor_id());
|
2021-08-16 16:28:54 +00:00
|
|
|
atomic64_add(size, &cgs->rcv_bytes);
|
|
|
|
|
atomic64_add(time, &cgs->rx_latency);
|
|
|
|
|
atomic_inc(&cgs->rx_io_cnt);
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_update_cmf_cmd - Adjust CMF counters for IO submission
|
|
|
|
|
* @phba: The HBA for which this call is being executed.
|
|
|
|
|
* @size: The size of the IO that will be issued
|
|
|
|
|
*
|
|
|
|
|
* The routine adjusts the various Burst and Bandwidth counters used in
|
|
|
|
|
* Congestion management and E2E.
|
|
|
|
|
*/
|
|
|
|
|
int
|
|
|
|
|
lpfc_update_cmf_cmd(struct lpfc_hba *phba, uint32_t size)
|
|
|
|
|
{
|
|
|
|
|
uint64_t total;
|
|
|
|
|
struct lpfc_cgn_stat *cgs;
|
|
|
|
|
int cpu;
|
|
|
|
|
|
|
|
|
|
/* At this point we are either LPFC_CFG_MANAGED or LPFC_CFG_MONITOR */
|
2021-09-10 23:31:56 +00:00
|
|
|
if (phba->cmf_active_mode == LPFC_CFG_MANAGED &&
|
|
|
|
|
phba->cmf_max_bytes_per_interval) {
|
2021-08-16 16:28:54 +00:00
|
|
|
total = 0;
|
|
|
|
|
for_each_present_cpu(cpu) {
|
|
|
|
|
cgs = per_cpu_ptr(phba->cmf_stat, cpu);
|
|
|
|
|
total += atomic64_read(&cgs->total_bytes);
|
|
|
|
|
}
|
|
|
|
|
if (total >= phba->cmf_max_bytes_per_interval) {
|
|
|
|
|
if (!atomic_xchg(&phba->cmf_bw_wait, 1)) {
|
|
|
|
|
lpfc_block_requests(phba);
|
|
|
|
|
phba->cmf_last_ts =
|
|
|
|
|
lpfc_calc_cmf_latency(phba);
|
|
|
|
|
}
|
|
|
|
|
atomic_inc(&phba->cmf_busy);
|
|
|
|
|
return -EBUSY;
|
|
|
|
|
}
|
2021-08-16 16:28:55 +00:00
|
|
|
if (size > atomic_read(&phba->rx_max_read_cnt))
|
|
|
|
|
atomic_set(&phba->rx_max_read_cnt, size);
|
2021-08-16 16:28:54 +00:00
|
|
|
}
|
|
|
|
|
|
2022-04-12 22:19:57 +00:00
|
|
|
cgs = per_cpu_ptr(phba->cmf_stat, raw_smp_processor_id());
|
2021-08-16 16:28:54 +00:00
|
|
|
atomic64_add(size, &cgs->total_bytes);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2021-03-12 09:47:11 +00:00
|
|
|
* lpfc_handle_fcp_err - FCP response handler
|
2008-12-05 03:39:02 +00:00
|
|
|
* @vport: The virtual port for which this call is being executed.
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
* @lpfc_cmd: Pointer to lpfc_io_buf data structure.
|
2021-03-12 09:47:11 +00:00
|
|
|
* @fcpi_parm: FCP Initiator parameter.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* This routine is called to process response IOCB with status field
|
|
|
|
|
* IOSTAT_FCP_RSP_ERROR. This routine sets result field of scsi command
|
|
|
|
|
* based upon SCSI and FCP error.
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static void
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_handle_fcp_err(struct lpfc_vport *vport, struct lpfc_io_buf *lpfc_cmd,
|
2020-11-15 19:26:43 +00:00
|
|
|
uint32_t fcpi_parm)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_cmnd *fcpcmd = lpfc_cmd->fcp_cmnd;
|
|
|
|
|
struct fcp_rsp *fcprsp = lpfc_cmd->fcp_rsp;
|
|
|
|
|
uint32_t resp_info = fcprsp->rspStatus2;
|
|
|
|
|
uint32_t scsi_status = fcprsp->rspStatus3;
|
2006-12-02 18:34:42 +00:00
|
|
|
uint32_t *lp;
|
2005-04-17 21:05:31 +00:00
|
|
|
uint32_t host_status = DID_OK;
|
|
|
|
|
uint32_t rsplen = 0;
|
2015-12-16 23:12:01 +00:00
|
|
|
uint32_t fcpDl;
|
2006-12-02 18:34:42 +00:00
|
|
|
uint32_t logit = LOG_FCP | LOG_FCP_ERROR;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2008-09-07 15:52:10 +00:00
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
|
|
|
|
* If this is a task management command, there is no
|
|
|
|
|
* scsi packet associated with this lpfc_cmd. The driver
|
|
|
|
|
* consumes it.
|
|
|
|
|
*/
|
|
|
|
|
if (fcpcmd->fcpCntl2) {
|
|
|
|
|
scsi_status = 0;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
2009-10-02 19:16:51 +00:00
|
|
|
if (resp_info & RSP_LEN_VALID) {
|
|
|
|
|
rsplen = be32_to_cpu(fcprsp->rspRspLen);
|
2010-02-26 19:13:54 +00:00
|
|
|
if (rsplen != 0 && rsplen != 4 && rsplen != 8) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"2719 Invalid response length: "
|
|
|
|
|
"tgt x%x lun x%llx cmnd x%x rsplen "
|
|
|
|
|
"x%x\n", cmnd->device->id,
|
|
|
|
|
cmnd->device->lun, cmnd->cmnd[0],
|
|
|
|
|
rsplen);
|
2009-10-02 19:16:51 +00:00
|
|
|
host_status = DID_ERROR;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
2010-02-26 19:13:54 +00:00
|
|
|
if (fcprsp->rspInfo3 != RSP_NO_FAILURE) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2010-02-26 19:13:54 +00:00
|
|
|
"2757 Protocol failure detected during "
|
|
|
|
|
"processing of FCP I/O op: "
|
2014-06-25 13:27:36 +00:00
|
|
|
"tgt x%x lun x%llx cmnd x%x rspInfo3 x%x\n",
|
2010-02-26 19:13:54 +00:00
|
|
|
cmnd->device->id,
|
|
|
|
|
cmnd->device->lun, cmnd->cmnd[0],
|
|
|
|
|
fcprsp->rspInfo3);
|
|
|
|
|
host_status = DID_ERROR;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
2009-10-02 19:16:51 +00:00
|
|
|
}
|
|
|
|
|
|
2006-12-02 18:34:42 +00:00
|
|
|
if ((resp_info & SNS_LEN_VALID) && fcprsp->rspSnsLen) {
|
|
|
|
|
uint32_t snslen = be32_to_cpu(fcprsp->rspSnsLen);
|
|
|
|
|
if (snslen > SCSI_SENSE_BUFFERSIZE)
|
|
|
|
|
snslen = SCSI_SENSE_BUFFERSIZE;
|
|
|
|
|
|
|
|
|
|
if (resp_info & RSP_LEN_VALID)
|
|
|
|
|
rsplen = be32_to_cpu(fcprsp->rspRspLen);
|
|
|
|
|
memcpy(cmnd->sense_buffer, &fcprsp->rspInfo0 + rsplen, snslen);
|
|
|
|
|
}
|
|
|
|
|
lp = (uint32_t *)cmnd->sense_buffer;
|
|
|
|
|
|
2012-08-14 18:26:06 +00:00
|
|
|
/* special handling for under run conditions */
|
|
|
|
|
if (!scsi_status && (resp_info & RESID_UNDER)) {
|
|
|
|
|
/* don't log under runs if fcp set... */
|
|
|
|
|
if (vport->cfg_log_verbose & LOG_FCP)
|
|
|
|
|
logit = LOG_FCP_ERROR;
|
|
|
|
|
/* unless operator says so */
|
|
|
|
|
if (vport->cfg_log_verbose & LOG_FCP_UNDER)
|
|
|
|
|
logit = LOG_FCP_UNDER;
|
|
|
|
|
}
|
2006-12-02 18:34:42 +00:00
|
|
|
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, logit,
|
2008-12-05 03:40:02 +00:00
|
|
|
"9024 FCP command x%x failed: x%x SNS x%x x%x "
|
2007-08-02 15:10:09 +00:00
|
|
|
"Data: x%x x%x x%x x%x x%x\n",
|
|
|
|
|
cmnd->cmnd[0], scsi_status,
|
|
|
|
|
be32_to_cpu(*lp), be32_to_cpu(*(lp + 3)), resp_info,
|
|
|
|
|
be32_to_cpu(fcprsp->rspResId),
|
|
|
|
|
be32_to_cpu(fcprsp->rspSnsLen),
|
|
|
|
|
be32_to_cpu(fcprsp->rspRspLen),
|
|
|
|
|
fcprsp->rspInfo3);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2007-06-18 00:56:39 +00:00
|
|
|
scsi_set_resid(cmnd, 0);
|
2015-12-16 23:12:01 +00:00
|
|
|
fcpDl = be32_to_cpu(fcpcmd->fcpDl);
|
2005-04-17 21:05:31 +00:00
|
|
|
if (resp_info & RESID_UNDER) {
|
2007-06-18 00:56:39 +00:00
|
|
|
scsi_set_resid(cmnd, be32_to_cpu(fcprsp->rspResId));
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2011-10-11 01:32:10 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP_UNDER,
|
2018-01-30 23:59:00 +00:00
|
|
|
"9025 FCP Underrun, expected %d, "
|
2007-08-02 15:10:09 +00:00
|
|
|
"residual %d Data: x%x x%x x%x\n",
|
2015-12-16 23:12:01 +00:00
|
|
|
fcpDl,
|
2007-08-02 15:10:09 +00:00
|
|
|
scsi_get_resid(cmnd), fcpi_parm, cmnd->cmnd[0],
|
|
|
|
|
cmnd->underflow);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2007-04-25 13:52:34 +00:00
|
|
|
/*
|
2018-01-30 23:59:00 +00:00
|
|
|
* If there is an under run, check if under run reported by
|
2007-04-25 13:52:34 +00:00
|
|
|
* storage array is same as the under run reported by HBA.
|
|
|
|
|
* If this is not same, there is a dropped frame.
|
|
|
|
|
*/
|
2018-01-30 23:59:00 +00:00
|
|
|
if (fcpi_parm && (scsi_get_resid(cmnd) != fcpi_parm)) {
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING,
|
|
|
|
|
LOG_FCP | LOG_FCP_ERROR,
|
2008-12-05 03:40:02 +00:00
|
|
|
"9026 FCP Read Check Error "
|
2007-08-02 15:10:09 +00:00
|
|
|
"and Underrun Data: x%x x%x x%x x%x\n",
|
2015-12-16 23:12:01 +00:00
|
|
|
fcpDl,
|
2007-08-02 15:10:09 +00:00
|
|
|
scsi_get_resid(cmnd), fcpi_parm,
|
|
|
|
|
cmnd->cmnd[0]);
|
2007-06-18 00:56:39 +00:00
|
|
|
scsi_set_resid(cmnd, scsi_bufflen(cmnd));
|
2007-04-25 13:52:34 +00:00
|
|
|
host_status = DID_ERROR;
|
|
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
|
|
|
|
* The cmnd->underflow is the minimum number of bytes that must
|
2011-03-31 01:57:33 +00:00
|
|
|
* be transferred for this command. Provided a sense condition
|
2005-04-17 21:05:31 +00:00
|
|
|
* is not present, make sure the actual amount transferred is at
|
|
|
|
|
* least the underflow value or fail.
|
|
|
|
|
*/
|
|
|
|
|
if (!(resp_info & SNS_LEN_VALID) &&
|
|
|
|
|
(scsi_status == SAM_STAT_GOOD) &&
|
2007-06-18 00:56:39 +00:00
|
|
|
(scsi_bufflen(cmnd) - scsi_get_resid(cmnd)
|
|
|
|
|
< cmnd->underflow)) {
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2008-12-05 03:40:02 +00:00
|
|
|
"9027 FCP command x%x residual "
|
2007-08-02 15:10:09 +00:00
|
|
|
"underrun converted to error "
|
|
|
|
|
"Data: x%x x%x x%x\n",
|
2007-08-05 10:08:38 +00:00
|
|
|
cmnd->cmnd[0], scsi_bufflen(cmnd),
|
2007-08-02 15:10:09 +00:00
|
|
|
scsi_get_resid(cmnd), cmnd->underflow);
|
2005-04-17 21:05:31 +00:00
|
|
|
host_status = DID_ERROR;
|
|
|
|
|
}
|
|
|
|
|
} else if (resp_info & RESID_OVER) {
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
2008-12-05 03:40:02 +00:00
|
|
|
"9028 FCP command x%x residual overrun error. "
|
2009-07-19 14:01:38 +00:00
|
|
|
"Data: x%x x%x\n", cmnd->cmnd[0],
|
2007-08-02 15:10:09 +00:00
|
|
|
scsi_bufflen(cmnd), scsi_get_resid(cmnd));
|
2005-04-17 21:05:31 +00:00
|
|
|
host_status = DID_ERROR;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check SLI validation that all the transfer was actually done
|
2013-09-06 16:19:17 +00:00
|
|
|
* (fcpi_parm should be zero). Apply check only to reads.
|
2005-04-17 21:05:31 +00:00
|
|
|
*/
|
2015-12-16 23:12:01 +00:00
|
|
|
} else if (fcpi_parm) {
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP | LOG_FCP_ERROR,
|
2020-11-15 19:26:43 +00:00
|
|
|
"9029 FCP %s Check Error Data: "
|
2010-09-29 15:19:08 +00:00
|
|
|
"x%x x%x x%x x%x x%x\n",
|
2015-12-16 23:12:01 +00:00
|
|
|
((cmnd->sc_data_direction == DMA_FROM_DEVICE) ?
|
|
|
|
|
"Read" : "Write"),
|
|
|
|
|
fcpDl, be32_to_cpu(fcprsp->rspResId),
|
2010-09-29 15:19:08 +00:00
|
|
|
fcpi_parm, cmnd->cmnd[0], scsi_status);
|
2015-12-16 23:12:01 +00:00
|
|
|
|
|
|
|
|
/* There is some issue with the LPe12000 that causes it
|
|
|
|
|
* to miscalculate the fcpi_parm and falsely trip this
|
|
|
|
|
* recovery logic. Detect this case and don't error when true.
|
|
|
|
|
*/
|
|
|
|
|
if (fcpi_parm > fcpDl)
|
|
|
|
|
goto out;
|
|
|
|
|
|
2010-09-29 15:19:08 +00:00
|
|
|
switch (scsi_status) {
|
|
|
|
|
case SAM_STAT_GOOD:
|
|
|
|
|
case SAM_STAT_CHECK_CONDITION:
|
|
|
|
|
/* Fabric dropped a data frame. Fail any successful
|
|
|
|
|
* command in which we detected dropped frames.
|
|
|
|
|
* A status of good or some check conditions could
|
|
|
|
|
* be considered a successful command.
|
|
|
|
|
*/
|
|
|
|
|
host_status = DID_ERROR;
|
|
|
|
|
break;
|
|
|
|
|
}
|
2007-06-18 00:56:39 +00:00
|
|
|
scsi_set_resid(cmnd, scsi_bufflen(cmnd));
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
out:
|
2018-07-05 11:01:39 +00:00
|
|
|
cmnd->result = host_status << 16 | scsi_status;
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_send_scsi_error_event(vport->phba, vport, lpfc_cmd, fcpi_parm);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_fcp_io_cmd_wqe_cmpl - Complete a FCP IO
|
|
|
|
|
* @phba: The hba for which this call is being executed.
|
|
|
|
|
* @pwqeIn: The command WQE for the scsi cmnd.
|
2022-02-25 02:22:52 +00:00
|
|
|
* @pwqeOut: Pointer to driver response WQE object.
|
2020-11-15 19:26:43 +00:00
|
|
|
*
|
|
|
|
|
* This routine assigns scsi command result by looking into response WQE
|
|
|
|
|
* status field appropriately. This routine handles QUEUE FULL condition as
|
|
|
|
|
* well by ramping down device queue depth.
|
|
|
|
|
**/
|
|
|
|
|
static void
|
|
|
|
|
lpfc_fcp_io_cmd_wqe_cmpl(struct lpfc_hba *phba, struct lpfc_iocbq *pwqeIn,
|
2022-02-25 02:22:52 +00:00
|
|
|
struct lpfc_iocbq *pwqeOut)
|
2020-11-15 19:26:43 +00:00
|
|
|
{
|
2022-04-12 22:20:02 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd = pwqeIn->io_buf;
|
2022-02-25 02:22:52 +00:00
|
|
|
struct lpfc_wcqe_complete *wcqe = &pwqeOut->wcqe_cmpl;
|
2020-11-15 19:26:43 +00:00
|
|
|
struct lpfc_vport *vport = pwqeIn->vport;
|
2020-11-18 13:13:45 +00:00
|
|
|
struct lpfc_rport_data *rdata;
|
|
|
|
|
struct lpfc_nodelist *ndlp;
|
2020-11-15 19:26:43 +00:00
|
|
|
struct scsi_cmnd *cmd;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
struct lpfc_fast_path_event *fast_path_evt;
|
|
|
|
|
struct Scsi_Host *shost;
|
|
|
|
|
u32 logit = LOG_FCP;
|
2023-05-23 18:32:03 +00:00
|
|
|
u32 idx;
|
2021-08-16 16:28:54 +00:00
|
|
|
u32 lat;
|
2021-05-14 19:55:57 +00:00
|
|
|
u8 wait_xb_clr = 0;
|
2020-11-15 19:26:43 +00:00
|
|
|
|
|
|
|
|
/* Sanity check on return of outstanding command */
|
|
|
|
|
if (!lpfc_cmd) {
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"9032 Null lpfc_cmd pointer. No "
|
|
|
|
|
"release, skip completion\n");
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-18 13:13:45 +00:00
|
|
|
rdata = lpfc_cmd->rdata;
|
|
|
|
|
ndlp = rdata->pnode;
|
|
|
|
|
|
2020-11-15 19:26:43 +00:00
|
|
|
/* Sanity check on return of outstanding command */
|
|
|
|
|
cmd = lpfc_cmd->pCmd;
|
2021-03-01 17:18:07 +00:00
|
|
|
if (!cmd) {
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
|
|
|
|
"9042 I/O completion: Not an active IO\n");
|
|
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
|
|
|
|
return;
|
|
|
|
|
}
|
2022-02-25 02:23:04 +00:00
|
|
|
/* Guard against abort handler being called at same time */
|
|
|
|
|
spin_lock(&lpfc_cmd->buf_lock);
|
2020-11-15 19:26:43 +00:00
|
|
|
idx = lpfc_cmd->cur_iocbq.hba_wqidx;
|
|
|
|
|
if (phba->sli4_hba.hdwq)
|
|
|
|
|
phba->sli4_hba.hdwq[idx].scsi_cstat.io_cmpls++;
|
|
|
|
|
|
|
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
if (unlikely(phba->hdwqstat_on & LPFC_CHECK_SCSI_IO))
|
|
|
|
|
this_cpu_inc(phba->sli4_hba.c_stat->cmpl_io);
|
|
|
|
|
#endif
|
|
|
|
|
shost = cmd->device->host;
|
|
|
|
|
|
2023-05-23 18:32:03 +00:00
|
|
|
lpfc_cmd->status = bf_get(lpfc_wcqe_c_status, wcqe);
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_cmd->result = (wcqe->parameter & IOERR_PARAM_MASK);
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->flags &= ~LPFC_SBUF_XBUSY;
|
2021-05-14 19:55:57 +00:00
|
|
|
if (bf_get(lpfc_wcqe_c_xb, wcqe)) {
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_cmd->flags |= LPFC_SBUF_XBUSY;
|
2021-05-14 19:55:57 +00:00
|
|
|
if (phba->cfg_fcp_wait_abts_rsp)
|
|
|
|
|
wait_xb_clr = 1;
|
|
|
|
|
}
|
2020-11-15 19:26:43 +00:00
|
|
|
|
|
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
if (lpfc_cmd->prot_data_type) {
|
|
|
|
|
struct scsi_dif_tuple *src = NULL;
|
|
|
|
|
|
|
|
|
|
src = (struct scsi_dif_tuple *)lpfc_cmd->prot_data_segment;
|
|
|
|
|
/*
|
|
|
|
|
* Used to restore any changes to protection
|
|
|
|
|
* data for error injection.
|
|
|
|
|
*/
|
|
|
|
|
switch (lpfc_cmd->prot_data_type) {
|
|
|
|
|
case LPFC_INJERR_REFTAG:
|
|
|
|
|
src->ref_tag =
|
|
|
|
|
lpfc_cmd->prot_data;
|
|
|
|
|
break;
|
|
|
|
|
case LPFC_INJERR_APPTAG:
|
|
|
|
|
src->app_tag =
|
|
|
|
|
(uint16_t)lpfc_cmd->prot_data;
|
|
|
|
|
break;
|
|
|
|
|
case LPFC_INJERR_GUARD:
|
|
|
|
|
src->guard_tag =
|
|
|
|
|
(uint16_t)lpfc_cmd->prot_data;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->prot_data = 0;
|
|
|
|
|
lpfc_cmd->prot_data_type = 0;
|
|
|
|
|
lpfc_cmd->prot_data_segment = NULL;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
if (unlikely(lpfc_cmd->status)) {
|
|
|
|
|
if (lpfc_cmd->status == IOSTAT_FCP_RSP_ERROR &&
|
|
|
|
|
!lpfc_cmd->fcp_rsp->rspStatus3 &&
|
|
|
|
|
(lpfc_cmd->fcp_rsp->rspStatus2 & RESID_UNDER) &&
|
|
|
|
|
!(vport->cfg_log_verbose & LOG_FCP_UNDER))
|
|
|
|
|
logit = 0;
|
|
|
|
|
else
|
|
|
|
|
logit = LOG_FCP | LOG_FCP_UNDER;
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, logit,
|
|
|
|
|
"9034 FCP cmd x%x failed <%d/%lld> "
|
|
|
|
|
"status: x%x result: x%x "
|
|
|
|
|
"sid: x%x did: x%x oxid: x%x "
|
|
|
|
|
"Data: x%x x%x x%x\n",
|
|
|
|
|
cmd->cmnd[0],
|
|
|
|
|
cmd->device ? cmd->device->id : 0xffff,
|
|
|
|
|
cmd->device ? cmd->device->lun : 0xffff,
|
|
|
|
|
lpfc_cmd->status, lpfc_cmd->result,
|
|
|
|
|
vport->fc_myDID,
|
|
|
|
|
(ndlp) ? ndlp->nlp_DID : 0,
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_xritag,
|
|
|
|
|
wcqe->parameter, wcqe->total_data_placed,
|
|
|
|
|
lpfc_cmd->cur_iocbq.iotag);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch (lpfc_cmd->status) {
|
2023-05-23 18:32:03 +00:00
|
|
|
case CQE_STATUS_SUCCESS:
|
2020-11-15 19:26:43 +00:00
|
|
|
cmd->result = DID_OK << 16;
|
|
|
|
|
break;
|
2023-05-23 18:32:03 +00:00
|
|
|
case CQE_STATUS_FCP_RSP_FAILURE:
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_handle_fcp_err(vport, lpfc_cmd,
|
|
|
|
|
pwqeIn->wqe.fcp_iread.total_xfer_len -
|
|
|
|
|
wcqe->total_data_placed);
|
|
|
|
|
break;
|
2023-05-23 18:32:03 +00:00
|
|
|
case CQE_STATUS_NPORT_BSY:
|
|
|
|
|
case CQE_STATUS_FABRIC_BSY:
|
2020-11-15 19:26:43 +00:00
|
|
|
cmd->result = DID_TRANSPORT_DISRUPTED << 16;
|
|
|
|
|
fast_path_evt = lpfc_alloc_fast_evt(phba);
|
|
|
|
|
if (!fast_path_evt)
|
|
|
|
|
break;
|
|
|
|
|
fast_path_evt->un.fabric_evt.event_type =
|
|
|
|
|
FC_REG_FABRIC_EVENT;
|
|
|
|
|
fast_path_evt->un.fabric_evt.subcategory =
|
|
|
|
|
(lpfc_cmd->status == IOSTAT_NPORT_BSY) ?
|
|
|
|
|
LPFC_EVENT_PORT_BUSY : LPFC_EVENT_FABRIC_BUSY;
|
|
|
|
|
if (ndlp) {
|
|
|
|
|
memcpy(&fast_path_evt->un.fabric_evt.wwpn,
|
|
|
|
|
&ndlp->nlp_portname,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(&fast_path_evt->un.fabric_evt.wwnn,
|
|
|
|
|
&ndlp->nlp_nodename,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
}
|
|
|
|
|
fast_path_evt->vport = vport;
|
|
|
|
|
fast_path_evt->work_evt.evt =
|
|
|
|
|
LPFC_EVT_FASTPATH_MGMT_EVT;
|
|
|
|
|
spin_lock_irqsave(&phba->hbalock, flags);
|
|
|
|
|
list_add_tail(&fast_path_evt->work_evt.evt_listp,
|
|
|
|
|
&phba->work_list);
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, flags);
|
|
|
|
|
lpfc_worker_wake_up(phba);
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, logit,
|
|
|
|
|
"9035 Fabric/Node busy FCP cmd x%x failed"
|
|
|
|
|
" <%d/%lld> "
|
|
|
|
|
"status: x%x result: x%x "
|
|
|
|
|
"sid: x%x did: x%x oxid: x%x "
|
|
|
|
|
"Data: x%x x%x x%x\n",
|
|
|
|
|
cmd->cmnd[0],
|
|
|
|
|
cmd->device ? cmd->device->id : 0xffff,
|
|
|
|
|
cmd->device ? cmd->device->lun : 0xffff,
|
|
|
|
|
lpfc_cmd->status, lpfc_cmd->result,
|
|
|
|
|
vport->fc_myDID,
|
|
|
|
|
(ndlp) ? ndlp->nlp_DID : 0,
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_xritag,
|
|
|
|
|
wcqe->parameter,
|
|
|
|
|
wcqe->total_data_placed,
|
|
|
|
|
lpfc_cmd->cur_iocbq.iocb.ulpIoTag);
|
|
|
|
|
break;
|
2023-05-23 18:32:03 +00:00
|
|
|
case CQE_STATUS_DI_ERROR:
|
|
|
|
|
if (bf_get(lpfc_wcqe_c_bg_edir, wcqe))
|
|
|
|
|
lpfc_cmd->result = IOERR_RX_DMA_FAILED;
|
|
|
|
|
else
|
|
|
|
|
lpfc_cmd->result = IOERR_TX_DMA_FAILED;
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP | LOG_BG,
|
|
|
|
|
"9048 DI Error xri x%x status x%x DI ext "
|
|
|
|
|
"status x%x data placed x%x\n",
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_xritag,
|
|
|
|
|
lpfc_cmd->status, wcqe->parameter,
|
|
|
|
|
wcqe->total_data_placed);
|
|
|
|
|
if (scsi_get_prot_op(cmd) != SCSI_PROT_NORMAL) {
|
|
|
|
|
/* BG enabled cmd. Parse BG error */
|
|
|
|
|
lpfc_parse_bg_err(phba, lpfc_cmd, pwqeOut);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
cmd->result = DID_ERROR << 16;
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_BG,
|
|
|
|
|
"9040 DI Error on unprotected cmd\n");
|
|
|
|
|
break;
|
|
|
|
|
case CQE_STATUS_REMOTE_STOP:
|
2020-11-15 19:26:43 +00:00
|
|
|
if (ndlp) {
|
|
|
|
|
/* This I/O was aborted by the target, we don't
|
|
|
|
|
* know the rxid and because we did not send the
|
|
|
|
|
* ABTS we cannot generate and RRQ.
|
|
|
|
|
*/
|
|
|
|
|
lpfc_set_rrq_active(phba, ndlp,
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_lxritag,
|
|
|
|
|
0, 0);
|
|
|
|
|
}
|
|
|
|
|
fallthrough;
|
2023-05-23 18:32:03 +00:00
|
|
|
case CQE_STATUS_LOCAL_REJECT:
|
2020-11-15 19:26:43 +00:00
|
|
|
if (lpfc_cmd->result & IOERR_DRVR_MASK)
|
|
|
|
|
lpfc_cmd->status = IOSTAT_DRIVER_REJECT;
|
|
|
|
|
if (lpfc_cmd->result == IOERR_ELXSEC_KEY_UNWRAP_ERROR ||
|
|
|
|
|
lpfc_cmd->result ==
|
|
|
|
|
IOERR_ELXSEC_KEY_UNWRAP_COMPARE_ERROR ||
|
|
|
|
|
lpfc_cmd->result == IOERR_ELXSEC_CRYPTO_ERROR ||
|
|
|
|
|
lpfc_cmd->result ==
|
|
|
|
|
IOERR_ELXSEC_CRYPTO_COMPARE_ERROR) {
|
|
|
|
|
cmd->result = DID_NO_CONNECT << 16;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (lpfc_cmd->result == IOERR_INVALID_RPI ||
|
2022-04-12 22:19:47 +00:00
|
|
|
lpfc_cmd->result == IOERR_LINK_DOWN ||
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_cmd->result == IOERR_NO_RESOURCES ||
|
|
|
|
|
lpfc_cmd->result == IOERR_ABORT_REQUESTED ||
|
2021-12-04 00:26:37 +00:00
|
|
|
lpfc_cmd->result == IOERR_RPI_SUSPENDED ||
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_cmd->result == IOERR_SLER_CMD_RCV_FAILURE) {
|
2022-08-24 06:00:33 +00:00
|
|
|
cmd->result = DID_TRANSPORT_DISRUPTED << 16;
|
2020-11-15 19:26:43 +00:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, logit,
|
|
|
|
|
"9036 Local Reject FCP cmd x%x failed"
|
|
|
|
|
" <%d/%lld> "
|
|
|
|
|
"status: x%x result: x%x "
|
|
|
|
|
"sid: x%x did: x%x oxid: x%x "
|
|
|
|
|
"Data: x%x x%x x%x\n",
|
|
|
|
|
cmd->cmnd[0],
|
|
|
|
|
cmd->device ? cmd->device->id : 0xffff,
|
|
|
|
|
cmd->device ? cmd->device->lun : 0xffff,
|
|
|
|
|
lpfc_cmd->status, lpfc_cmd->result,
|
|
|
|
|
vport->fc_myDID,
|
|
|
|
|
(ndlp) ? ndlp->nlp_DID : 0,
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_xritag,
|
|
|
|
|
wcqe->parameter,
|
|
|
|
|
wcqe->total_data_placed,
|
|
|
|
|
lpfc_cmd->cur_iocbq.iocb.ulpIoTag);
|
|
|
|
|
fallthrough;
|
|
|
|
|
default:
|
|
|
|
|
cmd->result = DID_ERROR << 16;
|
2023-05-23 18:32:03 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2020-11-15 19:26:43 +00:00
|
|
|
"9037 FCP Completion Error: xri %x "
|
|
|
|
|
"status x%x result x%x [x%x] "
|
|
|
|
|
"placed x%x\n",
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_xritag,
|
|
|
|
|
lpfc_cmd->status, lpfc_cmd->result,
|
|
|
|
|
wcqe->parameter,
|
|
|
|
|
wcqe->total_data_placed);
|
|
|
|
|
}
|
|
|
|
|
if (cmd->result || lpfc_cmd->fcp_rsp->rspSnsLen) {
|
|
|
|
|
u32 *lp = (u32 *)cmd->sense_buffer;
|
|
|
|
|
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2021-04-12 01:31:24 +00:00
|
|
|
"9039 Iodone <%d/%llu> cmd x%px, error "
|
2021-12-04 00:26:37 +00:00
|
|
|
"x%x SNS x%x x%x LBA x%llx Data: x%x x%x\n",
|
2020-11-15 19:26:43 +00:00
|
|
|
cmd->device->id, cmd->device->lun, cmd,
|
2021-12-04 00:26:37 +00:00
|
|
|
cmd->result, *lp, *(lp + 3),
|
2023-04-17 19:15:52 +00:00
|
|
|
(cmd->device->sector_size) ?
|
|
|
|
|
(u64)scsi_get_lba(cmd) : 0,
|
2021-12-04 00:26:37 +00:00
|
|
|
cmd->retries, scsi_get_resid(cmd));
|
2020-11-15 19:26:43 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (vport->cfg_max_scsicmpl_time &&
|
|
|
|
|
time_after(jiffies, lpfc_cmd->start_time +
|
|
|
|
|
msecs_to_jiffies(vport->cfg_max_scsicmpl_time))) {
|
|
|
|
|
spin_lock_irqsave(shost->host_lock, flags);
|
|
|
|
|
if (ndlp) {
|
|
|
|
|
if (ndlp->cmd_qdepth >
|
|
|
|
|
atomic_read(&ndlp->cmd_pending) &&
|
|
|
|
|
(atomic_read(&ndlp->cmd_pending) >
|
|
|
|
|
LPFC_MIN_TGT_QDEPTH) &&
|
|
|
|
|
(cmd->cmnd[0] == READ_10 ||
|
|
|
|
|
cmd->cmnd[0] == WRITE_10))
|
|
|
|
|
ndlp->cmd_qdepth =
|
|
|
|
|
atomic_read(&ndlp->cmd_pending);
|
|
|
|
|
|
|
|
|
|
ndlp->last_change_time = jiffies;
|
|
|
|
|
}
|
|
|
|
|
spin_unlock_irqrestore(shost->host_lock, flags);
|
|
|
|
|
}
|
|
|
|
|
lpfc_scsi_unprep_dma_buf(phba, lpfc_cmd);
|
|
|
|
|
|
|
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
if (lpfc_cmd->ts_cmd_start) {
|
|
|
|
|
lpfc_cmd->ts_isr_cmpl = lpfc_cmd->cur_iocbq.isr_timestamp;
|
|
|
|
|
lpfc_cmd->ts_data_io = ktime_get_ns();
|
|
|
|
|
phba->ktime_last_cmd = lpfc_cmd->ts_data_io;
|
|
|
|
|
lpfc_io_ktime(phba, lpfc_cmd);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
2021-08-16 16:28:54 +00:00
|
|
|
if (likely(!wait_xb_clr))
|
|
|
|
|
lpfc_cmd->pCmd = NULL;
|
|
|
|
|
spin_unlock(&lpfc_cmd->buf_lock);
|
|
|
|
|
|
|
|
|
|
/* Check if IO qualified for CMF */
|
|
|
|
|
if (phba->cmf_active_mode != LPFC_CFG_OFF &&
|
|
|
|
|
cmd->sc_data_direction == DMA_FROM_DEVICE &&
|
|
|
|
|
(scsi_sg_count(cmd))) {
|
|
|
|
|
/* Used when calculating average latency */
|
|
|
|
|
lat = ktime_get_ns() - lpfc_cmd->rx_cmd_start;
|
|
|
|
|
lpfc_update_cmf_cmpl(phba, lat, scsi_bufflen(cmd), shost);
|
|
|
|
|
}
|
|
|
|
|
|
2021-05-14 19:55:57 +00:00
|
|
|
if (wait_xb_clr)
|
|
|
|
|
goto out;
|
2020-11-15 19:26:43 +00:00
|
|
|
|
|
|
|
|
/* The sdev is not guaranteed to be valid post scsi_done upcall. */
|
2021-10-07 20:28:41 +00:00
|
|
|
scsi_done(cmd);
|
2020-11-15 19:26:43 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If there is an abort thread waiting for command completion
|
|
|
|
|
* wake up the thread.
|
|
|
|
|
*/
|
|
|
|
|
spin_lock(&lpfc_cmd->buf_lock);
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag &= ~LPFC_DRIVER_ABORTED;
|
2020-11-15 19:26:43 +00:00
|
|
|
if (lpfc_cmd->waitq)
|
|
|
|
|
wake_up(lpfc_cmd->waitq);
|
|
|
|
|
spin_unlock(&lpfc_cmd->buf_lock);
|
2021-08-16 16:28:54 +00:00
|
|
|
out:
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_scsi_cmd_iocb_cmpl - Scsi cmnd IOCB completion routine
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @pIocbIn: The command IOCBQ for the scsi cmnd.
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* @pIocbOut: The response IOCBQ for the scsi cmnd.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* This routine assigns scsi command result by looking into response IOCB
|
|
|
|
|
* status field appropriately. This routine handles QUEUE FULL condition as
|
|
|
|
|
* well by ramping down device queue depth.
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static void
|
|
|
|
|
lpfc_scsi_cmd_iocb_cmpl(struct lpfc_hba *phba, struct lpfc_iocbq *pIocbIn,
|
|
|
|
|
struct lpfc_iocbq *pIocbOut)
|
|
|
|
|
{
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd =
|
2022-04-12 22:20:02 +00:00
|
|
|
(struct lpfc_io_buf *) pIocbIn->io_buf;
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_vport *vport = pIocbIn->vport;
|
2005-04-17 21:05:31 +00:00
|
|
|
struct lpfc_rport_data *rdata = lpfc_cmd->rdata;
|
|
|
|
|
struct lpfc_nodelist *pnode = rdata->pnode;
|
2010-06-08 22:31:21 +00:00
|
|
|
struct scsi_cmnd *cmd;
|
2008-01-11 06:52:42 +00:00
|
|
|
unsigned long flags;
|
2008-09-07 15:52:10 +00:00
|
|
|
struct lpfc_fast_path_event *fast_path_evt;
|
2010-06-08 22:31:21 +00:00
|
|
|
struct Scsi_Host *shost;
|
2019-01-28 19:14:25 +00:00
|
|
|
int idx;
|
2011-10-11 01:32:10 +00:00
|
|
|
uint32_t logit = LOG_FCP;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
/* Guard against abort handler being called at same time */
|
|
|
|
|
spin_lock(&lpfc_cmd->buf_lock);
|
|
|
|
|
|
2010-06-08 22:31:21 +00:00
|
|
|
/* Sanity check on return of outstanding command */
|
|
|
|
|
cmd = lpfc_cmd->pCmd;
|
2019-09-22 03:58:57 +00:00
|
|
|
if (!cmd || !phba) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2019-01-28 19:14:37 +00:00
|
|
|
"2621 IO completion: Not an active IO\n");
|
|
|
|
|
spin_unlock(&lpfc_cmd->buf_lock);
|
2015-12-16 23:11:57 +00:00
|
|
|
return;
|
2019-01-28 19:14:37 +00:00
|
|
|
}
|
2019-01-28 19:14:25 +00:00
|
|
|
|
|
|
|
|
idx = lpfc_cmd->cur_iocbq.hba_wqidx;
|
|
|
|
|
if (phba->sli4_hba.hdwq)
|
|
|
|
|
phba->sli4_hba.hdwq[idx].scsi_cstat.io_cmpls++;
|
|
|
|
|
|
2019-01-28 19:14:31 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2020-03-22 18:13:00 +00:00
|
|
|
if (unlikely(phba->hdwqstat_on & LPFC_CHECK_SCSI_IO))
|
|
|
|
|
this_cpu_inc(phba->sli4_hba.c_stat->cmpl_io);
|
2019-01-28 19:14:31 +00:00
|
|
|
#endif
|
2010-06-08 22:31:21 +00:00
|
|
|
shost = cmd->device->host;
|
|
|
|
|
|
2012-08-14 18:25:43 +00:00
|
|
|
lpfc_cmd->result = (pIocbOut->iocb.un.ulpWord[4] & IOERR_PARAM_MASK);
|
2005-04-17 21:05:31 +00:00
|
|
|
lpfc_cmd->status = pIocbOut->iocb.ulpStatus;
|
2021-05-14 19:55:57 +00:00
|
|
|
/* pick up SLI4 exchange busy status from HBA */
|
|
|
|
|
lpfc_cmd->flags &= ~LPFC_SBUF_XBUSY;
|
2022-02-25 02:22:52 +00:00
|
|
|
if (pIocbOut->cmd_flag & LPFC_EXCHANGE_BUSY)
|
2019-10-18 21:18:21 +00:00
|
|
|
lpfc_cmd->flags |= LPFC_SBUF_XBUSY;
|
2010-01-27 04:07:37 +00:00
|
|
|
|
2012-03-02 03:37:42 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
if (lpfc_cmd->prot_data_type) {
|
|
|
|
|
struct scsi_dif_tuple *src = NULL;
|
|
|
|
|
|
|
|
|
|
src = (struct scsi_dif_tuple *)lpfc_cmd->prot_data_segment;
|
|
|
|
|
/*
|
|
|
|
|
* Used to restore any changes to protection
|
|
|
|
|
* data for error injection.
|
|
|
|
|
*/
|
|
|
|
|
switch (lpfc_cmd->prot_data_type) {
|
|
|
|
|
case LPFC_INJERR_REFTAG:
|
|
|
|
|
src->ref_tag =
|
|
|
|
|
lpfc_cmd->prot_data;
|
|
|
|
|
break;
|
|
|
|
|
case LPFC_INJERR_APPTAG:
|
|
|
|
|
src->app_tag =
|
|
|
|
|
(uint16_t)lpfc_cmd->prot_data;
|
|
|
|
|
break;
|
|
|
|
|
case LPFC_INJERR_GUARD:
|
|
|
|
|
src->guard_tag =
|
|
|
|
|
(uint16_t)lpfc_cmd->prot_data;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->prot_data = 0;
|
|
|
|
|
lpfc_cmd->prot_data_type = 0;
|
|
|
|
|
lpfc_cmd->prot_data_segment = NULL;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
2017-02-12 21:52:27 +00:00
|
|
|
|
2019-10-18 21:18:24 +00:00
|
|
|
if (unlikely(lpfc_cmd->status)) {
|
2005-04-17 21:05:31 +00:00
|
|
|
if (lpfc_cmd->status == IOSTAT_LOCAL_REJECT &&
|
|
|
|
|
(lpfc_cmd->result & IOERR_DRVR_MASK))
|
|
|
|
|
lpfc_cmd->status = IOSTAT_DRIVER_REJECT;
|
|
|
|
|
else if (lpfc_cmd->status >= IOSTAT_CNT)
|
|
|
|
|
lpfc_cmd->status = IOSTAT_DEFAULT;
|
2012-08-14 18:26:06 +00:00
|
|
|
if (lpfc_cmd->status == IOSTAT_FCP_RSP_ERROR &&
|
|
|
|
|
!lpfc_cmd->fcp_rsp->rspStatus3 &&
|
|
|
|
|
(lpfc_cmd->fcp_rsp->rspStatus2 & RESID_UNDER) &&
|
|
|
|
|
!(vport->cfg_log_verbose & LOG_FCP_UNDER))
|
2011-10-11 01:32:10 +00:00
|
|
|
logit = 0;
|
|
|
|
|
else
|
|
|
|
|
logit = LOG_FCP | LOG_FCP_UNDER;
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, logit,
|
2014-06-25 13:27:36 +00:00
|
|
|
"9030 FCP cmd x%x failed <%d/%lld> "
|
2012-05-10 01:18:20 +00:00
|
|
|
"status: x%x result: x%x "
|
|
|
|
|
"sid: x%x did: x%x oxid: x%x "
|
|
|
|
|
"Data: x%x x%x\n",
|
2011-10-11 01:32:10 +00:00
|
|
|
cmd->cmnd[0],
|
|
|
|
|
cmd->device ? cmd->device->id : 0xffff,
|
|
|
|
|
cmd->device ? cmd->device->lun : 0xffff,
|
|
|
|
|
lpfc_cmd->status, lpfc_cmd->result,
|
2013-05-31 21:03:18 +00:00
|
|
|
vport->fc_myDID,
|
|
|
|
|
(pnode) ? pnode->nlp_DID : 0,
|
2012-05-10 01:18:20 +00:00
|
|
|
phba->sli_rev == LPFC_SLI_REV4 ?
|
|
|
|
|
lpfc_cmd->cur_iocbq.sli4_xritag : 0xffff,
|
2011-10-11 01:32:10 +00:00
|
|
|
pIocbOut->iocb.ulpContext,
|
|
|
|
|
lpfc_cmd->cur_iocbq.iocb.ulpIoTag);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
switch (lpfc_cmd->status) {
|
|
|
|
|
case IOSTAT_FCP_RSP_ERROR:
|
|
|
|
|
/* Call FCP RSP handler to determine result */
|
2020-11-15 19:26:43 +00:00
|
|
|
lpfc_handle_fcp_err(vport, lpfc_cmd,
|
|
|
|
|
pIocbOut->iocb.un.fcpi.fcpi_parm);
|
2005-04-17 21:05:31 +00:00
|
|
|
break;
|
|
|
|
|
case IOSTAT_NPORT_BSY:
|
|
|
|
|
case IOSTAT_FABRIC_BSY:
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_TRANSPORT_DISRUPTED << 16;
|
2008-09-07 15:52:10 +00:00
|
|
|
fast_path_evt = lpfc_alloc_fast_evt(phba);
|
|
|
|
|
if (!fast_path_evt)
|
|
|
|
|
break;
|
|
|
|
|
fast_path_evt->un.fabric_evt.event_type =
|
|
|
|
|
FC_REG_FABRIC_EVENT;
|
|
|
|
|
fast_path_evt->un.fabric_evt.subcategory =
|
|
|
|
|
(lpfc_cmd->status == IOSTAT_NPORT_BSY) ?
|
|
|
|
|
LPFC_EVENT_PORT_BUSY : LPFC_EVENT_FABRIC_BUSY;
|
2020-11-15 19:26:30 +00:00
|
|
|
if (pnode) {
|
2008-09-07 15:52:10 +00:00
|
|
|
memcpy(&fast_path_evt->un.fabric_evt.wwpn,
|
|
|
|
|
&pnode->nlp_portname,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(&fast_path_evt->un.fabric_evt.wwnn,
|
|
|
|
|
&pnode->nlp_nodename,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
}
|
|
|
|
|
fast_path_evt->vport = vport;
|
|
|
|
|
fast_path_evt->work_evt.evt =
|
|
|
|
|
LPFC_EVT_FASTPATH_MGMT_EVT;
|
|
|
|
|
spin_lock_irqsave(&phba->hbalock, flags);
|
|
|
|
|
list_add_tail(&fast_path_evt->work_evt.evt_listp,
|
|
|
|
|
&phba->work_list);
|
|
|
|
|
spin_unlock_irqrestore(&phba->hbalock, flags);
|
|
|
|
|
lpfc_worker_wake_up(phba);
|
2005-04-17 21:05:31 +00:00
|
|
|
break;
|
2007-06-18 00:56:38 +00:00
|
|
|
case IOSTAT_LOCAL_REJECT:
|
2011-02-16 17:39:35 +00:00
|
|
|
case IOSTAT_REMOTE_STOP:
|
2011-02-16 17:39:57 +00:00
|
|
|
if (lpfc_cmd->result == IOERR_ELXSEC_KEY_UNWRAP_ERROR ||
|
|
|
|
|
lpfc_cmd->result ==
|
|
|
|
|
IOERR_ELXSEC_KEY_UNWRAP_COMPARE_ERROR ||
|
|
|
|
|
lpfc_cmd->result == IOERR_ELXSEC_CRYPTO_ERROR ||
|
|
|
|
|
lpfc_cmd->result ==
|
|
|
|
|
IOERR_ELXSEC_CRYPTO_COMPARE_ERROR) {
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_NO_CONNECT << 16;
|
2011-02-16 17:39:57 +00:00
|
|
|
break;
|
|
|
|
|
}
|
2008-08-25 01:50:00 +00:00
|
|
|
if (lpfc_cmd->result == IOERR_INVALID_RPI ||
|
2007-06-18 00:56:38 +00:00
|
|
|
lpfc_cmd->result == IOERR_NO_RESOURCES ||
|
2010-06-07 19:24:12 +00:00
|
|
|
lpfc_cmd->result == IOERR_ABORT_REQUESTED ||
|
|
|
|
|
lpfc_cmd->result == IOERR_SLER_CMD_RCV_FAILURE) {
|
2022-08-24 06:00:33 +00:00
|
|
|
cmd->result = DID_TRANSPORT_DISRUPTED << 16;
|
2008-04-07 14:15:56 +00:00
|
|
|
break;
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
if ((lpfc_cmd->result == IOERR_RX_DMA_FAILED ||
|
|
|
|
|
lpfc_cmd->result == IOERR_TX_DMA_FAILED) &&
|
|
|
|
|
pIocbOut->iocb.unsli3.sli3_bg.bgstat) {
|
|
|
|
|
if (scsi_get_prot_op(cmd) != SCSI_PROT_NORMAL) {
|
|
|
|
|
/*
|
|
|
|
|
* This is a response for a BG enabled
|
|
|
|
|
* cmd. Parse BG error
|
|
|
|
|
*/
|
|
|
|
|
lpfc_parse_bg_err(phba, lpfc_cmd,
|
|
|
|
|
pIocbOut);
|
|
|
|
|
break;
|
|
|
|
|
} else {
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING,
|
|
|
|
|
LOG_BG,
|
|
|
|
|
"9031 non-zero BGSTAT "
|
2009-10-02 19:16:51 +00:00
|
|
|
"on unprotected cmd\n");
|
2008-12-05 03:40:02 +00:00
|
|
|
}
|
|
|
|
|
}
|
2011-02-16 17:39:35 +00:00
|
|
|
if ((lpfc_cmd->status == IOSTAT_REMOTE_STOP)
|
|
|
|
|
&& (phba->sli_rev == LPFC_SLI_REV4)
|
2020-11-15 19:26:30 +00:00
|
|
|
&& pnode) {
|
2011-02-16 17:39:35 +00:00
|
|
|
/* This IO was aborted by the target, we don't
|
|
|
|
|
* know the rxid and because we did not send the
|
|
|
|
|
* ABTS we cannot generate and RRQ.
|
|
|
|
|
*/
|
|
|
|
|
lpfc_set_rrq_active(phba, pnode,
|
2012-05-10 01:19:14 +00:00
|
|
|
lpfc_cmd->cur_iocbq.sli4_lxritag,
|
|
|
|
|
0, 0);
|
2011-02-16 17:39:35 +00:00
|
|
|
}
|
2020-08-23 22:36:59 +00:00
|
|
|
fallthrough;
|
2005-04-17 21:05:31 +00:00
|
|
|
default:
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_ERROR << 16;
|
2005-04-17 21:05:31 +00:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!pnode || (pnode->nlp_state != NLP_STE_MAPPED_NODE))
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_TRANSPORT_DISRUPTED << 16 |
|
|
|
|
|
SAM_STAT_BUSY;
|
2011-02-16 17:39:57 +00:00
|
|
|
} else
|
2018-07-05 11:01:39 +00:00
|
|
|
cmd->result = DID_OK << 16;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
if (cmd->result || lpfc_cmd->fcp_rsp->rspSnsLen) {
|
|
|
|
|
uint32_t *lp = (uint32_t *)cmd->sense_buffer;
|
|
|
|
|
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2019-08-14 23:57:06 +00:00
|
|
|
"0710 Iodone <%d/%llu> cmd x%px, error "
|
2007-08-02 15:10:09 +00:00
|
|
|
"x%x SNS x%x x%x Data: x%x x%x\n",
|
|
|
|
|
cmd->device->id, cmd->device->lun, cmd,
|
|
|
|
|
cmd->result, *lp, *(lp + 3), cmd->retries,
|
|
|
|
|
scsi_get_resid(cmd));
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2008-09-07 15:52:04 +00:00
|
|
|
if (vport->cfg_max_scsicmpl_time &&
|
|
|
|
|
time_after(jiffies, lpfc_cmd->start_time +
|
|
|
|
|
msecs_to_jiffies(vport->cfg_max_scsicmpl_time))) {
|
2009-04-06 22:48:10 +00:00
|
|
|
spin_lock_irqsave(shost->host_lock, flags);
|
2020-11-15 19:26:30 +00:00
|
|
|
if (pnode) {
|
2008-12-05 03:39:08 +00:00
|
|
|
if (pnode->cmd_qdepth >
|
|
|
|
|
atomic_read(&pnode->cmd_pending) &&
|
|
|
|
|
(atomic_read(&pnode->cmd_pending) >
|
|
|
|
|
LPFC_MIN_TGT_QDEPTH) &&
|
|
|
|
|
((cmd->cmnd[0] == READ_10) ||
|
|
|
|
|
(cmd->cmnd[0] == WRITE_10)))
|
|
|
|
|
pnode->cmd_qdepth =
|
|
|
|
|
atomic_read(&pnode->cmd_pending);
|
|
|
|
|
|
|
|
|
|
pnode->last_change_time = jiffies;
|
|
|
|
|
}
|
2009-04-06 22:48:10 +00:00
|
|
|
spin_unlock_irqrestore(shost->host_lock, flags);
|
2008-09-07 15:52:04 +00:00
|
|
|
}
|
2007-04-25 13:51:30 +00:00
|
|
|
lpfc_scsi_unprep_dma_buf(phba, lpfc_cmd);
|
2009-04-06 22:48:10 +00:00
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
lpfc_cmd->pCmd = NULL;
|
|
|
|
|
spin_unlock(&lpfc_cmd->buf_lock);
|
2012-08-14 18:26:28 +00:00
|
|
|
|
2020-03-22 18:13:02 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
if (lpfc_cmd->ts_cmd_start) {
|
|
|
|
|
lpfc_cmd->ts_isr_cmpl = pIocbIn->isr_timestamp;
|
|
|
|
|
lpfc_cmd->ts_data_io = ktime_get_ns();
|
|
|
|
|
phba->ktime_last_cmd = lpfc_cmd->ts_data_io;
|
|
|
|
|
lpfc_io_ktime(phba, lpfc_cmd);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
2021-05-14 19:55:57 +00:00
|
|
|
|
2016-10-13 22:06:15 +00:00
|
|
|
/* The sdev is not guaranteed to be valid post scsi_done upcall. */
|
2021-10-07 20:28:41 +00:00
|
|
|
scsi_done(cmd);
|
2016-10-13 22:06:15 +00:00
|
|
|
|
2008-01-11 06:52:42 +00:00
|
|
|
/*
|
2019-01-28 19:14:37 +00:00
|
|
|
* If there is an abort thread waiting for command completion
|
2008-01-11 06:52:42 +00:00
|
|
|
* wake up the thread.
|
|
|
|
|
*/
|
2019-01-28 19:14:37 +00:00
|
|
|
spin_lock(&lpfc_cmd->buf_lock);
|
2022-02-25 02:22:52 +00:00
|
|
|
lpfc_cmd->cur_iocbq.cmd_flag &= ~LPFC_DRIVER_ABORTED;
|
2019-05-22 00:49:01 +00:00
|
|
|
if (lpfc_cmd->waitq)
|
2019-04-03 18:10:34 +00:00
|
|
|
wake_up(lpfc_cmd->waitq);
|
2019-01-28 19:14:37 +00:00
|
|
|
spin_unlock(&lpfc_cmd->buf_lock);
|
2008-01-11 06:52:42 +00:00
|
|
|
|
2005-10-29 00:30:02 +00:00
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2015-04-28 08:26:22 +00:00
|
|
|
/**
|
2020-11-15 19:26:42 +00:00
|
|
|
* lpfc_scsi_prep_cmnd_buf_s3 - SLI-3 IOCB init for the IO
|
2021-03-12 09:47:11 +00:00
|
|
|
* @vport: Pointer to vport object.
|
2020-11-15 19:26:42 +00:00
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be prep'ed.
|
|
|
|
|
* @tmo: timeout value for the IO
|
2015-04-28 08:26:22 +00:00
|
|
|
*
|
2020-11-15 19:26:42 +00:00
|
|
|
* Based on the data-direction of the command, initialize IOCB
|
|
|
|
|
* in the I/O buffer. Fill in the IOCB fields which are independent
|
|
|
|
|
* of the scsi buffer
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
2020-11-15 19:26:42 +00:00
|
|
|
* RETURNS 0 - SUCCESS,
|
2008-12-05 03:39:02 +00:00
|
|
|
**/
|
2020-11-15 19:26:42 +00:00
|
|
|
static int lpfc_scsi_prep_cmnd_buf_s3(struct lpfc_vport *vport,
|
|
|
|
|
struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
uint8_t tmo)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
2020-11-15 19:26:42 +00:00
|
|
|
IOCB_t *iocb_cmd = &lpfc_cmd->cur_iocbq.iocb;
|
|
|
|
|
struct lpfc_iocbq *piocbq = &lpfc_cmd->cur_iocbq;
|
2005-04-17 21:05:31 +00:00
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
2020-11-15 19:26:42 +00:00
|
|
|
struct lpfc_nodelist *pnode = lpfc_cmd->ndlp;
|
2005-04-17 21:05:31 +00:00
|
|
|
int datadir = scsi_cmnd->sc_data_direction;
|
2020-11-15 19:26:42 +00:00
|
|
|
u32 fcpdl;
|
2008-04-07 14:15:56 +00:00
|
|
|
|
2014-04-04 17:51:25 +00:00
|
|
|
piocbq->iocb.un.fcpi.fcpi_XRdy = 0;
|
2012-08-03 16:35:44 +00:00
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
|
|
|
|
* There are three possibilities here - use scatter-gather segment, use
|
|
|
|
|
* the single mapping, or neither. Start the lpfc command prep by
|
|
|
|
|
* bumping the bpl beyond the fcp_cmnd and fcp_rsp regions to the first
|
|
|
|
|
* data bde entry.
|
|
|
|
|
*/
|
2007-06-18 00:56:39 +00:00
|
|
|
if (scsi_sg_count(scsi_cmnd)) {
|
2005-04-17 21:05:31 +00:00
|
|
|
if (datadir == DMA_TO_DEVICE) {
|
|
|
|
|
iocb_cmd->ulpCommand = CMD_FCP_IWRITE64_CR;
|
2013-07-15 22:34:05 +00:00
|
|
|
iocb_cmd->ulpPU = PARM_READ_CHECK;
|
2013-07-15 22:35:04 +00:00
|
|
|
if (vport->cfg_first_burst_size &&
|
|
|
|
|
(pnode->nlp_flag & NLP_FIRSTBURST)) {
|
2020-11-15 19:26:42 +00:00
|
|
|
u32 xrdy_len;
|
|
|
|
|
|
2013-09-06 16:18:45 +00:00
|
|
|
fcpdl = scsi_bufflen(scsi_cmnd);
|
2020-11-15 19:26:42 +00:00
|
|
|
xrdy_len = min(fcpdl,
|
|
|
|
|
vport->cfg_first_burst_size);
|
|
|
|
|
piocbq->iocb.un.fcpi.fcpi_XRdy = xrdy_len;
|
2013-07-15 22:35:04 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
fcp_cmnd->fcpCntl3 = WRITE_DATA;
|
|
|
|
|
} else {
|
|
|
|
|
iocb_cmd->ulpCommand = CMD_FCP_IREAD64_CR;
|
|
|
|
|
iocb_cmd->ulpPU = PARM_READ_CHECK;
|
|
|
|
|
fcp_cmnd->fcpCntl3 = READ_DATA;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
iocb_cmd->ulpCommand = CMD_FCP_ICMND64_CR;
|
|
|
|
|
iocb_cmd->un.fcpi.fcpi_parm = 0;
|
|
|
|
|
iocb_cmd->ulpPU = 0;
|
|
|
|
|
fcp_cmnd->fcpCntl3 = 0;
|
|
|
|
|
}
|
2020-11-15 19:26:42 +00:00
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
|
|
|
|
* Finish initializing those IOCB fields that are independent
|
|
|
|
|
* of the scsi_cmnd request_buffer
|
|
|
|
|
*/
|
|
|
|
|
piocbq->iocb.ulpContext = pnode->nlp_rpi;
|
|
|
|
|
if (pnode->nlp_fcp_info & NLP_FCP_2_DEVICE)
|
|
|
|
|
piocbq->iocb.ulpFCP2Rcvy = 1;
|
2008-01-11 06:52:54 +00:00
|
|
|
else
|
|
|
|
|
piocbq->iocb.ulpFCP2Rcvy = 0;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
piocbq->iocb.ulpClass = (pnode->nlp_fcp_info & 0x0f);
|
2022-04-12 22:20:02 +00:00
|
|
|
piocbq->io_buf = lpfc_cmd;
|
2022-02-25 02:22:52 +00:00
|
|
|
if (!piocbq->cmd_cmpl)
|
|
|
|
|
piocbq->cmd_cmpl = lpfc_scsi_cmd_iocb_cmpl;
|
2020-11-15 19:26:42 +00:00
|
|
|
piocbq->iocb.ulpTimeout = tmo;
|
2007-06-18 00:56:36 +00:00
|
|
|
piocbq->vport = vport;
|
2020-11-15 19:26:42 +00:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_cmnd_buf_s4 - SLI-4 WQE init for the IO
|
2021-03-12 09:47:11 +00:00
|
|
|
* @vport: Pointer to vport object.
|
2020-11-15 19:26:42 +00:00
|
|
|
* @lpfc_cmd: The scsi buffer which is going to be prep'ed.
|
|
|
|
|
* @tmo: timeout value for the IO
|
|
|
|
|
*
|
|
|
|
|
* Based on the data-direction of the command copy WQE template
|
|
|
|
|
* to I/O buffer WQE. Fill in the WQE fields which are independent
|
|
|
|
|
* of the scsi buffer
|
|
|
|
|
*
|
|
|
|
|
* RETURNS 0 - SUCCESS,
|
|
|
|
|
**/
|
|
|
|
|
static int lpfc_scsi_prep_cmnd_buf_s4(struct lpfc_vport *vport,
|
|
|
|
|
struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
uint8_t tmo)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
|
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
|
|
|
|
struct lpfc_sli4_hdw_queue *hdwq = NULL;
|
|
|
|
|
struct lpfc_iocbq *pwqeq = &lpfc_cmd->cur_iocbq;
|
|
|
|
|
struct lpfc_nodelist *pnode = lpfc_cmd->ndlp;
|
|
|
|
|
union lpfc_wqe128 *wqe = &pwqeq->wqe;
|
|
|
|
|
u16 idx = lpfc_cmd->hdwq_no;
|
|
|
|
|
int datadir = scsi_cmnd->sc_data_direction;
|
|
|
|
|
|
|
|
|
|
hdwq = &phba->sli4_hba.hdwq[idx];
|
|
|
|
|
|
|
|
|
|
/* Initialize 64 bytes only */
|
|
|
|
|
memset(wqe, 0, sizeof(union lpfc_wqe128));
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* There are three possibilities here - use scatter-gather segment, use
|
|
|
|
|
* the single mapping, or neither.
|
|
|
|
|
*/
|
|
|
|
|
if (scsi_sg_count(scsi_cmnd)) {
|
|
|
|
|
if (datadir == DMA_TO_DEVICE) {
|
|
|
|
|
/* From the iwrite template, initialize words 7 - 11 */
|
|
|
|
|
memcpy(&wqe->words[7],
|
|
|
|
|
&lpfc_iwrite_cmd_template.words[7],
|
|
|
|
|
sizeof(uint32_t) * 5);
|
|
|
|
|
|
|
|
|
|
fcp_cmnd->fcpCntl3 = WRITE_DATA;
|
|
|
|
|
if (hdwq)
|
|
|
|
|
hdwq->scsi_cstat.output_requests++;
|
|
|
|
|
} else {
|
|
|
|
|
/* From the iread template, initialize words 7 - 11 */
|
|
|
|
|
memcpy(&wqe->words[7],
|
|
|
|
|
&lpfc_iread_cmd_template.words[7],
|
|
|
|
|
sizeof(uint32_t) * 5);
|
|
|
|
|
|
|
|
|
|
/* Word 7 */
|
|
|
|
|
bf_set(wqe_tmo, &wqe->fcp_iread.wqe_com, tmo);
|
|
|
|
|
|
|
|
|
|
fcp_cmnd->fcpCntl3 = READ_DATA;
|
|
|
|
|
if (hdwq)
|
|
|
|
|
hdwq->scsi_cstat.input_requests++;
|
2021-08-16 16:28:54 +00:00
|
|
|
|
|
|
|
|
/* For a CMF Managed port, iod must be zero'ed */
|
|
|
|
|
if (phba->cmf_active_mode == LPFC_CFG_MANAGED)
|
|
|
|
|
bf_set(wqe_iod, &wqe->fcp_iread.wqe_com,
|
|
|
|
|
LPFC_WQE_IOD_NONE);
|
2020-11-15 19:26:42 +00:00
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
/* From the icmnd template, initialize words 4 - 11 */
|
|
|
|
|
memcpy(&wqe->words[4], &lpfc_icmnd_cmd_template.words[4],
|
|
|
|
|
sizeof(uint32_t) * 8);
|
|
|
|
|
|
|
|
|
|
/* Word 7 */
|
|
|
|
|
bf_set(wqe_tmo, &wqe->fcp_icmd.wqe_com, tmo);
|
|
|
|
|
|
|
|
|
|
fcp_cmnd->fcpCntl3 = 0;
|
|
|
|
|
if (hdwq)
|
|
|
|
|
hdwq->scsi_cstat.control_requests++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Finish initializing those WQE fields that are independent
|
|
|
|
|
* of the request_buffer
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/* Word 3 */
|
|
|
|
|
bf_set(payload_offset_len, &wqe->fcp_icmd,
|
|
|
|
|
sizeof(struct fcp_cmnd) + sizeof(struct fcp_rsp));
|
|
|
|
|
|
|
|
|
|
/* Word 6 */
|
|
|
|
|
bf_set(wqe_ctxt_tag, &wqe->generic.wqe_com,
|
|
|
|
|
phba->sli4_hba.rpi_ids[pnode->nlp_rpi]);
|
|
|
|
|
bf_set(wqe_xri_tag, &wqe->generic.wqe_com, pwqeq->sli4_xritag);
|
|
|
|
|
|
|
|
|
|
/* Word 7*/
|
|
|
|
|
if (pnode->nlp_fcp_info & NLP_FCP_2_DEVICE)
|
|
|
|
|
bf_set(wqe_erp, &wqe->generic.wqe_com, 1);
|
|
|
|
|
|
|
|
|
|
bf_set(wqe_class, &wqe->generic.wqe_com,
|
|
|
|
|
(pnode->nlp_fcp_info & 0x0f));
|
|
|
|
|
|
|
|
|
|
/* Word 8 */
|
|
|
|
|
wqe->generic.wqe_com.abort_tag = pwqeq->iotag;
|
|
|
|
|
|
|
|
|
|
/* Word 9 */
|
|
|
|
|
bf_set(wqe_reqtag, &wqe->generic.wqe_com, pwqeq->iotag);
|
|
|
|
|
|
|
|
|
|
pwqeq->vport = vport;
|
2022-04-12 22:20:02 +00:00
|
|
|
pwqeq->io_buf = lpfc_cmd;
|
2020-11-15 19:26:42 +00:00
|
|
|
pwqeq->hba_wqidx = lpfc_cmd->hdwq_no;
|
2022-02-25 02:22:52 +00:00
|
|
|
pwqeq->cmd_cmpl = lpfc_fcp_io_cmd_wqe_cmpl;
|
2020-11-15 19:26:42 +00:00
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_cmnd - Wrapper func for convert scsi cmnd to FCP info unit
|
|
|
|
|
* @vport: The virtual port for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: The scsi command which needs to send.
|
|
|
|
|
* @pnode: Pointer to lpfc_nodelist.
|
|
|
|
|
*
|
|
|
|
|
* This routine initializes fcp_cmnd and iocb data structure from scsi command
|
|
|
|
|
* to transfer for device with SLI3 interface spec.
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_scsi_prep_cmnd(struct lpfc_vport *vport, struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
struct lpfc_nodelist *pnode)
|
|
|
|
|
{
|
|
|
|
|
struct scsi_cmnd *scsi_cmnd = lpfc_cmd->pCmd;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
|
|
|
|
u8 *ptr;
|
|
|
|
|
|
|
|
|
|
if (!pnode)
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
lpfc_cmd->fcp_rsp->rspSnsLen = 0;
|
|
|
|
|
/* clear task management bits */
|
|
|
|
|
lpfc_cmd->fcp_cmnd->fcpCntl2 = 0;
|
|
|
|
|
|
|
|
|
|
int_to_scsilun(lpfc_cmd->pCmd->device->lun,
|
|
|
|
|
&lpfc_cmd->fcp_cmnd->fcp_lun);
|
|
|
|
|
|
|
|
|
|
ptr = &fcp_cmnd->fcpCdb[0];
|
|
|
|
|
memcpy(ptr, scsi_cmnd->cmnd, scsi_cmnd->cmd_len);
|
|
|
|
|
if (scsi_cmnd->cmd_len < LPFC_FCP_CDB_LEN) {
|
|
|
|
|
ptr += scsi_cmnd->cmd_len;
|
|
|
|
|
memset(ptr, 0, (LPFC_FCP_CDB_LEN - scsi_cmnd->cmd_len));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
fcp_cmnd->fcpCntl1 = SIMPLE_Q;
|
|
|
|
|
|
|
|
|
|
lpfc_scsi_prep_cmnd_buf(vport, lpfc_cmd, lpfc_cmd->timeout);
|
|
|
|
|
|
|
|
|
|
return 0;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2009-05-22 18:51:39 +00:00
|
|
|
/**
|
2022-02-25 02:23:04 +00:00
|
|
|
* lpfc_scsi_prep_task_mgmt_cmd_s3 - Convert SLI3 scsi TM cmd to FCP info unit
|
2008-12-05 03:39:02 +00:00
|
|
|
* @vport: The virtual port for which this call is being executed.
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
* @lpfc_cmd: Pointer to lpfc_io_buf data structure.
|
2008-12-05 03:39:02 +00:00
|
|
|
* @lun: Logical unit number.
|
|
|
|
|
* @task_mgmt_cmd: SCSI task management command.
|
|
|
|
|
*
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* This routine creates FCP information unit corresponding to @task_mgmt_cmd
|
|
|
|
|
* for device with SLI-3 interface spec.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* 0 - Error
|
|
|
|
|
* 1 - Success
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
2022-02-25 02:23:04 +00:00
|
|
|
lpfc_scsi_prep_task_mgmt_cmd_s3(struct lpfc_vport *vport,
|
|
|
|
|
struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
u64 lun, u8 task_mgmt_cmd)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
|
|
|
|
struct lpfc_iocbq *piocbq;
|
|
|
|
|
IOCB_t *piocb;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd;
|
2006-05-02 01:50:40 +00:00
|
|
|
struct lpfc_rport_data *rdata = lpfc_cmd->rdata;
|
2005-04-17 21:05:31 +00:00
|
|
|
struct lpfc_nodelist *ndlp = rdata->pnode;
|
|
|
|
|
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!ndlp || ndlp->nlp_state != NLP_STE_MAPPED_NODE)
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
piocbq = &(lpfc_cmd->cur_iocbq);
|
2007-06-18 00:56:36 +00:00
|
|
|
piocbq->vport = vport;
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
piocb = &piocbq->iocb;
|
|
|
|
|
|
|
|
|
|
fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
2008-08-25 01:49:55 +00:00
|
|
|
/* Clear out any old data in the FCP command area */
|
|
|
|
|
memset(fcp_cmnd, 0, sizeof(struct fcp_cmnd));
|
|
|
|
|
int_to_scsilun(lun, &fcp_cmnd->fcp_lun);
|
2005-04-17 21:05:31 +00:00
|
|
|
fcp_cmnd->fcpCntl2 = task_mgmt_cmd;
|
2022-02-25 02:23:04 +00:00
|
|
|
if (!(vport->phba->sli3_options & LPFC_SLI3_BG_ENABLED))
|
2008-08-25 01:49:55 +00:00
|
|
|
lpfc_fcpcmd_to_iocb(piocb->unsli3.fcp_ext.icd, fcp_cmnd);
|
2005-04-17 21:05:31 +00:00
|
|
|
piocb->ulpCommand = CMD_FCP_ICMND64_CR;
|
|
|
|
|
piocb->ulpContext = ndlp->nlp_rpi;
|
2013-10-10 16:24:07 +00:00
|
|
|
piocb->ulpFCP2Rcvy = (ndlp->nlp_fcp_info & NLP_FCP_2_DEVICE) ? 1 : 0;
|
2005-04-17 21:05:31 +00:00
|
|
|
piocb->ulpClass = (ndlp->nlp_fcp_info & 0x0f);
|
2014-02-20 14:57:28 +00:00
|
|
|
piocb->ulpPU = 0;
|
|
|
|
|
piocb->un.fcpi.fcpi_parm = 0;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
|
|
|
|
/* ulpTimeout is only one byte */
|
|
|
|
|
if (lpfc_cmd->timeout > 0xff) {
|
|
|
|
|
/*
|
|
|
|
|
* Do not timeout the command at the firmware level.
|
|
|
|
|
* The driver will provide the timeout mechanism.
|
|
|
|
|
*/
|
|
|
|
|
piocb->ulpTimeout = 0;
|
2009-06-10 21:22:44 +00:00
|
|
|
} else
|
2005-04-17 21:05:31 +00:00
|
|
|
piocb->ulpTimeout = lpfc_cmd->timeout;
|
2009-05-22 18:51:39 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_scsi_prep_task_mgmt_cmd_s4 - Convert SLI4 scsi TM cmd to FCP info unit
|
|
|
|
|
* @vport: The virtual port for which this call is being executed.
|
|
|
|
|
* @lpfc_cmd: Pointer to lpfc_io_buf data structure.
|
|
|
|
|
* @lun: Logical unit number.
|
|
|
|
|
* @task_mgmt_cmd: SCSI task management command.
|
|
|
|
|
*
|
|
|
|
|
* This routine creates FCP information unit corresponding to @task_mgmt_cmd
|
|
|
|
|
* for device with SLI-4 interface spec.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* 0 - Error
|
|
|
|
|
* 1 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_scsi_prep_task_mgmt_cmd_s4(struct lpfc_vport *vport,
|
|
|
|
|
struct lpfc_io_buf *lpfc_cmd,
|
|
|
|
|
u64 lun, u8 task_mgmt_cmd)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_iocbq *pwqeq = &lpfc_cmd->cur_iocbq;
|
|
|
|
|
union lpfc_wqe128 *wqe = &pwqeq->wqe;
|
|
|
|
|
struct fcp_cmnd *fcp_cmnd;
|
|
|
|
|
struct lpfc_rport_data *rdata = lpfc_cmd->rdata;
|
|
|
|
|
struct lpfc_nodelist *ndlp = rdata->pnode;
|
|
|
|
|
|
|
|
|
|
if (!ndlp || ndlp->nlp_state != NLP_STE_MAPPED_NODE)
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
pwqeq->vport = vport;
|
|
|
|
|
/* Initialize 64 bytes only */
|
|
|
|
|
memset(wqe, 0, sizeof(union lpfc_wqe128));
|
|
|
|
|
|
|
|
|
|
/* From the icmnd template, initialize words 4 - 11 */
|
|
|
|
|
memcpy(&wqe->words[4], &lpfc_icmnd_cmd_template.words[4],
|
|
|
|
|
sizeof(uint32_t) * 8);
|
|
|
|
|
|
|
|
|
|
fcp_cmnd = lpfc_cmd->fcp_cmnd;
|
|
|
|
|
/* Clear out any old data in the FCP command area */
|
|
|
|
|
memset(fcp_cmnd, 0, sizeof(struct fcp_cmnd));
|
|
|
|
|
int_to_scsilun(lun, &fcp_cmnd->fcp_lun);
|
|
|
|
|
fcp_cmnd->fcpCntl3 = 0;
|
|
|
|
|
fcp_cmnd->fcpCntl2 = task_mgmt_cmd;
|
|
|
|
|
|
|
|
|
|
bf_set(payload_offset_len, &wqe->fcp_icmd,
|
|
|
|
|
sizeof(struct fcp_cmnd) + sizeof(struct fcp_rsp));
|
|
|
|
|
bf_set(cmd_buff_len, &wqe->fcp_icmd, 0);
|
|
|
|
|
bf_set(wqe_ctxt_tag, &wqe->generic.wqe_com, /* ulpContext */
|
|
|
|
|
vport->phba->sli4_hba.rpi_ids[ndlp->nlp_rpi]);
|
|
|
|
|
bf_set(wqe_erp, &wqe->fcp_icmd.wqe_com,
|
|
|
|
|
((ndlp->nlp_fcp_info & NLP_FCP_2_DEVICE) ? 1 : 0));
|
|
|
|
|
bf_set(wqe_class, &wqe->fcp_icmd.wqe_com,
|
|
|
|
|
(ndlp->nlp_fcp_info & 0x0f));
|
|
|
|
|
|
|
|
|
|
/* ulpTimeout is only one byte */
|
|
|
|
|
if (lpfc_cmd->timeout > 0xff) {
|
|
|
|
|
/*
|
|
|
|
|
* Do not timeout the command at the firmware level.
|
|
|
|
|
* The driver will provide the timeout mechanism.
|
|
|
|
|
*/
|
|
|
|
|
bf_set(wqe_tmo, &wqe->fcp_icmd.wqe_com, 0);
|
|
|
|
|
} else {
|
|
|
|
|
bf_set(wqe_tmo, &wqe->fcp_icmd.wqe_com, lpfc_cmd->timeout);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
lpfc_prep_embed_io(vport->phba, lpfc_cmd);
|
|
|
|
|
bf_set(wqe_xri_tag, &wqe->generic.wqe_com, pwqeq->sli4_xritag);
|
|
|
|
|
wqe->generic.wqe_com.abort_tag = pwqeq->iotag;
|
|
|
|
|
bf_set(wqe_reqtag, &wqe->generic.wqe_com, pwqeq->iotag);
|
|
|
|
|
|
|
|
|
|
lpfc_sli4_set_rsp_sgl_last(vport->phba, lpfc_cmd);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
|
2009-06-10 21:22:44 +00:00
|
|
|
return 1;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
2011-03-31 01:57:33 +00:00
|
|
|
* lpfc_scsi_api_table_setup - Set up scsi api function jump table
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
* @phba: The hba struct for which this call is being executed.
|
|
|
|
|
* @dev_grp: The HBA PCI-Device group number.
|
|
|
|
|
*
|
|
|
|
|
* This routine sets up the SCSI interface API function jump table in @phba
|
|
|
|
|
* struct.
|
|
|
|
|
* Returns: 0 - success, -ENODEV - failure.
|
|
|
|
|
**/
|
|
|
|
|
int
|
|
|
|
|
lpfc_scsi_api_table_setup(struct lpfc_hba *phba, uint8_t dev_grp)
|
|
|
|
|
{
|
|
|
|
|
|
2009-06-10 21:22:44 +00:00
|
|
|
phba->lpfc_scsi_unprep_dma_buf = lpfc_scsi_unprep_dma_buf;
|
|
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
switch (dev_grp) {
|
|
|
|
|
case LPFC_PCI_DEV_LP:
|
|
|
|
|
phba->lpfc_scsi_prep_dma_buf = lpfc_scsi_prep_dma_buf_s3;
|
2012-01-18 21:25:09 +00:00
|
|
|
phba->lpfc_bg_scsi_prep_dma_buf = lpfc_bg_scsi_prep_dma_buf_s3;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
phba->lpfc_release_scsi_buf = lpfc_release_scsi_buf_s3;
|
2010-11-21 04:11:55 +00:00
|
|
|
phba->lpfc_get_scsi_buf = lpfc_get_scsi_buf_s3;
|
2020-11-15 19:26:42 +00:00
|
|
|
phba->lpfc_scsi_prep_cmnd_buf = lpfc_scsi_prep_cmnd_buf_s3;
|
2022-02-25 02:23:04 +00:00
|
|
|
phba->lpfc_scsi_prep_task_mgmt_cmd =
|
|
|
|
|
lpfc_scsi_prep_task_mgmt_cmd_s3;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
break;
|
2009-05-22 18:51:39 +00:00
|
|
|
case LPFC_PCI_DEV_OC:
|
|
|
|
|
phba->lpfc_scsi_prep_dma_buf = lpfc_scsi_prep_dma_buf_s4;
|
2012-01-18 21:25:09 +00:00
|
|
|
phba->lpfc_bg_scsi_prep_dma_buf = lpfc_bg_scsi_prep_dma_buf_s4;
|
2009-05-22 18:51:39 +00:00
|
|
|
phba->lpfc_release_scsi_buf = lpfc_release_scsi_buf_s4;
|
2010-11-21 04:11:55 +00:00
|
|
|
phba->lpfc_get_scsi_buf = lpfc_get_scsi_buf_s4;
|
2020-11-15 19:26:42 +00:00
|
|
|
phba->lpfc_scsi_prep_cmnd_buf = lpfc_scsi_prep_cmnd_buf_s4;
|
2022-02-25 02:23:04 +00:00
|
|
|
phba->lpfc_scsi_prep_task_mgmt_cmd =
|
|
|
|
|
lpfc_scsi_prep_task_mgmt_cmd_s4;
|
2009-05-22 18:51:39 +00:00
|
|
|
break;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
default:
|
2021-10-20 21:14:10 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
"1418 Invalid HBA PCI-device group: 0x%x\n",
|
|
|
|
|
dev_grp);
|
|
|
|
|
return -ENODEV;
|
|
|
|
|
}
|
|
|
|
|
phba->lpfc_rampdown_queue_depth = lpfc_rampdown_queue_depth;
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2021-03-12 09:47:11 +00:00
|
|
|
* lpfc_tskmgmt_def_cmpl - IOCB completion routine for task management command
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
* @cmdiocbq: Pointer to lpfc_iocbq data structure.
|
|
|
|
|
* @rspiocbq: Pointer to lpfc_iocbq data structure.
|
|
|
|
|
*
|
|
|
|
|
* This routine is IOCB completion routine for device reset and target reset
|
|
|
|
|
* routine. This routine release scsi buffer associated with lpfc_cmd.
|
|
|
|
|
**/
|
2007-04-25 13:52:34 +00:00
|
|
|
static void
|
|
|
|
|
lpfc_tskmgmt_def_cmpl(struct lpfc_hba *phba,
|
|
|
|
|
struct lpfc_iocbq *cmdiocbq,
|
|
|
|
|
struct lpfc_iocbq *rspiocbq)
|
|
|
|
|
{
|
2022-04-12 22:20:02 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd = cmdiocbq->io_buf;
|
2007-04-25 13:52:34 +00:00
|
|
|
if (lpfc_cmd)
|
|
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2018-12-13 23:17:57 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_check_pci_resettable - Walks list of devices on pci_dev's bus to check
|
|
|
|
|
* if issuing a pci_bus_reset is possibly unsafe
|
|
|
|
|
* @phba: lpfc_hba pointer.
|
|
|
|
|
*
|
|
|
|
|
* Description:
|
|
|
|
|
* Walks the bus_list to ensure only PCI devices with Emulex
|
|
|
|
|
* vendor id, device ids that support hot reset, and only one occurrence
|
|
|
|
|
* of function 0.
|
|
|
|
|
*
|
|
|
|
|
* Returns:
|
|
|
|
|
* -EBADSLT, detected invalid device
|
|
|
|
|
* 0, successful
|
|
|
|
|
*/
|
|
|
|
|
int
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_check_pci_resettable(struct lpfc_hba *phba)
|
2018-12-13 23:17:57 +00:00
|
|
|
{
|
|
|
|
|
const struct pci_dev *pdev = phba->pcidev;
|
|
|
|
|
struct pci_dev *ptr = NULL;
|
|
|
|
|
u8 counter = 0;
|
|
|
|
|
|
|
|
|
|
/* Walk the list of devices on the pci_dev's bus */
|
|
|
|
|
list_for_each_entry(ptr, &pdev->bus->devices, bus_list) {
|
|
|
|
|
/* Check for Emulex Vendor ID */
|
|
|
|
|
if (ptr->vendor != PCI_VENDOR_ID_EMULEX) {
|
|
|
|
|
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
|
|
|
|
|
"8346 Non-Emulex vendor found: "
|
|
|
|
|
"0x%04x\n", ptr->vendor);
|
|
|
|
|
return -EBADSLT;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Check for valid Emulex Device ID */
|
2021-07-22 22:17:18 +00:00
|
|
|
if (phba->sli_rev != LPFC_SLI_REV4 ||
|
|
|
|
|
phba->hba_flag & HBA_FCOE_MODE) {
|
2018-12-13 23:17:57 +00:00
|
|
|
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
|
2021-03-01 17:18:17 +00:00
|
|
|
"8347 Incapable PCI reset device: "
|
2018-12-13 23:17:57 +00:00
|
|
|
"0x%04x\n", ptr->device);
|
|
|
|
|
return -EBADSLT;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Check for only one function 0 ID to ensure only one HBA on
|
|
|
|
|
* secondary bus
|
|
|
|
|
*/
|
|
|
|
|
if (ptr->devfn == 0) {
|
|
|
|
|
if (++counter > 1) {
|
|
|
|
|
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
|
|
|
|
|
"8348 More than one device on "
|
|
|
|
|
"secondary bus found\n");
|
|
|
|
|
return -EBADSLT;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_info - Info entry point of scsi_host_template data structure
|
2008-12-05 03:39:02 +00:00
|
|
|
* @host: The scsi host for which this call is being executed.
|
|
|
|
|
*
|
|
|
|
|
* This routine provides module information about hba.
|
|
|
|
|
*
|
|
|
|
|
* Reutrn code:
|
|
|
|
|
* Pointer to char - Success.
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
const char *
|
|
|
|
|
lpfc_info(struct Scsi_Host *host)
|
|
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) host->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
2018-12-13 23:17:57 +00:00
|
|
|
int link_speed = 0;
|
|
|
|
|
static char lpfcinfobuf[384];
|
|
|
|
|
char tmp[384] = {0};
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2018-12-13 23:17:57 +00:00
|
|
|
memset(lpfcinfobuf, 0, sizeof(lpfcinfobuf));
|
2005-04-17 21:05:31 +00:00
|
|
|
if (phba && phba->pcidev){
|
2018-12-13 23:17:57 +00:00
|
|
|
/* Model Description */
|
|
|
|
|
scnprintf(tmp, sizeof(tmp), phba->ModelDesc);
|
|
|
|
|
if (strlcat(lpfcinfobuf, tmp, sizeof(lpfcinfobuf)) >=
|
|
|
|
|
sizeof(lpfcinfobuf))
|
|
|
|
|
goto buffer_done;
|
|
|
|
|
|
|
|
|
|
/* PCI Info */
|
|
|
|
|
scnprintf(tmp, sizeof(tmp),
|
|
|
|
|
" on PCI bus %02x device %02x irq %d",
|
|
|
|
|
phba->pcidev->bus->number, phba->pcidev->devfn,
|
|
|
|
|
phba->pcidev->irq);
|
|
|
|
|
if (strlcat(lpfcinfobuf, tmp, sizeof(lpfcinfobuf)) >=
|
|
|
|
|
sizeof(lpfcinfobuf))
|
|
|
|
|
goto buffer_done;
|
|
|
|
|
|
|
|
|
|
/* Port Number */
|
2005-04-17 21:05:31 +00:00
|
|
|
if (phba->Port[0]) {
|
2018-12-13 23:17:57 +00:00
|
|
|
scnprintf(tmp, sizeof(tmp), " port %s", phba->Port);
|
|
|
|
|
if (strlcat(lpfcinfobuf, tmp, sizeof(lpfcinfobuf)) >=
|
|
|
|
|
sizeof(lpfcinfobuf))
|
|
|
|
|
goto buffer_done;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
2018-12-13 23:17:57 +00:00
|
|
|
|
|
|
|
|
/* Link Speed */
|
2015-12-16 23:12:02 +00:00
|
|
|
link_speed = lpfc_sli_port_speed_get(phba);
|
2018-12-13 23:17:57 +00:00
|
|
|
if (link_speed != 0) {
|
|
|
|
|
scnprintf(tmp, sizeof(tmp),
|
|
|
|
|
" Logical Link Speed: %d Mbps", link_speed);
|
|
|
|
|
if (strlcat(lpfcinfobuf, tmp, sizeof(lpfcinfobuf)) >=
|
|
|
|
|
sizeof(lpfcinfobuf))
|
|
|
|
|
goto buffer_done;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* PCI resettable */
|
|
|
|
|
if (!lpfc_check_pci_resettable(phba)) {
|
|
|
|
|
scnprintf(tmp, sizeof(tmp), " PCI resettable");
|
|
|
|
|
strlcat(lpfcinfobuf, tmp, sizeof(lpfcinfobuf));
|
|
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
2018-12-13 23:17:57 +00:00
|
|
|
|
|
|
|
|
buffer_done:
|
2005-04-17 21:05:31 +00:00
|
|
|
return lpfcinfobuf;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2021-03-12 09:47:11 +00:00
|
|
|
* lpfc_poll_rearm_timer - Routine to modify fcp_poll timer of hba
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
*
|
|
|
|
|
* This routine modifies fcp_poll_timer field of @phba by cfg_poll_tmo.
|
|
|
|
|
* The default value of cfg_poll_tmo is 10 milliseconds.
|
|
|
|
|
**/
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
static __inline__ void lpfc_poll_rearm_timer(struct lpfc_hba * phba)
|
|
|
|
|
{
|
|
|
|
|
unsigned long poll_tmo_expires =
|
|
|
|
|
(jiffies + msecs_to_jiffies(phba->cfg_poll_tmo));
|
|
|
|
|
|
2017-02-12 21:52:30 +00:00
|
|
|
if (!list_empty(&phba->sli.sli3_ring[LPFC_FCP_RING].txcmplq))
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
mod_timer(&phba->fcp_poll_timer,
|
|
|
|
|
poll_tmo_expires);
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_poll_start_timer - Routine to start fcp_poll_timer of HBA
|
2008-12-05 03:39:02 +00:00
|
|
|
* @phba: The Hba for which this call is being executed.
|
|
|
|
|
*
|
|
|
|
|
* This routine starts the fcp_poll_timer of @phba.
|
|
|
|
|
**/
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
void lpfc_poll_start_timer(struct lpfc_hba * phba)
|
|
|
|
|
{
|
|
|
|
|
lpfc_poll_rearm_timer(phba);
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_poll_timeout - Restart polling timer
|
2020-11-02 14:23:43 +00:00
|
|
|
* @t: Timer construct where lpfc_hba data structure pointer is obtained.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
|
|
|
|
* This routine restarts fcp_poll timer, when FCP ring polling is enable
|
|
|
|
|
* and FCP Ring interrupt is disable.
|
|
|
|
|
**/
|
2017-09-07 03:24:26 +00:00
|
|
|
void lpfc_poll_timeout(struct timer_list *t)
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
{
|
2017-09-07 03:24:26 +00:00
|
|
|
struct lpfc_hba *phba = from_timer(phba, t, fcp_poll_timer);
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
|
|
|
|
|
if (phba->cfg_poll & ENABLE_FCP_RING_POLLING) {
|
2009-10-02 19:17:02 +00:00
|
|
|
lpfc_sli_handle_fast_ring_event(phba,
|
2017-02-12 21:52:30 +00:00
|
|
|
&phba->sli.sli3_ring[LPFC_FCP_RING], HA_R0RE_REQ);
|
2009-10-02 19:17:02 +00:00
|
|
|
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
if (phba->cfg_poll & DISABLE_FCP_RING_INT)
|
|
|
|
|
lpfc_poll_rearm_timer(phba);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-08 04:35:56 +00:00
|
|
|
/*
|
|
|
|
|
* lpfc_is_command_vm_io - get the UUID from blk cgroup
|
|
|
|
|
* @cmd: Pointer to scsi_cmnd data structure
|
|
|
|
|
* Returns UUID if present, otherwise NULL
|
|
|
|
|
*/
|
|
|
|
|
static char *lpfc_is_command_vm_io(struct scsi_cmnd *cmd)
|
|
|
|
|
{
|
2021-08-09 23:03:30 +00:00
|
|
|
struct bio *bio = scsi_cmd_to_rq(cmd)->bio;
|
2021-06-08 04:35:56 +00:00
|
|
|
|
2022-04-20 04:27:12 +00:00
|
|
|
if (!IS_ENABLED(CONFIG_BLK_CGROUP_FC_APPID) || !bio)
|
|
|
|
|
return NULL;
|
|
|
|
|
return blkcg_get_fc_appid(bio);
|
2021-06-08 04:35:56 +00:00
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_queuecommand - scsi_host_template queuecommand entry point
|
2020-11-02 14:23:43 +00:00
|
|
|
* @shost: kernel scsi host pointer.
|
2008-12-05 03:39:02 +00:00
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
|
|
|
|
*
|
|
|
|
|
* Driver registers this routine to scsi midlayer to submit a @cmd to process.
|
|
|
|
|
* This routine prepares an IOCB from scsi command and provides to firmware.
|
|
|
|
|
* The @done callback is invoked after driver finished processing the command.
|
|
|
|
|
*
|
|
|
|
|
* Return value :
|
|
|
|
|
* 0 - Success
|
|
|
|
|
* SCSI_MLQUEUE_HOST_BUSY - Block all devices served by this host temporarily.
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
2012-08-03 16:35:24 +00:00
|
|
|
lpfc_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *cmnd)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) shost->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
2022-02-25 02:23:04 +00:00
|
|
|
struct lpfc_iocbq *cur_iocbq = NULL;
|
2014-02-20 14:56:45 +00:00
|
|
|
struct lpfc_rport_data *rdata;
|
2009-11-18 20:40:49 +00:00
|
|
|
struct lpfc_nodelist *ndlp;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd;
|
[SCSI] update fc_transport for removal of block/unblock functions
We recently went back to implement a board reset. When we perform the
reset, we wanted to tear down the internal data structures and rebuild
them. Unfortunately, when it came to the rport structure, things were
odd. If we deleted them, the scsi targets and sdevs would be
torn down. Not a good thing for a temporary reset. We could block the
rports, but we either maintain the internal structures to keep the
rport reference (perhaps even replicating what's in the transport),
or we have to fatten the fc transport with new search routines to find
the rport (and deal with a case of a dangling rport that the driver
forgets).
It dawned on me that we had actually reached this state incorrectly.
When the fc transport first started, we did the block/unblock first, then
added the rport interface. The purpose of block/unblock is to hide the
temporary disappearance of the rport (e.g. being deleted, then readded).
Why are we making the driver do the block/unblock ? We should be making
the transport have only an rport add/delete, and the let the transport
handle the block/unblock.
So... This patch removes the existing fc_remote_port_block/unblock
functions. It moves the block/unblock functionality into the
fc_remote_port_add/delete functions. Updates for the lpfc driver are
included. Qlogic driver updates are also enclosed, thanks to the
contributions of Andrew Vasquez. [Note: the qla2xxx changes are
relative to the scsi-misc-2.6 tree as of this morning - which does
not include the recent patches sent by Andrew]. The zfcp driver does
not use the block/unblock functions.
One last comment: The resulting behavior feels very clean. The LLDD is
concerned only with add/delete, which corresponds to the physical
disappearance. However, the fact that the scsi target and sdevs are
not immediately torn down after the LLDD calls delete causes an
interesting scenario... the midlayer can call the xxx_slave_alloc and
xxx_queuecommand functions with a sdev that is at the location the
rport used to be. The driver must validate the device exists when it
first enters these functions. In thinking about it, this has always
been the case for the LLDD and these routines. The existing drivers
already check for existence. However, this highlights that simple
validation via data structure dereferencing needs to be watched.
To deal with this, a new transport function, fc_remote_port_chkready()
was created that LLDDs should call when they first enter these two
routines. It validates the rport state, and returns a scsi result
which could be returned. In addition to solving the above, it also
creates consistent behavior from the LLDD's when the block and deletes
are occuring.
Rejections fixed up and
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-10-18 16:03:35 +00:00
|
|
|
struct fc_rport *rport = starget_to_rport(scsi_target(cmnd->device));
|
2019-01-28 19:14:25 +00:00
|
|
|
int err, idx;
|
2021-06-18 17:18:42 +00:00
|
|
|
u8 *uuid = NULL;
|
2021-09-08 05:09:27 +00:00
|
|
|
uint64_t start;
|
2020-03-22 18:13:02 +00:00
|
|
|
|
2021-08-16 16:28:54 +00:00
|
|
|
start = ktime_get_ns();
|
2014-02-20 14:56:45 +00:00
|
|
|
rdata = lpfc_rport_data_from_scsi_device(cmnd->device);
|
2018-06-26 15:24:29 +00:00
|
|
|
|
|
|
|
|
/* sanity check on references */
|
|
|
|
|
if (unlikely(!rdata) || unlikely(!rport))
|
|
|
|
|
goto out_fail_command;
|
|
|
|
|
|
[SCSI] update fc_transport for removal of block/unblock functions
We recently went back to implement a board reset. When we perform the
reset, we wanted to tear down the internal data structures and rebuild
them. Unfortunately, when it came to the rport structure, things were
odd. If we deleted them, the scsi targets and sdevs would be
torn down. Not a good thing for a temporary reset. We could block the
rports, but we either maintain the internal structures to keep the
rport reference (perhaps even replicating what's in the transport),
or we have to fatten the fc transport with new search routines to find
the rport (and deal with a case of a dangling rport that the driver
forgets).
It dawned on me that we had actually reached this state incorrectly.
When the fc transport first started, we did the block/unblock first, then
added the rport interface. The purpose of block/unblock is to hide the
temporary disappearance of the rport (e.g. being deleted, then readded).
Why are we making the driver do the block/unblock ? We should be making
the transport have only an rport add/delete, and the let the transport
handle the block/unblock.
So... This patch removes the existing fc_remote_port_block/unblock
functions. It moves the block/unblock functionality into the
fc_remote_port_add/delete functions. Updates for the lpfc driver are
included. Qlogic driver updates are also enclosed, thanks to the
contributions of Andrew Vasquez. [Note: the qla2xxx changes are
relative to the scsi-misc-2.6 tree as of this morning - which does
not include the recent patches sent by Andrew]. The zfcp driver does
not use the block/unblock functions.
One last comment: The resulting behavior feels very clean. The LLDD is
concerned only with add/delete, which corresponds to the physical
disappearance. However, the fact that the scsi target and sdevs are
not immediately torn down after the LLDD calls delete causes an
interesting scenario... the midlayer can call the xxx_slave_alloc and
xxx_queuecommand functions with a sdev that is at the location the
rport used to be. The driver must validate the device exists when it
first enters these functions. In thinking about it, this has always
been the case for the LLDD and these routines. The existing drivers
already check for existence. However, this highlights that simple
validation via data structure dereferencing needs to be watched.
To deal with this, a new transport function, fc_remote_port_chkready()
was created that LLDDs should call when they first enter these two
routines. It validates the rport state, and returns a scsi result
which could be returned. In addition to solving the above, it also
creates consistent behavior from the LLDD's when the block and deletes
are occuring.
Rejections fixed up and
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-10-18 16:03:35 +00:00
|
|
|
err = fc_remote_port_chkready(rport);
|
|
|
|
|
if (err) {
|
|
|
|
|
cmnd->result = err;
|
2005-04-17 21:05:31 +00:00
|
|
|
goto out_fail_command;
|
|
|
|
|
}
|
2009-11-18 20:40:49 +00:00
|
|
|
ndlp = rdata->pnode;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2011-08-22 01:48:13 +00:00
|
|
|
if ((scsi_get_prot_op(cmnd) != SCSI_PROT_NORMAL) &&
|
2012-01-18 21:25:09 +00:00
|
|
|
(!(phba->sli3_options & LPFC_SLI3_BG_ENABLED))) {
|
2008-12-05 03:40:02 +00:00
|
|
|
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_log(phba, KERN_ERR, LOG_TRACE_EVENT,
|
2009-10-02 19:16:51 +00:00
|
|
|
"9058 BLKGRD: ERROR: rcvd protected cmd:%02x"
|
|
|
|
|
" op:%02x str=%s without registering for"
|
|
|
|
|
" BlockGuard - Rejecting command\n",
|
2008-12-05 03:40:02 +00:00
|
|
|
cmnd->cmnd[0], scsi_get_prot_op(cmnd),
|
|
|
|
|
dif_op_str[scsi_get_prot_op(cmnd)]);
|
|
|
|
|
goto out_fail_command;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
[SCSI] update fc_transport for removal of block/unblock functions
We recently went back to implement a board reset. When we perform the
reset, we wanted to tear down the internal data structures and rebuild
them. Unfortunately, when it came to the rport structure, things were
odd. If we deleted them, the scsi targets and sdevs would be
torn down. Not a good thing for a temporary reset. We could block the
rports, but we either maintain the internal structures to keep the
rport reference (perhaps even replicating what's in the transport),
or we have to fatten the fc transport with new search routines to find
the rport (and deal with a case of a dangling rport that the driver
forgets).
It dawned on me that we had actually reached this state incorrectly.
When the fc transport first started, we did the block/unblock first, then
added the rport interface. The purpose of block/unblock is to hide the
temporary disappearance of the rport (e.g. being deleted, then readded).
Why are we making the driver do the block/unblock ? We should be making
the transport have only an rport add/delete, and the let the transport
handle the block/unblock.
So... This patch removes the existing fc_remote_port_block/unblock
functions. It moves the block/unblock functionality into the
fc_remote_port_add/delete functions. Updates for the lpfc driver are
included. Qlogic driver updates are also enclosed, thanks to the
contributions of Andrew Vasquez. [Note: the qla2xxx changes are
relative to the scsi-misc-2.6 tree as of this morning - which does
not include the recent patches sent by Andrew]. The zfcp driver does
not use the block/unblock functions.
One last comment: The resulting behavior feels very clean. The LLDD is
concerned only with add/delete, which corresponds to the physical
disappearance. However, the fact that the scsi target and sdevs are
not immediately torn down after the LLDD calls delete causes an
interesting scenario... the midlayer can call the xxx_slave_alloc and
xxx_queuecommand functions with a sdev that is at the location the
rport used to be. The driver must validate the device exists when it
first enters these functions. In thinking about it, this has always
been the case for the LLDD and these routines. The existing drivers
already check for existence. However, this highlights that simple
validation via data structure dereferencing needs to be watched.
To deal with this, a new transport function, fc_remote_port_chkready()
was created that LLDDs should call when they first enter these two
routines. It validates the rport state, and returns a scsi result
which could be returned. In addition to solving the above, it also
creates consistent behavior from the LLDD's when the block and deletes
are occuring.
Rejections fixed up and
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-10-18 16:03:35 +00:00
|
|
|
* Catch race where our node has transitioned, but the
|
|
|
|
|
* transport is still transitioning.
|
2005-04-17 21:05:31 +00:00
|
|
|
*/
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!ndlp)
|
2021-08-16 16:28:54 +00:00
|
|
|
goto out_tgt_busy1;
|
|
|
|
|
|
|
|
|
|
/* Check if IO qualifies for CMF */
|
|
|
|
|
if (phba->cmf_active_mode != LPFC_CFG_OFF &&
|
|
|
|
|
cmnd->sc_data_direction == DMA_FROM_DEVICE &&
|
|
|
|
|
(scsi_sg_count(cmnd))) {
|
|
|
|
|
/* Latency start time saved in rx_cmd_start later in routine */
|
|
|
|
|
err = lpfc_update_cmf_cmd(phba, scsi_bufflen(cmnd));
|
|
|
|
|
if (err)
|
|
|
|
|
goto out_tgt_busy1;
|
|
|
|
|
}
|
|
|
|
|
|
2018-08-01 00:23:22 +00:00
|
|
|
if (lpfc_ndlp_check_qdepth(phba, ndlp)) {
|
|
|
|
|
if (atomic_read(&ndlp->cmd_pending) >= ndlp->cmd_qdepth) {
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP_ERROR,
|
|
|
|
|
"3377 Target Queue Full, scsi Id:%d "
|
|
|
|
|
"Qdepth:%d Pending command:%d"
|
|
|
|
|
" WWNN:%02x:%02x:%02x:%02x:"
|
|
|
|
|
"%02x:%02x:%02x:%02x, "
|
|
|
|
|
" WWPN:%02x:%02x:%02x:%02x:"
|
|
|
|
|
"%02x:%02x:%02x:%02x",
|
|
|
|
|
ndlp->nlp_sid, ndlp->cmd_qdepth,
|
|
|
|
|
atomic_read(&ndlp->cmd_pending),
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[0],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[1],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[2],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[3],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[4],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[5],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[6],
|
|
|
|
|
ndlp->nlp_nodename.u.wwn[7],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[0],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[1],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[2],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[3],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[4],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[5],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[6],
|
|
|
|
|
ndlp->nlp_portname.u.wwn[7]);
|
2021-08-16 16:28:54 +00:00
|
|
|
goto out_tgt_busy2;
|
2018-08-01 00:23:22 +00:00
|
|
|
}
|
2018-01-30 23:58:53 +00:00
|
|
|
}
|
2018-04-09 21:24:22 +00:00
|
|
|
|
2019-01-28 19:14:27 +00:00
|
|
|
lpfc_cmd = lpfc_get_scsi_buf(phba, ndlp, cmnd);
|
2005-04-17 21:05:31 +00:00
|
|
|
if (lpfc_cmd == NULL) {
|
2008-12-05 03:39:29 +00:00
|
|
|
lpfc_rampdown_queue_depth(phba);
|
2007-06-18 00:56:38 +00:00
|
|
|
|
2017-02-12 21:52:30 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP_ERROR,
|
2007-08-02 15:10:09 +00:00
|
|
|
"0707 driver's buffer pool is empty, "
|
|
|
|
|
"IO busied\n");
|
2005-04-17 21:05:31 +00:00
|
|
|
goto out_host_busy;
|
|
|
|
|
}
|
2021-08-16 16:28:54 +00:00
|
|
|
lpfc_cmd->rx_cmd_start = start;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
cur_iocbq = &lpfc_cmd->cur_iocbq;
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
|
|
|
|
* Store the midlayer's command structure for the completion phase
|
|
|
|
|
* and complete the command initialization.
|
|
|
|
|
*/
|
|
|
|
|
lpfc_cmd->pCmd = cmnd;
|
|
|
|
|
lpfc_cmd->rdata = rdata;
|
2018-08-01 00:23:22 +00:00
|
|
|
lpfc_cmd->ndlp = ndlp;
|
2022-02-25 02:23:04 +00:00
|
|
|
cur_iocbq->cmd_cmpl = NULL;
|
2005-04-17 21:05:31 +00:00
|
|
|
cmnd->host_scribble = (unsigned char *)lpfc_cmd;
|
|
|
|
|
|
2020-11-15 19:26:42 +00:00
|
|
|
err = lpfc_scsi_prep_cmnd(vport, lpfc_cmd, ndlp);
|
|
|
|
|
if (err)
|
|
|
|
|
goto out_host_busy_release_buf;
|
|
|
|
|
|
2008-12-05 03:40:02 +00:00
|
|
|
if (scsi_get_prot_op(cmnd) != SCSI_PROT_NORMAL) {
|
2009-10-02 19:16:51 +00:00
|
|
|
if (vport->phba->cfg_enable_bg) {
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_vlog(vport,
|
|
|
|
|
KERN_INFO, LOG_SCSI_CMD,
|
2012-08-14 18:25:50 +00:00
|
|
|
"9033 BLKGRD: rcvd %s cmd:x%x "
|
2021-03-01 17:18:02 +00:00
|
|
|
"reftag x%x cnt %u pt %x\n",
|
2012-08-14 18:25:50 +00:00
|
|
|
dif_op_str[scsi_get_prot_op(cmnd)],
|
|
|
|
|
cmnd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmnd),
|
|
|
|
|
scsi_logical_block_count(cmnd),
|
2012-08-14 18:25:50 +00:00
|
|
|
(cmnd->cmnd[1]>>5));
|
2009-10-02 19:16:51 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
err = lpfc_bg_scsi_prep_dma_buf(phba, lpfc_cmd);
|
|
|
|
|
} else {
|
2009-10-02 19:16:51 +00:00
|
|
|
if (vport->phba->cfg_enable_bg) {
|
2013-04-18 00:14:49 +00:00
|
|
|
lpfc_printf_vlog(vport,
|
|
|
|
|
KERN_INFO, LOG_SCSI_CMD,
|
2012-08-14 18:25:50 +00:00
|
|
|
"9038 BLKGRD: rcvd PROT_NORMAL cmd: "
|
2021-03-01 17:18:02 +00:00
|
|
|
"x%x reftag x%x cnt %u pt %x\n",
|
2012-08-14 18:25:50 +00:00
|
|
|
cmnd->cmnd[0],
|
2021-08-17 02:50:14 +00:00
|
|
|
scsi_prot_ref_tag(cmnd),
|
|
|
|
|
scsi_logical_block_count(cmnd),
|
2012-08-14 18:25:50 +00:00
|
|
|
(cmnd->cmnd[1]>>5));
|
2009-10-02 19:16:51 +00:00
|
|
|
}
|
2008-12-05 03:40:02 +00:00
|
|
|
err = lpfc_scsi_prep_dma_buf(phba, lpfc_cmd);
|
|
|
|
|
}
|
|
|
|
|
|
2019-10-18 21:18:24 +00:00
|
|
|
if (unlikely(err)) {
|
|
|
|
|
if (err == 2) {
|
|
|
|
|
cmnd->result = DID_ERROR << 16;
|
|
|
|
|
goto out_fail_command_release_buf;
|
|
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
goto out_host_busy_free_buf;
|
2019-08-14 23:56:54 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2021-06-08 04:35:56 +00:00
|
|
|
/* check the necessary and sufficient condition to support VMID */
|
|
|
|
|
if (lpfc_is_vmid_enabled(phba) &&
|
|
|
|
|
(ndlp->vmid_support ||
|
|
|
|
|
phba->pport->vmid_priority_tagging ==
|
|
|
|
|
LPFC_VMID_PRIO_TAG_ALL_TARGETS)) {
|
|
|
|
|
/* is the I/O generated by a VM, get the associated virtual */
|
|
|
|
|
/* entity id */
|
|
|
|
|
uuid = lpfc_is_command_vm_io(cmnd);
|
|
|
|
|
|
|
|
|
|
if (uuid) {
|
2022-05-19 12:31:09 +00:00
|
|
|
err = lpfc_vmid_get_appid(vport, uuid,
|
|
|
|
|
cmnd->sc_data_direction,
|
|
|
|
|
(union lpfc_vmid_io_tag *)
|
|
|
|
|
&cur_iocbq->vmid_tag);
|
2021-06-08 04:35:56 +00:00
|
|
|
if (!err)
|
2022-02-25 02:23:04 +00:00
|
|
|
cur_iocbq->cmd_flag |= LPFC_IO_VMID;
|
2021-06-08 04:35:56 +00:00
|
|
|
}
|
|
|
|
|
}
|
2022-02-25 02:23:04 +00:00
|
|
|
|
2019-01-28 19:14:31 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
2020-03-22 18:13:00 +00:00
|
|
|
if (unlikely(phba->hdwqstat_on & LPFC_CHECK_SCSI_IO))
|
|
|
|
|
this_cpu_inc(phba->sli4_hba.c_stat->xmt_io);
|
2019-01-28 19:14:31 +00:00
|
|
|
#endif
|
2020-11-15 19:26:41 +00:00
|
|
|
/* Issue I/O to adapter */
|
2022-02-25 02:23:04 +00:00
|
|
|
err = lpfc_sli_issue_fcp_io(phba, LPFC_FCP_RING, cur_iocbq,
|
2020-11-15 19:26:41 +00:00
|
|
|
SLI_IOCB_RET_IOCB);
|
2020-03-22 18:13:02 +00:00
|
|
|
#ifdef CONFIG_SCSI_LPFC_DEBUG_FS
|
|
|
|
|
if (start) {
|
|
|
|
|
lpfc_cmd->ts_cmd_start = start;
|
|
|
|
|
lpfc_cmd->ts_last_cmd = phba->ktime_last_cmd;
|
|
|
|
|
lpfc_cmd->ts_cmd_wqput = ktime_get_ns();
|
|
|
|
|
} else {
|
|
|
|
|
lpfc_cmd->ts_cmd_start = 0;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
2008-12-05 03:39:29 +00:00
|
|
|
if (err) {
|
2013-12-18 01:29:01 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2022-02-25 02:23:04 +00:00
|
|
|
"3376 FCP could not issue iocb err %x "
|
|
|
|
|
"FCP cmd x%x <%d/%llu> "
|
|
|
|
|
"sid: x%x did: x%x oxid: x%x "
|
|
|
|
|
"Data: x%x x%x x%x x%x\n",
|
|
|
|
|
err, cmnd->cmnd[0],
|
|
|
|
|
cmnd->device ? cmnd->device->id : 0xffff,
|
|
|
|
|
cmnd->device ? cmnd->device->lun : (u64)-1,
|
|
|
|
|
vport->fc_myDID, ndlp->nlp_DID,
|
|
|
|
|
phba->sli_rev == LPFC_SLI_REV4 ?
|
|
|
|
|
cur_iocbq->sli4_xritag : 0xffff,
|
|
|
|
|
phba->sli_rev == LPFC_SLI_REV4 ?
|
|
|
|
|
phba->sli4_hba.rpi_ids[ndlp->nlp_rpi] :
|
|
|
|
|
cur_iocbq->iocb.ulpContext,
|
|
|
|
|
cur_iocbq->iotag,
|
|
|
|
|
phba->sli_rev == LPFC_SLI_REV4 ?
|
|
|
|
|
bf_get(wqe_tmo,
|
|
|
|
|
&cur_iocbq->wqe.generic.wqe_com) :
|
|
|
|
|
cur_iocbq->iocb.ulpTimeout,
|
|
|
|
|
(uint32_t)(scsi_cmd_to_rq(cmnd)->timeout / 1000));
|
2013-12-18 01:29:01 +00:00
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
goto out_host_busy_free_buf;
|
2008-12-05 03:39:29 +00:00
|
|
|
}
|
2020-11-15 19:26:42 +00:00
|
|
|
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
if (phba->cfg_poll & ENABLE_FCP_RING_POLLING) {
|
2009-10-02 19:17:02 +00:00
|
|
|
lpfc_sli_handle_fast_ring_event(phba,
|
2017-02-12 21:52:30 +00:00
|
|
|
&phba->sli.sli3_ring[LPFC_FCP_RING], HA_R0RE_REQ);
|
2009-10-02 19:17:02 +00:00
|
|
|
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
if (phba->cfg_poll & DISABLE_FCP_RING_INT)
|
|
|
|
|
lpfc_poll_rearm_timer(phba);
|
|
|
|
|
}
|
|
|
|
|
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
if (phba->cfg_xri_rebalancing)
|
|
|
|
|
lpfc_keep_pvt_pool_above_lowwm(phba, lpfc_cmd->hdwq_no);
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
out_host_busy_free_buf:
|
2019-01-28 19:14:26 +00:00
|
|
|
idx = lpfc_cmd->hdwq_no;
|
2006-07-06 19:50:08 +00:00
|
|
|
lpfc_scsi_unprep_dma_buf(phba, lpfc_cmd);
|
2019-01-28 19:14:25 +00:00
|
|
|
if (phba->sli4_hba.hdwq) {
|
|
|
|
|
switch (lpfc_cmd->fcp_cmnd->fcpCntl3) {
|
|
|
|
|
case WRITE_DATA:
|
|
|
|
|
phba->sli4_hba.hdwq[idx].scsi_cstat.output_requests--;
|
|
|
|
|
break;
|
|
|
|
|
case READ_DATA:
|
|
|
|
|
phba->sli4_hba.hdwq[idx].scsi_cstat.input_requests--;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
phba->sli4_hba.hdwq[idx].scsi_cstat.control_requests--;
|
|
|
|
|
}
|
|
|
|
|
}
|
2020-11-15 19:26:42 +00:00
|
|
|
out_host_busy_release_buf:
|
2005-10-29 00:30:02 +00:00
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
2005-04-17 21:05:31 +00:00
|
|
|
out_host_busy:
|
2021-08-16 16:28:54 +00:00
|
|
|
lpfc_update_cmf_cmpl(phba, LPFC_CGN_NOT_SENT, scsi_bufflen(cmnd),
|
|
|
|
|
shost);
|
2005-04-17 21:05:31 +00:00
|
|
|
return SCSI_MLQUEUE_HOST_BUSY;
|
|
|
|
|
|
2021-08-16 16:28:54 +00:00
|
|
|
out_tgt_busy2:
|
|
|
|
|
lpfc_update_cmf_cmpl(phba, LPFC_CGN_NOT_SENT, scsi_bufflen(cmnd),
|
|
|
|
|
shost);
|
|
|
|
|
out_tgt_busy1:
|
2011-02-25 20:04:28 +00:00
|
|
|
return SCSI_MLQUEUE_TARGET_BUSY;
|
|
|
|
|
|
2019-08-14 23:56:54 +00:00
|
|
|
out_fail_command_release_buf:
|
|
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
2021-08-16 16:28:54 +00:00
|
|
|
lpfc_update_cmf_cmpl(phba, LPFC_CGN_NOT_SENT, scsi_bufflen(cmnd),
|
|
|
|
|
shost);
|
2019-08-14 23:56:54 +00:00
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
out_fail_command:
|
2021-10-07 20:28:41 +00:00
|
|
|
scsi_done(cmnd);
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-08 04:35:49 +00:00
|
|
|
/*
|
|
|
|
|
* lpfc_vmid_vport_cleanup - cleans up the resources associated with a vport
|
|
|
|
|
* @vport: The virtual port for which this call is being executed.
|
|
|
|
|
*/
|
|
|
|
|
void lpfc_vmid_vport_cleanup(struct lpfc_vport *vport)
|
|
|
|
|
{
|
|
|
|
|
u32 bucket;
|
|
|
|
|
struct lpfc_vmid *cur;
|
|
|
|
|
|
|
|
|
|
if (vport->port_type == LPFC_PHYSICAL_PORT)
|
|
|
|
|
del_timer_sync(&vport->phba->inactive_vmid_poll);
|
|
|
|
|
|
|
|
|
|
kfree(vport->qfpa_res);
|
|
|
|
|
kfree(vport->vmid_priority.vmid_range);
|
|
|
|
|
kfree(vport->vmid);
|
|
|
|
|
|
|
|
|
|
if (!hash_empty(vport->hash_table))
|
|
|
|
|
hash_for_each(vport->hash_table, bucket, cur, hnode)
|
|
|
|
|
hash_del(&cur->hnode);
|
|
|
|
|
|
|
|
|
|
vport->qfpa_res = NULL;
|
|
|
|
|
vport->vmid_priority.vmid_range = NULL;
|
|
|
|
|
vport->vmid = NULL;
|
|
|
|
|
vport->cur_vmid_cnt = 0;
|
|
|
|
|
}
|
2010-11-16 07:10:29 +00:00
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_abort_handler - scsi_host_template eh_abort_handler entry point
|
2008-12-05 03:39:02 +00:00
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
|
|
|
|
*
|
|
|
|
|
* This routine aborts @cmnd pending in base driver.
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
2005-11-28 16:41:53 +00:00
|
|
|
lpfc_abort_handler(struct scsi_cmnd *cmnd)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct Scsi_Host *shost = cmnd->device->host;
|
2022-03-01 14:37:16 +00:00
|
|
|
struct fc_rport *rport = starget_to_rport(scsi_target(cmnd->device));
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) shost->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
2005-10-29 00:30:02 +00:00
|
|
|
struct lpfc_iocbq *iocb;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd;
|
2012-06-12 17:54:42 +00:00
|
|
|
int ret = SUCCESS, status = 0;
|
2018-08-01 00:23:20 +00:00
|
|
|
struct lpfc_sli_ring *pring_s4 = NULL;
|
2020-11-15 19:26:44 +00:00
|
|
|
struct lpfc_sli_ring *pring = NULL;
|
2017-02-12 21:52:30 +00:00
|
|
|
int ret_val;
|
2018-04-09 21:24:25 +00:00
|
|
|
unsigned long flags;
|
2008-01-11 06:52:42 +00:00
|
|
|
DECLARE_WAIT_QUEUE_HEAD_ONSTACK(waitq);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2022-03-01 14:37:16 +00:00
|
|
|
status = fc_block_rport(rport);
|
2012-08-03 16:34:54 +00:00
|
|
|
if (status != 0 && status != SUCCESS)
|
2012-06-12 17:54:42 +00:00
|
|
|
return status;
|
2012-05-10 01:17:07 +00:00
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
lpfc_cmd = (struct lpfc_io_buf *)cmnd->host_scribble;
|
|
|
|
|
if (!lpfc_cmd)
|
|
|
|
|
return ret;
|
|
|
|
|
|
2022-04-12 22:19:48 +00:00
|
|
|
/* Guard against IO completion being called at same time */
|
|
|
|
|
spin_lock_irqsave(&lpfc_cmd->buf_lock, flags);
|
|
|
|
|
|
|
|
|
|
spin_lock(&phba->hbalock);
|
2012-05-10 01:17:07 +00:00
|
|
|
/* driver queued commands are in process of being flushed */
|
2019-08-14 23:57:11 +00:00
|
|
|
if (phba->hba_flag & HBA_IOQ_FLUSH) {
|
2012-05-10 01:17:07 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"3168 SCSI Layer abort requested I/O has been "
|
|
|
|
|
"flushed by LLD.\n");
|
2019-01-28 19:14:37 +00:00
|
|
|
ret = FAILED;
|
2022-04-12 22:19:48 +00:00
|
|
|
goto out_unlock_hba;
|
2012-05-10 01:17:07 +00:00
|
|
|
}
|
|
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
if (!lpfc_cmd->pCmd) {
|
2010-09-29 15:19:08 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"2873 SCSI Layer I/O Abort Request IO CMPL Status "
|
2014-06-25 13:27:36 +00:00
|
|
|
"x%x ID %d LUN %llu\n",
|
2012-06-12 17:54:42 +00:00
|
|
|
SUCCESS, cmnd->device->id, cmnd->device->lun);
|
2022-04-12 22:19:48 +00:00
|
|
|
goto out_unlock_hba;
|
2010-09-29 15:19:08 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2012-05-10 01:17:07 +00:00
|
|
|
iocb = &lpfc_cmd->cur_iocbq;
|
2018-08-01 00:23:20 +00:00
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4) {
|
2019-08-14 23:57:11 +00:00
|
|
|
pring_s4 = phba->sli4_hba.hdwq[iocb->hba_wqidx].io_wq->pring;
|
2018-08-01 00:23:20 +00:00
|
|
|
if (!pring_s4) {
|
|
|
|
|
ret = FAILED;
|
2022-04-12 22:19:48 +00:00
|
|
|
goto out_unlock_hba;
|
2018-08-01 00:23:20 +00:00
|
|
|
}
|
|
|
|
|
spin_lock(&pring_s4->ring_lock);
|
|
|
|
|
}
|
2012-05-10 01:17:07 +00:00
|
|
|
/* the command is in process of being cancelled */
|
2022-02-25 02:22:52 +00:00
|
|
|
if (!(iocb->cmd_flag & LPFC_IO_ON_TXCMPLQ)) {
|
2012-05-10 01:17:07 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"3169 SCSI Layer abort requested I/O has been "
|
|
|
|
|
"cancelled by LLD.\n");
|
2019-01-28 19:14:37 +00:00
|
|
|
ret = FAILED;
|
|
|
|
|
goto out_unlock_ring;
|
2012-05-10 01:17:07 +00:00
|
|
|
}
|
2005-10-29 00:30:02 +00:00
|
|
|
/*
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
* If pCmd field of the corresponding lpfc_io_buf structure
|
2005-10-29 00:30:02 +00:00
|
|
|
* points to a different SCSI command, then the driver has
|
|
|
|
|
* already completed this command, but the midlayer did not
|
2012-05-10 01:17:07 +00:00
|
|
|
* see the completion before the eh fired. Just return SUCCESS.
|
2005-10-29 00:30:02 +00:00
|
|
|
*/
|
2012-05-10 01:17:07 +00:00
|
|
|
if (lpfc_cmd->pCmd != cmnd) {
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"3170 SCSI Layer abort requested I/O has been "
|
|
|
|
|
"completed by LLD.\n");
|
2019-01-28 19:14:37 +00:00
|
|
|
goto out_unlock_ring;
|
2012-05-10 01:17:07 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2022-04-12 22:20:02 +00:00
|
|
|
WARN_ON(iocb->io_buf != lpfc_cmd);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2014-04-04 17:51:53 +00:00
|
|
|
/* abort issued in recovery is still in progress */
|
2022-02-25 02:22:52 +00:00
|
|
|
if (iocb->cmd_flag & LPFC_DRIVER_ABORTED) {
|
2014-04-04 17:51:53 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"3389 SCSI Layer I/O Abort Request is pending\n");
|
2018-08-01 00:23:20 +00:00
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4)
|
|
|
|
|
spin_unlock(&pring_s4->ring_lock);
|
2022-04-12 22:19:48 +00:00
|
|
|
spin_unlock(&phba->hbalock);
|
|
|
|
|
spin_unlock_irqrestore(&lpfc_cmd->buf_lock, flags);
|
2014-04-04 17:51:53 +00:00
|
|
|
goto wait_for_cmpl;
|
|
|
|
|
}
|
|
|
|
|
|
2018-08-01 00:23:20 +00:00
|
|
|
lpfc_cmd->waitq = &waitq;
|
2022-03-23 20:55:44 +00:00
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4) {
|
2018-04-09 21:24:25 +00:00
|
|
|
spin_unlock(&pring_s4->ring_lock);
|
2022-03-23 20:55:44 +00:00
|
|
|
ret_val = lpfc_sli4_issue_abort_iotag(phba, iocb,
|
|
|
|
|
lpfc_sli_abort_fcp_cmpl);
|
|
|
|
|
} else {
|
2020-11-15 19:26:44 +00:00
|
|
|
pring = &phba->sli.sli3_ring[LPFC_FCP_RING];
|
2022-03-23 20:55:44 +00:00
|
|
|
ret_val = lpfc_sli_issue_abort_iotag(phba, pring, iocb,
|
|
|
|
|
lpfc_sli_abort_fcp_cmpl);
|
|
|
|
|
}
|
2012-05-10 01:17:07 +00:00
|
|
|
|
scsi: lpfc: Implement health checking when aborting I/O
Several errors have occurred where the adapter stops or fails but does not
raise the register values for the driver to detect failure. Thus driver is
unaware of the failure. The failure typically results in I/O timeouts, the
I/O timeout handler failing (after several seconds), and the error handler
escalating recovery policy and resulting in more errors. Eventually, the
driver is in a position where things have spiraled and it can't do recovery
because other recovery ops are still outstanding and it becomes unusable.
Resolve the situation by having the I/O timeout handler (actually a els,
SCSI I/O, NVMe ls, or NVMe I/O timeout), in addition to aborting the I/O,
perform a mailbox command and look for a response from the hardware. If
the mailbox command fails, it will mark the adapter offline and then invoke
the adapter reset handler to clean up.
The new I/O timeout test will be limited to a test every 5s. If there are
multiple I/O timeouts concurrently, only the 1st I/O timeout will generate
the mailbox command. Further testing will only occur once a timeout occurs
after a 5s delay from the last mailbox command has expired.
Link: https://lore.kernel.org/r/20210104180240.46824-14-jsmart2021@gmail.com
Co-developed-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2021-01-04 18:02:38 +00:00
|
|
|
/* Make sure HBA is alive */
|
|
|
|
|
lpfc_issue_hb_tmo(phba);
|
|
|
|
|
|
2020-11-15 19:26:44 +00:00
|
|
|
if (ret_val != IOCB_SUCCESS) {
|
2018-08-01 00:23:20 +00:00
|
|
|
/* Indicate the IO is not being aborted by the driver. */
|
|
|
|
|
lpfc_cmd->waitq = NULL;
|
2005-10-29 00:30:02 +00:00
|
|
|
ret = FAILED;
|
2022-04-12 22:19:48 +00:00
|
|
|
goto out_unlock_hba;
|
2005-10-29 00:30:02 +00:00
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2019-10-18 21:18:22 +00:00
|
|
|
/* no longer need the lock after this point */
|
2022-04-12 22:19:48 +00:00
|
|
|
spin_unlock(&phba->hbalock);
|
|
|
|
|
spin_unlock_irqrestore(&lpfc_cmd->buf_lock, flags);
|
2019-01-28 19:14:37 +00:00
|
|
|
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
if (phba->cfg_poll & DISABLE_FCP_RING_INT)
|
2009-10-02 19:17:02 +00:00
|
|
|
lpfc_sli_handle_fast_ring_event(phba,
|
2017-02-12 21:52:30 +00:00
|
|
|
&phba->sli.sli3_ring[LPFC_FCP_RING], HA_R0RE_REQ);
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
|
2014-04-04 17:51:53 +00:00
|
|
|
wait_for_cmpl:
|
2020-11-15 19:26:44 +00:00
|
|
|
/*
|
2022-02-25 02:22:52 +00:00
|
|
|
* cmd_flag is set to LPFC_DRIVER_ABORTED before we wait
|
2020-11-15 19:26:44 +00:00
|
|
|
* for abort to complete.
|
|
|
|
|
*/
|
2008-01-11 06:52:42 +00:00
|
|
|
wait_event_timeout(waitq,
|
|
|
|
|
(lpfc_cmd->pCmd != cmnd),
|
2013-04-18 00:14:58 +00:00
|
|
|
msecs_to_jiffies(2*vport->cfg_devloss_tmo*1000));
|
2014-04-04 17:51:53 +00:00
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
spin_lock(&lpfc_cmd->buf_lock);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2005-10-29 00:30:02 +00:00
|
|
|
if (lpfc_cmd->pCmd == cmnd) {
|
|
|
|
|
ret = FAILED;
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2007-08-02 15:10:09 +00:00
|
|
|
"0748 abort handler timed out waiting "
|
2016-10-13 22:06:02 +00:00
|
|
|
"for aborting I/O (xri:x%x) to complete: "
|
2014-06-25 13:27:36 +00:00
|
|
|
"ret %#x, ID %d, LUN %llu\n",
|
2012-08-14 18:26:13 +00:00
|
|
|
iocb->sli4_xritag, ret,
|
|
|
|
|
cmnd->device->id, cmnd->device->lun);
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
2019-05-22 00:49:01 +00:00
|
|
|
|
|
|
|
|
lpfc_cmd->waitq = NULL;
|
|
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
spin_unlock(&lpfc_cmd->buf_lock);
|
2012-05-10 01:17:07 +00:00
|
|
|
goto out;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2019-01-28 19:14:37 +00:00
|
|
|
out_unlock_ring:
|
|
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4)
|
|
|
|
|
spin_unlock(&pring_s4->ring_lock);
|
2022-04-12 22:19:48 +00:00
|
|
|
out_unlock_hba:
|
|
|
|
|
spin_unlock(&phba->hbalock);
|
|
|
|
|
spin_unlock_irqrestore(&lpfc_cmd->buf_lock, flags);
|
2012-05-10 01:17:07 +00:00
|
|
|
out:
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"0749 SCSI Layer I/O Abort Request Status x%x ID %d "
|
2014-06-25 13:27:36 +00:00
|
|
|
"LUN %llu\n", ret, cmnd->device->id,
|
2011-04-04 13:42:14 +00:00
|
|
|
cmnd->device->lun);
|
2005-11-28 16:41:53 +00:00
|
|
|
return ret;
|
2005-05-28 11:54:40 +00:00
|
|
|
}
|
|
|
|
|
|
2009-06-10 21:23:16 +00:00
|
|
|
static char *
|
|
|
|
|
lpfc_taskmgmt_name(uint8_t task_mgmt_cmd)
|
|
|
|
|
{
|
|
|
|
|
switch (task_mgmt_cmd) {
|
|
|
|
|
case FCP_ABORT_TASK_SET:
|
|
|
|
|
return "ABORT_TASK_SET";
|
|
|
|
|
case FCP_CLEAR_TASK_SET:
|
|
|
|
|
return "FCP_CLEAR_TASK_SET";
|
|
|
|
|
case FCP_BUS_RESET:
|
|
|
|
|
return "FCP_BUS_RESET";
|
|
|
|
|
case FCP_LUN_RESET:
|
|
|
|
|
return "FCP_LUN_RESET";
|
|
|
|
|
case FCP_TARGET_RESET:
|
|
|
|
|
return "FCP_TARGET_RESET";
|
|
|
|
|
case FCP_CLEAR_ACA:
|
|
|
|
|
return "FCP_CLEAR_ACA";
|
|
|
|
|
case FCP_TERMINATE_TASK:
|
|
|
|
|
return "FCP_TERMINATE_TASK";
|
|
|
|
|
default:
|
|
|
|
|
return "unknown";
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2013-10-10 16:24:07 +00:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_check_fcp_rsp - check the returned fcp_rsp to see if task failed
|
|
|
|
|
* @vport: The virtual port for which this call is being executed.
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
* @lpfc_cmd: Pointer to lpfc_io_buf data structure.
|
2013-10-10 16:24:07 +00:00
|
|
|
*
|
|
|
|
|
* This routine checks the FCP RSP INFO to see if the tsk mgmt command succeded
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
lpfc_check_fcp_rsp(struct lpfc_vport *vport, struct lpfc_io_buf *lpfc_cmd)
|
2013-10-10 16:24:07 +00:00
|
|
|
{
|
|
|
|
|
struct fcp_rsp *fcprsp = lpfc_cmd->fcp_rsp;
|
|
|
|
|
uint32_t rsp_info;
|
|
|
|
|
uint32_t rsp_len;
|
|
|
|
|
uint8_t rsp_info_code;
|
|
|
|
|
int ret = FAILED;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (fcprsp == NULL)
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
|
|
|
|
"0703 fcp_rsp is missing\n");
|
|
|
|
|
else {
|
|
|
|
|
rsp_info = fcprsp->rspStatus2;
|
|
|
|
|
rsp_len = be32_to_cpu(fcprsp->rspRspLen);
|
|
|
|
|
rsp_info_code = fcprsp->rspInfo3;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO,
|
|
|
|
|
LOG_FCP,
|
|
|
|
|
"0706 fcp_rsp valid 0x%x,"
|
|
|
|
|
" rsp len=%d code 0x%x\n",
|
|
|
|
|
rsp_info,
|
|
|
|
|
rsp_len, rsp_info_code);
|
|
|
|
|
|
2019-05-22 00:49:07 +00:00
|
|
|
/* If FCP_RSP_LEN_VALID bit is one, then the FCP_RSP_LEN
|
|
|
|
|
* field specifies the number of valid bytes of FCP_RSP_INFO.
|
|
|
|
|
* The FCP_RSP_LEN field shall be set to 0x04 or 0x08
|
|
|
|
|
*/
|
|
|
|
|
if ((fcprsp->rspStatus2 & RSP_LEN_VALID) &&
|
|
|
|
|
((rsp_len == 8) || (rsp_len == 4))) {
|
2013-10-10 16:24:07 +00:00
|
|
|
switch (rsp_info_code) {
|
|
|
|
|
case RSP_NO_FAILURE:
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
|
|
|
|
"0715 Task Mgmt No Failure\n");
|
|
|
|
|
ret = SUCCESS;
|
|
|
|
|
break;
|
|
|
|
|
case RSP_TM_NOT_SUPPORTED: /* TM rejected */
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
|
|
|
|
"0716 Task Mgmt Target "
|
|
|
|
|
"reject\n");
|
|
|
|
|
break;
|
|
|
|
|
case RSP_TM_NOT_COMPLETED: /* TM failed */
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
|
|
|
|
"0717 Task Mgmt Target "
|
|
|
|
|
"failed TM\n");
|
|
|
|
|
break;
|
|
|
|
|
case RSP_TM_INVALID_LU: /* TM to invalid LU! */
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
|
|
|
|
"0718 Task Mgmt to invalid "
|
|
|
|
|
"LUN\n");
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-06-10 21:23:16 +00:00
|
|
|
* lpfc_send_taskmgmt - Generic SCSI Task Mgmt Handler
|
|
|
|
|
* @vport: The virtual port for which this call is being executed.
|
2022-03-01 14:37:17 +00:00
|
|
|
* @rport: Pointer to remote port
|
2009-06-10 21:23:16 +00:00
|
|
|
* @tgt_id: Target ID of remote device.
|
|
|
|
|
* @lun_id: Lun number for the TMF
|
|
|
|
|
* @task_mgmt_cmd: type of TMF to send
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
2009-06-10 21:23:16 +00:00
|
|
|
* This routine builds and sends a TMF (SCSI Task Mgmt Function) to
|
|
|
|
|
* a remote port.
|
2008-12-05 03:39:02 +00:00
|
|
|
*
|
2009-06-10 21:23:16 +00:00
|
|
|
* Return Code:
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success.
|
2008-12-05 03:39:02 +00:00
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
2022-03-01 14:37:17 +00:00
|
|
|
lpfc_send_taskmgmt(struct lpfc_vport *vport, struct fc_rport *rport,
|
2016-10-13 22:06:04 +00:00
|
|
|
unsigned int tgt_id, uint64_t lun_id,
|
|
|
|
|
uint8_t task_mgmt_cmd)
|
2005-04-17 21:05:31 +00:00
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_hba *phba = vport->phba;
|
scsi: lpfc: Adapt partitioned XRI lists to efficient sharing
The XRI get/put lists were partitioned per hardware queue. However, the
adapter rarely had sufficient resources to give a large number of resources
per queue. As such, it became common for a cpu to encounter a lack of XRI
resource and request the upper io stack to retry after returning a BUSY
condition. This occurred even though other cpus were idle and not using
their resources.
Create as efficient a scheme as possible to move resources to the cpus that
need them. Each cpu maintains a small private pool which it allocates from
for io. There is a watermark that the cpu attempts to keep in the private
pool. The private pool, when empty, pulls from a global pool from the
cpu. When the cpu's global pool is empty it will pull from other cpu's
global pool. As there many cpu global pools (1 per cpu or hardware queue
count) and as each cpu selects what cpu to pull from at different rates and
at different times, it creates a radomizing effect that minimizes the
number of cpu's that will contend with each other when the steal XRI's from
another cpu's global pool.
On io completion, a cpu will push the XRI back on to its private pool. A
watermark level is maintained for the private pool such that when it is
exceeded it will move XRI's to the CPU global pool so that other cpu's may
allocate them.
On NVME, as heartbeat commands are critical to get placed on the wire, a
single expedite pool is maintained. When a heartbeat is to be sent, it will
allocate an XRI from the expedite pool rather than the normal cpu
private/global pools. On any io completion, if a reduction in the expedite
pools is seen, it will be replenished before the XRI is placed on the cpu
private pool.
Statistics are added to aid understanding the XRI levels on each cpu and
their behaviors.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2019-01-28 19:14:28 +00:00
|
|
|
struct lpfc_io_buf *lpfc_cmd;
|
2009-06-10 21:23:16 +00:00
|
|
|
struct lpfc_iocbq *iocbq;
|
|
|
|
|
struct lpfc_iocbq *iocbqrsp;
|
2016-10-13 22:06:04 +00:00
|
|
|
struct lpfc_rport_data *rdata;
|
|
|
|
|
struct lpfc_nodelist *pnode;
|
2009-06-10 21:23:16 +00:00
|
|
|
int ret;
|
2008-06-15 02:52:38 +00:00
|
|
|
int status;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2022-03-01 14:37:17 +00:00
|
|
|
rdata = rport->dd_data;
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!rdata || !rdata->pnode)
|
2008-06-15 02:52:38 +00:00
|
|
|
return FAILED;
|
2016-10-13 22:06:04 +00:00
|
|
|
pnode = rdata->pnode;
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
lpfc_cmd = lpfc_get_scsi_buf(phba, rdata->pnode, NULL);
|
2005-04-17 21:05:31 +00:00
|
|
|
if (lpfc_cmd == NULL)
|
2008-06-15 02:52:38 +00:00
|
|
|
return FAILED;
|
2013-09-06 16:22:46 +00:00
|
|
|
lpfc_cmd->timeout = phba->cfg_task_mgmt_tmo;
|
2006-05-02 01:50:40 +00:00
|
|
|
lpfc_cmd->rdata = rdata;
|
2022-03-01 14:37:17 +00:00
|
|
|
lpfc_cmd->pCmd = NULL;
|
2018-08-01 00:23:22 +00:00
|
|
|
lpfc_cmd->ndlp = pnode;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
status = phba->lpfc_scsi_prep_task_mgmt_cmd(vport, lpfc_cmd, lun_id,
|
|
|
|
|
task_mgmt_cmd);
|
2008-06-15 02:52:38 +00:00
|
|
|
if (!status) {
|
|
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
|
|
|
|
return FAILED;
|
|
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2009-06-10 21:23:16 +00:00
|
|
|
iocbq = &lpfc_cmd->cur_iocbq;
|
2005-10-29 00:30:02 +00:00
|
|
|
iocbqrsp = lpfc_sli_get_iocbq(phba);
|
2008-06-15 02:52:38 +00:00
|
|
|
if (iocbqrsp == NULL) {
|
|
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
|
|
|
|
return FAILED;
|
|
|
|
|
}
|
2022-02-25 02:22:52 +00:00
|
|
|
iocbq->cmd_cmpl = lpfc_tskmgmt_def_cmpl;
|
2022-02-25 02:23:04 +00:00
|
|
|
iocbq->vport = vport;
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2014-06-25 13:27:36 +00:00
|
|
|
"0702 Issue %s to TGT %d LUN %llu "
|
2011-05-24 15:44:12 +00:00
|
|
|
"rpi x%x nlp_flag x%x Data: x%x x%x\n",
|
2009-06-10 21:23:16 +00:00
|
|
|
lpfc_taskmgmt_name(task_mgmt_cmd), tgt_id, lun_id,
|
2011-05-24 15:44:12 +00:00
|
|
|
pnode->nlp_rpi, pnode->nlp_flag, iocbq->sli4_xritag,
|
2022-02-25 02:22:52 +00:00
|
|
|
iocbq->cmd_flag);
|
2009-06-10 21:23:16 +00:00
|
|
|
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
status = lpfc_sli_issue_iocb_wait(phba, LPFC_FCP_RING,
|
2008-06-15 02:52:38 +00:00
|
|
|
iocbq, iocbqrsp, lpfc_cmd->timeout);
|
2013-10-10 16:24:07 +00:00
|
|
|
if ((status != IOCB_SUCCESS) ||
|
2022-02-25 02:23:04 +00:00
|
|
|
(get_job_ulpstatus(phba, iocbqrsp) != IOSTAT_SUCCESS)) {
|
2015-05-22 14:42:41 +00:00
|
|
|
if (status != IOCB_SUCCESS ||
|
2022-02-25 02:23:04 +00:00
|
|
|
get_job_ulpstatus(phba, iocbqrsp) != IOSTAT_FCP_RSP_ERROR)
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2015-05-22 14:42:41 +00:00
|
|
|
"0727 TMF %s to TGT %d LUN %llu "
|
2022-02-25 02:22:52 +00:00
|
|
|
"failed (%d, %d) cmd_flag x%x\n",
|
2015-05-22 14:42:41 +00:00
|
|
|
lpfc_taskmgmt_name(task_mgmt_cmd),
|
|
|
|
|
tgt_id, lun_id,
|
2022-02-25 02:23:04 +00:00
|
|
|
get_job_ulpstatus(phba, iocbqrsp),
|
|
|
|
|
get_job_word4(phba, iocbqrsp),
|
2022-02-25 02:22:52 +00:00
|
|
|
iocbq->cmd_flag);
|
2013-10-10 16:24:07 +00:00
|
|
|
/* if ulpStatus != IOCB_SUCCESS, then status == IOCB_SUCCESS */
|
|
|
|
|
if (status == IOCB_SUCCESS) {
|
2022-02-25 02:23:04 +00:00
|
|
|
if (get_job_ulpstatus(phba, iocbqrsp) ==
|
|
|
|
|
IOSTAT_FCP_RSP_ERROR)
|
2013-10-10 16:24:07 +00:00
|
|
|
/* Something in the FCP_RSP was invalid.
|
|
|
|
|
* Check conditions */
|
|
|
|
|
ret = lpfc_check_fcp_rsp(vport, lpfc_cmd);
|
|
|
|
|
else
|
|
|
|
|
ret = FAILED;
|
2022-02-25 02:23:04 +00:00
|
|
|
} else if ((status == IOCB_TIMEDOUT) ||
|
|
|
|
|
(status == IOCB_ABORTED)) {
|
2013-10-10 16:24:07 +00:00
|
|
|
ret = TIMEOUT_ERROR;
|
|
|
|
|
} else {
|
|
|
|
|
ret = FAILED;
|
|
|
|
|
}
|
|
|
|
|
} else
|
2009-06-10 21:23:16 +00:00
|
|
|
ret = SUCCESS;
|
|
|
|
|
|
2005-11-28 16:42:05 +00:00
|
|
|
lpfc_sli_release_iocbq(phba, iocbqrsp);
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2022-02-25 02:23:04 +00:00
|
|
|
if (status != IOCB_TIMEDOUT)
|
2009-06-10 21:23:16 +00:00
|
|
|
lpfc_release_scsi_buf(phba, lpfc_cmd);
|
|
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_chk_tgt_mapped -
|
|
|
|
|
* @vport: The virtual port to check on
|
2022-03-01 14:37:18 +00:00
|
|
|
* @rport: Pointer to fc_rport data structure.
|
2009-06-10 21:23:16 +00:00
|
|
|
*
|
|
|
|
|
* This routine delays until the scsi target (aka rport) for the
|
|
|
|
|
* command exists (is present and logged in) or we declare it non-existent.
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
2022-03-01 14:37:18 +00:00
|
|
|
lpfc_chk_tgt_mapped(struct lpfc_vport *vport, struct fc_rport *rport)
|
2009-06-10 21:23:16 +00:00
|
|
|
{
|
2014-02-20 14:56:45 +00:00
|
|
|
struct lpfc_rport_data *rdata;
|
2022-03-01 14:37:18 +00:00
|
|
|
struct lpfc_nodelist *pnode = NULL;
|
2009-06-10 21:23:16 +00:00
|
|
|
unsigned long later;
|
|
|
|
|
|
2022-03-01 14:37:18 +00:00
|
|
|
rdata = rport->dd_data;
|
2009-11-18 20:40:49 +00:00
|
|
|
if (!rdata) {
|
|
|
|
|
lpfc_printf_vlog(vport, KERN_INFO, LOG_FCP,
|
2019-08-14 23:57:06 +00:00
|
|
|
"0797 Tgt Map rport failure: rdata x%px\n", rdata);
|
2009-11-18 20:40:49 +00:00
|
|
|
return FAILED;
|
|
|
|
|
}
|
|
|
|
|
pnode = rdata->pnode;
|
2022-03-01 14:37:18 +00:00
|
|
|
|
2009-06-10 21:23:16 +00:00
|
|
|
/*
|
|
|
|
|
* If target is not in a MAPPED state, delay until
|
|
|
|
|
* target is rediscovered or devloss timeout expires.
|
|
|
|
|
*/
|
|
|
|
|
later = msecs_to_jiffies(2 * vport->cfg_devloss_tmo * 1000) + jiffies;
|
|
|
|
|
while (time_after(later, jiffies)) {
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!pnode)
|
2009-06-10 21:23:16 +00:00
|
|
|
return FAILED;
|
|
|
|
|
if (pnode->nlp_state == NLP_STE_MAPPED_NODE)
|
|
|
|
|
return SUCCESS;
|
|
|
|
|
schedule_timeout_uninterruptible(msecs_to_jiffies(500));
|
2022-03-01 14:37:18 +00:00
|
|
|
rdata = rport->dd_data;
|
2009-06-10 21:23:16 +00:00
|
|
|
if (!rdata)
|
|
|
|
|
return FAILED;
|
|
|
|
|
pnode = rdata->pnode;
|
|
|
|
|
}
|
2020-11-15 19:26:30 +00:00
|
|
|
if (!pnode || (pnode->nlp_state != NLP_STE_MAPPED_NODE))
|
2009-06-10 21:23:16 +00:00
|
|
|
return FAILED;
|
|
|
|
|
return SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_reset_flush_io_context -
|
|
|
|
|
* @vport: The virtual port (scsi_host) for the flush context
|
|
|
|
|
* @tgt_id: If aborting by Target contect - specifies the target id
|
|
|
|
|
* @lun_id: If aborting by Lun context - specifies the lun id
|
|
|
|
|
* @context: specifies the context level to flush at.
|
|
|
|
|
*
|
|
|
|
|
* After a reset condition via TMF, we need to flush orphaned i/o
|
|
|
|
|
* contexts from the adapter. This routine aborts any contexts
|
|
|
|
|
* outstanding, then waits for their completions. The wait is
|
|
|
|
|
* bounded by devloss_tmo though.
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_reset_flush_io_context(struct lpfc_vport *vport, uint16_t tgt_id,
|
|
|
|
|
uint64_t lun_id, lpfc_ctx_cmd context)
|
|
|
|
|
{
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
|
|
|
|
unsigned long later;
|
|
|
|
|
int cnt;
|
|
|
|
|
|
|
|
|
|
cnt = lpfc_sli_sum_iocb(vport, tgt_id, lun_id, context);
|
2005-11-28 16:42:05 +00:00
|
|
|
if (cnt)
|
2014-04-04 17:52:31 +00:00
|
|
|
lpfc_sli_abort_taskmgmt(vport,
|
2017-02-12 21:52:30 +00:00
|
|
|
&phba->sli.sli3_ring[LPFC_FCP_RING],
|
2014-04-04 17:52:31 +00:00
|
|
|
tgt_id, lun_id, context);
|
2008-06-15 02:52:38 +00:00
|
|
|
later = msecs_to_jiffies(2 * vport->cfg_devloss_tmo * 1000) + jiffies;
|
|
|
|
|
while (time_after(later, jiffies) && cnt) {
|
|
|
|
|
schedule_timeout_uninterruptible(msecs_to_jiffies(20));
|
2009-06-10 21:23:16 +00:00
|
|
|
cnt = lpfc_sli_sum_iocb(vport, tgt_id, lun_id, context);
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
if (cnt) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2009-06-10 21:23:16 +00:00
|
|
|
"0724 I/O flush failure for context %s : cnt x%x\n",
|
|
|
|
|
((context == LPFC_CTX_LUN) ? "LUN" :
|
|
|
|
|
((context == LPFC_CTX_TGT) ? "TGT" :
|
|
|
|
|
((context == LPFC_CTX_HOST) ? "HOST" : "Unknown"))),
|
|
|
|
|
cnt);
|
|
|
|
|
return FAILED;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
2009-06-10 21:23:16 +00:00
|
|
|
return SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_device_reset_handler - scsi_host_template eh_device_reset entry point
|
|
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
|
|
|
|
*
|
|
|
|
|
* This routine does a device reset by sending a LUN_RESET task management
|
|
|
|
|
* command.
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_device_reset_handler(struct scsi_cmnd *cmnd)
|
|
|
|
|
{
|
|
|
|
|
struct Scsi_Host *shost = cmnd->device->host;
|
2022-03-01 14:37:16 +00:00
|
|
|
struct fc_rport *rport = starget_to_rport(scsi_target(cmnd->device));
|
2009-06-10 21:23:16 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) shost->hostdata;
|
2014-02-20 14:56:45 +00:00
|
|
|
struct lpfc_rport_data *rdata;
|
2009-11-18 20:40:49 +00:00
|
|
|
struct lpfc_nodelist *pnode;
|
2009-06-10 21:23:16 +00:00
|
|
|
unsigned tgt_id = cmnd->device->id;
|
2014-06-25 13:27:36 +00:00
|
|
|
uint64_t lun_id = cmnd->device->lun;
|
2009-06-10 21:23:16 +00:00
|
|
|
struct lpfc_scsi_event_header scsi_event;
|
2013-10-10 16:24:07 +00:00
|
|
|
int status;
|
2021-01-04 18:02:33 +00:00
|
|
|
u32 logit = LOG_FCP;
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2022-06-03 17:43:24 +00:00
|
|
|
if (!rport)
|
|
|
|
|
return FAILED;
|
|
|
|
|
|
2022-03-01 14:37:16 +00:00
|
|
|
rdata = rport->dd_data;
|
2015-04-07 19:07:26 +00:00
|
|
|
if (!rdata || !rdata->pnode) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:57:06 +00:00
|
|
|
"0798 Device Reset rdata failure: rdata x%px\n",
|
2015-04-07 19:07:26 +00:00
|
|
|
rdata);
|
2009-11-18 20:40:49 +00:00
|
|
|
return FAILED;
|
|
|
|
|
}
|
|
|
|
|
pnode = rdata->pnode;
|
2022-03-01 14:37:16 +00:00
|
|
|
status = fc_block_rport(rport);
|
2012-08-03 16:34:54 +00:00
|
|
|
if (status != 0 && status != SUCCESS)
|
2010-07-14 19:30:54 +00:00
|
|
|
return status;
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2022-03-01 14:37:18 +00:00
|
|
|
status = lpfc_chk_tgt_mapped(vport, rport);
|
2009-06-10 21:23:16 +00:00
|
|
|
if (status == FAILED) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:57:06 +00:00
|
|
|
"0721 Device Reset rport failure: rdata x%px\n", rdata);
|
2009-06-10 21:23:16 +00:00
|
|
|
return FAILED;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
scsi_event.event_type = FC_REG_SCSI_EVENT;
|
|
|
|
|
scsi_event.subcategory = LPFC_EVENT_LUNRESET;
|
|
|
|
|
scsi_event.lun = lun_id;
|
|
|
|
|
memcpy(scsi_event.wwpn, &pnode->nlp_portname, sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(scsi_event.wwnn, &pnode->nlp_nodename, sizeof(struct lpfc_name));
|
|
|
|
|
|
|
|
|
|
fc_host_post_vendor_event(shost, fc_get_event_number(),
|
|
|
|
|
sizeof(scsi_event), (char *)&scsi_event, LPFC_NL_VENDOR_ID);
|
|
|
|
|
|
2022-03-01 14:37:17 +00:00
|
|
|
status = lpfc_send_taskmgmt(vport, rport, tgt_id, lun_id,
|
2009-06-10 21:23:16 +00:00
|
|
|
FCP_LUN_RESET);
|
2021-01-04 18:02:33 +00:00
|
|
|
if (status != SUCCESS)
|
|
|
|
|
logit = LOG_TRACE_EVENT;
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2021-01-04 18:02:33 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, logit,
|
2014-06-25 13:27:36 +00:00
|
|
|
"0713 SCSI layer issued Device Reset (%d, %llu) "
|
2009-06-10 21:23:16 +00:00
|
|
|
"return x%x\n", tgt_id, lun_id, status);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We have to clean up i/o as : they may be orphaned by the TMF;
|
|
|
|
|
* or if the TMF failed, they may be in an indeterminate state.
|
|
|
|
|
* So, continue on.
|
|
|
|
|
* We will report success if all the i/o aborts successfully.
|
|
|
|
|
*/
|
2013-10-10 16:24:07 +00:00
|
|
|
if (status == SUCCESS)
|
|
|
|
|
status = lpfc_reset_flush_io_context(vport, tgt_id, lun_id,
|
2009-06-10 21:23:16 +00:00
|
|
|
LPFC_CTX_LUN);
|
2013-10-10 16:24:07 +00:00
|
|
|
|
|
|
|
|
return status;
|
2009-06-10 21:23:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_target_reset_handler - scsi_host_template eh_target_reset entry point
|
|
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
|
|
|
|
*
|
|
|
|
|
* This routine does a target reset by sending a TARGET_RESET task management
|
|
|
|
|
* command.
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_target_reset_handler(struct scsi_cmnd *cmnd)
|
|
|
|
|
{
|
|
|
|
|
struct Scsi_Host *shost = cmnd->device->host;
|
2022-03-01 14:37:16 +00:00
|
|
|
struct fc_rport *rport = starget_to_rport(scsi_target(cmnd->device));
|
2009-06-10 21:23:16 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) shost->hostdata;
|
2014-02-20 14:56:45 +00:00
|
|
|
struct lpfc_rport_data *rdata;
|
2009-11-18 20:40:49 +00:00
|
|
|
struct lpfc_nodelist *pnode;
|
2009-06-10 21:23:16 +00:00
|
|
|
unsigned tgt_id = cmnd->device->id;
|
2014-06-25 13:27:36 +00:00
|
|
|
uint64_t lun_id = cmnd->device->lun;
|
2009-06-10 21:23:16 +00:00
|
|
|
struct lpfc_scsi_event_header scsi_event;
|
2013-10-10 16:24:07 +00:00
|
|
|
int status;
|
2021-01-04 18:02:33 +00:00
|
|
|
u32 logit = LOG_FCP;
|
2021-07-07 18:43:37 +00:00
|
|
|
u32 dev_loss_tmo = vport->cfg_devloss_tmo;
|
2021-01-04 18:02:34 +00:00
|
|
|
unsigned long flags;
|
|
|
|
|
DECLARE_WAIT_QUEUE_HEAD_ONSTACK(waitq);
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2022-06-03 17:43:24 +00:00
|
|
|
if (!rport)
|
|
|
|
|
return FAILED;
|
|
|
|
|
|
2022-03-01 14:37:16 +00:00
|
|
|
rdata = rport->dd_data;
|
2019-03-12 23:30:17 +00:00
|
|
|
if (!rdata || !rdata->pnode) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:57:06 +00:00
|
|
|
"0799 Target Reset rdata failure: rdata x%px\n",
|
2019-03-12 23:30:17 +00:00
|
|
|
rdata);
|
2009-11-18 20:40:49 +00:00
|
|
|
return FAILED;
|
|
|
|
|
}
|
|
|
|
|
pnode = rdata->pnode;
|
2022-03-01 14:37:16 +00:00
|
|
|
status = fc_block_rport(rport);
|
2012-08-03 16:34:54 +00:00
|
|
|
if (status != 0 && status != SUCCESS)
|
2010-07-14 19:30:54 +00:00
|
|
|
return status;
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2022-03-01 14:37:18 +00:00
|
|
|
status = lpfc_chk_tgt_mapped(vport, rport);
|
2009-06-10 21:23:16 +00:00
|
|
|
if (status == FAILED) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:57:06 +00:00
|
|
|
"0722 Target Reset rport failure: rdata x%px\n", rdata);
|
2015-04-07 19:07:20 +00:00
|
|
|
if (pnode) {
|
2021-01-04 18:02:34 +00:00
|
|
|
spin_lock_irqsave(&pnode->lock, flags);
|
2015-04-07 19:07:20 +00:00
|
|
|
pnode->nlp_flag &= ~NLP_NPR_ADISC;
|
|
|
|
|
pnode->nlp_fcp_info &= ~NLP_FCP_2_DEVICE;
|
2021-01-04 18:02:34 +00:00
|
|
|
spin_unlock_irqrestore(&pnode->lock, flags);
|
2015-04-07 19:07:20 +00:00
|
|
|
}
|
2014-09-03 16:58:16 +00:00
|
|
|
lpfc_reset_flush_io_context(vport, tgt_id, lun_id,
|
|
|
|
|
LPFC_CTX_TGT);
|
|
|
|
|
return FAST_IO_FAIL;
|
2009-06-10 21:23:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
scsi_event.event_type = FC_REG_SCSI_EVENT;
|
|
|
|
|
scsi_event.subcategory = LPFC_EVENT_TGTRESET;
|
|
|
|
|
scsi_event.lun = 0;
|
|
|
|
|
memcpy(scsi_event.wwpn, &pnode->nlp_portname, sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(scsi_event.wwnn, &pnode->nlp_nodename, sizeof(struct lpfc_name));
|
|
|
|
|
|
|
|
|
|
fc_host_post_vendor_event(shost, fc_get_event_number(),
|
|
|
|
|
sizeof(scsi_event), (char *)&scsi_event, LPFC_NL_VENDOR_ID);
|
|
|
|
|
|
2022-03-01 14:37:17 +00:00
|
|
|
status = lpfc_send_taskmgmt(vport, rport, tgt_id, lun_id,
|
2009-06-10 21:23:16 +00:00
|
|
|
FCP_TARGET_RESET);
|
2021-07-07 18:43:37 +00:00
|
|
|
if (status != SUCCESS) {
|
|
|
|
|
logit = LOG_TRACE_EVENT;
|
|
|
|
|
|
|
|
|
|
/* Issue LOGO, if no LOGO is outstanding */
|
|
|
|
|
spin_lock_irqsave(&pnode->lock, flags);
|
2021-10-20 21:14:16 +00:00
|
|
|
if (!(pnode->save_flags & NLP_WAIT_FOR_LOGO) &&
|
2021-07-07 18:43:37 +00:00
|
|
|
!pnode->logo_waitq) {
|
|
|
|
|
pnode->logo_waitq = &waitq;
|
|
|
|
|
pnode->nlp_fcp_info &= ~NLP_FCP_2_DEVICE;
|
|
|
|
|
pnode->nlp_flag |= NLP_ISSUE_LOGO;
|
2021-10-20 21:14:16 +00:00
|
|
|
pnode->save_flags |= NLP_WAIT_FOR_LOGO;
|
2021-07-07 18:43:37 +00:00
|
|
|
spin_unlock_irqrestore(&pnode->lock, flags);
|
|
|
|
|
lpfc_unreg_rpi(vport, pnode);
|
|
|
|
|
wait_event_timeout(waitq,
|
2021-10-20 21:14:16 +00:00
|
|
|
(!(pnode->save_flags &
|
2021-07-07 18:43:37 +00:00
|
|
|
NLP_WAIT_FOR_LOGO)),
|
|
|
|
|
msecs_to_jiffies(dev_loss_tmo *
|
|
|
|
|
1000));
|
|
|
|
|
|
2021-10-20 21:14:16 +00:00
|
|
|
if (pnode->save_flags & NLP_WAIT_FOR_LOGO) {
|
2021-07-07 18:43:37 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, logit,
|
|
|
|
|
"0725 SCSI layer TGTRST "
|
|
|
|
|
"failed & LOGO TMO (%d, %llu) "
|
|
|
|
|
"return x%x\n",
|
|
|
|
|
tgt_id, lun_id, status);
|
|
|
|
|
spin_lock_irqsave(&pnode->lock, flags);
|
2021-10-20 21:14:16 +00:00
|
|
|
pnode->save_flags &= ~NLP_WAIT_FOR_LOGO;
|
2021-07-07 18:43:37 +00:00
|
|
|
} else {
|
|
|
|
|
spin_lock_irqsave(&pnode->lock, flags);
|
|
|
|
|
}
|
|
|
|
|
pnode->logo_waitq = NULL;
|
|
|
|
|
spin_unlock_irqrestore(&pnode->lock, flags);
|
|
|
|
|
status = SUCCESS;
|
|
|
|
|
|
2021-01-04 18:02:34 +00:00
|
|
|
} else {
|
2021-07-07 18:43:37 +00:00
|
|
|
spin_unlock_irqrestore(&pnode->lock, flags);
|
|
|
|
|
status = FAILED;
|
2021-01-04 18:02:34 +00:00
|
|
|
}
|
|
|
|
|
}
|
2009-06-10 21:23:16 +00:00
|
|
|
|
2021-01-04 18:02:33 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, logit,
|
2014-06-25 13:27:36 +00:00
|
|
|
"0723 SCSI layer issued Target Reset (%d, %llu) "
|
2009-06-10 21:23:16 +00:00
|
|
|
"return x%x\n", tgt_id, lun_id, status);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We have to clean up i/o as : they may be orphaned by the TMF;
|
|
|
|
|
* or if the TMF failed, they may be in an indeterminate state.
|
|
|
|
|
* So, continue on.
|
|
|
|
|
* We will report success if all the i/o aborts successfully.
|
|
|
|
|
*/
|
2013-10-10 16:24:07 +00:00
|
|
|
if (status == SUCCESS)
|
|
|
|
|
status = lpfc_reset_flush_io_context(vport, tgt_id, lun_id,
|
2012-06-12 17:54:42 +00:00
|
|
|
LPFC_CTX_TGT);
|
2013-10-10 16:24:07 +00:00
|
|
|
return status;
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
2012-05-10 01:19:44 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_host_reset_handler - scsi_host_template eh_host_reset_handler entry pt
|
|
|
|
|
* @cmnd: Pointer to scsi_cmnd data structure.
|
|
|
|
|
*
|
|
|
|
|
* This routine does host reset to the adaptor port. It brings the HBA
|
|
|
|
|
* offline, performs a board restart, and then brings the board back online.
|
|
|
|
|
* The lpfc_offline calls lpfc_sli_hba_down which will abort and local
|
|
|
|
|
* reject all outstanding SCSI commands to the host and error returned
|
|
|
|
|
* back to SCSI mid-level. As this will be SCSI mid-level's last resort
|
|
|
|
|
* of error handling, it will only return error if resetting of the adapter
|
|
|
|
|
* is not successful; in all other cases, will return success.
|
|
|
|
|
*
|
|
|
|
|
* Return code :
|
|
|
|
|
* 0x2003 - Error
|
|
|
|
|
* 0x2002 - Success
|
|
|
|
|
**/
|
|
|
|
|
static int
|
|
|
|
|
lpfc_host_reset_handler(struct scsi_cmnd *cmnd)
|
|
|
|
|
{
|
|
|
|
|
struct Scsi_Host *shost = cmnd->device->host;
|
|
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) shost->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
|
|
|
|
int rc, ret = SUCCESS;
|
|
|
|
|
|
2021-01-04 18:02:33 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_FCP,
|
2013-04-18 00:18:39 +00:00
|
|
|
"3172 SCSI layer issued Host Reset Data:\n");
|
|
|
|
|
|
2012-06-12 17:54:36 +00:00
|
|
|
lpfc_offline_prep(phba, LPFC_MBX_WAIT);
|
2012-05-10 01:19:44 +00:00
|
|
|
lpfc_offline(phba);
|
|
|
|
|
rc = lpfc_sli_brdrestart(phba);
|
|
|
|
|
if (rc)
|
2019-08-14 23:56:53 +00:00
|
|
|
goto error;
|
|
|
|
|
|
2021-10-20 21:14:11 +00:00
|
|
|
/* Wait for successful restart of adapter */
|
|
|
|
|
if (phba->sli_rev < LPFC_SLI_REV4) {
|
|
|
|
|
rc = lpfc_sli_chipset_init(phba);
|
|
|
|
|
if (rc)
|
|
|
|
|
goto error;
|
|
|
|
|
}
|
|
|
|
|
|
2013-04-18 00:18:39 +00:00
|
|
|
rc = lpfc_online(phba);
|
|
|
|
|
if (rc)
|
2019-08-14 23:56:53 +00:00
|
|
|
goto error;
|
|
|
|
|
|
2012-05-10 01:19:44 +00:00
|
|
|
lpfc_unblock_mgmt_io(phba);
|
|
|
|
|
|
|
|
|
|
return ret;
|
2019-08-14 23:56:53 +00:00
|
|
|
error:
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2019-08-14 23:56:53 +00:00
|
|
|
"3323 Failed host reset\n");
|
|
|
|
|
lpfc_unblock_mgmt_io(phba);
|
|
|
|
|
return FAILED;
|
2012-05-10 01:19:44 +00:00
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_slave_alloc - scsi_host_template slave_alloc entry point
|
2008-12-05 03:39:02 +00:00
|
|
|
* @sdev: Pointer to scsi_device.
|
|
|
|
|
*
|
|
|
|
|
* This routine populates the cmds_per_lun count + 2 scsi_bufs into this host's
|
|
|
|
|
* globally available list of scsi buffers. This routine also makes sure scsi
|
|
|
|
|
* buffer is not allocated more than HBA limit conveyed to midlayer. This list
|
|
|
|
|
* of scsi buffer exists for the lifetime of the driver.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* non-0 - Error
|
|
|
|
|
* 0 - Success
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_slave_alloc(struct scsi_device *sdev)
|
|
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) sdev->host->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
[SCSI] update fc_transport for removal of block/unblock functions
We recently went back to implement a board reset. When we perform the
reset, we wanted to tear down the internal data structures and rebuild
them. Unfortunately, when it came to the rport structure, things were
odd. If we deleted them, the scsi targets and sdevs would be
torn down. Not a good thing for a temporary reset. We could block the
rports, but we either maintain the internal structures to keep the
rport reference (perhaps even replicating what's in the transport),
or we have to fatten the fc transport with new search routines to find
the rport (and deal with a case of a dangling rport that the driver
forgets).
It dawned on me that we had actually reached this state incorrectly.
When the fc transport first started, we did the block/unblock first, then
added the rport interface. The purpose of block/unblock is to hide the
temporary disappearance of the rport (e.g. being deleted, then readded).
Why are we making the driver do the block/unblock ? We should be making
the transport have only an rport add/delete, and the let the transport
handle the block/unblock.
So... This patch removes the existing fc_remote_port_block/unblock
functions. It moves the block/unblock functionality into the
fc_remote_port_add/delete functions. Updates for the lpfc driver are
included. Qlogic driver updates are also enclosed, thanks to the
contributions of Andrew Vasquez. [Note: the qla2xxx changes are
relative to the scsi-misc-2.6 tree as of this morning - which does
not include the recent patches sent by Andrew]. The zfcp driver does
not use the block/unblock functions.
One last comment: The resulting behavior feels very clean. The LLDD is
concerned only with add/delete, which corresponds to the physical
disappearance. However, the fact that the scsi target and sdevs are
not immediately torn down after the LLDD calls delete causes an
interesting scenario... the midlayer can call the xxx_slave_alloc and
xxx_queuecommand functions with a sdev that is at the location the
rport used to be. The driver must validate the device exists when it
first enters these functions. In thinking about it, this has always
been the case for the LLDD and these routines. The existing drivers
already check for existence. However, this highlights that simple
validation via data structure dereferencing needs to be watched.
To deal with this, a new transport function, fc_remote_port_chkready()
was created that LLDDs should call when they first enter these two
routines. It validates the rport state, and returns a scsi result
which could be returned. In addition to solving the above, it also
creates consistent behavior from the LLDD's when the block and deletes
are occuring.
Rejections fixed up and
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-10-18 16:03:35 +00:00
|
|
|
struct fc_rport *rport = starget_to_rport(scsi_target(sdev));
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
uint32_t total = 0;
|
2005-04-17 21:05:31 +00:00
|
|
|
uint32_t num_to_alloc = 0;
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
int num_allocated = 0;
|
2010-06-08 22:31:54 +00:00
|
|
|
uint32_t sdev_cnt;
|
2014-02-20 14:56:45 +00:00
|
|
|
struct lpfc_device_data *device_data;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
struct lpfc_name target_wwpn;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
[SCSI] update fc_transport for removal of block/unblock functions
We recently went back to implement a board reset. When we perform the
reset, we wanted to tear down the internal data structures and rebuild
them. Unfortunately, when it came to the rport structure, things were
odd. If we deleted them, the scsi targets and sdevs would be
torn down. Not a good thing for a temporary reset. We could block the
rports, but we either maintain the internal structures to keep the
rport reference (perhaps even replicating what's in the transport),
or we have to fatten the fc transport with new search routines to find
the rport (and deal with a case of a dangling rport that the driver
forgets).
It dawned on me that we had actually reached this state incorrectly.
When the fc transport first started, we did the block/unblock first, then
added the rport interface. The purpose of block/unblock is to hide the
temporary disappearance of the rport (e.g. being deleted, then readded).
Why are we making the driver do the block/unblock ? We should be making
the transport have only an rport add/delete, and the let the transport
handle the block/unblock.
So... This patch removes the existing fc_remote_port_block/unblock
functions. It moves the block/unblock functionality into the
fc_remote_port_add/delete functions. Updates for the lpfc driver are
included. Qlogic driver updates are also enclosed, thanks to the
contributions of Andrew Vasquez. [Note: the qla2xxx changes are
relative to the scsi-misc-2.6 tree as of this morning - which does
not include the recent patches sent by Andrew]. The zfcp driver does
not use the block/unblock functions.
One last comment: The resulting behavior feels very clean. The LLDD is
concerned only with add/delete, which corresponds to the physical
disappearance. However, the fact that the scsi target and sdevs are
not immediately torn down after the LLDD calls delete causes an
interesting scenario... the midlayer can call the xxx_slave_alloc and
xxx_queuecommand functions with a sdev that is at the location the
rport used to be. The driver must validate the device exists when it
first enters these functions. In thinking about it, this has always
been the case for the LLDD and these routines. The existing drivers
already check for existence. However, this highlights that simple
validation via data structure dereferencing needs to be watched.
To deal with this, a new transport function, fc_remote_port_chkready()
was created that LLDDs should call when they first enter these two
routines. It validates the rport state, and returns a scsi result
which could be returned. In addition to solving the above, it also
creates consistent behavior from the LLDD's when the block and deletes
are occuring.
Rejections fixed up and
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-10-18 16:03:35 +00:00
|
|
|
if (!rport || fc_remote_port_chkready(rport))
|
2005-04-17 21:05:31 +00:00
|
|
|
return -ENXIO;
|
|
|
|
|
|
2014-04-04 17:52:21 +00:00
|
|
|
if (phba->cfg_fof) {
|
2014-02-20 14:56:45 +00:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check to see if the device data structure for the lun
|
|
|
|
|
* exists. If not, create one.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
u64_to_wwn(rport->port_name, target_wwpn.u.wwn);
|
|
|
|
|
spin_lock_irqsave(&phba->devicelock, flags);
|
|
|
|
|
device_data = __lpfc_get_device_data(phba,
|
|
|
|
|
&phba->luns,
|
|
|
|
|
&vport->fc_portname,
|
|
|
|
|
&target_wwpn,
|
|
|
|
|
sdev->lun);
|
|
|
|
|
if (!device_data) {
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
device_data = lpfc_create_device_data(phba,
|
|
|
|
|
&vport->fc_portname,
|
|
|
|
|
&target_wwpn,
|
2016-12-19 23:07:26 +00:00
|
|
|
sdev->lun,
|
|
|
|
|
phba->cfg_XLanePriority,
|
|
|
|
|
true);
|
2014-02-20 14:56:45 +00:00
|
|
|
if (!device_data)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
spin_lock_irqsave(&phba->devicelock, flags);
|
|
|
|
|
list_add_tail(&device_data->listentry, &phba->luns);
|
|
|
|
|
}
|
|
|
|
|
device_data->rport_data = rport->dd_data;
|
|
|
|
|
device_data->available = true;
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
sdev->hostdata = device_data;
|
|
|
|
|
} else {
|
|
|
|
|
sdev->hostdata = rport->dd_data;
|
|
|
|
|
}
|
2010-06-08 22:31:54 +00:00
|
|
|
sdev_cnt = atomic_inc_return(&phba->sdev_cnt);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2019-01-28 19:14:19 +00:00
|
|
|
/* For SLI4, all IO buffers are pre-allocated */
|
|
|
|
|
if (phba->sli_rev == LPFC_SLI_REV4)
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
/* This code path is now ONLY for SLI3 adapters */
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
/*
|
|
|
|
|
* Populate the cmds_per_lun count scsi_bufs into this host's globally
|
|
|
|
|
* available list of scsi buffers. Don't allocate more than the
|
2005-10-29 00:29:51 +00:00
|
|
|
* HBA limit conveyed to the midlayer via the host structure. The
|
|
|
|
|
* formula accounts for the lun_queue_depth + error handlers + 1
|
|
|
|
|
* extra. This list of scsi bufs exists for the lifetime of the driver.
|
2005-04-17 21:05:31 +00:00
|
|
|
*/
|
|
|
|
|
total = phba->total_scsi_bufs;
|
2007-08-02 15:09:59 +00:00
|
|
|
num_to_alloc = vport->cfg_lun_queue_depth + 2;
|
2007-06-18 00:56:38 +00:00
|
|
|
|
2010-06-08 22:31:54 +00:00
|
|
|
/* If allocated buffers are enough do nothing */
|
|
|
|
|
if ((sdev_cnt * (vport->cfg_lun_queue_depth + 2)) < total)
|
|
|
|
|
return 0;
|
|
|
|
|
|
2007-06-18 00:56:38 +00:00
|
|
|
/* Allow some exchanges to be available always to complete discovery */
|
|
|
|
|
if (total >= phba->cfg_hba_queue_depth - LPFC_DISC_IOCB_BUFF_COUNT ) {
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"0704 At limitation of %d preallocated "
|
|
|
|
|
"command buffers\n", total);
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
2007-06-18 00:56:38 +00:00
|
|
|
/* Allow some exchanges to be available always to complete discovery */
|
|
|
|
|
} else if (total + num_to_alloc >
|
|
|
|
|
phba->cfg_hba_queue_depth - LPFC_DISC_IOCB_BUFF_COUNT ) {
|
2007-08-02 15:10:09 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_WARNING, LOG_FCP,
|
|
|
|
|
"0705 Allocation request of %d "
|
|
|
|
|
"command buffers will exceed max of %d. "
|
|
|
|
|
"Reducing allocation request to %d.\n",
|
|
|
|
|
num_to_alloc, phba->cfg_hba_queue_depth,
|
|
|
|
|
(phba->cfg_hba_queue_depth - total));
|
2005-04-17 21:05:31 +00:00
|
|
|
num_to_alloc = phba->cfg_hba_queue_depth - total;
|
|
|
|
|
}
|
2019-01-28 19:14:19 +00:00
|
|
|
num_allocated = lpfc_new_scsi_buf_s3(vport, num_to_alloc);
|
[SCSI] lpfc 8.3.2 : Reorganization for SLI4
Preps the organization of the driver so that the bottom half, which
interacts with the hardware, can share common code sequences for
attachment, detachment, initialization, teardown, etc with new hardware.
For very common code sections, which become specific to the interface
type, the driver uses an indirect function call. The function is set at
initialization. For less common sections, such as initialization, the
driver looks at the interface type and calls the routines relative to
the interface.
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-05-22 18:50:54 +00:00
|
|
|
if (num_to_alloc != num_allocated) {
|
2020-06-30 21:50:00 +00:00
|
|
|
lpfc_printf_vlog(vport, KERN_ERR, LOG_TRACE_EVENT,
|
2013-04-18 00:16:15 +00:00
|
|
|
"0708 Allocation request of %d "
|
|
|
|
|
"command buffers did not succeed. "
|
|
|
|
|
"Allocated %d buffers.\n",
|
|
|
|
|
num_to_alloc, num_allocated);
|
2005-04-17 21:05:31 +00:00
|
|
|
}
|
2009-11-18 20:40:49 +00:00
|
|
|
if (num_allocated > 0)
|
|
|
|
|
phba->total_scsi_bufs += num_allocated;
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_slave_configure - scsi_host_template slave_configure entry point
|
2008-12-05 03:39:02 +00:00
|
|
|
* @sdev: Pointer to scsi_device.
|
|
|
|
|
*
|
|
|
|
|
* This routine configures following items
|
|
|
|
|
* - Tag command queuing support for @sdev if supported.
|
|
|
|
|
* - Enable SLI polling for fcp ring if ENABLE_FCP_RING_POLLING flag is set.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* 0 - Success
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_slave_configure(struct scsi_device *sdev)
|
|
|
|
|
{
|
2007-06-18 00:56:36 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) sdev->host->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
2005-04-17 21:05:31 +00:00
|
|
|
|
2014-11-13 14:08:42 +00:00
|
|
|
scsi_change_queue_depth(sdev, vport->cfg_lun_queue_depth);
|
2005-04-17 21:05:31 +00:00
|
|
|
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
if (phba->cfg_poll & ENABLE_FCP_RING_POLLING) {
|
2009-10-02 19:17:02 +00:00
|
|
|
lpfc_sli_handle_fast_ring_event(phba,
|
2017-02-12 21:52:30 +00:00
|
|
|
&phba->sli.sli3_ring[LPFC_FCP_RING], HA_R0RE_REQ);
|
[SCSI] lpfc 8.1.1 : Add polled-mode support
- Add functionality to run in polled mode only. Includes run time
attribute to enable mode.
- Enable runtime writable hba settings for coallescing and delay parameters
Customers have requested a mode in the driver to run strictly polled.
This is generally to support an environment where the server is extremely
loaded and is looking to reclaim some cpu cycles from adapter interrupt
handling.
This patch adds a new "poll" attribute, and the following behavior:
if value is 0 (default):
The driver uses the normal method for i/o completion. It uses the
firmware feature of interrupt coalesing. The firmware allows a
minimum number of i/o completions before an interrupt, or a maximum
time delay between interrupts. By default, the driver sets these
to no delay (disabled) or 1 i/o - meaning coalescing is disabled.
Attributes were provided to change the coalescing values, but it was
a module-load time only and global across all adapters.
This patch allows them to be writable on a per-adapter basis.
if value is 1 :
Interrupts are left enabled, expecting that the user has tuned the
interrupt coalescing values. When this setting is enabled, the driver
will attempt to service completed i/o whenever new i/o is submitted
to the adapter. If the coalescing values are large, and the i/o
generation rate steady, an interrupt will be avoided by servicing
completed i/o prior to the coalescing thresholds kicking in. However,
if the i/o completion load is high enough or i/o generation slow, the
coalescion values will ensure that completed i/o is serviced in a timely
fashion.
if value is 3 :
Turns off FCP i/o interrupts altogether. The coalescing values now have
no effect. A new attribute "poll_tmo" (default 10ms) exists to set
the polling interval for i/o completion. When this setting is enabled,
the driver will attempt to service completed i/o and restart the
interval timer whenever new i/o is submitted. This behavior allows for
servicing of completed i/o sooner than the interval timer, but ensures
that if no i/o is being issued, then the interval timer will kick in
to service the outstanding i/o.
Signed-off-by: James Smart <James.Smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2005-11-29 21:32:13 +00:00
|
|
|
if (phba->cfg_poll & DISABLE_FCP_RING_INT)
|
|
|
|
|
lpfc_poll_rearm_timer(phba);
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-05 03:39:02 +00:00
|
|
|
/**
|
2009-04-06 22:47:14 +00:00
|
|
|
* lpfc_slave_destroy - slave_destroy entry point of SHT data structure
|
2008-12-05 03:39:02 +00:00
|
|
|
* @sdev: Pointer to scsi_device.
|
|
|
|
|
*
|
|
|
|
|
* This routine sets @sdev hostatdata filed to null.
|
|
|
|
|
**/
|
2005-04-17 21:05:31 +00:00
|
|
|
static void
|
|
|
|
|
lpfc_slave_destroy(struct scsi_device *sdev)
|
|
|
|
|
{
|
2010-06-08 22:31:54 +00:00
|
|
|
struct lpfc_vport *vport = (struct lpfc_vport *) sdev->host->hostdata;
|
|
|
|
|
struct lpfc_hba *phba = vport->phba;
|
2014-02-20 14:56:45 +00:00
|
|
|
unsigned long flags;
|
|
|
|
|
struct lpfc_device_data *device_data = sdev->hostdata;
|
|
|
|
|
|
2010-06-08 22:31:54 +00:00
|
|
|
atomic_dec(&phba->sdev_cnt);
|
2014-04-04 17:52:21 +00:00
|
|
|
if ((phba->cfg_fof) && (device_data)) {
|
2014-02-20 14:56:45 +00:00
|
|
|
spin_lock_irqsave(&phba->devicelock, flags);
|
|
|
|
|
device_data->available = false;
|
|
|
|
|
if (!device_data->oas_enabled)
|
|
|
|
|
lpfc_delete_device_data(phba, device_data);
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
}
|
2005-04-17 21:05:31 +00:00
|
|
|
sdev->hostdata = NULL;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2014-02-20 14:56:45 +00:00
|
|
|
/**
|
|
|
|
|
* lpfc_create_device_data - creates and initializes device data structure for OAS
|
2020-11-02 14:23:43 +00:00
|
|
|
* @phba: Pointer to host bus adapter structure.
|
2014-02-20 14:56:45 +00:00
|
|
|
* @vport_wwpn: Pointer to vport's wwpn information
|
|
|
|
|
* @target_wwpn: Pointer to target's wwpn information
|
|
|
|
|
* @lun: Lun on target
|
2020-11-02 14:23:43 +00:00
|
|
|
* @pri: Priority
|
2014-02-20 14:56:45 +00:00
|
|
|
* @atomic_create: Flag to indicate if memory should be allocated using the
|
|
|
|
|
* GFP_ATOMIC flag or not.
|
|
|
|
|
*
|
|
|
|
|
* This routine creates a device data structure which will contain identifying
|
|
|
|
|
* information for the device (host wwpn, target wwpn, lun), state of OAS,
|
|
|
|
|
* whether or not the corresponding lun is available by the system,
|
|
|
|
|
* and pointer to the rport data.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* NULL - Error
|
|
|
|
|
* Pointer to lpfc_device_data - Success
|
|
|
|
|
**/
|
|
|
|
|
struct lpfc_device_data*
|
|
|
|
|
lpfc_create_device_data(struct lpfc_hba *phba, struct lpfc_name *vport_wwpn,
|
|
|
|
|
struct lpfc_name *target_wwpn, uint64_t lun,
|
2016-12-19 23:07:26 +00:00
|
|
|
uint32_t pri, bool atomic_create)
|
2014-02-20 14:56:45 +00:00
|
|
|
{
|
|
|
|
|
|
|
|
|
|
struct lpfc_device_data *lun_info;
|
|
|
|
|
int memory_flags;
|
|
|
|
|
|
|
|
|
|
if (unlikely(!phba) || !vport_wwpn || !target_wwpn ||
|
2014-04-04 17:52:21 +00:00
|
|
|
!(phba->cfg_fof))
|
2014-02-20 14:56:45 +00:00
|
|
|
return NULL;
|
|
|
|
|
|
|
|
|
|
/* Attempt to create the device data to contain lun info */
|
|
|
|
|
|
|
|
|
|
if (atomic_create)
|
|
|
|
|
memory_flags = GFP_ATOMIC;
|
|
|
|
|
else
|
|
|
|
|
memory_flags = GFP_KERNEL;
|
|
|
|
|
lun_info = mempool_alloc(phba->device_data_mem_pool, memory_flags);
|
|
|
|
|
if (!lun_info)
|
|
|
|
|
return NULL;
|
|
|
|
|
INIT_LIST_HEAD(&lun_info->listentry);
|
|
|
|
|
lun_info->rport_data = NULL;
|
|
|
|
|
memcpy(&lun_info->device_id.vport_wwpn, vport_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(&lun_info->device_id.target_wwpn, target_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
lun_info->device_id.lun = lun;
|
|
|
|
|
lun_info->oas_enabled = false;
|
2016-12-19 23:07:26 +00:00
|
|
|
lun_info->priority = pri;
|
2014-02-20 14:56:45 +00:00
|
|
|
lun_info->available = false;
|
|
|
|
|
return lun_info;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_delete_device_data - frees a device data structure for OAS
|
2020-11-02 14:23:43 +00:00
|
|
|
* @phba: Pointer to host bus adapter structure.
|
2014-02-20 14:56:45 +00:00
|
|
|
* @lun_info: Pointer to device data structure to free.
|
|
|
|
|
*
|
|
|
|
|
* This routine frees the previously allocated device data structure passed.
|
|
|
|
|
*
|
|
|
|
|
**/
|
|
|
|
|
void
|
|
|
|
|
lpfc_delete_device_data(struct lpfc_hba *phba,
|
|
|
|
|
struct lpfc_device_data *lun_info)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
if (unlikely(!phba) || !lun_info ||
|
2014-04-04 17:52:21 +00:00
|
|
|
!(phba->cfg_fof))
|
2014-02-20 14:56:45 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (!list_empty(&lun_info->listentry))
|
|
|
|
|
list_del(&lun_info->listentry);
|
|
|
|
|
mempool_free(lun_info, phba->device_data_mem_pool);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* __lpfc_get_device_data - returns the device data for the specified lun
|
2020-11-02 14:23:43 +00:00
|
|
|
* @phba: Pointer to host bus adapter structure.
|
2014-02-20 14:56:45 +00:00
|
|
|
* @list: Point to list to search.
|
|
|
|
|
* @vport_wwpn: Pointer to vport's wwpn information
|
|
|
|
|
* @target_wwpn: Pointer to target's wwpn information
|
|
|
|
|
* @lun: Lun on target
|
|
|
|
|
*
|
|
|
|
|
* This routine searches the list passed for the specified lun's device data.
|
|
|
|
|
* This function does not hold locks, it is the responsibility of the caller
|
|
|
|
|
* to ensure the proper lock is held before calling the function.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* NULL - Error
|
|
|
|
|
* Pointer to lpfc_device_data - Success
|
|
|
|
|
**/
|
|
|
|
|
struct lpfc_device_data*
|
|
|
|
|
__lpfc_get_device_data(struct lpfc_hba *phba, struct list_head *list,
|
|
|
|
|
struct lpfc_name *vport_wwpn,
|
|
|
|
|
struct lpfc_name *target_wwpn, uint64_t lun)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
struct lpfc_device_data *lun_info;
|
|
|
|
|
|
|
|
|
|
if (unlikely(!phba) || !list || !vport_wwpn || !target_wwpn ||
|
2014-04-04 17:52:21 +00:00
|
|
|
!phba->cfg_fof)
|
2014-02-20 14:56:45 +00:00
|
|
|
return NULL;
|
|
|
|
|
|
|
|
|
|
/* Check to see if the lun is already enabled for OAS. */
|
|
|
|
|
|
|
|
|
|
list_for_each_entry(lun_info, list, listentry) {
|
|
|
|
|
if ((memcmp(&lun_info->device_id.vport_wwpn, vport_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name)) == 0) &&
|
|
|
|
|
(memcmp(&lun_info->device_id.target_wwpn, target_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name)) == 0) &&
|
|
|
|
|
(lun_info->device_id.lun == lun))
|
|
|
|
|
return lun_info;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_find_next_oas_lun - searches for the next oas lun
|
2020-11-02 14:23:43 +00:00
|
|
|
* @phba: Pointer to host bus adapter structure.
|
2014-02-20 14:56:45 +00:00
|
|
|
* @vport_wwpn: Pointer to vport's wwpn information
|
|
|
|
|
* @target_wwpn: Pointer to target's wwpn information
|
|
|
|
|
* @starting_lun: Pointer to the lun to start searching for
|
|
|
|
|
* @found_vport_wwpn: Pointer to the found lun's vport wwpn information
|
|
|
|
|
* @found_target_wwpn: Pointer to the found lun's target wwpn information
|
|
|
|
|
* @found_lun: Pointer to the found lun.
|
|
|
|
|
* @found_lun_status: Pointer to status of the found lun.
|
2020-11-02 14:23:43 +00:00
|
|
|
* @found_lun_pri: Pointer to priority of the found lun.
|
2014-02-20 14:56:45 +00:00
|
|
|
*
|
|
|
|
|
* This routine searches the luns list for the specified lun
|
|
|
|
|
* or the first lun for the vport/target. If the vport wwpn contains
|
|
|
|
|
* a zero value then a specific vport is not specified. In this case
|
|
|
|
|
* any vport which contains the lun will be considered a match. If the
|
|
|
|
|
* target wwpn contains a zero value then a specific target is not specified.
|
|
|
|
|
* In this case any target which contains the lun will be considered a
|
|
|
|
|
* match. If the lun is found, the lun, vport wwpn, target wwpn and lun status
|
|
|
|
|
* are returned. The function will also return the next lun if available.
|
|
|
|
|
* If the next lun is not found, starting_lun parameter will be set to
|
|
|
|
|
* NO_MORE_OAS_LUN.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* non-0 - Error
|
|
|
|
|
* 0 - Success
|
|
|
|
|
**/
|
|
|
|
|
bool
|
|
|
|
|
lpfc_find_next_oas_lun(struct lpfc_hba *phba, struct lpfc_name *vport_wwpn,
|
|
|
|
|
struct lpfc_name *target_wwpn, uint64_t *starting_lun,
|
|
|
|
|
struct lpfc_name *found_vport_wwpn,
|
|
|
|
|
struct lpfc_name *found_target_wwpn,
|
|
|
|
|
uint64_t *found_lun,
|
2016-12-19 23:07:26 +00:00
|
|
|
uint32_t *found_lun_status,
|
|
|
|
|
uint32_t *found_lun_pri)
|
2014-02-20 14:56:45 +00:00
|
|
|
{
|
|
|
|
|
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
struct lpfc_device_data *lun_info;
|
|
|
|
|
struct lpfc_device_id *device_id;
|
|
|
|
|
uint64_t lun;
|
|
|
|
|
bool found = false;
|
|
|
|
|
|
|
|
|
|
if (unlikely(!phba) || !vport_wwpn || !target_wwpn ||
|
|
|
|
|
!starting_lun || !found_vport_wwpn ||
|
|
|
|
|
!found_target_wwpn || !found_lun || !found_lun_status ||
|
|
|
|
|
(*starting_lun == NO_MORE_OAS_LUN) ||
|
2014-04-04 17:52:21 +00:00
|
|
|
!phba->cfg_fof)
|
2014-02-20 14:56:45 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
|
|
lun = *starting_lun;
|
|
|
|
|
*found_lun = NO_MORE_OAS_LUN;
|
|
|
|
|
*starting_lun = NO_MORE_OAS_LUN;
|
|
|
|
|
|
|
|
|
|
/* Search for lun or the lun closet in value */
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&phba->devicelock, flags);
|
|
|
|
|
list_for_each_entry(lun_info, &phba->luns, listentry) {
|
|
|
|
|
if (((wwn_to_u64(vport_wwpn->u.wwn) == 0) ||
|
|
|
|
|
(memcmp(&lun_info->device_id.vport_wwpn, vport_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name)) == 0)) &&
|
|
|
|
|
((wwn_to_u64(target_wwpn->u.wwn) == 0) ||
|
|
|
|
|
(memcmp(&lun_info->device_id.target_wwpn, target_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name)) == 0)) &&
|
|
|
|
|
(lun_info->oas_enabled)) {
|
|
|
|
|
device_id = &lun_info->device_id;
|
|
|
|
|
if ((!found) &&
|
|
|
|
|
((lun == FIND_FIRST_OAS_LUN) ||
|
|
|
|
|
(device_id->lun == lun))) {
|
|
|
|
|
*found_lun = device_id->lun;
|
|
|
|
|
memcpy(found_vport_wwpn,
|
|
|
|
|
&device_id->vport_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(found_target_wwpn,
|
|
|
|
|
&device_id->target_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
if (lun_info->available)
|
|
|
|
|
*found_lun_status =
|
|
|
|
|
OAS_LUN_STATUS_EXISTS;
|
|
|
|
|
else
|
|
|
|
|
*found_lun_status = 0;
|
2016-12-19 23:07:26 +00:00
|
|
|
*found_lun_pri = lun_info->priority;
|
2014-02-20 14:56:45 +00:00
|
|
|
if (phba->cfg_oas_flags & OAS_FIND_ANY_VPORT)
|
|
|
|
|
memset(vport_wwpn, 0x0,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
if (phba->cfg_oas_flags & OAS_FIND_ANY_TARGET)
|
|
|
|
|
memset(target_wwpn, 0x0,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
found = true;
|
|
|
|
|
} else if (found) {
|
|
|
|
|
*starting_lun = device_id->lun;
|
|
|
|
|
memcpy(vport_wwpn, &device_id->vport_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
memcpy(target_wwpn, &device_id->target_wwpn,
|
|
|
|
|
sizeof(struct lpfc_name));
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
return found;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_enable_oas_lun - enables a lun for OAS operations
|
2020-11-02 14:23:43 +00:00
|
|
|
* @phba: Pointer to host bus adapter structure.
|
2014-02-20 14:56:45 +00:00
|
|
|
* @vport_wwpn: Pointer to vport's wwpn information
|
|
|
|
|
* @target_wwpn: Pointer to target's wwpn information
|
|
|
|
|
* @lun: Lun
|
2020-11-02 14:23:43 +00:00
|
|
|
* @pri: Priority
|
2014-02-20 14:56:45 +00:00
|
|
|
*
|
|
|
|
|
* This routine enables a lun for oas operations. The routines does so by
|
|
|
|
|
* doing the following :
|
|
|
|
|
*
|
|
|
|
|
* 1) Checks to see if the device data for the lun has been created.
|
|
|
|
|
* 2) If found, sets the OAS enabled flag if not set and returns.
|
|
|
|
|
* 3) Otherwise, creates a device data structure.
|
|
|
|
|
* 4) If successfully created, indicates the device data is for an OAS lun,
|
|
|
|
|
* indicates the lun is not available and add to the list of luns.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* false - Error
|
|
|
|
|
* true - Success
|
|
|
|
|
**/
|
|
|
|
|
bool
|
|
|
|
|
lpfc_enable_oas_lun(struct lpfc_hba *phba, struct lpfc_name *vport_wwpn,
|
2016-07-06 19:36:05 +00:00
|
|
|
struct lpfc_name *target_wwpn, uint64_t lun, uint8_t pri)
|
2014-02-20 14:56:45 +00:00
|
|
|
{
|
|
|
|
|
|
|
|
|
|
struct lpfc_device_data *lun_info;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
|
|
if (unlikely(!phba) || !vport_wwpn || !target_wwpn ||
|
2014-04-04 17:52:21 +00:00
|
|
|
!phba->cfg_fof)
|
2014-02-20 14:56:45 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&phba->devicelock, flags);
|
|
|
|
|
|
|
|
|
|
/* Check to see if the device data for the lun has been created */
|
|
|
|
|
lun_info = __lpfc_get_device_data(phba, &phba->luns, vport_wwpn,
|
|
|
|
|
target_wwpn, lun);
|
|
|
|
|
if (lun_info) {
|
|
|
|
|
if (!lun_info->oas_enabled)
|
|
|
|
|
lun_info->oas_enabled = true;
|
2016-12-19 23:07:26 +00:00
|
|
|
lun_info->priority = pri;
|
2014-02-20 14:56:45 +00:00
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Create an lun info structure and add to list of luns */
|
|
|
|
|
lun_info = lpfc_create_device_data(phba, vport_wwpn, target_wwpn, lun,
|
2019-05-22 00:48:51 +00:00
|
|
|
pri, true);
|
2014-02-20 14:56:45 +00:00
|
|
|
if (lun_info) {
|
|
|
|
|
lun_info->oas_enabled = true;
|
2016-07-06 19:36:05 +00:00
|
|
|
lun_info->priority = pri;
|
2014-02-20 14:56:45 +00:00
|
|
|
lun_info->available = false;
|
|
|
|
|
list_add_tail(&lun_info->listentry, &phba->luns);
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* lpfc_disable_oas_lun - disables a lun for OAS operations
|
2020-11-02 14:23:43 +00:00
|
|
|
* @phba: Pointer to host bus adapter structure.
|
2014-02-20 14:56:45 +00:00
|
|
|
* @vport_wwpn: Pointer to vport's wwpn information
|
|
|
|
|
* @target_wwpn: Pointer to target's wwpn information
|
|
|
|
|
* @lun: Lun
|
2020-11-02 14:23:43 +00:00
|
|
|
* @pri: Priority
|
2014-02-20 14:56:45 +00:00
|
|
|
*
|
|
|
|
|
* This routine disables a lun for oas operations. The routines does so by
|
|
|
|
|
* doing the following :
|
|
|
|
|
*
|
|
|
|
|
* 1) Checks to see if the device data for the lun is created.
|
|
|
|
|
* 2) If present, clears the flag indicating this lun is for OAS.
|
|
|
|
|
* 3) If the lun is not available by the system, the device data is
|
|
|
|
|
* freed.
|
|
|
|
|
*
|
|
|
|
|
* Return codes:
|
|
|
|
|
* false - Error
|
|
|
|
|
* true - Success
|
|
|
|
|
**/
|
|
|
|
|
bool
|
|
|
|
|
lpfc_disable_oas_lun(struct lpfc_hba *phba, struct lpfc_name *vport_wwpn,
|
2016-12-19 23:07:26 +00:00
|
|
|
struct lpfc_name *target_wwpn, uint64_t lun, uint8_t pri)
|
2014-02-20 14:56:45 +00:00
|
|
|
{
|
|
|
|
|
|
|
|
|
|
struct lpfc_device_data *lun_info;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
|
|
if (unlikely(!phba) || !vport_wwpn || !target_wwpn ||
|
2014-04-04 17:52:21 +00:00
|
|
|
!phba->cfg_fof)
|
2014-02-20 14:56:45 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&phba->devicelock, flags);
|
|
|
|
|
|
|
|
|
|
/* Check to see if the lun is available. */
|
|
|
|
|
lun_info = __lpfc_get_device_data(phba,
|
|
|
|
|
&phba->luns, vport_wwpn,
|
|
|
|
|
target_wwpn, lun);
|
|
|
|
|
if (lun_info) {
|
|
|
|
|
lun_info->oas_enabled = false;
|
2016-12-19 23:07:26 +00:00
|
|
|
lun_info->priority = pri;
|
2014-02-20 14:56:45 +00:00
|
|
|
if (!lun_info->available)
|
|
|
|
|
lpfc_delete_device_data(phba, lun_info);
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
spin_unlock_irqrestore(&phba->devicelock, flags);
|
|
|
|
|
return false;
|
|
|
|
|
}
|
2007-06-18 00:56:38 +00:00
|
|
|
|
2017-02-12 21:52:30 +00:00
|
|
|
static int
|
|
|
|
|
lpfc_no_command(struct Scsi_Host *shost, struct scsi_cmnd *cmnd)
|
|
|
|
|
{
|
|
|
|
|
return SCSI_MLQUEUE_HOST_BUSY;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
lpfc_no_slave(struct scsi_device *sdev)
|
|
|
|
|
{
|
|
|
|
|
return -ENODEV;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct scsi_host_template lpfc_template_nvme = {
|
|
|
|
|
.module = THIS_MODULE,
|
|
|
|
|
.name = LPFC_DRIVER_NAME,
|
|
|
|
|
.proc_name = LPFC_DRIVER_NAME,
|
|
|
|
|
.info = lpfc_info,
|
|
|
|
|
.queuecommand = lpfc_no_command,
|
|
|
|
|
.slave_alloc = lpfc_no_slave,
|
|
|
|
|
.slave_configure = lpfc_no_slave,
|
|
|
|
|
.scan_finished = lpfc_scan_finished,
|
|
|
|
|
.this_id = -1,
|
|
|
|
|
.sg_tablesize = 1,
|
|
|
|
|
.cmd_per_lun = 1,
|
2021-10-12 23:35:39 +00:00
|
|
|
.shost_groups = lpfc_hba_groups,
|
2020-10-20 20:27:16 +00:00
|
|
|
.max_sectors = 0xFFFFFFFF,
|
2017-02-12 21:52:30 +00:00
|
|
|
.vendor_id = LPFC_NL_VENDOR_ID,
|
|
|
|
|
.track_queue_depth = 0,
|
|
|
|
|
};
|
|
|
|
|
|
2005-04-17 21:05:31 +00:00
|
|
|
struct scsi_host_template lpfc_template = {
|
|
|
|
|
.module = THIS_MODULE,
|
|
|
|
|
.name = LPFC_DRIVER_NAME,
|
2016-07-06 19:35:59 +00:00
|
|
|
.proc_name = LPFC_DRIVER_NAME,
|
2005-04-17 21:05:31 +00:00
|
|
|
.info = lpfc_info,
|
|
|
|
|
.queuecommand = lpfc_queuecommand,
|
2017-01-30 12:18:58 +00:00
|
|
|
.eh_timed_out = fc_eh_timed_out,
|
2021-01-06 21:49:08 +00:00
|
|
|
.eh_should_retry_cmd = fc_eh_should_retry_cmd,
|
2005-04-17 21:05:31 +00:00
|
|
|
.eh_abort_handler = lpfc_abort_handler,
|
2009-06-10 21:23:16 +00:00
|
|
|
.eh_device_reset_handler = lpfc_device_reset_handler,
|
|
|
|
|
.eh_target_reset_handler = lpfc_target_reset_handler,
|
2012-05-10 01:19:44 +00:00
|
|
|
.eh_host_reset_handler = lpfc_host_reset_handler,
|
2005-04-17 21:05:31 +00:00
|
|
|
.slave_alloc = lpfc_slave_alloc,
|
|
|
|
|
.slave_configure = lpfc_slave_configure,
|
|
|
|
|
.slave_destroy = lpfc_slave_destroy,
|
2007-04-25 13:53:22 +00:00
|
|
|
.scan_finished = lpfc_scan_finished,
|
2005-04-17 21:05:31 +00:00
|
|
|
.this_id = -1,
|
2008-01-11 06:53:09 +00:00
|
|
|
.sg_tablesize = LPFC_DEFAULT_SG_SEG_CNT,
|
2005-04-17 21:05:31 +00:00
|
|
|
.cmd_per_lun = LPFC_CMD_PER_LUN,
|
2021-10-12 23:35:39 +00:00
|
|
|
.shost_groups = lpfc_hba_groups,
|
2020-10-20 20:27:16 +00:00
|
|
|
.max_sectors = 0xFFFFFFFF,
|
2009-07-19 14:01:32 +00:00
|
|
|
.vendor_id = LPFC_NL_VENDOR_ID,
|
2014-11-13 14:08:42 +00:00
|
|
|
.change_queue_depth = scsi_change_queue_depth,
|
2014-11-13 13:25:11 +00:00
|
|
|
.track_queue_depth = 1,
|
2005-04-17 21:05:31 +00:00
|
|
|
};
|
2022-09-11 22:14:58 +00:00
|
|
|
|
|
|
|
|
struct scsi_host_template lpfc_vport_template = {
|
|
|
|
|
.module = THIS_MODULE,
|
|
|
|
|
.name = LPFC_DRIVER_NAME,
|
|
|
|
|
.proc_name = LPFC_DRIVER_NAME,
|
|
|
|
|
.info = lpfc_info,
|
|
|
|
|
.queuecommand = lpfc_queuecommand,
|
|
|
|
|
.eh_timed_out = fc_eh_timed_out,
|
|
|
|
|
.eh_should_retry_cmd = fc_eh_should_retry_cmd,
|
|
|
|
|
.eh_abort_handler = lpfc_abort_handler,
|
|
|
|
|
.eh_device_reset_handler = lpfc_device_reset_handler,
|
|
|
|
|
.eh_target_reset_handler = lpfc_target_reset_handler,
|
|
|
|
|
.eh_bus_reset_handler = NULL,
|
|
|
|
|
.eh_host_reset_handler = NULL,
|
|
|
|
|
.slave_alloc = lpfc_slave_alloc,
|
|
|
|
|
.slave_configure = lpfc_slave_configure,
|
|
|
|
|
.slave_destroy = lpfc_slave_destroy,
|
|
|
|
|
.scan_finished = lpfc_scan_finished,
|
|
|
|
|
.this_id = -1,
|
|
|
|
|
.sg_tablesize = LPFC_DEFAULT_SG_SEG_CNT,
|
|
|
|
|
.cmd_per_lun = LPFC_CMD_PER_LUN,
|
|
|
|
|
.shost_groups = lpfc_vport_groups,
|
|
|
|
|
.max_sectors = 0xFFFFFFFF,
|
|
|
|
|
.vendor_id = 0,
|
|
|
|
|
.change_queue_depth = scsi_change_queue_depth,
|
|
|
|
|
.track_queue_depth = 1,
|
|
|
|
|
};
|