Do not rely on vendor version field of the CSPs to determine if we are in a
FA-PWWN environment. Instead, use the following procedure:
First, during HBA initialization, driver does a READ_CONFIG to determine if
FA-PWWN is configured on the HBA. A LPFC_FAWWPN_CONFIG hba_flag is set
accordingly.
Next, when the link comes up before the driver gets a link up event, the
firmware logs into the fabric with FA-PWWN. If the fabric port does not
support FA-PWWN, the driver will get a Misconfigured FA-WWN async event
before the link up. A LPFC_FAWWPN_FABRIC hba_flag will be set accordingly.
Finally, if the fabric supports FA-PWWN, the firmware will replace its CSPs
WWN with the Fabric Assigned ones. Then after link up, the driver will
retrieve the Fabric Assigned WWN when it does a READ_SPARAM mbox command.
Link: https://lore.kernel.org/r/20220412222008.126521-23-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>
Update copyrights to 2022 for files modified in the 14.2.0.0 patch set.
Link: https://lore.kernel.org/r/20220225022308.16486-18-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>
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>
Complete the enablement of the cm framework feature in the adapter. Perform
the following:
- Detect the presence of the congestion management framework feature.
When the cm framework is present:
- Issue the SET_FEATURE command to enable the feature.
- Register the cm statistics buffer with the adapter.
- Read the cm enablement buffer to determine the cm framework state for cm
management.
When cm management is enabled:
- Monitor all FPIN and congestion signalling events, incrementing
counters.
- Regularly sync with the adapter to communicate congestion events and to
receive an rx request limit.
- Monitor requests for rx data and ensure that no more than the
adapter prescribed limit is issued on the link. If the limit is
exceeded, SCSI and/or NVMe traffic is temporarily suspended.
- Maintain the minute, hourly, daily statistics buffer.
- Monitor for congestion enablement change events, causing a reread of the
enablement buffer and acting on any change in enablement.
And:
- Add teardown logic, including buffer deregistration, on adapter
detachment or reset.
Link: https://lore.kernel.org/r/20210816162901.121235-10-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>
Update copyrights for files modified by the 12.8.0.11 patch set.
Link: https://lore.kernel.org/r/20210707184351.67872-21-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>
Define additional status fields in mailbox commands to help provide
additional information when downloading new firmware.
Link: https://lore.kernel.org/r/20210707184351.67872-4-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>
Created new attribute lpfc_enable_mi, which by default is enabled.
Add command definition bits for SLI-4 parameters that recognize whether the
adapter has MIB information support and what revision of MIB data. Using
the adapter information, register vendor-specific MIB support with FDMI.
The registration will be done every link up.
During FDMI registration, encountered a couple of errors when reverting to
FDMI rev1. Code needed to exist once reverting. Fixed these.
Link: https://lore.kernel.org/r/20201020202719.54726-8-james.smart@broadcom.com
Co-developed-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The following call trace was seen during HBA reset testing:
BUG: scheduling while atomic: swapper/2/0/0x10000100
...
Call Trace:
dump_stack+0x19/0x1b
__schedule_bug+0x64/0x72
__schedule+0x782/0x840
__cond_resched+0x26/0x30
_cond_resched+0x3a/0x50
mempool_alloc+0xa0/0x170
lpfc_unreg_rpi+0x151/0x630 [lpfc]
lpfc_sli_abts_recover_port+0x171/0x190 [lpfc]
lpfc_sli4_abts_err_handler+0xb2/0x1f0 [lpfc]
lpfc_sli4_io_xri_aborted+0x256/0x300 [lpfc]
lpfc_sli4_sp_handle_abort_xri_wcqe.isra.51+0xa3/0x190 [lpfc]
lpfc_sli4_fp_handle_cqe+0x89/0x4d0 [lpfc]
__lpfc_sli4_process_cq+0xdb/0x2e0 [lpfc]
__lpfc_sli4_hba_process_cq+0x41/0x100 [lpfc]
lpfc_cq_poll_hdler+0x1a/0x30 [lpfc]
irq_poll_softirq+0xc7/0x100
__do_softirq+0xf5/0x280
call_softirq+0x1c/0x30
do_softirq+0x65/0xa0
irq_exit+0x105/0x110
do_IRQ+0x56/0xf0
common_interrupt+0x16a/0x16a
With the conversion to blk_io_poll for better interrupt latency in normal
cases, it introduced this code path, executed when I/O aborts or logouts
are seen, which attempts to allocate memory for a mailbox command to be
issued. The allocation is GFP_KERNEL, thus it could attempt to sleep.
Fix by creating a work element that performs the event handling for the
remote port. This will have the mailbox commands and other items performed
in the work element, not the irq. A much better method as the "irq" routine
does not stall while performing all this deep handling code.
Ensure that allocation failures are handled and send LOGO on failure.
Additionally, enlarge the mailbox memory pool to reduce the possibility of
additional allocation in this path.
Link: https://lore.kernel.org/r/20201020202719.54726-3-james.smart@broadcom.com
Fixes: 317aeb83c9 ("scsi: lpfc: Add blk_io_poll support for latency improvment")
Cc: <stable@vger.kernel.org> # v5.9+
Co-developed-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Although the existing implementation is very good at high I/O load, on
tests involving light load, especially on only a few hardware queues,
latency was a little higher than it can be due to using workqueue
scheduling. Other tasks in the system can delay handling.
Change the lower level to use irq_poll by default which uses a softirq for
I/O completion. This gives better latency as variance in when the cq is
processed is reduced over the workqueue interface. However, as high load is
better served by not being in softirq when the CPU is loaded, work queues
are still used under high I/O load.
Link: https://lore.kernel.org/r/20200630215001.70793-13-jsmart2021@gmail.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>
By default, the driver attempts to allocate a hdwq per logical cpu in order
to provide good cpu affinity. Some systems have extremely high cpu counts
and this can significantly raise memory consumption.
In testing on x86 platforms (non-AMD) it is found that sharing of a hdwq by
a physical cpu and its HT cpu can occur with little performance
degredation. By sharing, the hdwq count can be halved, significantly
reducing the memory overhead.
Change the default behavior of the driver on non-AMD x86 platforms to
share a hdwq by the cpu and its HT cpu.
Link: https://lore.kernel.org/r/20200501214310.91713-6-jsmart2021@gmail.com
Reviewed-by: Hannes Reinecke <hare@suse.de>
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>
The cpu io statistics were capped by a hard define limit of 128. This
effectively was a max number of CPUs, not an actual CPU count, nor actual
CPU numbers which can be even larger than both of those values. This made
stats off/misleading and on large CPU count systems, wrong.
Fix the stats so that all CPUs can have a stats struct. Fix the looping
such that it loops by hdwq, finds CPUs that used the hdwq, and sum the
stats, then display.
Link: https://lore.kernel.org/r/20200322181304.37655-9-jsmart2021@gmail.com
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The current driver attempts to allocate an interrupt vector per cpu using
the systems managed IRQ allocator (flag PCI_IRQ_AFFINITY). The system IRQ
allocator will either provide the per-cpu vector, or return fewer
vectors. When fewer vectors, they are evenly spread between the numa nodes
on the system. When run on an AMD architecture, if interrupts occur to a
cpu that is not in the same numa node as the adapter generating the
interrupt, there are extreme costs and overheads in performance. Thus, if
1:1 vector allocation is used, or the "balanced" vectors in the other numa
nodes, performance can be hit significantly.
A much more performant model is to allocate interrupts only on the cpus
that are in the numa node where the adapter resides. I/O completion is
still performed by the cpu where the I/O was generated. Unfortunately,
there is no flag to request the managed IRQ subsystem allocate vectors only
for the CPUs in the numa node as the adapter.
On AMD architecture, revert the irq allocation to the normal style
(non-managed) and then use irq_set_affinity_hint() to set the cpu
affinity and disable user-space rebalancing.
Tie the support into CPU offline/online. If the cpu being offlined owns a
vector, the vector is re-affinitized to one of the other CPUs on the same
numa node. If there are no more CPUs on the numa node, the vector has all
affinity removed and lets the system determine where it's serviced.
Similarly, when the cpu that owned a vector comes online, the vector is
reaffinitized to the cpu.
Link: https://lore.kernel.org/r/20191105005708.7399-10-jsmart2021@gmail.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>
The recent affinitization didn't address cpu offlining/onlining. If an
interrupt vector is shared and the low order cpu owning the vector is
offlined, as interrupts are managed, the vector is taken offline. This
causes the other CPUs sharing the vector will hang as they can't get io
completions.
Correct by registering callbacks with the system for Offline/Online
events. When a cpu is taken offline, its eq, which is tied to an interrupt
vector is found. If the cpu is the "owner" of the vector and if the
eq/vector is shared by other CPUs, the eq is placed into a polled mode.
Additionally, code paths that perform io submission on the "sharing CPUs"
will check the eq state and poll for completion after submission of new io
to a wq that uses the eq.
Similarly, when a cpu comes back online and owns an offlined vector, the eq
is taken out of polled mode and rearmed to start driving interrupts for eq.
Link: https://lore.kernel.org/r/20191105005708.7399-9-jsmart2021@gmail.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>
In the past, the lpe32000 models, based their main support being for 32G,
and as FC-AL is not supported in the FC standards past 8G, did not support
FC-AL operation.
This patch adds private-loop FC-AL support for the LPE32000 adapters
when a link is 8G or below. To avoid conditions where link rate may
change, which would cause non-connectivity to the AL device, FC-AL
mode must become a persistent setting and the link kept at a speed
supporting FC-AL.
The patch:
- Adds a pls attribute indicating whether the adapter properly supports
FC-AL.
- Adds support for the adapter to indicate that topology should be fixed
and the topology types to be configured.
- Adds a pt attribute to report the persistent topology if present.
Link: https://lore.kernel.org/r/20191018211832.7917-15-jsmart2021@gmail.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>
The existing "auto eq delay" mechanism was sometimes skipping over an EQ,
not ramping the coalescing down under light load fast enough, and in other
cases never kicked in as cpu sharing by multiple vectors didn't quite add
up right.
Tweak the interrupt mechanism such that:
- Add a flag to the EQ to force checking for colaescing values when being
serviced in the interrupt handler. The flag will be set by any CQ bound
to the EQ whenever the number of CQ elements process in a single scan
meets or exceeds the hardware queue notify level. E.g. there's a
significant number of completions happening.
- In the heartbeat work item that checks coalescing:
- Replace the structure that was counting the number of EQs that
interrupted on a single cpu with a new structure that looks at the EQ
to see whether EQ currently has a coalescing value (thus it should be
re-evaluate) or was marked by the new flag indicating heavy
completions.
- When a cpu, which may be servicing multiple vectors, had at least 1 EQ
that should be checked, a new coalescing delay is calculated based on
the number of interrupts that occurred on the cpu.
- The new coalescing value is then applied to the EQs that had
interrupted on the cpu.
Link: https://lore.kernel.org/r/20191018211832.7917-11-jsmart2021@gmail.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>
This is mostly update of the usual drivers: qla2xxx, ufs, smartpqi,
lpfc, hisi_sas, qedf, mpt3sas; plus a whole load of minor updates.
The only core change this time around is the addition of request
batching for virtio. Since batching requires an additional flag to
use, it should be invisible to the rest of the drivers.
Signed-off-by: James E.J. Bottomley <jejb@linux.ibm.com>
-----BEGIN PGP SIGNATURE-----
iJwEABMIAEQWIQTnYEDbdso9F2cI+arnQslM7pishQUCXYQE/yYcamFtZXMuYm90
dG9tbGV5QGhhbnNlbnBhcnRuZXJzaGlwLmNvbQAKCRDnQslM7pishXs9AP4usPY5
OpMlF6OiKFNeJrCdhCScVghf9uHbc7UA6cP+EgD/bCtRgcDe1ZjOTYWdeTwvwWqA
ltWYonnv6Lg3b1f9yqI=
=jRC/
-----END PGP SIGNATURE-----
Merge tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
Pull SCSI updates from James Bottomley:
"This is mostly update of the usual drivers: qla2xxx, ufs, smartpqi,
lpfc, hisi_sas, qedf, mpt3sas; plus a whole load of minor updates. The
only core change this time around is the addition of request batching
for virtio. Since batching requires an additional flag to use, it
should be invisible to the rest of the drivers"
* tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi: (264 commits)
scsi: hisi_sas: Fix the conflict between device gone and host reset
scsi: hisi_sas: Add BIST support for phy loopback
scsi: hisi_sas: Add hisi_sas_debugfs_alloc() to centralise allocation
scsi: hisi_sas: Remove some unused function arguments
scsi: hisi_sas: Remove redundant work declaration
scsi: hisi_sas: Remove hisi_sas_hw.slot_complete
scsi: hisi_sas: Assign NCQ tag for all NCQ commands
scsi: hisi_sas: Update all the registers after suspend and resume
scsi: hisi_sas: Retry 3 times TMF IO for SAS disks when init device
scsi: hisi_sas: Remove sleep after issue phy reset if sas_smp_phy_control() fails
scsi: hisi_sas: Directly return when running I_T_nexus reset if phy disabled
scsi: hisi_sas: Use true/false as input parameter of sas_phy_reset()
scsi: hisi_sas: add debugfs auto-trigger for internal abort time out
scsi: virtio_scsi: unplug LUNs when events missed
scsi: scsi_dh_rdac: zero cdb in send_mode_select()
scsi: fcoe: fix null-ptr-deref Read in fc_release_transport
scsi: ufs-hisi: use devm_platform_ioremap_resource() to simplify code
scsi: ufshcd: use devm_platform_ioremap_resource() to simplify code
scsi: hisi_sas: use devm_platform_ioremap_resource() to simplify code
scsi: ufs: Use kmemdup in ufshcd_read_string_desc()
...
Raise the config max for lpfc_fcp_mq_threshold variable to 256.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
CC: Hannes Reinecke <hare@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Per Dan Carpenter:
The patch d79c9e9d4b: "scsi: lpfc: Support dynamic unbounded SGL lists on
G7 hardware." from Aug 14, 2019, leads to the following static checker
warning:
drivers/scsi/lpfc/lpfc_init.c:4107 lpfc_new_io_buf()
error: not allocating enough data 784 vs 768
There was no need to compare sizes nor to allocate size based on a define.
Change allocation to use actual structure length
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
CC: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Currently, each hardware queue, typically allocated per-cpu, consists of a
WQ/CQ pair per protocol. Meaning if both SCSI and NVMe are supported 2
WQ/CQ pairs will exist for the hardware queue. Separate queues are
unnecessary. The current implementation wastes memory backing the 2nd set
of queues, and the use of double the SLI-4 WQ/CQ's means less hardware
queues can be supported which means there may not always be enough to have
a pair per cpu. If there is only 1 pair per cpu, more cpu's may get their
own WQ/CQ.
Rework the implementation to use a single WQ/CQ pair by both protocols.
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>
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>
When SCSI-MQ is enabled, the SCSI-MQ layers will do pre-allocation of MQ
resources based on shost values set by the driver. In newer cases of the
driver, which attempts to set nr_hw_queues to the cpu count, the
multipliers become excessive, with a single shost having SCSI-MQ
pre-allocation reaching into the multiple GBytes range. NPIV, which
creates additional shosts, only multiply this overhead. On lower-memory
systems, this can exhaust system memory very quickly, resulting in a system
crash or failures in the driver or elsewhere due to low memory conditions.
After testing several scenarios, the situation can be mitigated by limiting
the value set in shost->nr_hw_queues to 4. Although the shost values were
changed, the driver still had per-cpu hardware queues of its own that
allowed parallelization per-cpu. Testing revealed that even with the
smallish number for nr_hw_queues for SCSI-MQ, performance levels remained
near maximum with the within-driver affiinitization.
A module parameter was created to allow the value set for the nr_hw_queues
to be tunable.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Ewan D. Milne <emilne@redhat.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
While fixing the resources per socket, realized the driver was not using
hardware queues (up to 1 per cpu) if there were fewer interrupt
vectors. The driver was only using the hardware queue assigned to the cpu
with the vector.
Rework the affinity map check to use the additional hardware queue elements
that had been allocated. If the cpu count exceeds the hardware queue count
- share, but choose what is shared with by: hyperthread peer, core peer,
socket peer, or finally similar cpu in a different socket.
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>
The driver was coded expecting enough hardware queues and interrupt vectors
such that at least there was one per socket. In the case where there were
fewer than sockets, cpus were left unassigned thus null pointers.
Rework the affinity mappings. Map settings for the cpu's that are in the
irq cpu mask. For each cpu not in the mask, map to another cpu that does
have a mask. Choice of the "other" cpu will attempt to map to the same cpu
but differing hyperthread, or cpu within in same core, or cpu within same
socket, or finally cpu in the base socket.
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>
The driver unconditionally says fw doesn't support nvme when in
truth it was a driver parameter settings that disabled nvme support.
Rework the code validating nvme support to accurately report what
condition is disabling nvme support. Save state on whether nvme
fw supports nvme in case sysfs attributes change dynamically.
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>
The driver currently is relying on firmware to match ABTSs to existing
exchanges. This works fine as long as an exchange has been assigned to the
io and work posted to it. However, for unmapped frames (rxid=0xFFFF), the
driver has yet to assign an xri. The driver was blindly saying it couldn't
match the ABTS and sending the BA_xxx. However, the command frame may have
been in queues waiting on xri's before posting to the nvmet_fc layer. When
xri's became available, the command frame would still be pushed to the
transport and that io would execute, even though the io had been killed by
ABTS. The initiator, seeing the io ABTS'd, would reuse the exchange for a
different io which would be received on the target and pushed up. If the
"zombie" io then came back down and started transmitting, the initiator
would match the oxid and accept erroneous data. Bad things happened.
Add tracking of active exchanges in the target to allow matching of a
received ABTS against active or pending IO requests. If the ABTS is matched
to a pending or active IO, the drive initiates cleanup and conditionally
notifies the transport.
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>
Currently the driver is notified of new command frame receipt by CQEs. As
part of the CQE processing, the driver upcalls the nvmet_fc transport to
deliver the command. nvmet_fc, as part of receiving the command builds out
a context for it, where one of the first steps is to allocate memory for
the io.
When running with tests that do large ios (1MB), it was found on some
systems, the total number of outstanding I/O's, at 1MB per, completely
consumed the system's memory. Thus additional ios were getting blocked in
the memory allocator. Given that this blocked the lpfc thread processing
CQEs, there were lots of other commands that were received and which are
then held up, and given CQEs are serially processed, the aggregate delays
for an IO waiting behind the others became cummulative - enough so that the
initiator hit timeouts for the ios.
The basic fix is to avoid the direct upcall and instead schedule a work
item for each io as it is received. This allows the cq processing to
complete very quickly, and each io can then run or block on it's own.
However, this general solution hurts latency when there are few ios. As
such, implemented the fix such that the driver watches how many CQEs it has
processed sequentially in one run. As long as the count is below a
threshold, the direct nvmet_fc upcall will be made. Only when the count is
exceeded will it revert to work scheduling.
Given that debug of this showed a surprisingly long delay in cq processing,
the io timer stats were updated to better reflect the processing of the
different points.
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>
Declaring interrupt clear routines as inline is bogus as they are used as
an indirect pointer.
Remove the inline references.
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>
You can't declare a function inline in a header if it doesn't have a body
available to the compiler. So realistically you either don't declare it
inline or you make it a static inline in the header. I think the latter
applies in this case, so this should be the fix
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Acked-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Change the SLI4 queue creation code to use NUMA node based memory
allocation based on the cpu the queues will be related to.
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>
Currently the driver maintains a sideband structure which has a pointer for
each queue element. However, at 8 bytes per pointer, and up to 4k elements
per queue, and 100s of queues, this can take up a lot of memory.
Convert the driver to using an access routine that calculates the element
address based on its index rather than using the pointer table.
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>
When in trunking mode, the adapter can be placed into diagnostic mode and
each link in the trunk tested via loopback.
Add support to the driver to perform per-link loopback testing when in
trunking mode.
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>
Driver is using uint16_t and is encountering an overflow of the 16bits when
calculating link speed.
Fix by using a u32 type.
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>
For files modified as part of 12.2.0.0 patches, update copyright to 2019
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>
The work done to date utilized the number of present cpus when sizing
per-cpu structures. Structures should have been sized based on the max
possible cpu count.
Convert the driver over to possible cpu count for sizing allocation.
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>
When driving high iop counts, auto_imax coalescing kicks in and drives the
performance to extremely small iops levels.
There are two issues:
1) auto_imax is enabled by default. The auto algorithm, when iops gets
high, divides the iops by the hdwq count and uses that value to
calculate EQ_Delay. The EQ_Delay is set uniformly on all EQs whether
they have load or not. The EQ_delay is only manipulated every 5s (a
long time). Thus there were large 5s swings of no interrupt delay
followed by large/maximum delay, before repeating.
2) When processing a CQ, the driver got mixed up on the rate of when
to ring the doorbell to keep the chip appraised of the eqe or cqe
consumption as well as how how long to sit in the thread and
process queue entries. Currently, the driver capped its work at
64 entries (very small) and exited/rearmed the CQ. Thus, on heavy
loads, additional overheads were taken to exit and re-enter the
interrupt handler. Worse, if in the large/maximum coalescing
windows,k it could be a while before getting back to servicing.
The issues are corrected by the following:
- A change in defaults. Auto_imax is turned OFF and fcp_imax is set
to 0. Thus all interrupts are immediate.
- Cleanup of field names and their meanings. Existing names were
non-intuitive or used for duplicate things.
- Added max_proc_limit field, to control the length of time the
handlers would service completions.
- Reworked EQ handling:
Added common routine that walks eq, applying notify interval and max
processing limits. Use queue_claimed to claim ownership of the queue
while processing. Always rearm the queue whenever the common routine
is called.
Rework queue element processing, namely to eliminate hba_index vs
host_index. Only one index is necessary. The queue entry can be
marked invalid and the host_index updated immediately after eqe
processing.
After rework, xx_release routines are now DB write functions. Renamed
the routines as such.
Moved lpfc_sli4_eq_flush(), which does similar action, to same area.
Replaced the 2 individual loops that walk an eq with a call to the
common routine.
Slightly revised lpfc_sli4_hba_handle_eqe() calling syntax.
Added per-cpu counters to detect interrupt rates and scale
interrupt coalescing values.
- Reworked CQ handling:
Added common routine that walks cq, applying notify interval and max
processing limits. Use queue_claimed to claim ownership of the queue
while processing. Always rearm the queue whenever the common routine
is called.
Rework queue element processing, namely to eliminate hba_index vs
host_index. Only one index is necessary. The queue entry can be
marked invalid and the host_index updated immediately after cqe
processing.
After rework, xx_release routines are now DB write functions. Renamed
the routines as such.
Replaced the 3 individual loops that walk a cq with a call to the
common routine.
Redefined lpfc_sli4_sp_handle_mcqe() to commong handler definition with
queue reference. Add increment for mbox completion to handler.
- Added a new module/sysfs attribute: lpfc_cq_max_proc_limit To allow
dynamic changing of the CQ max_proc_limit value being used.
Although this leaves an EQ as an immediate interrupt, that interrupt will
only occur if a CQ bound to it is in an armed state and has cqe's to
process. By staying in the cq processing routine longer, high loads will
avoid generating more interrupts as they will only rearm as the processing
thread exits. The immediately interrupt is also beneficial to idle or
lower-processing CQ's as they get serviced immediately without being
penalized by sharing an EQ with a more loaded CQ.
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>
Review of the eq coalescing logic showed the code was a bit fragmented.
Sometimes it would save/set via an interrupt max value, while in others it
would do so via a usdelay. There were also two places changing eq delay,
one place that issued mailbox commands, and another that changed via
register writes if supported.
Clean this up by:
- Standardizing the operation of lpfc_modify_hba_eq_delay() routine so
that it is always told of a us delay to impose. The routine then chooses
the best way to set that - via register or via mbx.
- Rather than two value types stored in eq->q_mode (usdelay if change via
register, imax if change via mbox) - q_mode always contains usdelay.
Before any value change, old vs new value is compared and only if
different is a change done.
- Revised the dmult calculation. dmult is not set based on overall imax
divided by hardware queues - instead imax applies to a single cpu and
the value will be replicated to all cpus.
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>
So far MSIX vector allocation assumed it would be 1:1 with hardware
queues. However, there are several reasons why fewer MSIX vectors may be
allocated than hardware queues such as the platform being out of vectors or
adapter limits being less than cpu count.
This patch reworks the MSIX/EQ relationships with the per-cpu hardware
queues so they can function independently. MSIX vectors will be equitably
split been cpu sockets/cores and then the per-cpu hardware queues will be
mapped to the vectors most efficient for them.
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>
The desired affinity for the hardware queue behavior is for hdwq 0 to be
affinitized with cpu 0, hdwq 1 to cpu 1, and so on. The implementation so
far does not do this if the number of cpus is greater than the number of
hardware queues (e.g. hardware queue allocation was administratively
reduced or hardware queue resources could not scale to the cpu count).
Correct the queue affinitization logic when queue count is less than
cpu count.
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>
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>
Many io statistics were being sampled and saved using adapter-based data
structures. This was creating a lot of contention and cache thrashing in
the I/O path.
Move the statistics to the hardware queue data structures. Given the
per-queue data structures, use of atomic types is lessened.
Add new sysfs and debugfs stat routines to collate the per hardware queue
values and report at an adapter level.
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>
Similar to the io execution path that reports cpu context information, the
debugfs routines for cpu information needs to be aligned with new hardware
queue implementation.
Convert debugfs cnd nvme cpucheck statistics to report information per
Hardware Queue.
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>
Once the IO buff allocations were made shared, there was a single XRI
buffer list shared by all hardware queues. A single list isn't great for
performance when shared across the per-cpu hardware queues.
Create a separate XRI IO buffer get/put list for each Hardware Queue. As
SGLs and associated IO buffers get allocated/posted to the firmware; round
robin their assignment across all available hardware Queues so that there
is an equitable assignment.
Modify SCSI and NVME IO submit code paths to use the Hardware Queue logic
for XRI allocation.
Add a debugfs interface to display hardware queue statistics
Added new empty_io_bufs counter to track if a cpu runs out of XRIs.
Replace common_ variables/names with io_ to make meanings clearer.
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>
Currently, both nvme and fcp each have their own concept of an io_channel,
which is a combination wq/cq and associated msix. Different cpus would
share an io_channel.
The driver is now moving to per-cpu wq/cq pairs and msix vectors. The
driver will still use separate wq/cq pairs per protocol on each cpu, but
the protocols will share the msix vector.
Given the elimination of the nvme and fcp io channels, the module
parameters will be removed. A new parameter, lpfc_hdw_queue is added which
allows the wq/cq pair allocation per cpu to be overridden and allocated to
lesser value. If lpfc_hdw_queue is zero, the number of pairs allocated will
be based on the number of cpus. If non-zero, the parameter specifies the
number of queues to allocate. At this time, the maximum non-zero value is
64.
To manage this new paradigm, a new hardware queue structure is created to
track queue activity and relationships.
As MSIX vector allocation must be known before setting up the
relationships, msix allocation now occurs before queue datastructures are
allocated. If the number of vectors allocated is less than the desired
hardware queues, the hardware queue counts will be reduced to the number of
vectors
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>
There is a extra queue and msix vector for expresslane. Now that the driver
will be doing queues per cpu, this oddball queue is no longer needed.
Expresslane will utilize the normal per-cpu queues.
Updated debugfs sli4 queue output to go along with the change
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>
Currently, both NVME and SCSI get their IO buffers from separate
pools. XRI's are associated 1:1 with IO buffers, so XRI's are also split
between protocols.
Eliminate the independent pools and use a single pool. Each buffer
structure now has a common section and a protocol section. Per protocol
routines for SGL initialization are removed and replaced by common
routines. Initialization of the buffers is only done on the common area.
All other fields, which are protocol specific, are initialized when the
buffer is allocated for use in the per-protocol allocation routine.
In the past, the SCSI side allocated IO buffers as part of slave_alloc
calls until the maximum XRIs for SCSI was reached. As all XRIs are now
common and may be used for either protocol, allocation for everything is
done as part of adapter initialization and the scsi side has no action in
slave alloc.
As XRI's are no longer split, the lpfc_xri_split module parameter is
removed.
Adapters based on SLI3 will continue to use the older scsi_buf_list_get/put
routines. All SLI4 adapters utilize the new IO buffer scheme
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>
Add trunking support to the driver. Trunking is found on more recent
asics. In general, trunking appears as a single "port" to the driver
and overall behavior doesn't differ. Link speed is reported as an
aggregate value, while link speed control is done on a per-physical
link basis with all links in the trunk symmetrical. Some commands
returning port information are updated to additionally provide
trunking information. And new ACQEs are generated to report physical
link events relative to the trunk.
This patch contains the following modifications:
- Added link speed settings of 128GB and 256GB.
- Added handling of trunk-related ACQEs, mainly logging and trapping
of physical link statuses.
- Added additional bsg interface to query trunk state by applications.
- Augment link_state sysfs attribtute to display trunk link status
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>
On FCoE adapters, when running link bounce test in a loop, initiator
failed to login with switch switch and required driver reload to
recover. Switch reached a point where all subsequent FLOGIs would be
LS_RJT'd. Further testing showed the condition to be related to not
performing FCF discovery between FLOGI's.
Fix by monitoring FLOGI failures and once a repeated error is seen
repeat FCF discovery.
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>
This patch adds the ability to read firmware logs from the adapter. The driver
registers a buffer with the adapter that is then written to by the adapter.
The adapter posts CQEs to indicate content updates in the buffer. While the
adapter is writing to the buffer in a circular fashion, an application will
poll the driver to read the next amount of log data from the buffer.
Driver log buffer size is configurable via the ras_fwlog_buffsize sysfs
attribute. Verbosity to be used by firmware when logging to host memory is
controlled through the ras_fwlog_level attribute. The ras_fwlog_func
attribute enables or disables loggy by firmware.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Change references from "Broadcom Limited" to "Broadcom Inc." in the
copyright message. Update copyright duration if not yet updated for 2018.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Current implementation missed setting the duration field. Correct the code
to set the field.
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
