linux-stable/drivers/infiniband/hw/hfi1/uc.c

543 lines
14 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
/*
* Copyright(c) 2015 - 2018 Intel Corporation.
*/
#include "hfi.h"
#include "verbs_txreq.h"
#include "qp.h"
/* cut down ridiculously long IB macro names */
#define OP(x) UC_OP(x)
staging/rdma/hfi1: Adaptive PIO for short messages The change requires a new pio_busy field in the iowait structure to track the number of outstanding pios. The new counter together with the sdma counter serve as the basis for a packet by packet decision as to which egress mechanism to use. Since packets given to different egress mechanisms are not ordered, this scheme will preserve the order. The iowait drain/wait mechanisms are extended for a pio case. An additional qp wait flag is added for the PIO drain wait case. Currently the only pio wait is for buffers, so the no_bufs_available() routine name is changed to pio_wait() and a third argument is passed with one of the two pio wait flags to generalize the routine. A module parameter is added to hold a configurable threshold. For now, the module parameter is zero. A heuristic routine is added to return the func pointer of the proper egress routine to use. The heuristic is as follows: - SMI always uses pio - GSI,UD qps <= threshold use pio - UD qps > threadhold use sdma o No coordination with sdma is required because order is not required and this qp pio count is not maintained for UD - RC/UC ONLY packets <= threshold chose as follows: o If sdmas pending, use SDMA o Otherwise use pio and enable the pio tracking count at the time the pio buffer is allocated - RC/UC ONLY packets > threshold use SDMA o If pio's are pending the pio_wait with the new wait flag is called to delay for pios to drain The threshold is potentially reduced by the QP's mtu. The sc_buffer_alloc() has two additional args (a callback, a void *) which are exploited by the RC/UC cases to pass a new complete routine and a qp *. When the shadow ring completes the credit associated with a packet, the new complete routine is called. The verbs_pio_complete() will then decrement the busy count and trigger any drain waiters in qp destroy or reset. Reviewed-by: Jubin John <jubin.john@intel.com> Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-02-14 20:45:36 +00:00
/**
* hfi1_make_uc_req - construct a request packet (SEND, RDMA write)
* @qp: a pointer to the QP
* @ps: the current packet state
*
* Assume s_lock is held.
*
* Return 1 if constructed; otherwise, return 0.
*/
int hfi1_make_uc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
{
struct hfi1_qp_priv *priv = qp->priv;
struct ib_other_headers *ohdr;
struct rvt_swqe *wqe;
u32 hwords;
u32 bth0 = 0;
u32 len;
u32 pmtu = qp->pmtu;
int middle = 0;
ps->s_txreq = get_txreq(ps->dev, qp);
if (!ps->s_txreq)
goto bail_no_tx;
if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
goto bail;
/* We are in the error state, flush the work request. */
locking/atomics: COCCINELLE/treewide: Convert trivial ACCESS_ONCE() patterns to READ_ONCE()/WRITE_ONCE() Please do not apply this to mainline directly, instead please re-run the coccinelle script shown below and apply its output. For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't harmful, and changing them results in churn. However, for some features, the read/write distinction is critical to correct operation. To distinguish these cases, separate read/write accessors must be used. This patch migrates (most) remaining ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following coccinelle script: ---- // Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and // WRITE_ONCE() // $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: tj@kernel.org Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 21:07:29 +00:00
if (qp->s_last == READ_ONCE(qp->s_head))
goto bail;
/* If DMAs are in progress, we can't flush immediately. */
staging/rdma/hfi1: Adaptive PIO for short messages The change requires a new pio_busy field in the iowait structure to track the number of outstanding pios. The new counter together with the sdma counter serve as the basis for a packet by packet decision as to which egress mechanism to use. Since packets given to different egress mechanisms are not ordered, this scheme will preserve the order. The iowait drain/wait mechanisms are extended for a pio case. An additional qp wait flag is added for the PIO drain wait case. Currently the only pio wait is for buffers, so the no_bufs_available() routine name is changed to pio_wait() and a third argument is passed with one of the two pio wait flags to generalize the routine. A module parameter is added to hold a configurable threshold. For now, the module parameter is zero. A heuristic routine is added to return the func pointer of the proper egress routine to use. The heuristic is as follows: - SMI always uses pio - GSI,UD qps <= threshold use pio - UD qps > threadhold use sdma o No coordination with sdma is required because order is not required and this qp pio count is not maintained for UD - RC/UC ONLY packets <= threshold chose as follows: o If sdmas pending, use SDMA o Otherwise use pio and enable the pio tracking count at the time the pio buffer is allocated - RC/UC ONLY packets > threshold use SDMA o If pio's are pending the pio_wait with the new wait flag is called to delay for pios to drain The threshold is potentially reduced by the QP's mtu. The sc_buffer_alloc() has two additional args (a callback, a void *) which are exploited by the RC/UC cases to pass a new complete routine and a qp *. When the shadow ring completes the credit associated with a packet, the new complete routine is called. The verbs_pio_complete() will then decrement the busy count and trigger any drain waiters in qp destroy or reset. Reviewed-by: Jubin John <jubin.john@intel.com> Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-02-14 20:45:36 +00:00
if (iowait_sdma_pending(&priv->s_iowait)) {
qp->s_flags |= RVT_S_WAIT_DMA;
goto bail;
}
clear_ahg(qp);
wqe = rvt_get_swqe_ptr(qp, qp->s_last);
rvt_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
goto done_free_tx;
}
if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
/* header size in 32-bit words LRH+BTH = (8+12)/4. */
hwords = 5;
if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
else
ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
} else {
/* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
hwords = 7;
if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
(hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))))
ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
else
ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
}
/* Get the next send request. */
wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
qp->s_wqe = NULL;
switch (qp->s_state) {
default:
if (!(ib_rvt_state_ops[qp->state] &
RVT_PROCESS_NEXT_SEND_OK))
goto bail;
/* Check if send work queue is empty. */
locking/atomics: COCCINELLE/treewide: Convert trivial ACCESS_ONCE() patterns to READ_ONCE()/WRITE_ONCE() Please do not apply this to mainline directly, instead please re-run the coccinelle script shown below and apply its output. For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't harmful, and changing them results in churn. However, for some features, the read/write distinction is critical to correct operation. To distinguish these cases, separate read/write accessors must be used. This patch migrates (most) remaining ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following coccinelle script: ---- // Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and // WRITE_ONCE() // $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: tj@kernel.org Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 21:07:29 +00:00
if (qp->s_cur == READ_ONCE(qp->s_head)) {
clear_ahg(qp);
goto bail;
}
/*
* Local operations are processed immediately
* after all prior requests have completed.
*/
if (wqe->wr.opcode == IB_WR_REG_MR ||
wqe->wr.opcode == IB_WR_LOCAL_INV) {
IB/rdmavt, hfi1: Fix NFSoRDMA failure with FRMR enabled Hanging has been observed while writing a file over NFSoRDMA. Dmesg on the server contains messages like these: [ 931.992501] svcrdma: Error -22 posting RDMA_READ [ 952.076879] svcrdma: Error -22 posting RDMA_READ [ 982.154127] svcrdma: Error -22 posting RDMA_READ [ 1012.235884] svcrdma: Error -22 posting RDMA_READ [ 1042.319194] svcrdma: Error -22 posting RDMA_READ Here is why: With the base memory management extension enabled, FRMR is used instead of FMR. The xprtrdma server issues each RDMA read request as the following bundle: (1)IB_WR_REG_MR, signaled; (2)IB_WR_RDMA_READ, signaled; (3)IB_WR_LOCAL_INV, signaled & fencing. These requests are signaled. In order to generate completion, the fast register work request is processed by the hfi1 send engine after being posted to the work queue, and the corresponding lkey is not valid until the request is processed. However, the rdmavt driver validates lkey when the RDMA read request is posted and thus it fails immediately with error -EINVAL (-22). This patch changes the work flow of local operations (fast register and local invalidate) so that fast register work requests are always processed immediately to ensure that the corresponding lkey is valid when subsequent work requests are posted. Local invalidate requests are processed immediately if fencing is not required and no previous local invalidate request is pending. To allow completion generation for signaled local operations that have been processed before posting to the work queue, an internal send flag RVT_SEND_COMPLETION_ONLY is added. The hfi1 send engine checks this flag and only generates completion for such requests. Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com> Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-07-25 20:39:45 +00:00
int local_ops = 0;
int err = 0;
if (qp->s_last != qp->s_cur)
goto bail;
if (++qp->s_cur == qp->s_size)
qp->s_cur = 0;
IB/rdmavt, hfi1: Fix NFSoRDMA failure with FRMR enabled Hanging has been observed while writing a file over NFSoRDMA. Dmesg on the server contains messages like these: [ 931.992501] svcrdma: Error -22 posting RDMA_READ [ 952.076879] svcrdma: Error -22 posting RDMA_READ [ 982.154127] svcrdma: Error -22 posting RDMA_READ [ 1012.235884] svcrdma: Error -22 posting RDMA_READ [ 1042.319194] svcrdma: Error -22 posting RDMA_READ Here is why: With the base memory management extension enabled, FRMR is used instead of FMR. The xprtrdma server issues each RDMA read request as the following bundle: (1)IB_WR_REG_MR, signaled; (2)IB_WR_RDMA_READ, signaled; (3)IB_WR_LOCAL_INV, signaled & fencing. These requests are signaled. In order to generate completion, the fast register work request is processed by the hfi1 send engine after being posted to the work queue, and the corresponding lkey is not valid until the request is processed. However, the rdmavt driver validates lkey when the RDMA read request is posted and thus it fails immediately with error -EINVAL (-22). This patch changes the work flow of local operations (fast register and local invalidate) so that fast register work requests are always processed immediately to ensure that the corresponding lkey is valid when subsequent work requests are posted. Local invalidate requests are processed immediately if fencing is not required and no previous local invalidate request is pending. To allow completion generation for signaled local operations that have been processed before posting to the work queue, an internal send flag RVT_SEND_COMPLETION_ONLY is added. The hfi1 send engine checks this flag and only generates completion for such requests. Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com> Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-07-25 20:39:45 +00:00
if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
err = rvt_invalidate_rkey(
qp, wqe->wr.ex.invalidate_rkey);
IB/rdmavt, hfi1: Fix NFSoRDMA failure with FRMR enabled Hanging has been observed while writing a file over NFSoRDMA. Dmesg on the server contains messages like these: [ 931.992501] svcrdma: Error -22 posting RDMA_READ [ 952.076879] svcrdma: Error -22 posting RDMA_READ [ 982.154127] svcrdma: Error -22 posting RDMA_READ [ 1012.235884] svcrdma: Error -22 posting RDMA_READ [ 1042.319194] svcrdma: Error -22 posting RDMA_READ Here is why: With the base memory management extension enabled, FRMR is used instead of FMR. The xprtrdma server issues each RDMA read request as the following bundle: (1)IB_WR_REG_MR, signaled; (2)IB_WR_RDMA_READ, signaled; (3)IB_WR_LOCAL_INV, signaled & fencing. These requests are signaled. In order to generate completion, the fast register work request is processed by the hfi1 send engine after being posted to the work queue, and the corresponding lkey is not valid until the request is processed. However, the rdmavt driver validates lkey when the RDMA read request is posted and thus it fails immediately with error -EINVAL (-22). This patch changes the work flow of local operations (fast register and local invalidate) so that fast register work requests are always processed immediately to ensure that the corresponding lkey is valid when subsequent work requests are posted. Local invalidate requests are processed immediately if fencing is not required and no previous local invalidate request is pending. To allow completion generation for signaled local operations that have been processed before posting to the work queue, an internal send flag RVT_SEND_COMPLETION_ONLY is added. The hfi1 send engine checks this flag and only generates completion for such requests. Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com> Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-07-25 20:39:45 +00:00
local_ops = 1;
}
rvt_send_complete(qp, wqe, err ? IB_WC_LOC_PROT_ERR
: IB_WC_SUCCESS);
IB/rdmavt, hfi1: Fix NFSoRDMA failure with FRMR enabled Hanging has been observed while writing a file over NFSoRDMA. Dmesg on the server contains messages like these: [ 931.992501] svcrdma: Error -22 posting RDMA_READ [ 952.076879] svcrdma: Error -22 posting RDMA_READ [ 982.154127] svcrdma: Error -22 posting RDMA_READ [ 1012.235884] svcrdma: Error -22 posting RDMA_READ [ 1042.319194] svcrdma: Error -22 posting RDMA_READ Here is why: With the base memory management extension enabled, FRMR is used instead of FMR. The xprtrdma server issues each RDMA read request as the following bundle: (1)IB_WR_REG_MR, signaled; (2)IB_WR_RDMA_READ, signaled; (3)IB_WR_LOCAL_INV, signaled & fencing. These requests are signaled. In order to generate completion, the fast register work request is processed by the hfi1 send engine after being posted to the work queue, and the corresponding lkey is not valid until the request is processed. However, the rdmavt driver validates lkey when the RDMA read request is posted and thus it fails immediately with error -EINVAL (-22). This patch changes the work flow of local operations (fast register and local invalidate) so that fast register work requests are always processed immediately to ensure that the corresponding lkey is valid when subsequent work requests are posted. Local invalidate requests are processed immediately if fencing is not required and no previous local invalidate request is pending. To allow completion generation for signaled local operations that have been processed before posting to the work queue, an internal send flag RVT_SEND_COMPLETION_ONLY is added. The hfi1 send engine checks this flag and only generates completion for such requests. Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com> Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-07-25 20:39:45 +00:00
if (local_ops)
atomic_dec(&qp->local_ops_pending);
goto done_free_tx;
}
/*
* Start a new request.
*/
qp->s_psn = wqe->psn;
qp->s_sge.sge = wqe->sg_list[0];
qp->s_sge.sg_list = wqe->sg_list + 1;
qp->s_sge.num_sge = wqe->wr.num_sge;
qp->s_sge.total_len = wqe->length;
len = wqe->length;
qp->s_len = len;
switch (wqe->wr.opcode) {
case IB_WR_SEND:
case IB_WR_SEND_WITH_IMM:
if (len > pmtu) {
qp->s_state = OP(SEND_FIRST);
len = pmtu;
break;
}
if (wqe->wr.opcode == IB_WR_SEND) {
qp->s_state = OP(SEND_ONLY);
} else {
qp->s_state =
OP(SEND_ONLY_WITH_IMMEDIATE);
/* Immediate data comes after the BTH */
ohdr->u.imm_data = wqe->wr.ex.imm_data;
hwords += 1;
}
if (wqe->wr.send_flags & IB_SEND_SOLICITED)
bth0 |= IB_BTH_SOLICITED;
qp->s_wqe = wqe;
if (++qp->s_cur >= qp->s_size)
qp->s_cur = 0;
break;
case IB_WR_RDMA_WRITE:
case IB_WR_RDMA_WRITE_WITH_IMM:
ohdr->u.rc.reth.vaddr =
cpu_to_be64(wqe->rdma_wr.remote_addr);
ohdr->u.rc.reth.rkey =
cpu_to_be32(wqe->rdma_wr.rkey);
ohdr->u.rc.reth.length = cpu_to_be32(len);
hwords += sizeof(struct ib_reth) / 4;
if (len > pmtu) {
qp->s_state = OP(RDMA_WRITE_FIRST);
len = pmtu;
break;
}
if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
qp->s_state = OP(RDMA_WRITE_ONLY);
} else {
qp->s_state =
OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
/* Immediate data comes after the RETH */
ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
hwords += 1;
if (wqe->wr.send_flags & IB_SEND_SOLICITED)
bth0 |= IB_BTH_SOLICITED;
}
qp->s_wqe = wqe;
if (++qp->s_cur >= qp->s_size)
qp->s_cur = 0;
break;
default:
goto bail;
}
break;
case OP(SEND_FIRST):
qp->s_state = OP(SEND_MIDDLE);
fallthrough;
case OP(SEND_MIDDLE):
len = qp->s_len;
if (len > pmtu) {
len = pmtu;
middle = HFI1_CAP_IS_KSET(SDMA_AHG);
break;
}
if (wqe->wr.opcode == IB_WR_SEND) {
qp->s_state = OP(SEND_LAST);
} else {
qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
/* Immediate data comes after the BTH */
ohdr->u.imm_data = wqe->wr.ex.imm_data;
hwords += 1;
}
if (wqe->wr.send_flags & IB_SEND_SOLICITED)
bth0 |= IB_BTH_SOLICITED;
qp->s_wqe = wqe;
if (++qp->s_cur >= qp->s_size)
qp->s_cur = 0;
break;
case OP(RDMA_WRITE_FIRST):
qp->s_state = OP(RDMA_WRITE_MIDDLE);
fallthrough;
case OP(RDMA_WRITE_MIDDLE):
len = qp->s_len;
if (len > pmtu) {
len = pmtu;
middle = HFI1_CAP_IS_KSET(SDMA_AHG);
break;
}
if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
qp->s_state = OP(RDMA_WRITE_LAST);
} else {
qp->s_state =
OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
/* Immediate data comes after the BTH */
ohdr->u.imm_data = wqe->wr.ex.imm_data;
hwords += 1;
if (wqe->wr.send_flags & IB_SEND_SOLICITED)
bth0 |= IB_BTH_SOLICITED;
}
qp->s_wqe = wqe;
if (++qp->s_cur >= qp->s_size)
qp->s_cur = 0;
break;
}
qp->s_len -= len;
ps->s_txreq->hdr_dwords = hwords;
staging/rdma/hfi1: Adaptive PIO for short messages The change requires a new pio_busy field in the iowait structure to track the number of outstanding pios. The new counter together with the sdma counter serve as the basis for a packet by packet decision as to which egress mechanism to use. Since packets given to different egress mechanisms are not ordered, this scheme will preserve the order. The iowait drain/wait mechanisms are extended for a pio case. An additional qp wait flag is added for the PIO drain wait case. Currently the only pio wait is for buffers, so the no_bufs_available() routine name is changed to pio_wait() and a third argument is passed with one of the two pio wait flags to generalize the routine. A module parameter is added to hold a configurable threshold. For now, the module parameter is zero. A heuristic routine is added to return the func pointer of the proper egress routine to use. The heuristic is as follows: - SMI always uses pio - GSI,UD qps <= threshold use pio - UD qps > threadhold use sdma o No coordination with sdma is required because order is not required and this qp pio count is not maintained for UD - RC/UC ONLY packets <= threshold chose as follows: o If sdmas pending, use SDMA o Otherwise use pio and enable the pio tracking count at the time the pio buffer is allocated - RC/UC ONLY packets > threshold use SDMA o If pio's are pending the pio_wait with the new wait flag is called to delay for pios to drain The threshold is potentially reduced by the QP's mtu. The sc_buffer_alloc() has two additional args (a callback, a void *) which are exploited by the RC/UC cases to pass a new complete routine and a qp *. When the shadow ring completes the credit associated with a packet, the new complete routine is called. The verbs_pio_complete() will then decrement the busy count and trigger any drain waiters in qp destroy or reset. Reviewed-by: Jubin John <jubin.john@intel.com> Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com> Signed-off-by: Doug Ledford <dledford@redhat.com>
2016-02-14 20:45:36 +00:00
ps->s_txreq->sde = priv->s_sde;
ps->s_txreq->ss = &qp->s_sge;
ps->s_txreq->s_cur_size = len;
hfi1_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24),
qp->remote_qpn, mask_psn(qp->s_psn++),
middle, ps);
return 1;
done_free_tx:
hfi1_put_txreq(ps->s_txreq);
ps->s_txreq = NULL;
return 1;
bail:
hfi1_put_txreq(ps->s_txreq);
bail_no_tx:
ps->s_txreq = NULL;
qp->s_flags &= ~RVT_S_BUSY;
return 0;
}
/**
* hfi1_uc_rcv - handle an incoming UC packet
* @packet: the packet structure
*
* This is called from qp_rcv() to process an incoming UC packet
* for the given QP.
* Called at interrupt level.
*/
void hfi1_uc_rcv(struct hfi1_packet *packet)
{
struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
void *data = packet->payload;
u32 tlen = packet->tlen;
struct rvt_qp *qp = packet->qp;
struct ib_other_headers *ohdr = packet->ohdr;
u32 opcode = packet->opcode;
u32 hdrsize = packet->hlen;
u32 psn;
u32 pad = packet->pad;
struct ib_wc wc;
u32 pmtu = qp->pmtu;
struct ib_reth *reth;
int ret;
u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
if (hfi1_ruc_check_hdr(ibp, packet))
return;
process_ecn(qp, packet);
psn = ib_bth_get_psn(ohdr);
/* Compare the PSN verses the expected PSN. */
if (unlikely(cmp_psn(psn, qp->r_psn) != 0)) {
/*
* Handle a sequence error.
* Silently drop any current message.
*/
qp->r_psn = psn;
inv:
if (qp->r_state == OP(SEND_FIRST) ||
qp->r_state == OP(SEND_MIDDLE)) {
set_bit(RVT_R_REWIND_SGE, &qp->r_aflags);
qp->r_sge.num_sge = 0;
} else {
rvt_put_ss(&qp->r_sge);
}
qp->r_state = OP(SEND_LAST);
switch (opcode) {
case OP(SEND_FIRST):
case OP(SEND_ONLY):
case OP(SEND_ONLY_WITH_IMMEDIATE):
goto send_first;
case OP(RDMA_WRITE_FIRST):
case OP(RDMA_WRITE_ONLY):
case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
goto rdma_first;
default:
goto drop;
}
}
/* Check for opcode sequence errors. */
switch (qp->r_state) {
case OP(SEND_FIRST):
case OP(SEND_MIDDLE):
if (opcode == OP(SEND_MIDDLE) ||
opcode == OP(SEND_LAST) ||
opcode == OP(SEND_LAST_WITH_IMMEDIATE))
break;
goto inv;
case OP(RDMA_WRITE_FIRST):
case OP(RDMA_WRITE_MIDDLE):
if (opcode == OP(RDMA_WRITE_MIDDLE) ||
opcode == OP(RDMA_WRITE_LAST) ||
opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
break;
goto inv;
default:
if (opcode == OP(SEND_FIRST) ||
opcode == OP(SEND_ONLY) ||
opcode == OP(SEND_ONLY_WITH_IMMEDIATE) ||
opcode == OP(RDMA_WRITE_FIRST) ||
opcode == OP(RDMA_WRITE_ONLY) ||
opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
break;
goto inv;
}
if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
rvt_comm_est(qp);
/* OK, process the packet. */
switch (opcode) {
case OP(SEND_FIRST):
case OP(SEND_ONLY):
case OP(SEND_ONLY_WITH_IMMEDIATE):
send_first:
if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) {
qp->r_sge = qp->s_rdma_read_sge;
} else {
ret = rvt_get_rwqe(qp, false);
if (ret < 0)
goto op_err;
if (!ret)
goto drop;
/*
* qp->s_rdma_read_sge will be the owner
* of the mr references.
*/
qp->s_rdma_read_sge = qp->r_sge;
}
qp->r_rcv_len = 0;
if (opcode == OP(SEND_ONLY))
goto no_immediate_data;
else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
goto send_last_imm;
fallthrough;
case OP(SEND_MIDDLE):
/* Check for invalid length PMTU or posted rwqe len. */
/*
* There will be no padding for 9B packet but 16B packets
* will come in with some padding since we always add
* CRC and LT bytes which will need to be flit aligned
*/
if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
goto rewind;
qp->r_rcv_len += pmtu;
if (unlikely(qp->r_rcv_len > qp->r_len))
goto rewind;
rvt_copy_sge(qp, &qp->r_sge, data, pmtu, false, false);
break;
case OP(SEND_LAST_WITH_IMMEDIATE):
send_last_imm:
wc.ex.imm_data = ohdr->u.imm_data;
wc.wc_flags = IB_WC_WITH_IMM;
goto send_last;
case OP(SEND_LAST):
no_immediate_data:
wc.ex.imm_data = 0;
wc.wc_flags = 0;
send_last:
/* Check for invalid length. */
/* LAST len should be >= 1 */
if (unlikely(tlen < (hdrsize + extra_bytes)))
goto rewind;
/* Don't count the CRC. */
tlen -= (hdrsize + extra_bytes);
wc.byte_len = tlen + qp->r_rcv_len;
if (unlikely(wc.byte_len > qp->r_len))
goto rewind;
wc.opcode = IB_WC_RECV;
rvt_copy_sge(qp, &qp->r_sge, data, tlen, false, false);
rvt_put_ss(&qp->s_rdma_read_sge);
last_imm:
wc.wr_id = qp->r_wr_id;
wc.status = IB_WC_SUCCESS;
wc.qp = &qp->ibqp;
wc.src_qp = qp->remote_qpn;
wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
/*
* It seems that IB mandates the presence of an SL in a
* work completion only for the UD transport (see section
* 11.4.2 of IBTA Vol. 1).
*
* However, the way the SL is chosen below is consistent
* with the way that IB/qib works and is trying avoid
* introducing incompatibilities.
*
* See also OPA Vol. 1, section 9.7.6, and table 9-17.
*/
wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
/* zero fields that are N/A */
wc.vendor_err = 0;
wc.pkey_index = 0;
wc.dlid_path_bits = 0;
wc.port_num = 0;
/* Signal completion event if the solicited bit is set. */
rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr));
break;
case OP(RDMA_WRITE_FIRST):
case OP(RDMA_WRITE_ONLY):
case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): /* consume RWQE */
rdma_first:
if (unlikely(!(qp->qp_access_flags &
IB_ACCESS_REMOTE_WRITE))) {
goto drop;
}
reth = &ohdr->u.rc.reth;
qp->r_len = be32_to_cpu(reth->length);
qp->r_rcv_len = 0;
qp->r_sge.sg_list = NULL;
if (qp->r_len != 0) {
u32 rkey = be32_to_cpu(reth->rkey);
u64 vaddr = be64_to_cpu(reth->vaddr);
int ok;
/* Check rkey */
ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len,
vaddr, rkey, IB_ACCESS_REMOTE_WRITE);
if (unlikely(!ok))
goto drop;
qp->r_sge.num_sge = 1;
} else {
qp->r_sge.num_sge = 0;
qp->r_sge.sge.mr = NULL;
qp->r_sge.sge.vaddr = NULL;
qp->r_sge.sge.length = 0;
qp->r_sge.sge.sge_length = 0;
}
if (opcode == OP(RDMA_WRITE_ONLY)) {
goto rdma_last;
} else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) {
wc.ex.imm_data = ohdr->u.rc.imm_data;
goto rdma_last_imm;
}
fallthrough;
case OP(RDMA_WRITE_MIDDLE):
/* Check for invalid length PMTU or posted rwqe len. */
if (unlikely(tlen != (hdrsize + pmtu + 4)))
goto drop;
qp->r_rcv_len += pmtu;
if (unlikely(qp->r_rcv_len > qp->r_len))
goto drop;
rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
break;
case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
wc.ex.imm_data = ohdr->u.imm_data;
rdma_last_imm:
wc.wc_flags = IB_WC_WITH_IMM;
/* Check for invalid length. */
/* LAST len should be >= 1 */
if (unlikely(tlen < (hdrsize + pad + 4)))
goto drop;
/* Don't count the CRC. */
tlen -= (hdrsize + extra_bytes);
if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
goto drop;
if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) {
rvt_put_ss(&qp->s_rdma_read_sge);
} else {
ret = rvt_get_rwqe(qp, true);
if (ret < 0)
goto op_err;
if (!ret)
goto drop;
}
wc.byte_len = qp->r_len;
wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
rvt_put_ss(&qp->r_sge);
goto last_imm;
case OP(RDMA_WRITE_LAST):
rdma_last:
/* Check for invalid length. */
/* LAST len should be >= 1 */
if (unlikely(tlen < (hdrsize + pad + 4)))
goto drop;
/* Don't count the CRC. */
tlen -= (hdrsize + extra_bytes);
if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
goto drop;
rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
rvt_put_ss(&qp->r_sge);
break;
default:
/* Drop packet for unknown opcodes. */
goto drop;
}
qp->r_psn++;
qp->r_state = opcode;
return;
rewind:
set_bit(RVT_R_REWIND_SGE, &qp->r_aflags);
qp->r_sge.num_sge = 0;
drop:
ibp->rvp.n_pkt_drops++;
return;
op_err:
rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
}