2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/err.h>
49 #include <linux/vmalloc.h>
50 #include <linux/hash.h>
51 #include <linux/module.h>
52 #include <linux/random.h>
53 #include <linux/seq_file.h>
54 #include <rdma/rdma_vt.h>
55 #include <rdma/rdmavt_qp.h>
60 #include "verbs_txreq.h"
62 unsigned int hfi1_qp_table_size = 256;
63 module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
64 MODULE_PARM_DESC(qp_table_size, "QP table size");
66 static void flush_tx_list(struct rvt_qp *qp);
67 static int iowait_sleep(
68 struct sdma_engine *sde,
70 struct sdma_txreq *stx,
72 static void iowait_wakeup(struct iowait *wait, int reason);
73 static void iowait_sdma_drained(struct iowait *wait);
74 static void qp_pio_drain(struct rvt_qp *qp);
76 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
77 struct rvt_qpn_map *map, unsigned off)
79 return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
83 * Convert the AETH credit code into the number of credits.
85 static const u16 credit_table[31] = {
119 static void flush_tx_list(struct rvt_qp *qp)
121 struct hfi1_qp_priv *priv = qp->priv;
123 while (!list_empty(&priv->s_iowait.tx_head)) {
124 struct sdma_txreq *tx;
126 tx = list_first_entry(
127 &priv->s_iowait.tx_head,
130 list_del_init(&tx->list);
132 container_of(tx, struct verbs_txreq, txreq));
136 static void flush_iowait(struct rvt_qp *qp)
138 struct hfi1_qp_priv *priv = qp->priv;
139 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
142 write_seqlock_irqsave(&dev->iowait_lock, flags);
143 if (!list_empty(&priv->s_iowait.list)) {
144 list_del_init(&priv->s_iowait.list);
145 if (atomic_dec_and_test(&qp->refcount))
148 write_sequnlock_irqrestore(&dev->iowait_lock, flags);
151 static inline int opa_mtu_enum_to_int(int mtu)
154 case OPA_MTU_8192: return 8192;
155 case OPA_MTU_10240: return 10240;
161 * This function is what we would push to the core layer if we wanted to be a
162 * "first class citizen". Instead we hide this here and rely on Verbs ULPs
163 * to blindly pass the MTU enum value from the PathRecord to us.
165 * The actual flag used to determine "8k MTU" will change and is currently
168 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
170 int val = opa_mtu_enum_to_int((int)mtu);
174 return ib_mtu_enum_to_int(mtu);
177 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
178 int attr_mask, struct ib_udata *udata)
180 struct ib_qp *ibqp = &qp->ibqp;
181 struct hfi1_ibdev *dev = to_idev(ibqp->device);
182 struct hfi1_devdata *dd = dd_from_dev(dev);
185 if (attr_mask & IB_QP_AV) {
186 sc = ah_to_sc(ibqp->device, &attr->ah_attr);
190 if (!qp_to_sdma_engine(qp, sc) &&
191 dd->flags & HFI1_HAS_SEND_DMA)
194 if (!qp_to_send_context(qp, sc))
198 if (attr_mask & IB_QP_ALT_PATH) {
199 sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
203 if (!qp_to_sdma_engine(qp, sc) &&
204 dd->flags & HFI1_HAS_SEND_DMA)
207 if (!qp_to_send_context(qp, sc))
214 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
215 int attr_mask, struct ib_udata *udata)
217 struct ib_qp *ibqp = &qp->ibqp;
218 struct hfi1_qp_priv *priv = qp->priv;
220 if (attr_mask & IB_QP_AV) {
221 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
222 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
223 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
226 if (attr_mask & IB_QP_PATH_MIG_STATE &&
227 attr->path_mig_state == IB_MIG_MIGRATED &&
228 qp->s_mig_state == IB_MIG_ARMED) {
229 qp->s_flags |= RVT_S_AHG_CLEAR;
230 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
231 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
232 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
237 * hfi1_check_send_wqe - validate wqe
239 * @wqe - The built wqe
241 * validate wqe. This is called
242 * prior to inserting the wqe into
243 * the ring but after the wqe has been
246 * Returns 0 on success, -EINVAL on failure
249 int hfi1_check_send_wqe(struct rvt_qp *qp,
250 struct rvt_swqe *wqe)
252 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
255 switch (qp->ibqp.qp_type) {
258 if (wqe->length > 0x80000000U)
262 ah = ibah_to_rvtah(wqe->ud_wr.ah);
263 if (wqe->length > (1 << ah->log_pmtu))
268 ah = ibah_to_rvtah(wqe->ud_wr.ah);
269 if (wqe->length > (1 << ah->log_pmtu))
271 if (ibp->sl_to_sc[ah->attr.sl] == 0xf)
276 return wqe->length <= piothreshold;
280 * hfi1_compute_aeth - compute the AETH (syndrome + MSN)
281 * @qp: the queue pair to compute the AETH for
285 __be32 hfi1_compute_aeth(struct rvt_qp *qp)
287 u32 aeth = qp->r_msn & HFI1_MSN_MASK;
291 * Shared receive queues don't generate credits.
292 * Set the credit field to the invalid value.
294 aeth |= HFI1_AETH_CREDIT_INVAL << HFI1_AETH_CREDIT_SHIFT;
298 struct rvt_rwq *wq = qp->r_rq.wq;
302 /* sanity check pointers before trusting them */
304 if (head >= qp->r_rq.size)
307 if (tail >= qp->r_rq.size)
310 * Compute the number of credits available (RWQEs).
311 * There is a small chance that the pair of reads are
312 * not atomic, which is OK, since the fuzziness is
313 * resolved as further ACKs go out.
315 credits = head - tail;
316 if ((int)credits < 0)
317 credits += qp->r_rq.size;
319 * Binary search the credit table to find the code to
326 if (credit_table[x] == credits)
328 if (credit_table[x] > credits) {
336 aeth |= x << HFI1_AETH_CREDIT_SHIFT;
338 return cpu_to_be32(aeth);
342 * _hfi1_schedule_send - schedule progress
345 * This schedules qp progress w/o regard to the s_flags.
347 * It is only used in the post send, which doesn't hold
350 void _hfi1_schedule_send(struct rvt_qp *qp)
352 struct hfi1_qp_priv *priv = qp->priv;
353 struct hfi1_ibport *ibp =
354 to_iport(qp->ibqp.device, qp->port_num);
355 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
356 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
358 iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
361 cpumask_first(cpumask_of_node(dd->node)));
364 static void qp_pio_drain(struct rvt_qp *qp)
366 struct hfi1_ibdev *dev;
367 struct hfi1_qp_priv *priv = qp->priv;
369 if (!priv->s_sendcontext)
371 dev = to_idev(qp->ibqp.device);
372 while (iowait_pio_pending(&priv->s_iowait)) {
373 write_seqlock_irq(&dev->iowait_lock);
374 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
375 write_sequnlock_irq(&dev->iowait_lock);
376 iowait_pio_drain(&priv->s_iowait);
377 write_seqlock_irq(&dev->iowait_lock);
378 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
379 write_sequnlock_irq(&dev->iowait_lock);
384 * hfi1_schedule_send - schedule progress
387 * This schedules qp progress and caller should hold
390 void hfi1_schedule_send(struct rvt_qp *qp)
392 if (hfi1_send_ok(qp))
393 _hfi1_schedule_send(qp);
397 * hfi1_get_credit - flush the send work queue of a QP
398 * @qp: the qp who's send work queue to flush
399 * @aeth: the Acknowledge Extended Transport Header
401 * The QP s_lock should be held.
403 void hfi1_get_credit(struct rvt_qp *qp, u32 aeth)
405 u32 credit = (aeth >> HFI1_AETH_CREDIT_SHIFT) & HFI1_AETH_CREDIT_MASK;
408 * If the credit is invalid, we can send
409 * as many packets as we like. Otherwise, we have to
410 * honor the credit field.
412 if (credit == HFI1_AETH_CREDIT_INVAL) {
413 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
414 qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
415 if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
416 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
417 hfi1_schedule_send(qp);
420 } else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
421 /* Compute new LSN (i.e., MSN + credit) */
422 credit = (aeth + credit_table[credit]) & HFI1_MSN_MASK;
423 if (cmp_msn(credit, qp->s_lsn) > 0) {
425 if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
426 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
427 hfi1_schedule_send(qp);
433 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
437 spin_lock_irqsave(&qp->s_lock, flags);
438 if (qp->s_flags & flag) {
439 qp->s_flags &= ~flag;
440 trace_hfi1_qpwakeup(qp, flag);
441 hfi1_schedule_send(qp);
443 spin_unlock_irqrestore(&qp->s_lock, flags);
444 /* Notify hfi1_destroy_qp() if it is waiting. */
445 if (atomic_dec_and_test(&qp->refcount))
449 static int iowait_sleep(
450 struct sdma_engine *sde,
452 struct sdma_txreq *stx,
455 struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
457 struct hfi1_qp_priv *priv;
460 struct hfi1_ibdev *dev;
465 spin_lock_irqsave(&qp->s_lock, flags);
466 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
468 * If we couldn't queue the DMA request, save the info
469 * and try again later rather than destroying the
470 * buffer and undoing the side effects of the copy.
472 /* Make a common routine? */
473 dev = &sde->dd->verbs_dev;
474 list_add_tail(&stx->list, &wait->tx_head);
475 write_seqlock(&dev->iowait_lock);
476 if (sdma_progress(sde, seq, stx))
478 if (list_empty(&priv->s_iowait.list)) {
479 struct hfi1_ibport *ibp =
480 to_iport(qp->ibqp.device, qp->port_num);
482 ibp->rvp.n_dmawait++;
483 qp->s_flags |= RVT_S_WAIT_DMA_DESC;
484 list_add_tail(&priv->s_iowait.list, &sde->dmawait);
485 trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
486 atomic_inc(&qp->refcount);
488 write_sequnlock(&dev->iowait_lock);
489 qp->s_flags &= ~RVT_S_BUSY;
490 spin_unlock_irqrestore(&qp->s_lock, flags);
493 spin_unlock_irqrestore(&qp->s_lock, flags);
498 write_sequnlock(&dev->iowait_lock);
499 spin_unlock_irqrestore(&qp->s_lock, flags);
500 list_del_init(&stx->list);
504 static void iowait_wakeup(struct iowait *wait, int reason)
506 struct rvt_qp *qp = iowait_to_qp(wait);
508 WARN_ON(reason != SDMA_AVAIL_REASON);
509 hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
512 static void iowait_sdma_drained(struct iowait *wait)
514 struct rvt_qp *qp = iowait_to_qp(wait);
517 * This happens when the send engine notes
518 * a QP in the error state and cannot
519 * do the flush work until that QP's
520 * sdma work has finished.
522 if (qp->s_flags & RVT_S_WAIT_DMA) {
523 qp->s_flags &= ~RVT_S_WAIT_DMA;
524 hfi1_schedule_send(qp);
530 * qp_to_sdma_engine - map a qp to a send engine
535 * A send engine for the qp or NULL for SMI type qp.
537 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
539 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
540 struct sdma_engine *sde;
542 if (!(dd->flags & HFI1_HAS_SEND_DMA))
544 switch (qp->ibqp.qp_type) {
550 sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
555 * qp_to_send_context - map a qp to a send context
560 * A send context for the qp
562 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
564 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
566 switch (qp->ibqp.qp_type) {
568 /* SMA packets to VL15 */
569 return dd->vld[15].sc;
574 return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
579 struct hfi1_ibdev *dev;
585 struct qp_iter *qp_iter_init(struct hfi1_ibdev *dev)
587 struct qp_iter *iter;
589 iter = kzalloc(sizeof(*iter), GFP_KERNEL);
594 iter->specials = dev->rdi.ibdev.phys_port_cnt * 2;
595 if (qp_iter_next(iter)) {
603 int qp_iter_next(struct qp_iter *iter)
605 struct hfi1_ibdev *dev = iter->dev;
608 struct rvt_qp *pqp = iter->qp;
612 * The approach is to consider the special qps
613 * as an additional table entries before the
614 * real hash table. Since the qp code sets
615 * the qp->next hash link to NULL, this works just fine.
617 * iter->specials is 2 * # ports
619 * n = 0..iter->specials is the special qp indices
621 * n = iter->specials..dev->rdi.qp_dev->qp_table_size+iter->specials are
622 * the potential hash bucket entries
625 for (; n < dev->rdi.qp_dev->qp_table_size + iter->specials; n++) {
627 qp = rcu_dereference(pqp->next);
629 if (n < iter->specials) {
630 struct hfi1_pportdata *ppd;
631 struct hfi1_ibport *ibp;
634 pidx = n % dev->rdi.ibdev.phys_port_cnt;
635 ppd = &dd_from_dev(dev)->pport[pidx];
636 ibp = &ppd->ibport_data;
639 qp = rcu_dereference(ibp->rvp.qp[0]);
641 qp = rcu_dereference(ibp->rvp.qp[1]);
643 qp = rcu_dereference(
644 dev->rdi.qp_dev->qp_table[
645 (n - iter->specials)]);
658 static const char * const qp_type_str[] = {
659 "SMI", "GSI", "RC", "UC", "UD",
662 static int qp_idle(struct rvt_qp *qp)
665 qp->s_last == qp->s_acked &&
666 qp->s_acked == qp->s_cur &&
667 qp->s_cur == qp->s_tail &&
668 qp->s_tail == qp->s_head;
671 void qp_iter_print(struct seq_file *s, struct qp_iter *iter)
673 struct rvt_swqe *wqe;
674 struct rvt_qp *qp = iter->qp;
675 struct hfi1_qp_priv *priv = qp->priv;
676 struct sdma_engine *sde;
677 struct send_context *send_context;
679 sde = qp_to_sdma_engine(qp, priv->s_sc);
680 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
681 send_context = qp_to_send_context(qp, priv->s_sc);
683 "N %d %s QP %x R %u %s %u %u %u f=%x %u %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d\n",
685 qp_idle(qp) ? "I" : "B",
687 atomic_read(&qp->refcount),
688 qp_type_str[qp->ibqp.qp_type],
690 wqe ? wqe->wr.opcode : 0,
693 iowait_sdma_pending(&priv->s_iowait),
694 iowait_pio_pending(&priv->s_iowait),
695 !list_empty(&priv->s_iowait.list),
700 qp->s_psn, qp->s_next_psn,
701 qp->s_sending_psn, qp->s_sending_hpsn,
702 qp->s_last, qp->s_acked, qp->s_cur,
703 qp->s_tail, qp->s_head, qp->s_size,
706 qp->remote_ah_attr.dlid,
707 qp->remote_ah_attr.sl,
713 sde ? sde->this_idx : 0,
715 send_context ? send_context->sw_index : 0,
716 ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
717 ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
721 void qp_comm_est(struct rvt_qp *qp)
723 qp->r_flags |= RVT_R_COMM_EST;
724 if (qp->ibqp.event_handler) {
727 ev.device = qp->ibqp.device;
728 ev.element.qp = &qp->ibqp;
729 ev.event = IB_EVENT_COMM_EST;
730 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
734 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp,
737 struct hfi1_qp_priv *priv;
739 priv = kzalloc_node(sizeof(*priv), gfp, rdi->dparms.node);
741 return ERR_PTR(-ENOMEM);
745 priv->s_hdr = kzalloc_node(sizeof(*priv->s_hdr), gfp, rdi->dparms.node);
748 return ERR_PTR(-ENOMEM);
750 setup_timer(&priv->s_rnr_timer, hfi1_rc_rnr_retry, (unsigned long)qp);
751 qp->s_timer.function = hfi1_rc_timeout;
755 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
757 struct hfi1_qp_priv *priv = qp->priv;
763 unsigned free_all_qps(struct rvt_dev_info *rdi)
765 struct hfi1_ibdev *verbs_dev = container_of(rdi,
768 struct hfi1_devdata *dd = container_of(verbs_dev,
772 unsigned qp_inuse = 0;
774 for (n = 0; n < dd->num_pports; n++) {
775 struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
778 if (rcu_dereference(ibp->rvp.qp[0]))
780 if (rcu_dereference(ibp->rvp.qp[1]))
788 void flush_qp_waiters(struct rvt_qp *qp)
791 hfi1_stop_rc_timers(qp);
794 void stop_send_queue(struct rvt_qp *qp)
796 struct hfi1_qp_priv *priv = qp->priv;
798 cancel_work_sync(&priv->s_iowait.iowork);
799 hfi1_del_timers_sync(qp);
802 void quiesce_qp(struct rvt_qp *qp)
804 struct hfi1_qp_priv *priv = qp->priv;
806 iowait_sdma_drain(&priv->s_iowait);
811 void notify_qp_reset(struct rvt_qp *qp)
813 struct hfi1_qp_priv *priv = qp->priv;
821 iowait_sdma_drained);
822 priv->r_adefered = 0;
827 * Switch to alternate path.
828 * The QP s_lock should be held and interrupts disabled.
830 void hfi1_migrate_qp(struct rvt_qp *qp)
832 struct hfi1_qp_priv *priv = qp->priv;
835 qp->s_mig_state = IB_MIG_MIGRATED;
836 qp->remote_ah_attr = qp->alt_ah_attr;
837 qp->port_num = qp->alt_ah_attr.port_num;
838 qp->s_pkey_index = qp->s_alt_pkey_index;
839 qp->s_flags |= RVT_S_AHG_CLEAR;
840 priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
841 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
843 ev.device = qp->ibqp.device;
844 ev.element.qp = &qp->ibqp;
845 ev.event = IB_EVENT_PATH_MIG;
846 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
849 int mtu_to_path_mtu(u32 mtu)
851 return mtu_to_enum(mtu, OPA_MTU_8192);
854 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
857 struct hfi1_ibdev *verbs_dev = container_of(rdi,
860 struct hfi1_devdata *dd = container_of(verbs_dev,
863 struct hfi1_ibport *ibp;
866 ibp = &dd->pport[qp->port_num - 1].ibport_data;
867 sc = ibp->sl_to_sc[qp->remote_ah_attr.sl];
868 vl = sc_to_vlt(dd, sc);
870 mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
871 if (vl < PER_VL_SEND_CONTEXTS)
872 mtu = min_t(u32, mtu, dd->vld[vl].mtu);
876 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
877 struct ib_qp_attr *attr)
879 int mtu, pidx = qp->port_num - 1;
880 struct hfi1_ibdev *verbs_dev = container_of(rdi,
883 struct hfi1_devdata *dd = container_of(verbs_dev,
886 mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
888 return -1; /* values less than 0 are error */
890 if (mtu > dd->pport[pidx].ibmtu)
891 return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
893 return attr->path_mtu;
896 void notify_error_qp(struct rvt_qp *qp)
898 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
899 struct hfi1_qp_priv *priv = qp->priv;
901 write_seqlock(&dev->iowait_lock);
902 if (!list_empty(&priv->s_iowait.list) && !(qp->s_flags & RVT_S_BUSY)) {
903 qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
904 list_del_init(&priv->s_iowait.list);
905 if (atomic_dec_and_test(&qp->refcount))
908 write_sequnlock(&dev->iowait_lock);
910 if (!(qp->s_flags & RVT_S_BUSY)) {
913 rvt_put_mr(qp->s_rdma_mr);
914 qp->s_rdma_mr = NULL;
921 * hfi1_error_port_qps - put a port's RC/UC qps into error state
923 * @sl: the service level.
925 * This function places all RC/UC qps with a given service level into error
926 * state. It is generally called to force upper lay apps to abandon stale qps
927 * after an sl->sc mapping change.
929 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
931 struct rvt_qp *qp = NULL;
932 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
933 struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
940 /* Deal only with RC/UC qps that use the given SL. */
941 for (n = 0; n < dev->rdi.qp_dev->qp_table_size; n++) {
942 for (qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]); qp;
943 qp = rcu_dereference(qp->next)) {
944 if (qp->port_num == ppd->port &&
945 (qp->ibqp.qp_type == IB_QPT_UC ||
946 qp->ibqp.qp_type == IB_QPT_RC) &&
947 qp->remote_ah_attr.sl == sl &&
948 (ib_rvt_state_ops[qp->state] &
950 spin_lock_irq(&qp->r_lock);
951 spin_lock(&qp->s_hlock);
952 spin_lock(&qp->s_lock);
953 lastwqe = rvt_error_qp(qp,
955 spin_unlock(&qp->s_lock);
956 spin_unlock(&qp->s_hlock);
957 spin_unlock_irq(&qp->r_lock);
959 ev.device = qp->ibqp.device;
960 ev.element.qp = &qp->ibqp;
962 IB_EVENT_QP_LAST_WQE_REACHED;
963 qp->ibqp.event_handler(&ev,
964 qp->ibqp.qp_context);