2 * Copyright(c) 2015 - 2018 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/seq_file.h>
53 #include <rdma/rdma_vt.h>
54 #include <rdma/rdmavt_qp.h>
55 #include <rdma/ib_verbs.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,
69 struct iowait_work *wait,
70 struct sdma_txreq *stx,
73 static void iowait_wakeup(struct iowait *wait, int reason);
74 static void iowait_sdma_drained(struct iowait *wait);
75 static void qp_pio_drain(struct rvt_qp *qp);
77 const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
78 [IB_WR_RDMA_WRITE] = {
79 .length = sizeof(struct ib_rdma_wr),
80 .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
84 .length = sizeof(struct ib_rdma_wr),
85 .qpt_support = BIT(IB_QPT_RC),
86 .flags = RVT_OPERATION_ATOMIC,
89 [IB_WR_ATOMIC_CMP_AND_SWP] = {
90 .length = sizeof(struct ib_atomic_wr),
91 .qpt_support = BIT(IB_QPT_RC),
92 .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
95 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
96 .length = sizeof(struct ib_atomic_wr),
97 .qpt_support = BIT(IB_QPT_RC),
98 .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
101 [IB_WR_RDMA_WRITE_WITH_IMM] = {
102 .length = sizeof(struct ib_rdma_wr),
103 .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
107 .length = sizeof(struct ib_send_wr),
108 .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
109 BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
112 [IB_WR_SEND_WITH_IMM] = {
113 .length = sizeof(struct ib_send_wr),
114 .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
115 BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
119 .length = sizeof(struct ib_reg_wr),
120 .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
121 .flags = RVT_OPERATION_LOCAL,
124 [IB_WR_LOCAL_INV] = {
125 .length = sizeof(struct ib_send_wr),
126 .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
127 .flags = RVT_OPERATION_LOCAL,
130 [IB_WR_SEND_WITH_INV] = {
131 .length = sizeof(struct ib_send_wr),
132 .qpt_support = BIT(IB_QPT_RC),
136 .length = sizeof(struct ib_atomic_wr),
137 .qpt_support = BIT(IB_QPT_RC),
138 .flags = RVT_OPERATION_USE_RESERVE,
141 [IB_WR_TID_RDMA_WRITE] = {
142 .length = sizeof(struct ib_rdma_wr),
143 .qpt_support = BIT(IB_QPT_RC),
144 .flags = RVT_OPERATION_IGN_RNR_CNT,
149 static void flush_list_head(struct list_head *l)
151 while (!list_empty(l)) {
152 struct sdma_txreq *tx;
154 tx = list_first_entry(
158 list_del_init(&tx->list);
160 container_of(tx, struct verbs_txreq, txreq));
164 static void flush_tx_list(struct rvt_qp *qp)
166 struct hfi1_qp_priv *priv = qp->priv;
168 flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
169 flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
172 static void flush_iowait(struct rvt_qp *qp)
174 struct hfi1_qp_priv *priv = qp->priv;
176 seqlock_t *lock = priv->s_iowait.lock;
180 write_seqlock_irqsave(lock, flags);
181 if (!list_empty(&priv->s_iowait.list)) {
182 list_del_init(&priv->s_iowait.list);
183 priv->s_iowait.lock = NULL;
186 write_sequnlock_irqrestore(lock, flags);
189 static inline int opa_mtu_enum_to_int(int mtu)
192 case OPA_MTU_8192: return 8192;
193 case OPA_MTU_10240: return 10240;
199 * This function is what we would push to the core layer if we wanted to be a
200 * "first class citizen". Instead we hide this here and rely on Verbs ULPs
201 * to blindly pass the MTU enum value from the PathRecord to us.
203 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
207 /* Constraining 10KB packets to 8KB packets */
208 if (mtu == (enum ib_mtu)OPA_MTU_10240)
210 val = opa_mtu_enum_to_int((int)mtu);
213 return ib_mtu_enum_to_int(mtu);
216 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
217 int attr_mask, struct ib_udata *udata)
219 struct ib_qp *ibqp = &qp->ibqp;
220 struct hfi1_ibdev *dev = to_idev(ibqp->device);
221 struct hfi1_devdata *dd = dd_from_dev(dev);
224 if (attr_mask & IB_QP_AV) {
225 sc = ah_to_sc(ibqp->device, &attr->ah_attr);
229 if (!qp_to_sdma_engine(qp, sc) &&
230 dd->flags & HFI1_HAS_SEND_DMA)
233 if (!qp_to_send_context(qp, sc))
237 if (attr_mask & IB_QP_ALT_PATH) {
238 sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
242 if (!qp_to_sdma_engine(qp, sc) &&
243 dd->flags & HFI1_HAS_SEND_DMA)
246 if (!qp_to_send_context(qp, sc))
254 * qp_set_16b - Set the hdr_type based on whether the slid or the
255 * dlid in the connection is extended. Only applicable for RC and UC
256 * QPs. UD QPs determine this on the fly from the ah in the wqe
258 static inline void qp_set_16b(struct rvt_qp *qp)
260 struct hfi1_pportdata *ppd;
261 struct hfi1_ibport *ibp;
262 struct hfi1_qp_priv *priv = qp->priv;
264 /* Update ah_attr to account for extended LIDs */
265 hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
267 /* Create 32 bit LIDs */
268 hfi1_make_opa_lid(&qp->remote_ah_attr);
270 if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
273 ibp = to_iport(qp->ibqp.device, qp->port_num);
274 ppd = ppd_from_ibp(ibp);
275 priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
278 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
279 int attr_mask, struct ib_udata *udata)
281 struct ib_qp *ibqp = &qp->ibqp;
282 struct hfi1_qp_priv *priv = qp->priv;
284 if (attr_mask & IB_QP_AV) {
285 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
286 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
287 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
291 if (attr_mask & IB_QP_PATH_MIG_STATE &&
292 attr->path_mig_state == IB_MIG_MIGRATED &&
293 qp->s_mig_state == IB_MIG_ARMED) {
294 qp->s_flags |= HFI1_S_AHG_CLEAR;
295 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
296 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
297 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
301 opfn_qp_init(qp, attr, attr_mask);
305 * hfi1_setup_wqe - set up the wqe
307 * @wqe - The built wqe
308 * @call_send - Determine if the send should be posted or scheduled.
310 * Perform setup of the wqe. This is called
311 * prior to inserting the wqe into the ring but after
312 * the wqe has been setup by RDMAVT. This function
313 * allows the driver the opportunity to perform
314 * validation and additional setup of the wqe.
316 * Returns 0 on success, -EINVAL on failure
319 int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
321 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
323 struct hfi1_pportdata *ppd;
324 struct hfi1_devdata *dd;
326 switch (qp->ibqp.qp_type) {
328 hfi1_setup_tid_rdma_wqe(qp, wqe);
331 if (wqe->length > 0x80000000U)
333 if (wqe->length > qp->pmtu)
338 * SM packets should exclusively use VL15 and their SL is
339 * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
340 * is created, SL is 0 in most cases and as a result some
341 * fields (vl and pmtu) in ah may not be set correctly,
342 * depending on the SL2SC and SC2VL tables at the time.
344 ppd = ppd_from_ibp(ibp);
345 dd = dd_from_ppd(ppd);
346 if (wqe->length > dd->vld[15].mtu)
351 ah = ibah_to_rvtah(wqe->ud_wr.ah);
352 if (wqe->length > (1 << ah->log_pmtu))
354 if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
361 * System latency between send and schedule is large enough that
362 * forcing call_send to true for piothreshold packets is necessary.
364 if (wqe->length <= piothreshold)
370 * _hfi1_schedule_send - schedule progress
373 * This schedules qp progress w/o regard to the s_flags.
375 * It is only used in the post send, which doesn't hold
378 bool _hfi1_schedule_send(struct rvt_qp *qp)
380 struct hfi1_qp_priv *priv = qp->priv;
381 struct hfi1_ibport *ibp =
382 to_iport(qp->ibqp.device, qp->port_num);
383 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
384 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
386 return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
389 cpumask_first(cpumask_of_node(dd->node)));
392 static void qp_pio_drain(struct rvt_qp *qp)
394 struct hfi1_qp_priv *priv = qp->priv;
396 if (!priv->s_sendcontext)
398 while (iowait_pio_pending(&priv->s_iowait)) {
399 write_seqlock_irq(&priv->s_sendcontext->waitlock);
400 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
401 write_sequnlock_irq(&priv->s_sendcontext->waitlock);
402 iowait_pio_drain(&priv->s_iowait);
403 write_seqlock_irq(&priv->s_sendcontext->waitlock);
404 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
405 write_sequnlock_irq(&priv->s_sendcontext->waitlock);
410 * hfi1_schedule_send - schedule progress
413 * This schedules qp progress and caller should hold
415 * @return true if the first leg is scheduled;
416 * false if the first leg is not scheduled.
418 bool hfi1_schedule_send(struct rvt_qp *qp)
420 lockdep_assert_held(&qp->s_lock);
421 if (hfi1_send_ok(qp)) {
422 _hfi1_schedule_send(qp);
425 if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
426 iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
431 static void hfi1_qp_schedule(struct rvt_qp *qp)
433 struct hfi1_qp_priv *priv = qp->priv;
436 if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
437 ret = hfi1_schedule_send(qp);
439 iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
441 if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
442 ret = hfi1_schedule_tid_send(qp);
444 iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
448 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
452 spin_lock_irqsave(&qp->s_lock, flags);
453 if (qp->s_flags & flag) {
454 qp->s_flags &= ~flag;
455 trace_hfi1_qpwakeup(qp, flag);
456 hfi1_qp_schedule(qp);
458 spin_unlock_irqrestore(&qp->s_lock, flags);
459 /* Notify hfi1_destroy_qp() if it is waiting. */
463 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
465 struct hfi1_qp_priv *priv = qp->priv;
467 if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
468 qp->s_flags &= ~RVT_S_BUSY;
470 * If we are sending a first-leg packet from the second leg,
471 * we need to clear the busy flag from priv->s_flags to
472 * avoid a race condition when the qp wakes up before
473 * the call to hfi1_verbs_send() returns to the second
474 * leg. In that case, the second leg will terminate without
475 * being re-scheduled, resulting in failure to send TID RDMA
476 * WRITE DATA and TID RDMA ACK packets.
478 if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
479 priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
481 iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
484 priv->s_flags &= ~RVT_S_BUSY;
488 static int iowait_sleep(
489 struct sdma_engine *sde,
490 struct iowait_work *wait,
491 struct sdma_txreq *stx,
495 struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
497 struct hfi1_qp_priv *priv;
504 spin_lock_irqsave(&qp->s_lock, flags);
505 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
507 * If we couldn't queue the DMA request, save the info
508 * and try again later rather than destroying the
509 * buffer and undoing the side effects of the copy.
511 /* Make a common routine? */
512 list_add_tail(&stx->list, &wait->tx_head);
513 write_seqlock(&sde->waitlock);
514 if (sdma_progress(sde, seq, stx))
516 if (list_empty(&priv->s_iowait.list)) {
517 struct hfi1_ibport *ibp =
518 to_iport(qp->ibqp.device, qp->port_num);
520 ibp->rvp.n_dmawait++;
521 qp->s_flags |= RVT_S_WAIT_DMA_DESC;
522 iowait_get_priority(&priv->s_iowait);
523 iowait_queue(pkts_sent, &priv->s_iowait,
525 priv->s_iowait.lock = &sde->waitlock;
526 trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
529 write_sequnlock(&sde->waitlock);
530 hfi1_qp_unbusy(qp, wait);
531 spin_unlock_irqrestore(&qp->s_lock, flags);
534 spin_unlock_irqrestore(&qp->s_lock, flags);
539 write_sequnlock(&sde->waitlock);
540 spin_unlock_irqrestore(&qp->s_lock, flags);
541 list_del_init(&stx->list);
545 static void iowait_wakeup(struct iowait *wait, int reason)
547 struct rvt_qp *qp = iowait_to_qp(wait);
549 WARN_ON(reason != SDMA_AVAIL_REASON);
550 hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
553 static void iowait_sdma_drained(struct iowait *wait)
555 struct rvt_qp *qp = iowait_to_qp(wait);
559 * This happens when the send engine notes
560 * a QP in the error state and cannot
561 * do the flush work until that QP's
562 * sdma work has finished.
564 spin_lock_irqsave(&qp->s_lock, flags);
565 if (qp->s_flags & RVT_S_WAIT_DMA) {
566 qp->s_flags &= ~RVT_S_WAIT_DMA;
567 hfi1_schedule_send(qp);
569 spin_unlock_irqrestore(&qp->s_lock, flags);
572 static void hfi1_init_priority(struct iowait *w)
574 struct rvt_qp *qp = iowait_to_qp(w);
575 struct hfi1_qp_priv *priv = qp->priv;
577 if (qp->s_flags & RVT_S_ACK_PENDING)
579 if (priv->s_flags & RVT_S_ACK_PENDING)
584 * qp_to_sdma_engine - map a qp to a send engine
589 * A send engine for the qp or NULL for SMI type qp.
591 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
593 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
594 struct sdma_engine *sde;
596 if (!(dd->flags & HFI1_HAS_SEND_DMA))
598 switch (qp->ibqp.qp_type) {
604 sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
609 * qp_to_send_context - map a qp to a send context
614 * A send context for the qp
616 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
618 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
620 switch (qp->ibqp.qp_type) {
622 /* SMA packets to VL15 */
623 return dd->vld[15].sc;
628 return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
632 static const char * const qp_type_str[] = {
633 "SMI", "GSI", "RC", "UC", "UD",
636 static int qp_idle(struct rvt_qp *qp)
639 qp->s_last == qp->s_acked &&
640 qp->s_acked == qp->s_cur &&
641 qp->s_cur == qp->s_tail &&
642 qp->s_tail == qp->s_head;
646 * qp_iter_print - print the qp information to seq_file
647 * @s: the seq_file to emit the qp information on
648 * @iter: the iterator for the qp hash list
650 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
652 struct rvt_swqe *wqe;
653 struct rvt_qp *qp = iter->qp;
654 struct hfi1_qp_priv *priv = qp->priv;
655 struct sdma_engine *sde;
656 struct send_context *send_context;
657 struct rvt_ack_entry *e = NULL;
658 struct rvt_srq *srq = qp->ibqp.srq ?
659 ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
661 sde = qp_to_sdma_engine(qp, priv->s_sc);
662 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
663 send_context = qp_to_send_context(qp, priv->s_sc);
665 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
667 "N %d %s QP %x R %u %s %u %u f=%x %u %u %u %u %u %u SPSN %x %x %x %x %x RPSN %x S(%u %u %u %u %u %u %u) R(%u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d OS %x %x E %x %x %x RNR %d %s %d\n",
669 qp_idle(qp) ? "I" : "B",
671 atomic_read(&qp->refcount),
672 qp_type_str[qp->ibqp.qp_type],
674 wqe ? wqe->wr.opcode : 0,
676 iowait_sdma_pending(&priv->s_iowait),
677 iowait_pio_pending(&priv->s_iowait),
678 !list_empty(&priv->s_iowait.list),
683 qp->s_psn, qp->s_next_psn,
684 qp->s_sending_psn, qp->s_sending_hpsn,
686 qp->s_last, qp->s_acked, qp->s_cur,
687 qp->s_tail, qp->s_head, qp->s_size,
689 /* ack_queue ring pointers, size */
690 qp->s_tail_ack_queue, qp->r_head_ack_queue,
691 rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
694 rdma_ah_get_dlid(&qp->remote_ah_attr),
695 rdma_ah_get_sl(&qp->remote_ah_attr),
702 sde ? sde->this_idx : 0,
704 send_context ? send_context->sw_index : 0,
705 ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
706 ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
710 /* ack queue information */
716 srq ? srq->rq.size : qp->r_rq.size
720 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
722 struct hfi1_qp_priv *priv;
724 priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
726 return ERR_PTR(-ENOMEM);
730 priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
734 return ERR_PTR(-ENOMEM);
748 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
750 struct hfi1_qp_priv *priv = qp->priv;
752 hfi1_qp_priv_tid_free(rdi, qp);
757 unsigned free_all_qps(struct rvt_dev_info *rdi)
759 struct hfi1_ibdev *verbs_dev = container_of(rdi,
762 struct hfi1_devdata *dd = container_of(verbs_dev,
766 unsigned qp_inuse = 0;
768 for (n = 0; n < dd->num_pports; n++) {
769 struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
772 if (rcu_dereference(ibp->rvp.qp[0]))
774 if (rcu_dereference(ibp->rvp.qp[1]))
782 void flush_qp_waiters(struct rvt_qp *qp)
784 lockdep_assert_held(&qp->s_lock);
786 hfi1_tid_rdma_flush_wait(qp);
789 void stop_send_queue(struct rvt_qp *qp)
791 struct hfi1_qp_priv *priv = qp->priv;
793 iowait_cancel_work(&priv->s_iowait);
794 if (cancel_work_sync(&priv->tid_rdma.trigger_work))
798 void quiesce_qp(struct rvt_qp *qp)
800 struct hfi1_qp_priv *priv = qp->priv;
802 hfi1_del_tid_reap_timer(qp);
803 hfi1_del_tid_retry_timer(qp);
804 iowait_sdma_drain(&priv->s_iowait);
809 void notify_qp_reset(struct rvt_qp *qp)
811 hfi1_qp_kern_exp_rcv_clear_all(qp);
815 /* Clear any OPFN state */
816 if (qp->ibqp.qp_type == IB_QPT_RC)
821 * Switch to alternate path.
822 * The QP s_lock should be held and interrupts disabled.
824 void hfi1_migrate_qp(struct rvt_qp *qp)
826 struct hfi1_qp_priv *priv = qp->priv;
829 qp->s_mig_state = IB_MIG_MIGRATED;
830 qp->remote_ah_attr = qp->alt_ah_attr;
831 qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
832 qp->s_pkey_index = qp->s_alt_pkey_index;
833 qp->s_flags |= HFI1_S_AHG_CLEAR;
834 priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
835 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
838 ev.device = qp->ibqp.device;
839 ev.element.qp = &qp->ibqp;
840 ev.event = IB_EVENT_PATH_MIG;
841 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
844 int mtu_to_path_mtu(u32 mtu)
846 return mtu_to_enum(mtu, OPA_MTU_8192);
849 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
852 struct hfi1_ibdev *verbs_dev = container_of(rdi,
855 struct hfi1_devdata *dd = container_of(verbs_dev,
858 struct hfi1_ibport *ibp;
861 ibp = &dd->pport[qp->port_num - 1].ibport_data;
862 sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
863 vl = sc_to_vlt(dd, sc);
865 mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
866 if (vl < PER_VL_SEND_CONTEXTS)
867 mtu = min_t(u32, mtu, dd->vld[vl].mtu);
871 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
872 struct ib_qp_attr *attr)
874 int mtu, pidx = qp->port_num - 1;
875 struct hfi1_ibdev *verbs_dev = container_of(rdi,
878 struct hfi1_devdata *dd = container_of(verbs_dev,
881 mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
883 return -1; /* values less than 0 are error */
885 if (mtu > dd->pport[pidx].ibmtu)
886 return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
888 return attr->path_mtu;
891 void notify_error_qp(struct rvt_qp *qp)
893 struct hfi1_qp_priv *priv = qp->priv;
894 seqlock_t *lock = priv->s_iowait.lock;
898 if (!list_empty(&priv->s_iowait.list) &&
899 !(qp->s_flags & RVT_S_BUSY) &&
900 !(priv->s_flags & RVT_S_BUSY)) {
901 qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
902 iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
903 iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
904 list_del_init(&priv->s_iowait.list);
905 priv->s_iowait.lock = NULL;
908 write_sequnlock(lock);
911 if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
914 rvt_put_mr(qp->s_rdma_mr);
915 qp->s_rdma_mr = NULL;
922 * hfi1_qp_iter_cb - callback for iterator
924 * @v - the sl in low bits of v
926 * This is called from the iterator callback to work
927 * on an individual qp.
929 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
933 struct hfi1_ibport *ibp =
934 to_iport(qp->ibqp.device, qp->port_num);
935 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
938 if (qp->port_num != ppd->port ||
939 (qp->ibqp.qp_type != IB_QPT_UC &&
940 qp->ibqp.qp_type != IB_QPT_RC) ||
941 rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
942 !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
945 spin_lock_irq(&qp->r_lock);
946 spin_lock(&qp->s_hlock);
947 spin_lock(&qp->s_lock);
948 lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
949 spin_unlock(&qp->s_lock);
950 spin_unlock(&qp->s_hlock);
951 spin_unlock_irq(&qp->r_lock);
953 ev.device = qp->ibqp.device;
954 ev.element.qp = &qp->ibqp;
955 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
956 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
961 * hfi1_error_port_qps - put a port's RC/UC qps into error state
963 * @sl: the service level.
965 * This function places all RC/UC qps with a given service level into error
966 * state. It is generally called to force upper lay apps to abandon stale qps
967 * after an sl->sc mapping change.
969 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
971 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
972 struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
974 rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);