2 * Copyright(c) 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/hash.h>
49 #include <linux/bitops.h>
50 #include <linux/lockdep.h>
51 #include <linux/vmalloc.h>
52 #include <linux/slab.h>
53 #include <rdma/ib_verbs.h>
59 * Note that it is OK to post send work requests in the SQE and ERR
60 * states; rvt_do_send() will process them and generate error
61 * completions as per IB 1.2 C10-96.
63 const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
65 [IB_QPS_INIT] = RVT_POST_RECV_OK,
66 [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK,
67 [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
68 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK |
69 RVT_PROCESS_NEXT_SEND_OK,
70 [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
71 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK,
72 [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
73 RVT_POST_SEND_OK | RVT_FLUSH_SEND,
74 [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV |
75 RVT_POST_SEND_OK | RVT_FLUSH_SEND,
77 EXPORT_SYMBOL(ib_rvt_state_ops);
79 static void get_map_page(struct rvt_qpn_table *qpt,
80 struct rvt_qpn_map *map,
83 unsigned long page = get_zeroed_page(gfp);
86 * Free the page if someone raced with us installing it.
89 spin_lock(&qpt->lock);
93 map->page = (void *)page;
94 spin_unlock(&qpt->lock);
98 * init_qpn_table - initialize the QP number table for a device
101 static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt)
104 struct rvt_qpn_map *map;
107 if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start))
110 spin_lock_init(&qpt->lock);
112 qpt->last = rdi->dparms.qpn_start;
113 qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift;
116 * Drivers may want some QPs beyond what we need for verbs let them use
117 * our qpn table. No need for two. Lets go ahead and mark the bitmaps
118 * for those. The reserved range must be *after* the range which verbs
122 /* Figure out number of bit maps needed before reserved range */
123 qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE;
125 /* This should always be zero */
126 offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK;
128 /* Starting with the first reserved bit map */
129 map = &qpt->map[qpt->nmaps];
131 rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n",
132 rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end);
133 for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) {
135 get_map_page(qpt, map, GFP_KERNEL);
141 set_bit(offset, map->page);
143 if (offset == RVT_BITS_PER_PAGE) {
154 * free_qpn_table - free the QP number table for a device
155 * @qpt: the QPN table
157 static void free_qpn_table(struct rvt_qpn_table *qpt)
161 for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
162 free_page((unsigned long)qpt->map[i].page);
166 * rvt_driver_qp_init - Init driver qp resources
167 * @rdi: rvt dev strucutre
169 * Return: 0 on success
171 int rvt_driver_qp_init(struct rvt_dev_info *rdi)
176 if (!rdi->dparms.qp_table_size)
180 * If driver is not doing any QP allocation then make sure it is
181 * providing the necessary QP functions.
183 if (!rdi->driver_f.free_all_qps ||
184 !rdi->driver_f.qp_priv_alloc ||
185 !rdi->driver_f.qp_priv_free ||
186 !rdi->driver_f.notify_qp_reset)
189 /* allocate parent object */
190 rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL,
195 /* allocate hash table */
196 rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size;
197 rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size);
198 rdi->qp_dev->qp_table =
199 kmalloc_node(rdi->qp_dev->qp_table_size *
200 sizeof(*rdi->qp_dev->qp_table),
201 GFP_KERNEL, rdi->dparms.node);
202 if (!rdi->qp_dev->qp_table)
205 for (i = 0; i < rdi->qp_dev->qp_table_size; i++)
206 RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL);
208 spin_lock_init(&rdi->qp_dev->qpt_lock);
210 /* initialize qpn map */
211 if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table))
214 spin_lock_init(&rdi->n_qps_lock);
219 kfree(rdi->qp_dev->qp_table);
220 free_qpn_table(&rdi->qp_dev->qpn_table);
229 * free_all_qps - check for QPs still in use
230 * @qpt: the QP table to empty
232 * There should not be any QPs still in use.
233 * Free memory for table.
235 static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi)
239 unsigned n, qp_inuse = 0;
240 spinlock_t *ql; /* work around too long line below */
242 if (rdi->driver_f.free_all_qps)
243 qp_inuse = rdi->driver_f.free_all_qps(rdi);
245 qp_inuse += rvt_mcast_tree_empty(rdi);
250 ql = &rdi->qp_dev->qpt_lock;
251 spin_lock_irqsave(ql, flags);
252 for (n = 0; n < rdi->qp_dev->qp_table_size; n++) {
253 qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n],
254 lockdep_is_held(ql));
255 RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL);
257 for (; qp; qp = rcu_dereference_protected(qp->next,
258 lockdep_is_held(ql)))
261 spin_unlock_irqrestore(ql, flags);
267 * rvt_qp_exit - clean up qps on device exit
268 * @rdi: rvt dev structure
270 * Check for qp leaks and free resources.
272 void rvt_qp_exit(struct rvt_dev_info *rdi)
274 u32 qps_inuse = rvt_free_all_qps(rdi);
277 rvt_pr_err(rdi, "QP memory leak! %u still in use\n",
282 kfree(rdi->qp_dev->qp_table);
283 free_qpn_table(&rdi->qp_dev->qpn_table);
287 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
288 struct rvt_qpn_map *map, unsigned off)
290 return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
294 * alloc_qpn - Allocate the next available qpn or zero/one for QP type
295 * IB_QPT_SMI/IB_QPT_GSI
296 *@rdi: rvt device info structure
297 *@qpt: queue pair number table pointer
298 *@port_num: IB port number, 1 based, comes from core
300 * Return: The queue pair number
302 static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
303 enum ib_qp_type type, u8 port_num, gfp_t gfp)
305 u32 i, offset, max_scan, qpn;
306 struct rvt_qpn_map *map;
309 if (rdi->driver_f.alloc_qpn)
310 return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num, gfp);
312 if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
315 ret = type == IB_QPT_GSI;
316 n = 1 << (ret + 2 * (port_num - 1));
317 spin_lock(&qpt->lock);
322 spin_unlock(&qpt->lock);
326 qpn = qpt->last + qpt->incr;
327 if (qpn >= RVT_QPN_MAX)
328 qpn = qpt->incr | ((qpt->last & 1) ^ 1);
329 /* offset carries bit 0 */
330 offset = qpn & RVT_BITS_PER_PAGE_MASK;
331 map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
332 max_scan = qpt->nmaps - !offset;
334 if (unlikely(!map->page)) {
335 get_map_page(qpt, map, gfp);
336 if (unlikely(!map->page))
340 if (!test_and_set_bit(offset, map->page)) {
347 * This qpn might be bogus if offset >= BITS_PER_PAGE.
348 * That is OK. It gets re-assigned below
350 qpn = mk_qpn(qpt, map, offset);
351 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
353 * In order to keep the number of pages allocated to a
354 * minimum, we scan the all existing pages before increasing
355 * the size of the bitmap table.
357 if (++i > max_scan) {
358 if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
360 map = &qpt->map[qpt->nmaps++];
361 /* start at incr with current bit 0 */
362 offset = qpt->incr | (offset & 1);
363 } else if (map < &qpt->map[qpt->nmaps]) {
365 /* start at incr with current bit 0 */
366 offset = qpt->incr | (offset & 1);
369 /* wrap to first map page, invert bit 0 */
370 offset = qpt->incr | ((offset & 1) ^ 1);
372 /* there can be no set bits in low-order QoS bits */
373 WARN_ON(rdi->dparms.qos_shift > 1 &&
374 offset & ((BIT(rdi->dparms.qos_shift - 1) - 1) << 1));
375 qpn = mk_qpn(qpt, map, offset);
384 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
386 struct rvt_qpn_map *map;
388 map = qpt->map + qpn / RVT_BITS_PER_PAGE;
390 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
394 * rvt_clear_mr_refs - Drop help mr refs
395 * @qp: rvt qp data structure
396 * @clr_sends: If shoudl clear send side or not
398 static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends)
401 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
403 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
404 rvt_put_ss(&qp->s_rdma_read_sge);
406 rvt_put_ss(&qp->r_sge);
409 while (qp->s_last != qp->s_head) {
410 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
413 for (i = 0; i < wqe->wr.num_sge; i++) {
414 struct rvt_sge *sge = &wqe->sg_list[i];
418 if (qp->ibqp.qp_type == IB_QPT_UD ||
419 qp->ibqp.qp_type == IB_QPT_SMI ||
420 qp->ibqp.qp_type == IB_QPT_GSI)
421 atomic_dec(&ibah_to_rvtah(
422 wqe->ud_wr.ah)->refcount);
423 if (++qp->s_last >= qp->s_size)
425 smp_wmb(); /* see qp_set_savail */
428 rvt_put_mr(qp->s_rdma_mr);
429 qp->s_rdma_mr = NULL;
433 if (qp->ibqp.qp_type != IB_QPT_RC)
436 for (n = 0; n < rvt_max_atomic(rdi); n++) {
437 struct rvt_ack_entry *e = &qp->s_ack_queue[n];
439 if (e->rdma_sge.mr) {
440 rvt_put_mr(e->rdma_sge.mr);
441 e->rdma_sge.mr = NULL;
447 * rvt_remove_qp - remove qp form table
448 * @rdi: rvt dev struct
451 * Remove the QP from the table so it can't be found asynchronously by
452 * the receive routine.
454 static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
456 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
457 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
461 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
463 if (rcu_dereference_protected(rvp->qp[0],
464 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
465 RCU_INIT_POINTER(rvp->qp[0], NULL);
466 } else if (rcu_dereference_protected(rvp->qp[1],
467 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
468 RCU_INIT_POINTER(rvp->qp[1], NULL);
471 struct rvt_qp __rcu **qpp;
474 qpp = &rdi->qp_dev->qp_table[n];
475 for (; (q = rcu_dereference_protected(*qpp,
476 lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL;
479 RCU_INIT_POINTER(*qpp,
480 rcu_dereference_protected(qp->next,
481 lockdep_is_held(&rdi->qp_dev->qpt_lock)));
483 trace_rvt_qpremove(qp, n);
489 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
497 * rvt_init_qp - initialize the QP state to the reset state
498 * @qp: the QP to init or reinit
501 * This function is called from both rvt_create_qp() and
502 * rvt_reset_qp(). The difference is that the reset
503 * patch the necessary locks to protect against concurent
506 static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
507 enum ib_qp_type type)
511 qp->qp_access_flags = 0;
512 qp->s_flags &= RVT_S_SIGNAL_REQ_WR;
518 qp->s_sending_psn = 0;
519 qp->s_sending_hpsn = 0;
523 if (type == IB_QPT_RC) {
524 qp->s_state = IB_OPCODE_RC_SEND_LAST;
525 qp->r_state = IB_OPCODE_RC_SEND_LAST;
527 qp->s_state = IB_OPCODE_UC_SEND_LAST;
528 qp->r_state = IB_OPCODE_UC_SEND_LAST;
530 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
541 qp->s_mig_state = IB_MIG_MIGRATED;
542 qp->r_head_ack_queue = 0;
543 qp->s_tail_ack_queue = 0;
544 qp->s_num_rd_atomic = 0;
546 qp->r_rq.wq->head = 0;
547 qp->r_rq.wq->tail = 0;
549 qp->r_sge.num_sge = 0;
550 atomic_set(&qp->s_reserved_used, 0);
554 * rvt_reset_qp - initialize the QP state to the reset state
555 * @qp: the QP to reset
558 * r_lock, s_hlock, and s_lock are required to be held by the caller
560 static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
561 enum ib_qp_type type)
562 __must_hold(&qp->s_lock)
563 __must_hold(&qp->s_hlock)
564 __must_hold(&qp->r_lock)
566 lockdep_assert_held(&qp->r_lock);
567 lockdep_assert_held(&qp->s_hlock);
568 lockdep_assert_held(&qp->s_lock);
569 if (qp->state != IB_QPS_RESET) {
570 qp->state = IB_QPS_RESET;
572 /* Let drivers flush their waitlist */
573 rdi->driver_f.flush_qp_waiters(qp);
574 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
575 spin_unlock(&qp->s_lock);
576 spin_unlock(&qp->s_hlock);
577 spin_unlock_irq(&qp->r_lock);
579 /* Stop the send queue and the retry timer */
580 rdi->driver_f.stop_send_queue(qp);
582 /* Wait for things to stop */
583 rdi->driver_f.quiesce_qp(qp);
585 /* take qp out the hash and wait for it to be unused */
586 rvt_remove_qp(rdi, qp);
587 wait_event(qp->wait, !atomic_read(&qp->refcount));
589 /* grab the lock b/c it was locked at call time */
590 spin_lock_irq(&qp->r_lock);
591 spin_lock(&qp->s_hlock);
592 spin_lock(&qp->s_lock);
594 rvt_clear_mr_refs(qp, 1);
596 * Let the driver do any tear down or re-init it needs to for
597 * a qp that has been reset
599 rdi->driver_f.notify_qp_reset(qp);
601 rvt_init_qp(rdi, qp, type);
602 lockdep_assert_held(&qp->r_lock);
603 lockdep_assert_held(&qp->s_hlock);
604 lockdep_assert_held(&qp->s_lock);
608 * rvt_create_qp - create a queue pair for a device
609 * @ibpd: the protection domain who's device we create the queue pair for
610 * @init_attr: the attributes of the queue pair
611 * @udata: user data for libibverbs.so
613 * Queue pair creation is mostly an rvt issue. However, drivers have their own
614 * unique idea of what queue pair numbers mean. For instance there is a reserved
617 * Return: the queue pair on success, otherwise returns an errno.
619 * Called by the ib_create_qp() core verbs function.
621 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
622 struct ib_qp_init_attr *init_attr,
623 struct ib_udata *udata)
627 struct rvt_swqe *swq = NULL;
630 struct ib_qp *ret = ERR_PTR(-ENOMEM);
631 struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
637 return ERR_PTR(-EINVAL);
639 if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge ||
640 init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr ||
641 init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
642 return ERR_PTR(-EINVAL);
644 /* GFP_NOIO is applicable to RC QP's only */
646 if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
647 init_attr->qp_type != IB_QPT_RC)
648 return ERR_PTR(-EINVAL);
650 gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
651 GFP_NOIO : GFP_KERNEL;
653 /* Check receive queue parameters if no SRQ is specified. */
654 if (!init_attr->srq) {
655 if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge ||
656 init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
657 return ERR_PTR(-EINVAL);
659 if (init_attr->cap.max_send_sge +
660 init_attr->cap.max_send_wr +
661 init_attr->cap.max_recv_sge +
662 init_attr->cap.max_recv_wr == 0)
663 return ERR_PTR(-EINVAL);
666 init_attr->cap.max_send_wr + 1 +
667 rdi->dparms.reserved_operations;
668 switch (init_attr->qp_type) {
671 if (init_attr->port_num == 0 ||
672 init_attr->port_num > ibpd->device->phys_port_cnt)
673 return ERR_PTR(-EINVAL);
677 sz = sizeof(struct rvt_sge) *
678 init_attr->cap.max_send_sge +
679 sizeof(struct rvt_swqe);
683 gfp | __GFP_ZERO, PAGE_KERNEL);
689 return ERR_PTR(-ENOMEM);
693 if (init_attr->srq) {
694 struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
696 if (srq->rq.max_sge > 1)
697 sg_list_sz = sizeof(*qp->r_sg_list) *
698 (srq->rq.max_sge - 1);
699 } else if (init_attr->cap.max_recv_sge > 1)
700 sg_list_sz = sizeof(*qp->r_sg_list) *
701 (init_attr->cap.max_recv_sge - 1);
702 qp = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node);
706 RCU_INIT_POINTER(qp->next, NULL);
707 if (init_attr->qp_type == IB_QPT_RC) {
710 sizeof(*qp->s_ack_queue) *
714 if (!qp->s_ack_queue)
719 * Driver needs to set up it's private QP structure and do any
720 * initialization that is needed.
722 priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp);
728 qp->timeout_jiffies =
729 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
731 if (init_attr->srq) {
734 qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
735 qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
736 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
737 sizeof(struct rvt_rwqe);
739 qp->r_rq.wq = vmalloc_user(
740 sizeof(struct rvt_rwq) +
742 else if (gfp == GFP_NOIO)
743 qp->r_rq.wq = __vmalloc(
744 sizeof(struct rvt_rwq) +
746 gfp | __GFP_ZERO, PAGE_KERNEL);
748 qp->r_rq.wq = vzalloc_node(
749 sizeof(struct rvt_rwq) +
753 goto bail_driver_priv;
757 * ib_create_qp() will initialize qp->ibqp
758 * except for qp->ibqp.qp_num.
760 spin_lock_init(&qp->r_lock);
761 spin_lock_init(&qp->s_hlock);
762 spin_lock_init(&qp->s_lock);
763 spin_lock_init(&qp->r_rq.lock);
764 atomic_set(&qp->refcount, 0);
765 atomic_set(&qp->local_ops_pending, 0);
766 init_waitqueue_head(&qp->wait);
767 init_timer(&qp->s_timer);
768 qp->s_timer.data = (unsigned long)qp;
769 INIT_LIST_HEAD(&qp->rspwait);
770 qp->state = IB_QPS_RESET;
773 qp->s_avail = init_attr->cap.max_send_wr;
774 qp->s_max_sge = init_attr->cap.max_send_sge;
775 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
776 qp->s_flags = RVT_S_SIGNAL_REQ_WR;
778 err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
780 init_attr->port_num, gfp);
785 qp->ibqp.qp_num = err;
786 qp->port_num = init_attr->port_num;
787 rvt_init_qp(rdi, qp, init_attr->qp_type);
791 /* Don't support raw QPs */
792 return ERR_PTR(-EINVAL);
795 init_attr->cap.max_inline_data = 0;
798 * Return the address of the RWQ as the offset to mmap.
799 * See rvt_mmap() for details.
801 if (udata && udata->outlen >= sizeof(__u64)) {
805 err = ib_copy_to_udata(udata, &offset,
812 u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
814 qp->ip = rvt_create_mmap_info(rdi, s,
815 ibpd->uobject->context,
818 ret = ERR_PTR(-ENOMEM);
822 err = ib_copy_to_udata(udata, &qp->ip->offset,
823 sizeof(qp->ip->offset));
829 qp->pid = current->pid;
832 spin_lock(&rdi->n_qps_lock);
833 if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
834 spin_unlock(&rdi->n_qps_lock);
835 ret = ERR_PTR(-ENOMEM);
839 rdi->n_qps_allocated++;
841 * Maintain a busy_jiffies variable that will be added to the timeout
842 * period in mod_retry_timer and add_retry_timer. This busy jiffies
843 * is scaled by the number of rc qps created for the device to reduce
844 * the number of timeouts occurring when there is a large number of
845 * qps. busy_jiffies is incremented every rc qp scaling interval.
846 * The scaling interval is selected based on extensive performance
847 * evaluation of targeted workloads.
849 if (init_attr->qp_type == IB_QPT_RC) {
851 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
853 spin_unlock(&rdi->n_qps_lock);
856 spin_lock_irq(&rdi->pending_lock);
857 list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
858 spin_unlock_irq(&rdi->pending_lock);
864 * We have our QP and its good, now keep track of what types of opcodes
865 * can be processed on this QP. We do this by keeping track of what the
866 * 3 high order bits of the opcode are.
868 switch (init_attr->qp_type) {
872 qp->allowed_ops = IB_OPCODE_UD;
875 qp->allowed_ops = IB_OPCODE_RC;
878 qp->allowed_ops = IB_OPCODE_UC;
881 ret = ERR_PTR(-EINVAL);
888 kref_put(&qp->ip->ref, rvt_release_mmap_info);
891 free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
898 rdi->driver_f.qp_priv_free(rdi, qp);
901 kfree(qp->s_ack_queue);
911 * rvt_error_qp - put a QP into the error state
912 * @qp: the QP to put into the error state
913 * @err: the receive completion error to signal if a RWQE is active
915 * Flushes both send and receive work queues.
917 * Return: true if last WQE event should be generated.
918 * The QP r_lock and s_lock should be held and interrupts disabled.
919 * If we are already in error state, just return.
921 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
925 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
927 lockdep_assert_held(&qp->r_lock);
928 lockdep_assert_held(&qp->s_lock);
929 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
932 qp->state = IB_QPS_ERR;
934 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
935 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
936 del_timer(&qp->s_timer);
939 if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
940 qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
942 rdi->driver_f.notify_error_qp(qp);
944 /* Schedule the sending tasklet to drain the send work queue. */
945 if (ACCESS_ONCE(qp->s_last) != qp->s_head)
946 rdi->driver_f.schedule_send(qp);
948 rvt_clear_mr_refs(qp, 0);
950 memset(&wc, 0, sizeof(wc));
952 wc.opcode = IB_WC_RECV;
954 if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
955 wc.wr_id = qp->r_wr_id;
957 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
959 wc.status = IB_WC_WR_FLUSH_ERR;
966 spin_lock(&qp->r_rq.lock);
968 /* sanity check pointers before trusting them */
971 if (head >= qp->r_rq.size)
974 if (tail >= qp->r_rq.size)
976 while (tail != head) {
977 wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
978 if (++tail >= qp->r_rq.size)
980 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
984 spin_unlock(&qp->r_rq.lock);
985 } else if (qp->ibqp.event_handler) {
992 EXPORT_SYMBOL(rvt_error_qp);
995 * Put the QP into the hash table.
996 * The hash table holds a reference to the QP.
998 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
1000 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
1001 unsigned long flags;
1004 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
1006 if (qp->ibqp.qp_num <= 1) {
1007 rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
1009 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
1011 qp->next = rdi->qp_dev->qp_table[n];
1012 rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
1013 trace_rvt_qpinsert(qp, n);
1016 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
1020 * rvt_modify_qp - modify the attributes of a queue pair
1021 * @ibqp: the queue pair who's attributes we're modifying
1022 * @attr: the new attributes
1023 * @attr_mask: the mask of attributes to modify
1024 * @udata: user data for libibverbs.so
1026 * Return: 0 on success, otherwise returns an errno.
1028 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1029 int attr_mask, struct ib_udata *udata)
1031 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1032 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1033 enum ib_qp_state cur_state, new_state;
1037 int pmtu = 0; /* for gcc warning only */
1038 enum rdma_link_layer link;
1040 link = rdma_port_get_link_layer(ibqp->device, qp->port_num);
1042 spin_lock_irq(&qp->r_lock);
1043 spin_lock(&qp->s_hlock);
1044 spin_lock(&qp->s_lock);
1046 cur_state = attr_mask & IB_QP_CUR_STATE ?
1047 attr->cur_qp_state : qp->state;
1048 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1050 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1054 if (rdi->driver_f.check_modify_qp &&
1055 rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata))
1058 if (attr_mask & IB_QP_AV) {
1059 if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1061 if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
1065 if (attr_mask & IB_QP_ALT_PATH) {
1066 if (attr->alt_ah_attr.dlid >=
1067 be16_to_cpu(IB_MULTICAST_LID_BASE))
1069 if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
1071 if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
1075 if (attr_mask & IB_QP_PKEY_INDEX)
1076 if (attr->pkey_index >= rvt_get_npkeys(rdi))
1079 if (attr_mask & IB_QP_MIN_RNR_TIMER)
1080 if (attr->min_rnr_timer > 31)
1083 if (attr_mask & IB_QP_PORT)
1084 if (qp->ibqp.qp_type == IB_QPT_SMI ||
1085 qp->ibqp.qp_type == IB_QPT_GSI ||
1086 attr->port_num == 0 ||
1087 attr->port_num > ibqp->device->phys_port_cnt)
1090 if (attr_mask & IB_QP_DEST_QPN)
1091 if (attr->dest_qp_num > RVT_QPN_MASK)
1094 if (attr_mask & IB_QP_RETRY_CNT)
1095 if (attr->retry_cnt > 7)
1098 if (attr_mask & IB_QP_RNR_RETRY)
1099 if (attr->rnr_retry > 7)
1103 * Don't allow invalid path_mtu values. OK to set greater
1104 * than the active mtu (or even the max_cap, if we have tuned
1105 * that to a small mtu. We'll set qp->path_mtu
1106 * to the lesser of requested attribute mtu and active,
1107 * for packetizing messages.
1108 * Note that the QP port has to be set in INIT and MTU in RTR.
1110 if (attr_mask & IB_QP_PATH_MTU) {
1111 pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
1116 if (attr_mask & IB_QP_PATH_MIG_STATE) {
1117 if (attr->path_mig_state == IB_MIG_REARM) {
1118 if (qp->s_mig_state == IB_MIG_ARMED)
1120 if (new_state != IB_QPS_RTS)
1122 } else if (attr->path_mig_state == IB_MIG_MIGRATED) {
1123 if (qp->s_mig_state == IB_MIG_REARM)
1125 if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
1127 if (qp->s_mig_state == IB_MIG_ARMED)
1134 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1135 if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
1138 switch (new_state) {
1140 if (qp->state != IB_QPS_RESET)
1141 rvt_reset_qp(rdi, qp, ibqp->qp_type);
1145 /* Allow event to re-trigger if QP set to RTR more than once */
1146 qp->r_flags &= ~RVT_R_COMM_EST;
1147 qp->state = new_state;
1151 qp->s_draining = qp->s_last != qp->s_cur;
1152 qp->state = new_state;
1156 if (qp->ibqp.qp_type == IB_QPT_RC)
1158 qp->state = new_state;
1162 lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1166 qp->state = new_state;
1170 if (attr_mask & IB_QP_PKEY_INDEX)
1171 qp->s_pkey_index = attr->pkey_index;
1173 if (attr_mask & IB_QP_PORT)
1174 qp->port_num = attr->port_num;
1176 if (attr_mask & IB_QP_DEST_QPN)
1177 qp->remote_qpn = attr->dest_qp_num;
1179 if (attr_mask & IB_QP_SQ_PSN) {
1180 qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
1181 qp->s_psn = qp->s_next_psn;
1182 qp->s_sending_psn = qp->s_next_psn;
1183 qp->s_last_psn = qp->s_next_psn - 1;
1184 qp->s_sending_hpsn = qp->s_last_psn;
1187 if (attr_mask & IB_QP_RQ_PSN)
1188 qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
1190 if (attr_mask & IB_QP_ACCESS_FLAGS)
1191 qp->qp_access_flags = attr->qp_access_flags;
1193 if (attr_mask & IB_QP_AV) {
1194 qp->remote_ah_attr = attr->ah_attr;
1195 qp->s_srate = attr->ah_attr.static_rate;
1196 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
1199 if (attr_mask & IB_QP_ALT_PATH) {
1200 qp->alt_ah_attr = attr->alt_ah_attr;
1201 qp->s_alt_pkey_index = attr->alt_pkey_index;
1204 if (attr_mask & IB_QP_PATH_MIG_STATE) {
1205 qp->s_mig_state = attr->path_mig_state;
1207 qp->remote_ah_attr = qp->alt_ah_attr;
1208 qp->port_num = qp->alt_ah_attr.port_num;
1209 qp->s_pkey_index = qp->s_alt_pkey_index;
1213 if (attr_mask & IB_QP_PATH_MTU) {
1214 qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
1215 qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
1216 qp->log_pmtu = ilog2(qp->pmtu);
1219 if (attr_mask & IB_QP_RETRY_CNT) {
1220 qp->s_retry_cnt = attr->retry_cnt;
1221 qp->s_retry = attr->retry_cnt;
1224 if (attr_mask & IB_QP_RNR_RETRY) {
1225 qp->s_rnr_retry_cnt = attr->rnr_retry;
1226 qp->s_rnr_retry = attr->rnr_retry;
1229 if (attr_mask & IB_QP_MIN_RNR_TIMER)
1230 qp->r_min_rnr_timer = attr->min_rnr_timer;
1232 if (attr_mask & IB_QP_TIMEOUT) {
1233 qp->timeout = attr->timeout;
1234 qp->timeout_jiffies =
1235 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1239 if (attr_mask & IB_QP_QKEY)
1240 qp->qkey = attr->qkey;
1242 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1243 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
1245 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
1246 qp->s_max_rd_atomic = attr->max_rd_atomic;
1248 if (rdi->driver_f.modify_qp)
1249 rdi->driver_f.modify_qp(qp, attr, attr_mask, udata);
1251 spin_unlock(&qp->s_lock);
1252 spin_unlock(&qp->s_hlock);
1253 spin_unlock_irq(&qp->r_lock);
1255 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1256 rvt_insert_qp(rdi, qp);
1259 ev.device = qp->ibqp.device;
1260 ev.element.qp = &qp->ibqp;
1261 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1262 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1265 ev.device = qp->ibqp.device;
1266 ev.element.qp = &qp->ibqp;
1267 ev.event = IB_EVENT_PATH_MIG;
1268 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1273 spin_unlock(&qp->s_lock);
1274 spin_unlock(&qp->s_hlock);
1275 spin_unlock_irq(&qp->r_lock);
1279 /** rvt_free_qpn - Free a qpn from the bit map
1281 * @qpn: queue pair number to free
1283 static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
1285 struct rvt_qpn_map *map;
1287 map = qpt->map + qpn / RVT_BITS_PER_PAGE;
1289 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
1293 * rvt_destroy_qp - destroy a queue pair
1294 * @ibqp: the queue pair to destroy
1296 * Note that this can be called while the QP is actively sending or
1299 * Return: 0 on success.
1301 int rvt_destroy_qp(struct ib_qp *ibqp)
1303 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1304 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1306 spin_lock_irq(&qp->r_lock);
1307 spin_lock(&qp->s_hlock);
1308 spin_lock(&qp->s_lock);
1309 rvt_reset_qp(rdi, qp, ibqp->qp_type);
1310 spin_unlock(&qp->s_lock);
1311 spin_unlock(&qp->s_hlock);
1312 spin_unlock_irq(&qp->r_lock);
1314 /* qpn is now available for use again */
1315 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1317 spin_lock(&rdi->n_qps_lock);
1318 rdi->n_qps_allocated--;
1319 if (qp->ibqp.qp_type == IB_QPT_RC) {
1321 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1323 spin_unlock(&rdi->n_qps_lock);
1326 kref_put(&qp->ip->ref, rvt_release_mmap_info);
1330 rdi->driver_f.qp_priv_free(rdi, qp);
1331 kfree(qp->s_ack_queue);
1337 * rvt_query_qp - query an ipbq
1338 * @ibqp: IB qp to query
1339 * @attr: attr struct to fill in
1340 * @attr_mask: attr mask ignored
1341 * @init_attr: struct to fill in
1345 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1346 int attr_mask, struct ib_qp_init_attr *init_attr)
1348 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1349 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1351 attr->qp_state = qp->state;
1352 attr->cur_qp_state = attr->qp_state;
1353 attr->path_mtu = qp->path_mtu;
1354 attr->path_mig_state = qp->s_mig_state;
1355 attr->qkey = qp->qkey;
1356 attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask;
1357 attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask;
1358 attr->dest_qp_num = qp->remote_qpn;
1359 attr->qp_access_flags = qp->qp_access_flags;
1360 attr->cap.max_send_wr = qp->s_size - 1 -
1361 rdi->dparms.reserved_operations;
1362 attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
1363 attr->cap.max_send_sge = qp->s_max_sge;
1364 attr->cap.max_recv_sge = qp->r_rq.max_sge;
1365 attr->cap.max_inline_data = 0;
1366 attr->ah_attr = qp->remote_ah_attr;
1367 attr->alt_ah_attr = qp->alt_ah_attr;
1368 attr->pkey_index = qp->s_pkey_index;
1369 attr->alt_pkey_index = qp->s_alt_pkey_index;
1370 attr->en_sqd_async_notify = 0;
1371 attr->sq_draining = qp->s_draining;
1372 attr->max_rd_atomic = qp->s_max_rd_atomic;
1373 attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
1374 attr->min_rnr_timer = qp->r_min_rnr_timer;
1375 attr->port_num = qp->port_num;
1376 attr->timeout = qp->timeout;
1377 attr->retry_cnt = qp->s_retry_cnt;
1378 attr->rnr_retry = qp->s_rnr_retry_cnt;
1379 attr->alt_port_num = qp->alt_ah_attr.port_num;
1380 attr->alt_timeout = qp->alt_timeout;
1382 init_attr->event_handler = qp->ibqp.event_handler;
1383 init_attr->qp_context = qp->ibqp.qp_context;
1384 init_attr->send_cq = qp->ibqp.send_cq;
1385 init_attr->recv_cq = qp->ibqp.recv_cq;
1386 init_attr->srq = qp->ibqp.srq;
1387 init_attr->cap = attr->cap;
1388 if (qp->s_flags & RVT_S_SIGNAL_REQ_WR)
1389 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
1391 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1392 init_attr->qp_type = qp->ibqp.qp_type;
1393 init_attr->port_num = qp->port_num;
1398 * rvt_post_receive - post a receive on a QP
1399 * @ibqp: the QP to post the receive on
1400 * @wr: the WR to post
1401 * @bad_wr: the first bad WR is put here
1403 * This may be called from interrupt context.
1405 * Return: 0 on success otherwise errno
1407 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1408 struct ib_recv_wr **bad_wr)
1410 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1411 struct rvt_rwq *wq = qp->r_rq.wq;
1412 unsigned long flags;
1413 int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) &&
1416 /* Check that state is OK to post receive. */
1417 if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
1422 for (; wr; wr = wr->next) {
1423 struct rvt_rwqe *wqe;
1427 if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
1432 spin_lock_irqsave(&qp->r_rq.lock, flags);
1433 next = wq->head + 1;
1434 if (next >= qp->r_rq.size)
1436 if (next == wq->tail) {
1437 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1441 if (unlikely(qp_err_flush)) {
1444 memset(&wc, 0, sizeof(wc));
1446 wc.opcode = IB_WC_RECV;
1447 wc.wr_id = wr->wr_id;
1448 wc.status = IB_WC_WR_FLUSH_ERR;
1449 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1451 wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
1452 wqe->wr_id = wr->wr_id;
1453 wqe->num_sge = wr->num_sge;
1454 for (i = 0; i < wr->num_sge; i++)
1455 wqe->sg_list[i] = wr->sg_list[i];
1457 * Make sure queue entry is written
1458 * before the head index.
1463 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1469 * rvt_qp_valid_operation - validate post send wr request
1471 * @post-parms - the post send table for the driver
1472 * @wr - the work request
1474 * The routine validates the operation based on the
1475 * validation table an returns the length of the operation
1476 * which can extend beyond the ib_send_bw. Operation
1477 * dependent flags key atomic operation validation.
1479 * There is an exception for UD qps that validates the pd and
1480 * overrides the length to include the additional UD specific
1483 * Returns a negative error or the length of the work request
1484 * for building the swqe.
1486 static inline int rvt_qp_valid_operation(
1488 const struct rvt_operation_params *post_parms,
1489 struct ib_send_wr *wr)
1493 if (wr->opcode >= RVT_OPERATION_MAX || !post_parms[wr->opcode].length)
1495 if (!(post_parms[wr->opcode].qpt_support & BIT(qp->ibqp.qp_type)))
1497 if ((post_parms[wr->opcode].flags & RVT_OPERATION_PRIV) &&
1498 ibpd_to_rvtpd(qp->ibqp.pd)->user)
1500 if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC_SGE &&
1501 (wr->num_sge == 0 ||
1502 wr->sg_list[0].length < sizeof(u64) ||
1503 wr->sg_list[0].addr & (sizeof(u64) - 1)))
1505 if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC &&
1506 !qp->s_max_rd_atomic)
1508 len = post_parms[wr->opcode].length;
1510 if (qp->ibqp.qp_type != IB_QPT_UC &&
1511 qp->ibqp.qp_type != IB_QPT_RC) {
1512 if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
1514 len = sizeof(struct ib_ud_wr);
1520 * rvt_qp_is_avail - determine queue capacity
1522 * @rdi - the rdmavt device
1523 * @reserved_op - is reserved operation
1525 * This assumes the s_hlock is held but the s_last
1526 * qp variable is uncontrolled.
1528 * For non reserved operations, the qp->s_avail
1531 * The return value is zero or a -ENOMEM.
1533 static inline int rvt_qp_is_avail(
1535 struct rvt_dev_info *rdi,
1542 /* see rvt_qp_wqe_unreserve() */
1543 smp_mb__before_atomic();
1544 reserved_used = atomic_read(&qp->s_reserved_used);
1545 if (unlikely(reserved_op)) {
1546 /* see rvt_qp_wqe_unreserve() */
1547 smp_mb__before_atomic();
1548 if (reserved_used >= rdi->dparms.reserved_operations)
1552 /* non-reserved operations */
1553 if (likely(qp->s_avail))
1555 smp_read_barrier_depends(); /* see rc.c */
1556 slast = ACCESS_ONCE(qp->s_last);
1557 if (qp->s_head >= slast)
1558 avail = qp->s_size - (qp->s_head - slast);
1560 avail = slast - qp->s_head;
1562 /* see rvt_qp_wqe_unreserve() */
1563 smp_mb__before_atomic();
1564 reserved_used = atomic_read(&qp->s_reserved_used);
1566 (rdi->dparms.reserved_operations - reserved_used);
1567 /* insure we don't assign a negative s_avail */
1568 if ((s32)avail <= 0)
1570 qp->s_avail = avail;
1571 if (WARN_ON(qp->s_avail >
1572 (qp->s_size - 1 - rdi->dparms.reserved_operations)))
1574 "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
1575 qp->ibqp.qp_num, qp->s_size, qp->s_avail,
1576 qp->s_head, qp->s_tail, qp->s_cur,
1577 qp->s_acked, qp->s_last);
1582 * rvt_post_one_wr - post one RC, UC, or UD send work request
1583 * @qp: the QP to post on
1584 * @wr: the work request to send
1586 static int rvt_post_one_wr(struct rvt_qp *qp,
1587 struct ib_send_wr *wr,
1590 struct rvt_swqe *wqe;
1595 struct rvt_lkey_table *rkt;
1597 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1602 int local_ops_delayed = 0;
1604 BUILD_BUG_ON(IB_QPT_MAX >= (sizeof(u32) * BITS_PER_BYTE));
1606 /* IB spec says that num_sge == 0 is OK. */
1607 if (unlikely(wr->num_sge > qp->s_max_sge))
1610 ret = rvt_qp_valid_operation(qp, rdi->post_parms, wr);
1616 * Local operations include fast register and local invalidate.
1617 * Fast register needs to be processed immediately because the
1618 * registered lkey may be used by following work requests and the
1619 * lkey needs to be valid at the time those requests are posted.
1620 * Local invalidate can be processed immediately if fencing is
1621 * not required and no previous local invalidate ops are pending.
1622 * Signaled local operations that have been processed immediately
1623 * need to have requests with "completion only" flags set posted
1624 * to the send queue in order to generate completions.
1626 if ((rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) {
1627 switch (wr->opcode) {
1629 ret = rvt_fast_reg_mr(qp,
1632 reg_wr(wr)->access);
1633 if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
1636 case IB_WR_LOCAL_INV:
1637 if ((wr->send_flags & IB_SEND_FENCE) ||
1638 atomic_read(&qp->local_ops_pending)) {
1639 local_ops_delayed = 1;
1641 ret = rvt_invalidate_rkey(
1642 qp, wr->ex.invalidate_rkey);
1643 if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
1652 reserved_op = rdi->post_parms[wr->opcode].flags &
1653 RVT_OPERATION_USE_RESERVE;
1654 /* check for avail */
1655 ret = rvt_qp_is_avail(qp, rdi, reserved_op);
1658 next = qp->s_head + 1;
1659 if (next >= qp->s_size)
1662 rkt = &rdi->lkey_table;
1663 pd = ibpd_to_rvtpd(qp->ibqp.pd);
1664 wqe = rvt_get_swqe_ptr(qp, qp->s_head);
1666 /* cplen has length from above */
1667 memcpy(&wqe->wr, wr, cplen);
1672 acc = wr->opcode >= IB_WR_RDMA_READ ?
1673 IB_ACCESS_LOCAL_WRITE : 0;
1674 for (i = 0; i < wr->num_sge; i++) {
1675 u32 length = wr->sg_list[i].length;
1680 ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j],
1681 &wr->sg_list[i], acc);
1684 goto bail_inval_free;
1686 wqe->length += length;
1689 wqe->wr.num_sge = j;
1692 /* general part of wqe valid - allow for driver checks */
1693 if (rdi->driver_f.check_send_wqe) {
1694 ret = rdi->driver_f.check_send_wqe(qp, wqe);
1696 goto bail_inval_free;
1701 log_pmtu = qp->log_pmtu;
1702 if (qp->ibqp.qp_type != IB_QPT_UC &&
1703 qp->ibqp.qp_type != IB_QPT_RC) {
1704 struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah);
1706 log_pmtu = ah->log_pmtu;
1707 atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
1710 if (rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL) {
1711 if (local_ops_delayed)
1712 atomic_inc(&qp->local_ops_pending);
1714 wqe->wr.send_flags |= RVT_SEND_COMPLETION_ONLY;
1719 wqe->ssn = qp->s_ssn++;
1720 wqe->psn = qp->s_next_psn;
1721 wqe->lpsn = wqe->psn +
1723 ((wqe->length - 1) >> log_pmtu) :
1725 qp->s_next_psn = wqe->lpsn + 1;
1727 trace_rvt_post_one_wr(qp, wqe);
1728 if (unlikely(reserved_op))
1729 rvt_qp_wqe_reserve(qp, wqe);
1732 smp_wmb(); /* see request builders */
1738 /* release mr holds */
1740 struct rvt_sge *sge = &wqe->sg_list[--j];
1742 rvt_put_mr(sge->mr);
1748 * rvt_post_send - post a send on a QP
1749 * @ibqp: the QP to post the send on
1750 * @wr: the list of work requests to post
1751 * @bad_wr: the first bad WR is put here
1753 * This may be called from interrupt context.
1755 * Return: 0 on success else errno
1757 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1758 struct ib_send_wr **bad_wr)
1760 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1761 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1762 unsigned long flags = 0;
1767 spin_lock_irqsave(&qp->s_hlock, flags);
1770 * Ensure QP state is such that we can send. If not bail out early,
1771 * there is no need to do this every time we post a send.
1773 if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
1774 spin_unlock_irqrestore(&qp->s_hlock, flags);
1779 * If the send queue is empty, and we only have a single WR then just go
1780 * ahead and kick the send engine into gear. Otherwise we will always
1781 * just schedule the send to happen later.
1783 call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next;
1785 for (; wr; wr = wr->next) {
1786 err = rvt_post_one_wr(qp, wr, &call_send);
1787 if (unlikely(err)) {
1794 spin_unlock_irqrestore(&qp->s_hlock, flags);
1797 rdi->driver_f.do_send(qp);
1799 rdi->driver_f.schedule_send_no_lock(qp);
1805 * rvt_post_srq_receive - post a receive on a shared receive queue
1806 * @ibsrq: the SRQ to post the receive on
1807 * @wr: the list of work requests to post
1808 * @bad_wr: A pointer to the first WR to cause a problem is put here
1810 * This may be called from interrupt context.
1812 * Return: 0 on success else errno
1814 int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1815 struct ib_recv_wr **bad_wr)
1817 struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
1819 unsigned long flags;
1821 for (; wr; wr = wr->next) {
1822 struct rvt_rwqe *wqe;
1826 if ((unsigned)wr->num_sge > srq->rq.max_sge) {
1831 spin_lock_irqsave(&srq->rq.lock, flags);
1833 next = wq->head + 1;
1834 if (next >= srq->rq.size)
1836 if (next == wq->tail) {
1837 spin_unlock_irqrestore(&srq->rq.lock, flags);
1842 wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head);
1843 wqe->wr_id = wr->wr_id;
1844 wqe->num_sge = wr->num_sge;
1845 for (i = 0; i < wr->num_sge; i++)
1846 wqe->sg_list[i] = wr->sg_list[i];
1847 /* Make sure queue entry is written before the head index. */
1850 spin_unlock_irqrestore(&srq->rq.lock, flags);