2 * Copyright (c) 2012-2016 VMware, Inc. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of EITHER the GNU General Public License
6 * version 2 as published by the Free Software Foundation or the BSD
7 * 2-Clause License. This program is distributed in the hope that it
8 * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
9 * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
10 * See the GNU General Public License version 2 for more details at
11 * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
13 * You should have received a copy of the GNU General Public License
14 * along with this program available in the file COPYING in the main
15 * directory of this source tree.
17 * The BSD 2-Clause License
19 * Redistribution and use in source and binary forms, with or
20 * without modification, are permitted provided that the following
23 * - Redistributions of source code must retain the above
24 * copyright notice, this list of conditions and the following
27 * - Redistributions in binary form must reproduce the above
28 * copyright notice, this list of conditions and the following
29 * disclaimer in the documentation and/or other materials
30 * provided with the distribution.
32 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
33 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
34 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
35 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
36 * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
37 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
38 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
39 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
41 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
42 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
43 * OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/wait.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_smi.h>
51 #include <rdma/ib_user_verbs.h>
55 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
56 struct pvrdma_cq **recv_cq)
58 *send_cq = to_vcq(qp->ibqp.send_cq);
59 *recv_cq = to_vcq(qp->ibqp.recv_cq);
62 static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
63 unsigned long *scq_flags,
64 unsigned long *rcq_flags)
65 __acquires(scq->cq_lock) __acquires(rcq->cq_lock)
68 spin_lock_irqsave(&scq->cq_lock, *scq_flags);
69 __acquire(rcq->cq_lock);
70 } else if (scq->cq_handle < rcq->cq_handle) {
71 spin_lock_irqsave(&scq->cq_lock, *scq_flags);
72 spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags,
73 SINGLE_DEPTH_NESTING);
75 spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
76 spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
77 SINGLE_DEPTH_NESTING);
81 static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
82 unsigned long *scq_flags,
83 unsigned long *rcq_flags)
84 __releases(scq->cq_lock) __releases(rcq->cq_lock)
87 __release(rcq->cq_lock);
88 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
89 } else if (scq->cq_handle < rcq->cq_handle) {
90 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
91 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
93 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
94 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
98 static void pvrdma_reset_qp(struct pvrdma_qp *qp)
100 struct pvrdma_cq *scq, *rcq;
101 unsigned long scq_flags, rcq_flags;
104 get_cqs(qp, &scq, &rcq);
105 pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
107 _pvrdma_flush_cqe(qp, scq);
109 _pvrdma_flush_cqe(qp, rcq);
111 pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
114 * Reset queuepair. The checks are because usermode queuepairs won't
115 * have kernel ringstates.
118 atomic_set(&qp->rq.ring->cons_head, 0);
119 atomic_set(&qp->rq.ring->prod_tail, 0);
122 atomic_set(&qp->sq.ring->cons_head, 0);
123 atomic_set(&qp->sq.ring->prod_tail, 0);
127 static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
128 struct ib_qp_cap *req_cap,
129 struct pvrdma_qp *qp)
131 if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr ||
132 req_cap->max_recv_sge > dev->dsr->caps.max_sge) {
133 dev_warn(&dev->pdev->dev, "recv queue size invalid\n");
137 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr));
138 qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge));
141 req_cap->max_recv_wr = qp->rq.wqe_cnt;
142 req_cap->max_recv_sge = qp->rq.max_sg;
144 qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
145 sizeof(struct pvrdma_sge) *
147 qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
153 static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
154 struct pvrdma_qp *qp)
156 if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr ||
157 req_cap->max_send_sge > dev->dsr->caps.max_sge) {
158 dev_warn(&dev->pdev->dev, "send queue size invalid\n");
162 qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr));
163 qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge));
166 req_cap->max_send_wr = qp->sq.wqe_cnt;
167 req_cap->max_send_sge = qp->sq.max_sg;
169 qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
170 sizeof(struct pvrdma_sge) *
172 /* Note: one extra page for the header. */
173 qp->npages_send = 1 + (qp->sq.wqe_cnt * qp->sq.wqe_size +
174 PAGE_SIZE - 1) / PAGE_SIZE;
180 * pvrdma_create_qp - create queue pair
181 * @pd: protection domain
182 * @init_attr: queue pair attributes
185 * @return: the ib_qp pointer on success, otherwise returns an errno.
187 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
188 struct ib_qp_init_attr *init_attr,
189 struct ib_udata *udata)
191 struct pvrdma_qp *qp = NULL;
192 struct pvrdma_dev *dev = to_vdev(pd->device);
193 union pvrdma_cmd_req req;
194 union pvrdma_cmd_resp rsp;
195 struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
196 struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
197 struct pvrdma_create_qp ucmd;
201 if (init_attr->create_flags) {
202 dev_warn(&dev->pdev->dev,
203 "invalid create queuepair flags %#x\n",
204 init_attr->create_flags);
205 return ERR_PTR(-EINVAL);
208 if (init_attr->qp_type != IB_QPT_RC &&
209 init_attr->qp_type != IB_QPT_UD &&
210 init_attr->qp_type != IB_QPT_GSI) {
211 dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
213 return ERR_PTR(-EINVAL);
216 if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
217 return ERR_PTR(-ENOMEM);
219 switch (init_attr->qp_type) {
221 if (init_attr->port_num == 0 ||
222 init_attr->port_num > pd->device->phys_port_cnt ||
224 dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
231 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
237 spin_lock_init(&qp->sq.lock);
238 spin_lock_init(&qp->rq.lock);
239 mutex_init(&qp->mutex);
240 atomic_set(&qp->refcnt, 1);
241 init_waitqueue_head(&qp->wait);
243 qp->state = IB_QPS_RESET;
245 if (pd->uobject && udata) {
246 dev_dbg(&dev->pdev->dev,
247 "create queuepair from user space\n");
249 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
254 /* set qp->sq.wqe_cnt, shift, buf_size.. */
255 qp->rumem = ib_umem_get(pd->uobject->context,
257 ucmd.rbuf_size, 0, 0);
258 if (IS_ERR(qp->rumem)) {
259 ret = PTR_ERR(qp->rumem);
263 qp->sumem = ib_umem_get(pd->uobject->context,
265 ucmd.sbuf_size, 0, 0);
266 if (IS_ERR(qp->sumem)) {
267 ib_umem_release(qp->rumem);
268 ret = PTR_ERR(qp->sumem);
272 qp->npages_send = ib_umem_page_count(qp->sumem);
273 qp->npages_recv = ib_umem_page_count(qp->rumem);
274 qp->npages = qp->npages_send + qp->npages_recv;
276 qp->is_kernel = true;
278 ret = pvrdma_set_sq_size(to_vdev(pd->device),
279 &init_attr->cap, qp);
283 ret = pvrdma_set_rq_size(to_vdev(pd->device),
284 &init_attr->cap, qp);
288 qp->npages = qp->npages_send + qp->npages_recv;
290 /* Skip header page. */
291 qp->sq.offset = PAGE_SIZE;
293 /* Recv queue pages are after send pages. */
294 qp->rq.offset = qp->npages_send * PAGE_SIZE;
297 if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
298 dev_warn(&dev->pdev->dev,
299 "overflow pages in queuepair\n");
304 ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
307 dev_warn(&dev->pdev->dev,
308 "could not allocate page directory\n");
312 if (!qp->is_kernel) {
313 pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
314 pvrdma_page_dir_insert_umem(&qp->pdir, qp->rumem,
317 /* Ring state is always the first page. */
318 qp->sq.ring = qp->pdir.pages[0];
319 qp->rq.ring = &qp->sq.ring[1];
328 init_attr->cap.max_inline_data = 0;
330 memset(cmd, 0, sizeof(*cmd));
331 cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
332 cmd->pd_handle = to_vpd(pd)->pd_handle;
333 cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
334 cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
335 cmd->max_send_wr = init_attr->cap.max_send_wr;
336 cmd->max_recv_wr = init_attr->cap.max_recv_wr;
337 cmd->max_send_sge = init_attr->cap.max_send_sge;
338 cmd->max_recv_sge = init_attr->cap.max_recv_sge;
339 cmd->max_inline_data = init_attr->cap.max_inline_data;
340 cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
341 cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
342 cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
343 cmd->total_chunks = qp->npages;
344 cmd->send_chunks = qp->npages_send - 1;
345 cmd->pdir_dma = qp->pdir.dir_dma;
347 dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
348 cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
351 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
353 dev_warn(&dev->pdev->dev,
354 "could not create queuepair, error: %d\n", ret);
358 /* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
359 qp->qp_handle = resp->qpn;
360 qp->port = init_attr->port_num;
361 qp->ibqp.qp_num = resp->qpn;
362 spin_lock_irqsave(&dev->qp_tbl_lock, flags);
363 dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
364 spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
369 pvrdma_page_dir_cleanup(dev, &qp->pdir);
371 if (pd->uobject && udata) {
373 ib_umem_release(qp->rumem);
375 ib_umem_release(qp->sumem);
379 atomic_dec(&dev->num_qps);
384 static void pvrdma_free_qp(struct pvrdma_qp *qp)
386 struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
387 struct pvrdma_cq *scq;
388 struct pvrdma_cq *rcq;
389 unsigned long flags, scq_flags, rcq_flags;
391 /* In case cq is polling */
392 get_cqs(qp, &scq, &rcq);
393 pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
395 _pvrdma_flush_cqe(qp, scq);
397 _pvrdma_flush_cqe(qp, rcq);
399 spin_lock_irqsave(&dev->qp_tbl_lock, flags);
400 dev->qp_tbl[qp->qp_handle] = NULL;
401 spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
403 pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
405 atomic_dec(&qp->refcnt);
406 wait_event(qp->wait, !atomic_read(&qp->refcnt));
408 pvrdma_page_dir_cleanup(dev, &qp->pdir);
412 atomic_dec(&dev->num_qps);
416 * pvrdma_destroy_qp - destroy a queue pair
417 * @qp: the queue pair to destroy
419 * @return: 0 on success.
421 int pvrdma_destroy_qp(struct ib_qp *qp)
423 struct pvrdma_qp *vqp = to_vqp(qp);
424 union pvrdma_cmd_req req;
425 struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
428 memset(cmd, 0, sizeof(*cmd));
429 cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
430 cmd->qp_handle = vqp->qp_handle;
432 ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
434 dev_warn(&to_vdev(qp->device)->pdev->dev,
435 "destroy queuepair failed, error: %d\n", ret);
443 * pvrdma_modify_qp - modify queue pair attributes
444 * @ibqp: the queue pair
445 * @attr: the new queue pair's attributes
446 * @attr_mask: attributes mask
449 * @returns 0 on success, otherwise returns an errno.
451 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
452 int attr_mask, struct ib_udata *udata)
454 struct pvrdma_dev *dev = to_vdev(ibqp->device);
455 struct pvrdma_qp *qp = to_vqp(ibqp);
456 union pvrdma_cmd_req req;
457 union pvrdma_cmd_resp rsp;
458 struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
459 int cur_state, next_state;
462 /* Sanity checking. Should need lock here */
463 mutex_lock(&qp->mutex);
464 cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
466 next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
468 if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
469 attr_mask, IB_LINK_LAYER_ETHERNET)) {
474 if (attr_mask & IB_QP_PORT) {
475 if (attr->port_num == 0 ||
476 attr->port_num > ibqp->device->phys_port_cnt) {
482 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
483 if (attr->min_rnr_timer > 31) {
489 if (attr_mask & IB_QP_PKEY_INDEX) {
490 if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
496 if (attr_mask & IB_QP_QKEY)
497 qp->qkey = attr->qkey;
499 if (cur_state == next_state && cur_state == IB_QPS_RESET) {
504 qp->state = next_state;
505 memset(cmd, 0, sizeof(*cmd));
506 cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
507 cmd->qp_handle = qp->qp_handle;
508 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
509 cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
510 cmd->attrs.cur_qp_state =
511 ib_qp_state_to_pvrdma(attr->cur_qp_state);
512 cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
513 cmd->attrs.path_mig_state =
514 ib_mig_state_to_pvrdma(attr->path_mig_state);
515 cmd->attrs.qkey = attr->qkey;
516 cmd->attrs.rq_psn = attr->rq_psn;
517 cmd->attrs.sq_psn = attr->sq_psn;
518 cmd->attrs.dest_qp_num = attr->dest_qp_num;
519 cmd->attrs.qp_access_flags =
520 ib_access_flags_to_pvrdma(attr->qp_access_flags);
521 cmd->attrs.pkey_index = attr->pkey_index;
522 cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
523 cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
524 cmd->attrs.sq_draining = attr->sq_draining;
525 cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
526 cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
527 cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
528 cmd->attrs.port_num = attr->port_num;
529 cmd->attrs.timeout = attr->timeout;
530 cmd->attrs.retry_cnt = attr->retry_cnt;
531 cmd->attrs.rnr_retry = attr->rnr_retry;
532 cmd->attrs.alt_port_num = attr->alt_port_num;
533 cmd->attrs.alt_timeout = attr->alt_timeout;
534 ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
535 ib_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
536 ib_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
538 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
540 dev_warn(&dev->pdev->dev,
541 "could not modify queuepair, error: %d\n", ret);
542 } else if (rsp.hdr.err > 0) {
543 dev_warn(&dev->pdev->dev,
544 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
548 if (ret == 0 && next_state == IB_QPS_RESET)
552 mutex_unlock(&qp->mutex);
557 static inline void *get_sq_wqe(struct pvrdma_qp *qp, int n)
559 return pvrdma_page_dir_get_ptr(&qp->pdir,
560 qp->sq.offset + n * qp->sq.wqe_size);
563 static inline void *get_rq_wqe(struct pvrdma_qp *qp, int n)
565 return pvrdma_page_dir_get_ptr(&qp->pdir,
566 qp->rq.offset + n * qp->rq.wqe_size);
569 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_wr *wr)
571 struct pvrdma_user_mr *mr = to_vmr(wr->mr);
573 wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
574 wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
575 wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
576 wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
577 wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
578 wqe_hdr->wr.fast_reg.access_flags = wr->access;
579 wqe_hdr->wr.fast_reg.rkey = wr->key;
581 return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
586 * pvrdma_post_send - post send work request entries on a QP
588 * @wr: work request list to post
589 * @bad_wr: the first bad WR returned
591 * @return: 0 on success, otherwise errno returned.
593 int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
594 struct ib_send_wr **bad_wr)
596 struct pvrdma_qp *qp = to_vqp(ibqp);
597 struct pvrdma_dev *dev = to_vdev(ibqp->device);
599 struct pvrdma_sq_wqe_hdr *wqe_hdr;
600 struct pvrdma_sge *sge;
606 * In states lower than RTS, we can fail immediately. In other states,
607 * just post and let the device figure it out.
609 if (qp->state < IB_QPS_RTS) {
614 spin_lock_irqsave(&qp->sq.lock, flags);
616 index = pvrdma_idx(&qp->sq.ring->prod_tail, qp->sq.wqe_cnt);
617 for (nreq = 0; wr; nreq++, wr = wr->next) {
620 if (unlikely(!pvrdma_idx_ring_has_space(
621 qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
622 dev_warn_ratelimited(&dev->pdev->dev,
623 "send queue is full\n");
629 if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
630 dev_warn_ratelimited(&dev->pdev->dev,
631 "send SGE overflow\n");
637 if (unlikely(wr->opcode < 0)) {
638 dev_warn_ratelimited(&dev->pdev->dev,
639 "invalid send opcode\n");
646 * Only support UD, RC.
647 * Need to check opcode table for thorough checking.
650 * _SEND_WITH_IMM x x x
652 * _RDMA_WRITE_WITH_IMM x x
656 * _ATOMIC_CMP_AND_SWP x
657 * _ATOMIC_FETCH_AND_ADD x
658 * _MASK_ATOMIC_CMP_AND_SWP x
659 * _MASK_ATOMIC_FETCH_AND_ADD x
663 if (qp->ibqp.qp_type != IB_QPT_UD &&
664 qp->ibqp.qp_type != IB_QPT_RC &&
665 wr->opcode != IB_WR_SEND) {
666 dev_warn_ratelimited(&dev->pdev->dev,
667 "unsupported queuepair type\n");
671 } else if (qp->ibqp.qp_type == IB_QPT_UD ||
672 qp->ibqp.qp_type == IB_QPT_GSI) {
673 if (wr->opcode != IB_WR_SEND &&
674 wr->opcode != IB_WR_SEND_WITH_IMM) {
675 dev_warn_ratelimited(&dev->pdev->dev,
676 "invalid send opcode\n");
683 wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, index);
684 memset(wqe_hdr, 0, sizeof(*wqe_hdr));
685 wqe_hdr->wr_id = wr->wr_id;
686 wqe_hdr->num_sge = wr->num_sge;
687 wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
688 wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
689 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
690 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
691 wqe_hdr->ex.imm_data = wr->ex.imm_data;
693 switch (qp->ibqp.qp_type) {
696 if (unlikely(!ud_wr(wr)->ah)) {
697 dev_warn_ratelimited(&dev->pdev->dev,
698 "invalid address handle\n");
705 * Use qkey from qp context if high order bit set,
706 * otherwise from work request.
708 wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
709 wqe_hdr->wr.ud.remote_qkey =
710 ud_wr(wr)->remote_qkey & 0x80000000 ?
711 qp->qkey : ud_wr(wr)->remote_qkey;
712 wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
716 switch (wr->opcode) {
717 case IB_WR_RDMA_READ:
718 case IB_WR_RDMA_WRITE:
719 case IB_WR_RDMA_WRITE_WITH_IMM:
720 wqe_hdr->wr.rdma.remote_addr =
721 rdma_wr(wr)->remote_addr;
722 wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
724 case IB_WR_LOCAL_INV:
725 case IB_WR_SEND_WITH_INV:
726 wqe_hdr->ex.invalidate_rkey =
727 wr->ex.invalidate_rkey;
729 case IB_WR_ATOMIC_CMP_AND_SWP:
730 case IB_WR_ATOMIC_FETCH_AND_ADD:
731 wqe_hdr->wr.atomic.remote_addr =
732 atomic_wr(wr)->remote_addr;
733 wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
734 wqe_hdr->wr.atomic.compare_add =
735 atomic_wr(wr)->compare_add;
736 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
737 wqe_hdr->wr.atomic.swap =
741 ret = set_reg_seg(wqe_hdr, reg_wr(wr));
743 dev_warn_ratelimited(&dev->pdev->dev,
744 "Failed to set fast register work request\n");
755 dev_warn_ratelimited(&dev->pdev->dev,
756 "invalid queuepair type\n");
762 sge = (struct pvrdma_sge *)(wqe_hdr + 1);
763 for (i = 0; i < wr->num_sge; i++) {
764 /* Need to check wqe_size 0 or max size */
765 sge->addr = wr->sg_list[i].addr;
766 sge->length = wr->sg_list[i].length;
767 sge->lkey = wr->sg_list[i].lkey;
771 /* Make sure wqe is written before index update */
775 if (unlikely(index >= qp->sq.wqe_cnt))
777 /* Update shared sq ring */
778 pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
785 spin_unlock_irqrestore(&qp->sq.lock, flags);
788 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
794 * pvrdma_post_receive - post receive work request entries on a QP
796 * @wr: the work request list to post
797 * @bad_wr: the first bad WR returned
799 * @return: 0 on success, otherwise errno returned.
801 int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
802 struct ib_recv_wr **bad_wr)
804 struct pvrdma_dev *dev = to_vdev(ibqp->device);
806 struct pvrdma_qp *qp = to_vqp(ibqp);
807 struct pvrdma_rq_wqe_hdr *wqe_hdr;
808 struct pvrdma_sge *sge;
814 * In the RESET state, we can fail immediately. For other states,
815 * just post and let the device figure it out.
817 if (qp->state == IB_QPS_RESET) {
822 spin_lock_irqsave(&qp->rq.lock, flags);
824 index = pvrdma_idx(&qp->rq.ring->prod_tail, qp->rq.wqe_cnt);
825 for (nreq = 0; wr; nreq++, wr = wr->next) {
828 if (unlikely(wr->num_sge > qp->rq.max_sg ||
832 dev_warn_ratelimited(&dev->pdev->dev,
833 "recv SGE overflow\n");
837 if (unlikely(!pvrdma_idx_ring_has_space(
838 qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
841 dev_warn_ratelimited(&dev->pdev->dev,
842 "recv queue full\n");
846 wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, index);
847 wqe_hdr->wr_id = wr->wr_id;
848 wqe_hdr->num_sge = wr->num_sge;
849 wqe_hdr->total_len = 0;
851 sge = (struct pvrdma_sge *)(wqe_hdr + 1);
852 for (i = 0; i < wr->num_sge; i++) {
853 sge->addr = wr->sg_list[i].addr;
854 sge->length = wr->sg_list[i].length;
855 sge->lkey = wr->sg_list[i].lkey;
859 /* Make sure wqe is written before index update */
863 if (unlikely(index >= qp->rq.wqe_cnt))
865 /* Update shared rq ring */
866 pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
870 spin_unlock_irqrestore(&qp->rq.lock, flags);
872 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
877 spin_unlock_irqrestore(&qp->rq.lock, flags);
883 * pvrdma_query_qp - query a queue pair's attributes
884 * @ibqp: the queue pair to query
885 * @attr: the queue pair's attributes
886 * @attr_mask: attributes mask
887 * @init_attr: initial queue pair attributes
889 * @returns 0 on success, otherwise returns an errno.
891 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
892 int attr_mask, struct ib_qp_init_attr *init_attr)
894 struct pvrdma_dev *dev = to_vdev(ibqp->device);
895 struct pvrdma_qp *qp = to_vqp(ibqp);
896 union pvrdma_cmd_req req;
897 union pvrdma_cmd_resp rsp;
898 struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
899 struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
902 mutex_lock(&qp->mutex);
904 if (qp->state == IB_QPS_RESET) {
905 attr->qp_state = IB_QPS_RESET;
909 memset(cmd, 0, sizeof(*cmd));
910 cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
911 cmd->qp_handle = qp->qp_handle;
912 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
914 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
916 dev_warn(&dev->pdev->dev,
917 "could not query queuepair, error: %d\n", ret);
921 attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
923 pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
924 attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
925 attr->path_mig_state =
926 pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
927 attr->qkey = resp->attrs.qkey;
928 attr->rq_psn = resp->attrs.rq_psn;
929 attr->sq_psn = resp->attrs.sq_psn;
930 attr->dest_qp_num = resp->attrs.dest_qp_num;
931 attr->qp_access_flags =
932 pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
933 attr->pkey_index = resp->attrs.pkey_index;
934 attr->alt_pkey_index = resp->attrs.alt_pkey_index;
935 attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
936 attr->sq_draining = resp->attrs.sq_draining;
937 attr->max_rd_atomic = resp->attrs.max_rd_atomic;
938 attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
939 attr->min_rnr_timer = resp->attrs.min_rnr_timer;
940 attr->port_num = resp->attrs.port_num;
941 attr->timeout = resp->attrs.timeout;
942 attr->retry_cnt = resp->attrs.retry_cnt;
943 attr->rnr_retry = resp->attrs.rnr_retry;
944 attr->alt_port_num = resp->attrs.alt_port_num;
945 attr->alt_timeout = resp->attrs.alt_timeout;
946 pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
947 pvrdma_ah_attr_to_ib(&attr->ah_attr, &resp->attrs.ah_attr);
948 pvrdma_ah_attr_to_ib(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
950 qp->state = attr->qp_state;
955 attr->cur_qp_state = attr->qp_state;
957 init_attr->event_handler = qp->ibqp.event_handler;
958 init_attr->qp_context = qp->ibqp.qp_context;
959 init_attr->send_cq = qp->ibqp.send_cq;
960 init_attr->recv_cq = qp->ibqp.recv_cq;
961 init_attr->srq = qp->ibqp.srq;
962 init_attr->xrcd = NULL;
963 init_attr->cap = attr->cap;
964 init_attr->sq_sig_type = 0;
965 init_attr->qp_type = qp->ibqp.qp_type;
966 init_attr->create_flags = 0;
967 init_attr->port_num = qp->port;
969 mutex_unlock(&qp->mutex);