1 /*
2 * Copyright (c) 2012-2016 VMware, Inc. All rights reserved.
3 *
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.
12 *
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.
16 *
17 * The BSD 2-Clause License
18 *
19 * Redistribution and use in source and binary forms, with or
20 * without modification, are permitted provided that the following
21 * conditions are met:
22 *
23 * - Redistributions of source code must retain the above
24 * copyright notice, this list of conditions and the following
25 * disclaimer.
26 *
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.
31 *
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.
44 */
45
46 #include <asm/page.h>
47 #include <linux/io.h>
48 #include <linux/wait.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_smi.h>
51 #include <rdma/ib_user_verbs.h>
52
53 #include "pvrdma.h"
54
get_cqs(struct pvrdma_qp * qp,struct pvrdma_cq ** send_cq,struct pvrdma_cq ** recv_cq)55 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
56 struct pvrdma_cq **recv_cq)
57 {
58 *send_cq = to_vcq(qp->ibqp.send_cq);
59 *recv_cq = to_vcq(qp->ibqp.recv_cq);
60 }
61
pvrdma_lock_cqs(struct pvrdma_cq * scq,struct pvrdma_cq * rcq,unsigned long * scq_flags,unsigned long * rcq_flags)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)
66 {
67 if (scq == rcq) {
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);
74 } else {
75 spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
76 spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
77 SINGLE_DEPTH_NESTING);
78 }
79 }
80
pvrdma_unlock_cqs(struct pvrdma_cq * scq,struct pvrdma_cq * rcq,unsigned long * scq_flags,unsigned long * rcq_flags)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)
85 {
86 if (scq == rcq) {
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);
92 } else {
93 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
94 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
95 }
96 }
97
pvrdma_reset_qp(struct pvrdma_qp * qp)98 static void pvrdma_reset_qp(struct pvrdma_qp *qp)
99 {
100 struct pvrdma_cq *scq, *rcq;
101 unsigned long scq_flags, rcq_flags;
102
103 /* Clean up cqes */
104 get_cqs(qp, &scq, &rcq);
105 pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
106
107 _pvrdma_flush_cqe(qp, scq);
108 if (scq != rcq)
109 _pvrdma_flush_cqe(qp, rcq);
110
111 pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
112
113 /*
114 * Reset queuepair. The checks are because usermode queuepairs won't
115 * have kernel ringstates.
116 */
117 if (qp->rq.ring) {
118 atomic_set(&qp->rq.ring->cons_head, 0);
119 atomic_set(&qp->rq.ring->prod_tail, 0);
120 }
121 if (qp->sq.ring) {
122 atomic_set(&qp->sq.ring->cons_head, 0);
123 atomic_set(&qp->sq.ring->prod_tail, 0);
124 }
125 }
126
pvrdma_set_rq_size(struct pvrdma_dev * dev,struct ib_qp_cap * req_cap,struct pvrdma_qp * qp)127 static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
128 struct ib_qp_cap *req_cap,
129 struct pvrdma_qp *qp)
130 {
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");
134 return -EINVAL;
135 }
136
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));
139
140 /* Write back */
141 req_cap->max_recv_wr = qp->rq.wqe_cnt;
142 req_cap->max_recv_sge = qp->rq.max_sg;
143
144 qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
145 sizeof(struct pvrdma_sge) *
146 qp->rq.max_sg);
147 qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
148 PAGE_SIZE;
149
150 return 0;
151 }
152
pvrdma_set_sq_size(struct pvrdma_dev * dev,struct ib_qp_cap * req_cap,struct pvrdma_qp * qp)153 static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
154 struct pvrdma_qp *qp)
155 {
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");
159 return -EINVAL;
160 }
161
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));
164
165 /* Write back */
166 req_cap->max_send_wr = qp->sq.wqe_cnt;
167 req_cap->max_send_sge = qp->sq.max_sg;
168
169 qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
170 sizeof(struct pvrdma_sge) *
171 qp->sq.max_sg);
172 /* Note: one extra page for the header. */
173 qp->npages_send = PVRDMA_QP_NUM_HEADER_PAGES +
174 (qp->sq.wqe_cnt * qp->sq.wqe_size + PAGE_SIZE - 1) /
175 PAGE_SIZE;
176
177 return 0;
178 }
179
180 /**
181 * pvrdma_create_qp - create queue pair
182 * @pd: protection domain
183 * @init_attr: queue pair attributes
184 * @udata: user data
185 *
186 * @return: the ib_qp pointer on success, otherwise returns an errno.
187 */
pvrdma_create_qp(struct ib_pd * pd,struct ib_qp_init_attr * init_attr,struct ib_udata * udata)188 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
189 struct ib_qp_init_attr *init_attr,
190 struct ib_udata *udata)
191 {
192 struct pvrdma_qp *qp = NULL;
193 struct pvrdma_dev *dev = to_vdev(pd->device);
194 union pvrdma_cmd_req req;
195 union pvrdma_cmd_resp rsp;
196 struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
197 struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
198 struct pvrdma_create_qp ucmd;
199 unsigned long flags;
200 int ret;
201
202 if (init_attr->create_flags) {
203 dev_warn(&dev->pdev->dev,
204 "invalid create queuepair flags %#x\n",
205 init_attr->create_flags);
206 return ERR_PTR(-EINVAL);
207 }
208
209 if (init_attr->qp_type != IB_QPT_RC &&
210 init_attr->qp_type != IB_QPT_UD &&
211 init_attr->qp_type != IB_QPT_GSI) {
212 dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
213 init_attr->qp_type);
214 return ERR_PTR(-EINVAL);
215 }
216
217 if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
218 return ERR_PTR(-ENOMEM);
219
220 switch (init_attr->qp_type) {
221 case IB_QPT_GSI:
222 if (init_attr->port_num == 0 ||
223 init_attr->port_num > pd->device->phys_port_cnt ||
224 udata) {
225 dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
226 ret = -EINVAL;
227 goto err_qp;
228 }
229 /* fall through */
230 case IB_QPT_RC:
231 case IB_QPT_UD:
232 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
233 if (!qp) {
234 ret = -ENOMEM;
235 goto err_qp;
236 }
237
238 spin_lock_init(&qp->sq.lock);
239 spin_lock_init(&qp->rq.lock);
240 mutex_init(&qp->mutex);
241 atomic_set(&qp->refcnt, 1);
242 init_waitqueue_head(&qp->wait);
243
244 qp->state = IB_QPS_RESET;
245
246 if (pd->uobject && udata) {
247 dev_dbg(&dev->pdev->dev,
248 "create queuepair from user space\n");
249
250 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
251 ret = -EFAULT;
252 goto err_qp;
253 }
254
255 /* set qp->sq.wqe_cnt, shift, buf_size.. */
256 qp->rumem = ib_umem_get(pd->uobject->context,
257 ucmd.rbuf_addr,
258 ucmd.rbuf_size, 0, 0);
259 if (IS_ERR(qp->rumem)) {
260 ret = PTR_ERR(qp->rumem);
261 goto err_qp;
262 }
263
264 qp->sumem = ib_umem_get(pd->uobject->context,
265 ucmd.sbuf_addr,
266 ucmd.sbuf_size, 0, 0);
267 if (IS_ERR(qp->sumem)) {
268 ib_umem_release(qp->rumem);
269 ret = PTR_ERR(qp->sumem);
270 goto err_qp;
271 }
272
273 qp->npages_send = ib_umem_page_count(qp->sumem);
274 qp->npages_recv = ib_umem_page_count(qp->rumem);
275 qp->npages = qp->npages_send + qp->npages_recv;
276 } else {
277 qp->is_kernel = true;
278
279 ret = pvrdma_set_sq_size(to_vdev(pd->device),
280 &init_attr->cap, qp);
281 if (ret)
282 goto err_qp;
283
284 ret = pvrdma_set_rq_size(to_vdev(pd->device),
285 &init_attr->cap, qp);
286 if (ret)
287 goto err_qp;
288
289 qp->npages = qp->npages_send + qp->npages_recv;
290
291 /* Skip header page. */
292 qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE;
293
294 /* Recv queue pages are after send pages. */
295 qp->rq.offset = qp->npages_send * PAGE_SIZE;
296 }
297
298 if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
299 dev_warn(&dev->pdev->dev,
300 "overflow pages in queuepair\n");
301 ret = -EINVAL;
302 goto err_umem;
303 }
304
305 ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
306 qp->is_kernel);
307 if (ret) {
308 dev_warn(&dev->pdev->dev,
309 "could not allocate page directory\n");
310 goto err_umem;
311 }
312
313 if (!qp->is_kernel) {
314 pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
315 pvrdma_page_dir_insert_umem(&qp->pdir, qp->rumem,
316 qp->npages_send);
317 } else {
318 /* Ring state is always the first page. */
319 qp->sq.ring = qp->pdir.pages[0];
320 qp->rq.ring = &qp->sq.ring[1];
321 }
322 break;
323 default:
324 ret = -EINVAL;
325 goto err_qp;
326 }
327
328 /* Not supported */
329 init_attr->cap.max_inline_data = 0;
330
331 memset(cmd, 0, sizeof(*cmd));
332 cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
333 cmd->pd_handle = to_vpd(pd)->pd_handle;
334 cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
335 cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
336 cmd->max_send_wr = init_attr->cap.max_send_wr;
337 cmd->max_recv_wr = init_attr->cap.max_recv_wr;
338 cmd->max_send_sge = init_attr->cap.max_send_sge;
339 cmd->max_recv_sge = init_attr->cap.max_recv_sge;
340 cmd->max_inline_data = init_attr->cap.max_inline_data;
341 cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
342 cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
343 cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
344 cmd->total_chunks = qp->npages;
345 cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
346 cmd->pdir_dma = qp->pdir.dir_dma;
347
348 dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
349 cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
350 cmd->max_recv_sge);
351
352 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
353 if (ret < 0) {
354 dev_warn(&dev->pdev->dev,
355 "could not create queuepair, error: %d\n", ret);
356 goto err_pdir;
357 }
358
359 /* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
360 qp->qp_handle = resp->qpn;
361 qp->port = init_attr->port_num;
362 qp->ibqp.qp_num = resp->qpn;
363 spin_lock_irqsave(&dev->qp_tbl_lock, flags);
364 dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
365 spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
366
367 return &qp->ibqp;
368
369 err_pdir:
370 pvrdma_page_dir_cleanup(dev, &qp->pdir);
371 err_umem:
372 if (pd->uobject && udata) {
373 if (qp->rumem)
374 ib_umem_release(qp->rumem);
375 if (qp->sumem)
376 ib_umem_release(qp->sumem);
377 }
378 err_qp:
379 kfree(qp);
380 atomic_dec(&dev->num_qps);
381
382 return ERR_PTR(ret);
383 }
384
pvrdma_free_qp(struct pvrdma_qp * qp)385 static void pvrdma_free_qp(struct pvrdma_qp *qp)
386 {
387 struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
388 struct pvrdma_cq *scq;
389 struct pvrdma_cq *rcq;
390 unsigned long flags, scq_flags, rcq_flags;
391
392 /* In case cq is polling */
393 get_cqs(qp, &scq, &rcq);
394 pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
395
396 _pvrdma_flush_cqe(qp, scq);
397 if (scq != rcq)
398 _pvrdma_flush_cqe(qp, rcq);
399
400 spin_lock_irqsave(&dev->qp_tbl_lock, flags);
401 dev->qp_tbl[qp->qp_handle] = NULL;
402 spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
403
404 pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
405
406 atomic_dec(&qp->refcnt);
407 wait_event(qp->wait, !atomic_read(&qp->refcnt));
408
409 if (!qp->is_kernel) {
410 if (qp->rumem)
411 ib_umem_release(qp->rumem);
412 if (qp->sumem)
413 ib_umem_release(qp->sumem);
414 }
415
416 pvrdma_page_dir_cleanup(dev, &qp->pdir);
417
418 kfree(qp);
419
420 atomic_dec(&dev->num_qps);
421 }
422
423 /**
424 * pvrdma_destroy_qp - destroy a queue pair
425 * @qp: the queue pair to destroy
426 *
427 * @return: 0 on success.
428 */
pvrdma_destroy_qp(struct ib_qp * qp)429 int pvrdma_destroy_qp(struct ib_qp *qp)
430 {
431 struct pvrdma_qp *vqp = to_vqp(qp);
432 union pvrdma_cmd_req req;
433 struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
434 int ret;
435
436 memset(cmd, 0, sizeof(*cmd));
437 cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
438 cmd->qp_handle = vqp->qp_handle;
439
440 ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
441 if (ret < 0)
442 dev_warn(&to_vdev(qp->device)->pdev->dev,
443 "destroy queuepair failed, error: %d\n", ret);
444
445 pvrdma_free_qp(vqp);
446
447 return 0;
448 }
449
450 /**
451 * pvrdma_modify_qp - modify queue pair attributes
452 * @ibqp: the queue pair
453 * @attr: the new queue pair's attributes
454 * @attr_mask: attributes mask
455 * @udata: user data
456 *
457 * @returns 0 on success, otherwise returns an errno.
458 */
pvrdma_modify_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int attr_mask,struct ib_udata * udata)459 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
460 int attr_mask, struct ib_udata *udata)
461 {
462 struct pvrdma_dev *dev = to_vdev(ibqp->device);
463 struct pvrdma_qp *qp = to_vqp(ibqp);
464 union pvrdma_cmd_req req;
465 union pvrdma_cmd_resp rsp;
466 struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
467 int cur_state, next_state;
468 int ret;
469
470 /* Sanity checking. Should need lock here */
471 mutex_lock(&qp->mutex);
472 cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
473 qp->state;
474 next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
475
476 if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
477 attr_mask, IB_LINK_LAYER_ETHERNET)) {
478 ret = -EINVAL;
479 goto out;
480 }
481
482 if (attr_mask & IB_QP_PORT) {
483 if (attr->port_num == 0 ||
484 attr->port_num > ibqp->device->phys_port_cnt) {
485 ret = -EINVAL;
486 goto out;
487 }
488 }
489
490 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
491 if (attr->min_rnr_timer > 31) {
492 ret = -EINVAL;
493 goto out;
494 }
495 }
496
497 if (attr_mask & IB_QP_PKEY_INDEX) {
498 if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
499 ret = -EINVAL;
500 goto out;
501 }
502 }
503
504 if (attr_mask & IB_QP_QKEY)
505 qp->qkey = attr->qkey;
506
507 if (cur_state == next_state && cur_state == IB_QPS_RESET) {
508 ret = 0;
509 goto out;
510 }
511
512 qp->state = next_state;
513 memset(cmd, 0, sizeof(*cmd));
514 cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
515 cmd->qp_handle = qp->qp_handle;
516 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
517 cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
518 cmd->attrs.cur_qp_state =
519 ib_qp_state_to_pvrdma(attr->cur_qp_state);
520 cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
521 cmd->attrs.path_mig_state =
522 ib_mig_state_to_pvrdma(attr->path_mig_state);
523 cmd->attrs.qkey = attr->qkey;
524 cmd->attrs.rq_psn = attr->rq_psn;
525 cmd->attrs.sq_psn = attr->sq_psn;
526 cmd->attrs.dest_qp_num = attr->dest_qp_num;
527 cmd->attrs.qp_access_flags =
528 ib_access_flags_to_pvrdma(attr->qp_access_flags);
529 cmd->attrs.pkey_index = attr->pkey_index;
530 cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
531 cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
532 cmd->attrs.sq_draining = attr->sq_draining;
533 cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
534 cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
535 cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
536 cmd->attrs.port_num = attr->port_num;
537 cmd->attrs.timeout = attr->timeout;
538 cmd->attrs.retry_cnt = attr->retry_cnt;
539 cmd->attrs.rnr_retry = attr->rnr_retry;
540 cmd->attrs.alt_port_num = attr->alt_port_num;
541 cmd->attrs.alt_timeout = attr->alt_timeout;
542 ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
543 rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
544 rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
545
546 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
547 if (ret < 0) {
548 dev_warn(&dev->pdev->dev,
549 "could not modify queuepair, error: %d\n", ret);
550 } else if (rsp.hdr.err > 0) {
551 dev_warn(&dev->pdev->dev,
552 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
553 ret = -EINVAL;
554 }
555
556 if (ret == 0 && next_state == IB_QPS_RESET)
557 pvrdma_reset_qp(qp);
558
559 out:
560 mutex_unlock(&qp->mutex);
561
562 return ret;
563 }
564
get_sq_wqe(struct pvrdma_qp * qp,unsigned int n)565 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n)
566 {
567 return pvrdma_page_dir_get_ptr(&qp->pdir,
568 qp->sq.offset + n * qp->sq.wqe_size);
569 }
570
get_rq_wqe(struct pvrdma_qp * qp,unsigned int n)571 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n)
572 {
573 return pvrdma_page_dir_get_ptr(&qp->pdir,
574 qp->rq.offset + n * qp->rq.wqe_size);
575 }
576
set_reg_seg(struct pvrdma_sq_wqe_hdr * wqe_hdr,struct ib_reg_wr * wr)577 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_wr *wr)
578 {
579 struct pvrdma_user_mr *mr = to_vmr(wr->mr);
580
581 wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
582 wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
583 wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
584 wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
585 wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
586 wqe_hdr->wr.fast_reg.access_flags = wr->access;
587 wqe_hdr->wr.fast_reg.rkey = wr->key;
588
589 return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
590 mr->npages);
591 }
592
593 /**
594 * pvrdma_post_send - post send work request entries on a QP
595 * @ibqp: the QP
596 * @wr: work request list to post
597 * @bad_wr: the first bad WR returned
598 *
599 * @return: 0 on success, otherwise errno returned.
600 */
pvrdma_post_send(struct ib_qp * ibqp,struct ib_send_wr * wr,struct ib_send_wr ** bad_wr)601 int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
602 struct ib_send_wr **bad_wr)
603 {
604 struct pvrdma_qp *qp = to_vqp(ibqp);
605 struct pvrdma_dev *dev = to_vdev(ibqp->device);
606 unsigned long flags;
607 struct pvrdma_sq_wqe_hdr *wqe_hdr;
608 struct pvrdma_sge *sge;
609 int i, ret;
610
611 /*
612 * In states lower than RTS, we can fail immediately. In other states,
613 * just post and let the device figure it out.
614 */
615 if (qp->state < IB_QPS_RTS) {
616 *bad_wr = wr;
617 return -EINVAL;
618 }
619
620 spin_lock_irqsave(&qp->sq.lock, flags);
621
622 while (wr) {
623 unsigned int tail = 0;
624
625 if (unlikely(!pvrdma_idx_ring_has_space(
626 qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
627 dev_warn_ratelimited(&dev->pdev->dev,
628 "send queue is full\n");
629 *bad_wr = wr;
630 ret = -ENOMEM;
631 goto out;
632 }
633
634 if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
635 dev_warn_ratelimited(&dev->pdev->dev,
636 "send SGE overflow\n");
637 *bad_wr = wr;
638 ret = -EINVAL;
639 goto out;
640 }
641
642 if (unlikely(wr->opcode < 0)) {
643 dev_warn_ratelimited(&dev->pdev->dev,
644 "invalid send opcode\n");
645 *bad_wr = wr;
646 ret = -EINVAL;
647 goto out;
648 }
649
650 /*
651 * Only support UD, RC.
652 * Need to check opcode table for thorough checking.
653 * opcode _UD _UC _RC
654 * _SEND x x x
655 * _SEND_WITH_IMM x x x
656 * _RDMA_WRITE x x
657 * _RDMA_WRITE_WITH_IMM x x
658 * _LOCAL_INV x x
659 * _SEND_WITH_INV x x
660 * _RDMA_READ x
661 * _ATOMIC_CMP_AND_SWP x
662 * _ATOMIC_FETCH_AND_ADD x
663 * _MASK_ATOMIC_CMP_AND_SWP x
664 * _MASK_ATOMIC_FETCH_AND_ADD x
665 * _REG_MR x
666 *
667 */
668 if (qp->ibqp.qp_type != IB_QPT_UD &&
669 qp->ibqp.qp_type != IB_QPT_RC &&
670 wr->opcode != IB_WR_SEND) {
671 dev_warn_ratelimited(&dev->pdev->dev,
672 "unsupported queuepair type\n");
673 *bad_wr = wr;
674 ret = -EINVAL;
675 goto out;
676 } else if (qp->ibqp.qp_type == IB_QPT_UD ||
677 qp->ibqp.qp_type == IB_QPT_GSI) {
678 if (wr->opcode != IB_WR_SEND &&
679 wr->opcode != IB_WR_SEND_WITH_IMM) {
680 dev_warn_ratelimited(&dev->pdev->dev,
681 "invalid send opcode\n");
682 *bad_wr = wr;
683 ret = -EINVAL;
684 goto out;
685 }
686 }
687
688 wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail);
689 memset(wqe_hdr, 0, sizeof(*wqe_hdr));
690 wqe_hdr->wr_id = wr->wr_id;
691 wqe_hdr->num_sge = wr->num_sge;
692 wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
693 wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
694 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
695 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
696 wqe_hdr->ex.imm_data = wr->ex.imm_data;
697
698 if (unlikely(wqe_hdr->opcode == PVRDMA_WR_ERROR)) {
699 *bad_wr = wr;
700 ret = -EINVAL;
701 goto out;
702 }
703
704 switch (qp->ibqp.qp_type) {
705 case IB_QPT_GSI:
706 case IB_QPT_UD:
707 if (unlikely(!ud_wr(wr)->ah)) {
708 dev_warn_ratelimited(&dev->pdev->dev,
709 "invalid address handle\n");
710 *bad_wr = wr;
711 ret = -EINVAL;
712 goto out;
713 }
714
715 /*
716 * Use qkey from qp context if high order bit set,
717 * otherwise from work request.
718 */
719 wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
720 wqe_hdr->wr.ud.remote_qkey =
721 ud_wr(wr)->remote_qkey & 0x80000000 ?
722 qp->qkey : ud_wr(wr)->remote_qkey;
723 wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
724
725 break;
726 case IB_QPT_RC:
727 switch (wr->opcode) {
728 case IB_WR_RDMA_READ:
729 case IB_WR_RDMA_WRITE:
730 case IB_WR_RDMA_WRITE_WITH_IMM:
731 wqe_hdr->wr.rdma.remote_addr =
732 rdma_wr(wr)->remote_addr;
733 wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
734 break;
735 case IB_WR_LOCAL_INV:
736 case IB_WR_SEND_WITH_INV:
737 wqe_hdr->ex.invalidate_rkey =
738 wr->ex.invalidate_rkey;
739 break;
740 case IB_WR_ATOMIC_CMP_AND_SWP:
741 case IB_WR_ATOMIC_FETCH_AND_ADD:
742 wqe_hdr->wr.atomic.remote_addr =
743 atomic_wr(wr)->remote_addr;
744 wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
745 wqe_hdr->wr.atomic.compare_add =
746 atomic_wr(wr)->compare_add;
747 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
748 wqe_hdr->wr.atomic.swap =
749 atomic_wr(wr)->swap;
750 break;
751 case IB_WR_REG_MR:
752 ret = set_reg_seg(wqe_hdr, reg_wr(wr));
753 if (ret < 0) {
754 dev_warn_ratelimited(&dev->pdev->dev,
755 "Failed to set fast register work request\n");
756 *bad_wr = wr;
757 goto out;
758 }
759 break;
760 default:
761 break;
762 }
763
764 break;
765 default:
766 dev_warn_ratelimited(&dev->pdev->dev,
767 "invalid queuepair type\n");
768 ret = -EINVAL;
769 *bad_wr = wr;
770 goto out;
771 }
772
773 sge = (struct pvrdma_sge *)(wqe_hdr + 1);
774 for (i = 0; i < wr->num_sge; i++) {
775 /* Need to check wqe_size 0 or max size */
776 sge->addr = wr->sg_list[i].addr;
777 sge->length = wr->sg_list[i].length;
778 sge->lkey = wr->sg_list[i].lkey;
779 sge++;
780 }
781
782 /* Make sure wqe is written before index update */
783 smp_wmb();
784
785 /* Update shared sq ring */
786 pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
787 qp->sq.wqe_cnt);
788
789 wr = wr->next;
790 }
791
792 ret = 0;
793
794 out:
795 spin_unlock_irqrestore(&qp->sq.lock, flags);
796
797 if (!ret)
798 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
799
800 return ret;
801 }
802
803 /**
804 * pvrdma_post_receive - post receive work request entries on a QP
805 * @ibqp: the QP
806 * @wr: the work request list to post
807 * @bad_wr: the first bad WR returned
808 *
809 * @return: 0 on success, otherwise errno returned.
810 */
pvrdma_post_recv(struct ib_qp * ibqp,struct ib_recv_wr * wr,struct ib_recv_wr ** bad_wr)811 int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
812 struct ib_recv_wr **bad_wr)
813 {
814 struct pvrdma_dev *dev = to_vdev(ibqp->device);
815 unsigned long flags;
816 struct pvrdma_qp *qp = to_vqp(ibqp);
817 struct pvrdma_rq_wqe_hdr *wqe_hdr;
818 struct pvrdma_sge *sge;
819 int ret = 0;
820 int i;
821
822 /*
823 * In the RESET state, we can fail immediately. For other states,
824 * just post and let the device figure it out.
825 */
826 if (qp->state == IB_QPS_RESET) {
827 *bad_wr = wr;
828 return -EINVAL;
829 }
830
831 spin_lock_irqsave(&qp->rq.lock, flags);
832
833 while (wr) {
834 unsigned int tail = 0;
835
836 if (unlikely(wr->num_sge > qp->rq.max_sg ||
837 wr->num_sge < 0)) {
838 ret = -EINVAL;
839 *bad_wr = wr;
840 dev_warn_ratelimited(&dev->pdev->dev,
841 "recv SGE overflow\n");
842 goto out;
843 }
844
845 if (unlikely(!pvrdma_idx_ring_has_space(
846 qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
847 ret = -ENOMEM;
848 *bad_wr = wr;
849 dev_warn_ratelimited(&dev->pdev->dev,
850 "recv queue full\n");
851 goto out;
852 }
853
854 wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail);
855 wqe_hdr->wr_id = wr->wr_id;
856 wqe_hdr->num_sge = wr->num_sge;
857 wqe_hdr->total_len = 0;
858
859 sge = (struct pvrdma_sge *)(wqe_hdr + 1);
860 for (i = 0; i < wr->num_sge; i++) {
861 sge->addr = wr->sg_list[i].addr;
862 sge->length = wr->sg_list[i].length;
863 sge->lkey = wr->sg_list[i].lkey;
864 sge++;
865 }
866
867 /* Make sure wqe is written before index update */
868 smp_wmb();
869
870 /* Update shared rq ring */
871 pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
872 qp->rq.wqe_cnt);
873
874 wr = wr->next;
875 }
876
877 spin_unlock_irqrestore(&qp->rq.lock, flags);
878
879 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
880
881 return ret;
882
883 out:
884 spin_unlock_irqrestore(&qp->rq.lock, flags);
885
886 return ret;
887 }
888
889 /**
890 * pvrdma_query_qp - query a queue pair's attributes
891 * @ibqp: the queue pair to query
892 * @attr: the queue pair's attributes
893 * @attr_mask: attributes mask
894 * @init_attr: initial queue pair attributes
895 *
896 * @returns 0 on success, otherwise returns an errno.
897 */
pvrdma_query_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int attr_mask,struct ib_qp_init_attr * init_attr)898 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
899 int attr_mask, struct ib_qp_init_attr *init_attr)
900 {
901 struct pvrdma_dev *dev = to_vdev(ibqp->device);
902 struct pvrdma_qp *qp = to_vqp(ibqp);
903 union pvrdma_cmd_req req;
904 union pvrdma_cmd_resp rsp;
905 struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
906 struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
907 int ret = 0;
908
909 mutex_lock(&qp->mutex);
910
911 if (qp->state == IB_QPS_RESET) {
912 attr->qp_state = IB_QPS_RESET;
913 goto out;
914 }
915
916 memset(cmd, 0, sizeof(*cmd));
917 cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
918 cmd->qp_handle = qp->qp_handle;
919 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
920
921 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
922 if (ret < 0) {
923 dev_warn(&dev->pdev->dev,
924 "could not query queuepair, error: %d\n", ret);
925 goto out;
926 }
927
928 attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
929 attr->cur_qp_state =
930 pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
931 attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
932 attr->path_mig_state =
933 pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
934 attr->qkey = resp->attrs.qkey;
935 attr->rq_psn = resp->attrs.rq_psn;
936 attr->sq_psn = resp->attrs.sq_psn;
937 attr->dest_qp_num = resp->attrs.dest_qp_num;
938 attr->qp_access_flags =
939 pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
940 attr->pkey_index = resp->attrs.pkey_index;
941 attr->alt_pkey_index = resp->attrs.alt_pkey_index;
942 attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
943 attr->sq_draining = resp->attrs.sq_draining;
944 attr->max_rd_atomic = resp->attrs.max_rd_atomic;
945 attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
946 attr->min_rnr_timer = resp->attrs.min_rnr_timer;
947 attr->port_num = resp->attrs.port_num;
948 attr->timeout = resp->attrs.timeout;
949 attr->retry_cnt = resp->attrs.retry_cnt;
950 attr->rnr_retry = resp->attrs.rnr_retry;
951 attr->alt_port_num = resp->attrs.alt_port_num;
952 attr->alt_timeout = resp->attrs.alt_timeout;
953 pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
954 pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr);
955 pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
956
957 qp->state = attr->qp_state;
958
959 ret = 0;
960
961 out:
962 attr->cur_qp_state = attr->qp_state;
963
964 init_attr->event_handler = qp->ibqp.event_handler;
965 init_attr->qp_context = qp->ibqp.qp_context;
966 init_attr->send_cq = qp->ibqp.send_cq;
967 init_attr->recv_cq = qp->ibqp.recv_cq;
968 init_attr->srq = qp->ibqp.srq;
969 init_attr->xrcd = NULL;
970 init_attr->cap = attr->cap;
971 init_attr->sq_sig_type = 0;
972 init_attr->qp_type = qp->ibqp.qp_type;
973 init_attr->create_flags = 0;
974 init_attr->port_num = qp->port;
975
976 mutex_unlock(&qp->mutex);
977 return ret;
978 }
979