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 55static void __pvrdma_destroy_qp(struct pvrdma_dev *dev, 56 struct pvrdma_qp *qp); 57 58static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq, 59 struct pvrdma_cq **recv_cq) 60{ 61 *send_cq = to_vcq(qp->ibqp.send_cq); 62 *recv_cq = to_vcq(qp->ibqp.recv_cq); 63} 64 65static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq, 66 unsigned long *scq_flags, 67 unsigned long *rcq_flags) 68 __acquires(scq->cq_lock) __acquires(rcq->cq_lock) 69{ 70 if (scq == rcq) { 71 spin_lock_irqsave(&scq->cq_lock, *scq_flags); 72 __acquire(rcq->cq_lock); 73 } else if (scq->cq_handle < rcq->cq_handle) { 74 spin_lock_irqsave(&scq->cq_lock, *scq_flags); 75 spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags, 76 SINGLE_DEPTH_NESTING); 77 } else { 78 spin_lock_irqsave(&rcq->cq_lock, *rcq_flags); 79 spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags, 80 SINGLE_DEPTH_NESTING); 81 } 82} 83 84static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq, 85 unsigned long *scq_flags, 86 unsigned long *rcq_flags) 87 __releases(scq->cq_lock) __releases(rcq->cq_lock) 88{ 89 if (scq == rcq) { 90 __release(rcq->cq_lock); 91 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags); 92 } else if (scq->cq_handle < rcq->cq_handle) { 93 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags); 94 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags); 95 } else { 96 spin_unlock_irqrestore(&scq->cq_lock, *scq_flags); 97 spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags); 98 } 99} 100 101static void pvrdma_reset_qp(struct pvrdma_qp *qp) 102{ 103 struct pvrdma_cq *scq, *rcq; 104 unsigned long scq_flags, rcq_flags; 105 106 /* Clean up cqes */ 107 get_cqs(qp, &scq, &rcq); 108 pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags); 109 110 _pvrdma_flush_cqe(qp, scq); 111 if (scq != rcq) 112 _pvrdma_flush_cqe(qp, rcq); 113 114 pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags); 115 116 /* 117 * Reset queuepair. The checks are because usermode queuepairs won't 118 * have kernel ringstates. 119 */ 120 if (qp->rq.ring) { 121 atomic_set(&qp->rq.ring->cons_head, 0); 122 atomic_set(&qp->rq.ring->prod_tail, 0); 123 } 124 if (qp->sq.ring) { 125 atomic_set(&qp->sq.ring->cons_head, 0); 126 atomic_set(&qp->sq.ring->prod_tail, 0); 127 } 128} 129 130static int pvrdma_set_rq_size(struct pvrdma_dev *dev, 131 struct ib_qp_cap *req_cap, 132 struct pvrdma_qp *qp) 133{ 134 if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr || 135 req_cap->max_recv_sge > dev->dsr->caps.max_sge) { 136 dev_warn(&dev->pdev->dev, "recv queue size invalid\n"); 137 return -EINVAL; 138 } 139 140 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr)); 141 qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge)); 142 143 /* Write back */ 144 req_cap->max_recv_wr = qp->rq.wqe_cnt; 145 req_cap->max_recv_sge = qp->rq.max_sg; 146 147 qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) + 148 sizeof(struct pvrdma_sge) * 149 qp->rq.max_sg); 150 qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) / 151 PAGE_SIZE; 152 153 return 0; 154} 155 156static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap, 157 struct pvrdma_qp *qp) 158{ 159 if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr || 160 req_cap->max_send_sge > dev->dsr->caps.max_sge) { 161 dev_warn(&dev->pdev->dev, "send queue size invalid\n"); 162 return -EINVAL; 163 } 164 165 qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr)); 166 qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge)); 167 168 /* Write back */ 169 req_cap->max_send_wr = qp->sq.wqe_cnt; 170 req_cap->max_send_sge = qp->sq.max_sg; 171 172 qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) + 173 sizeof(struct pvrdma_sge) * 174 qp->sq.max_sg); 175 /* Note: one extra page for the header. */ 176 qp->npages_send = PVRDMA_QP_NUM_HEADER_PAGES + 177 (qp->sq.wqe_cnt * qp->sq.wqe_size + PAGE_SIZE - 1) / 178 PAGE_SIZE; 179 180 return 0; 181} 182 183/** 184 * pvrdma_create_qp - create queue pair 185 * @pd: protection domain 186 * @init_attr: queue pair attributes 187 * @udata: user data 188 * 189 * @return: the ib_qp pointer on success, otherwise returns an errno. 190 */ 191struct ib_qp *pvrdma_create_qp(struct ib_pd *pd, 192 struct ib_qp_init_attr *init_attr, 193 struct ib_udata *udata) 194{ 195 struct pvrdma_qp *qp = NULL; 196 struct pvrdma_dev *dev = to_vdev(pd->device); 197 union pvrdma_cmd_req req; 198 union pvrdma_cmd_resp rsp; 199 struct pvrdma_cmd_create_qp *cmd = &req.create_qp; 200 struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp; 201 struct pvrdma_cmd_create_qp_resp_v2 *resp_v2 = &rsp.create_qp_resp_v2; 202 struct pvrdma_create_qp ucmd; 203 struct pvrdma_create_qp_resp qp_resp = {}; 204 unsigned long flags; 205 int ret; 206 bool is_srq = !!init_attr->srq; 207 208 if (init_attr->create_flags) { 209 dev_warn(&dev->pdev->dev, 210 "invalid create queuepair flags %#x\n", 211 init_attr->create_flags); 212 return ERR_PTR(-EINVAL); 213 } 214 215 if (init_attr->qp_type != IB_QPT_RC && 216 init_attr->qp_type != IB_QPT_UD && 217 init_attr->qp_type != IB_QPT_GSI) { 218 dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n", 219 init_attr->qp_type); 220 return ERR_PTR(-EOPNOTSUPP); 221 } 222 223 if (is_srq && !dev->dsr->caps.max_srq) { 224 dev_warn(&dev->pdev->dev, 225 "SRQs not supported by device\n"); 226 return ERR_PTR(-EINVAL); 227 } 228 229 if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp)) 230 return ERR_PTR(-ENOMEM); 231 232 switch (init_attr->qp_type) { 233 case IB_QPT_GSI: 234 if (init_attr->port_num == 0 || 235 init_attr->port_num > pd->device->phys_port_cnt) { 236 dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n"); 237 ret = -EINVAL; 238 goto err_qp; 239 } 240 fallthrough; 241 case IB_QPT_RC: 242 case IB_QPT_UD: 243 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 244 if (!qp) { 245 ret = -ENOMEM; 246 goto err_qp; 247 } 248 249 spin_lock_init(&qp->sq.lock); 250 spin_lock_init(&qp->rq.lock); 251 mutex_init(&qp->mutex); 252 refcount_set(&qp->refcnt, 1); 253 init_completion(&qp->free); 254 255 qp->state = IB_QPS_RESET; 256 qp->is_kernel = !udata; 257 258 if (!qp->is_kernel) { 259 dev_dbg(&dev->pdev->dev, 260 "create queuepair from user space\n"); 261 262 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) { 263 ret = -EFAULT; 264 goto err_qp; 265 } 266 267 /* Userspace supports qpn and qp handles? */ 268 if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION && 269 udata->outlen < sizeof(qp_resp)) { 270 dev_warn(&dev->pdev->dev, 271 "create queuepair not supported\n"); 272 ret = -EOPNOTSUPP; 273 goto err_qp; 274 } 275 276 if (!is_srq) { 277 /* set qp->sq.wqe_cnt, shift, buf_size.. */ 278 qp->rumem = 279 ib_umem_get(pd->device, ucmd.rbuf_addr, 280 ucmd.rbuf_size, 0); 281 if (IS_ERR(qp->rumem)) { 282 ret = PTR_ERR(qp->rumem); 283 goto err_qp; 284 } 285 qp->srq = NULL; 286 } else { 287 qp->rumem = NULL; 288 qp->srq = to_vsrq(init_attr->srq); 289 } 290 291 qp->sumem = ib_umem_get(pd->device, ucmd.sbuf_addr, 292 ucmd.sbuf_size, 0); 293 if (IS_ERR(qp->sumem)) { 294 if (!is_srq) 295 ib_umem_release(qp->rumem); 296 ret = PTR_ERR(qp->sumem); 297 goto err_qp; 298 } 299 300 qp->npages_send = 301 ib_umem_num_dma_blocks(qp->sumem, PAGE_SIZE); 302 if (!is_srq) 303 qp->npages_recv = ib_umem_num_dma_blocks( 304 qp->rumem, PAGE_SIZE); 305 else 306 qp->npages_recv = 0; 307 qp->npages = qp->npages_send + qp->npages_recv; 308 } else { 309 ret = pvrdma_set_sq_size(to_vdev(pd->device), 310 &init_attr->cap, qp); 311 if (ret) 312 goto err_qp; 313 314 ret = pvrdma_set_rq_size(to_vdev(pd->device), 315 &init_attr->cap, qp); 316 if (ret) 317 goto err_qp; 318 319 qp->npages = qp->npages_send + qp->npages_recv; 320 321 /* Skip header page. */ 322 qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE; 323 324 /* Recv queue pages are after send pages. */ 325 qp->rq.offset = qp->npages_send * PAGE_SIZE; 326 } 327 328 if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) { 329 dev_warn(&dev->pdev->dev, 330 "overflow pages in queuepair\n"); 331 ret = -EINVAL; 332 goto err_umem; 333 } 334 335 ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages, 336 qp->is_kernel); 337 if (ret) { 338 dev_warn(&dev->pdev->dev, 339 "could not allocate page directory\n"); 340 goto err_umem; 341 } 342 343 if (!qp->is_kernel) { 344 pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0); 345 if (!is_srq) 346 pvrdma_page_dir_insert_umem(&qp->pdir, 347 qp->rumem, 348 qp->npages_send); 349 } else { 350 /* Ring state is always the first page. */ 351 qp->sq.ring = qp->pdir.pages[0]; 352 qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1]; 353 } 354 break; 355 default: 356 ret = -EINVAL; 357 goto err_qp; 358 } 359 360 /* Not supported */ 361 init_attr->cap.max_inline_data = 0; 362 363 memset(cmd, 0, sizeof(*cmd)); 364 cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP; 365 cmd->pd_handle = to_vpd(pd)->pd_handle; 366 cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle; 367 cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle; 368 if (is_srq) 369 cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle; 370 else 371 cmd->srq_handle = 0; 372 cmd->max_send_wr = init_attr->cap.max_send_wr; 373 cmd->max_recv_wr = init_attr->cap.max_recv_wr; 374 cmd->max_send_sge = init_attr->cap.max_send_sge; 375 cmd->max_recv_sge = init_attr->cap.max_recv_sge; 376 cmd->max_inline_data = init_attr->cap.max_inline_data; 377 cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0; 378 cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type); 379 cmd->is_srq = is_srq; 380 cmd->lkey = 0; 381 cmd->access_flags = IB_ACCESS_LOCAL_WRITE; 382 cmd->total_chunks = qp->npages; 383 cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES; 384 cmd->pdir_dma = qp->pdir.dir_dma; 385 386 dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n", 387 cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge, 388 cmd->max_recv_sge); 389 390 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP); 391 if (ret < 0) { 392 dev_warn(&dev->pdev->dev, 393 "could not create queuepair, error: %d\n", ret); 394 goto err_pdir; 395 } 396 397 /* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */ 398 qp->port = init_attr->port_num; 399 400 if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION) { 401 qp->ibqp.qp_num = resp_v2->qpn; 402 qp->qp_handle = resp_v2->qp_handle; 403 } else { 404 qp->ibqp.qp_num = resp->qpn; 405 qp->qp_handle = resp->qpn; 406 } 407 408 spin_lock_irqsave(&dev->qp_tbl_lock, flags); 409 dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp; 410 spin_unlock_irqrestore(&dev->qp_tbl_lock, flags); 411 412 if (udata) { 413 qp_resp.qpn = qp->ibqp.qp_num; 414 qp_resp.qp_handle = qp->qp_handle; 415 416 if (ib_copy_to_udata(udata, &qp_resp, 417 min(udata->outlen, sizeof(qp_resp)))) { 418 dev_warn(&dev->pdev->dev, 419 "failed to copy back udata\n"); 420 __pvrdma_destroy_qp(dev, qp); 421 return ERR_PTR(-EINVAL); 422 } 423 } 424 425 return &qp->ibqp; 426 427err_pdir: 428 pvrdma_page_dir_cleanup(dev, &qp->pdir); 429err_umem: 430 ib_umem_release(qp->rumem); 431 ib_umem_release(qp->sumem); 432err_qp: 433 kfree(qp); 434 atomic_dec(&dev->num_qps); 435 436 return ERR_PTR(ret); 437} 438 439static void _pvrdma_free_qp(struct pvrdma_qp *qp) 440{ 441 unsigned long flags; 442 struct pvrdma_dev *dev = to_vdev(qp->ibqp.device); 443 444 spin_lock_irqsave(&dev->qp_tbl_lock, flags); 445 dev->qp_tbl[qp->qp_handle] = NULL; 446 spin_unlock_irqrestore(&dev->qp_tbl_lock, flags); 447 448 if (refcount_dec_and_test(&qp->refcnt)) 449 complete(&qp->free); 450 wait_for_completion(&qp->free); 451 452 ib_umem_release(qp->rumem); 453 ib_umem_release(qp->sumem); 454 455 pvrdma_page_dir_cleanup(dev, &qp->pdir); 456 457 kfree(qp); 458 459 atomic_dec(&dev->num_qps); 460} 461 462static void pvrdma_free_qp(struct pvrdma_qp *qp) 463{ 464 struct pvrdma_cq *scq; 465 struct pvrdma_cq *rcq; 466 unsigned long scq_flags, rcq_flags; 467 468 /* In case cq is polling */ 469 get_cqs(qp, &scq, &rcq); 470 pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags); 471 472 _pvrdma_flush_cqe(qp, scq); 473 if (scq != rcq) 474 _pvrdma_flush_cqe(qp, rcq); 475 476 /* 477 * We're now unlocking the CQs before clearing out the qp handle this 478 * should still be safe. We have destroyed the backend QP and flushed 479 * the CQEs so there should be no other completions for this QP. 480 */ 481 pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags); 482 483 _pvrdma_free_qp(qp); 484} 485 486static inline void _pvrdma_destroy_qp_work(struct pvrdma_dev *dev, 487 u32 qp_handle) 488{ 489 union pvrdma_cmd_req req; 490 struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp; 491 int ret; 492 493 memset(cmd, 0, sizeof(*cmd)); 494 cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP; 495 cmd->qp_handle = qp_handle; 496 497 ret = pvrdma_cmd_post(dev, &req, NULL, 0); 498 if (ret < 0) 499 dev_warn(&dev->pdev->dev, 500 "destroy queuepair failed, error: %d\n", ret); 501} 502 503/** 504 * pvrdma_destroy_qp - destroy a queue pair 505 * @qp: the queue pair to destroy 506 * @udata: user data or null for kernel object 507 * 508 * @return: always 0. 509 */ 510int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata) 511{ 512 struct pvrdma_qp *vqp = to_vqp(qp); 513 514 _pvrdma_destroy_qp_work(to_vdev(qp->device), vqp->qp_handle); 515 pvrdma_free_qp(vqp); 516 517 return 0; 518} 519 520static void __pvrdma_destroy_qp(struct pvrdma_dev *dev, 521 struct pvrdma_qp *qp) 522{ 523 _pvrdma_destroy_qp_work(dev, qp->qp_handle); 524 _pvrdma_free_qp(qp); 525} 526 527/** 528 * pvrdma_modify_qp - modify queue pair attributes 529 * @ibqp: the queue pair 530 * @attr: the new queue pair's attributes 531 * @attr_mask: attributes mask 532 * @udata: user data 533 * 534 * @returns 0 on success, otherwise returns an errno. 535 */ 536int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 537 int attr_mask, struct ib_udata *udata) 538{ 539 struct pvrdma_dev *dev = to_vdev(ibqp->device); 540 struct pvrdma_qp *qp = to_vqp(ibqp); 541 union pvrdma_cmd_req req; 542 union pvrdma_cmd_resp rsp; 543 struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp; 544 enum ib_qp_state cur_state, next_state; 545 int ret; 546 547 /* Sanity checking. Should need lock here */ 548 mutex_lock(&qp->mutex); 549 cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state : 550 qp->state; 551 next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state; 552 553 if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type, 554 attr_mask)) { 555 ret = -EINVAL; 556 goto out; 557 } 558 559 if (attr_mask & IB_QP_PORT) { 560 if (attr->port_num == 0 || 561 attr->port_num > ibqp->device->phys_port_cnt) { 562 ret = -EINVAL; 563 goto out; 564 } 565 } 566 567 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 568 if (attr->min_rnr_timer > 31) { 569 ret = -EINVAL; 570 goto out; 571 } 572 } 573 574 if (attr_mask & IB_QP_PKEY_INDEX) { 575 if (attr->pkey_index >= dev->dsr->caps.max_pkeys) { 576 ret = -EINVAL; 577 goto out; 578 } 579 } 580 581 if (attr_mask & IB_QP_QKEY) 582 qp->qkey = attr->qkey; 583 584 if (cur_state == next_state && cur_state == IB_QPS_RESET) { 585 ret = 0; 586 goto out; 587 } 588 589 qp->state = next_state; 590 memset(cmd, 0, sizeof(*cmd)); 591 cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP; 592 cmd->qp_handle = qp->qp_handle; 593 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask); 594 cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state); 595 cmd->attrs.cur_qp_state = 596 ib_qp_state_to_pvrdma(attr->cur_qp_state); 597 cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu); 598 cmd->attrs.path_mig_state = 599 ib_mig_state_to_pvrdma(attr->path_mig_state); 600 cmd->attrs.qkey = attr->qkey; 601 cmd->attrs.rq_psn = attr->rq_psn; 602 cmd->attrs.sq_psn = attr->sq_psn; 603 cmd->attrs.dest_qp_num = attr->dest_qp_num; 604 cmd->attrs.qp_access_flags = 605 ib_access_flags_to_pvrdma(attr->qp_access_flags); 606 cmd->attrs.pkey_index = attr->pkey_index; 607 cmd->attrs.alt_pkey_index = attr->alt_pkey_index; 608 cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify; 609 cmd->attrs.sq_draining = attr->sq_draining; 610 cmd->attrs.max_rd_atomic = attr->max_rd_atomic; 611 cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic; 612 cmd->attrs.min_rnr_timer = attr->min_rnr_timer; 613 cmd->attrs.port_num = attr->port_num; 614 cmd->attrs.timeout = attr->timeout; 615 cmd->attrs.retry_cnt = attr->retry_cnt; 616 cmd->attrs.rnr_retry = attr->rnr_retry; 617 cmd->attrs.alt_port_num = attr->alt_port_num; 618 cmd->attrs.alt_timeout = attr->alt_timeout; 619 ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap); 620 rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr); 621 rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr); 622 623 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP); 624 if (ret < 0) { 625 dev_warn(&dev->pdev->dev, 626 "could not modify queuepair, error: %d\n", ret); 627 } else if (rsp.hdr.err > 0) { 628 dev_warn(&dev->pdev->dev, 629 "cannot modify queuepair, error: %d\n", rsp.hdr.err); 630 ret = -EINVAL; 631 } 632 633 if (ret == 0 && next_state == IB_QPS_RESET) 634 pvrdma_reset_qp(qp); 635 636out: 637 mutex_unlock(&qp->mutex); 638 639 return ret; 640} 641 642static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n) 643{ 644 return pvrdma_page_dir_get_ptr(&qp->pdir, 645 qp->sq.offset + n * qp->sq.wqe_size); 646} 647 648static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n) 649{ 650 return pvrdma_page_dir_get_ptr(&qp->pdir, 651 qp->rq.offset + n * qp->rq.wqe_size); 652} 653 654static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, 655 const struct ib_reg_wr *wr) 656{ 657 struct pvrdma_user_mr *mr = to_vmr(wr->mr); 658 659 wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova; 660 wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma; 661 wqe_hdr->wr.fast_reg.page_shift = mr->page_shift; 662 wqe_hdr->wr.fast_reg.page_list_len = mr->npages; 663 wqe_hdr->wr.fast_reg.length = mr->ibmr.length; 664 wqe_hdr->wr.fast_reg.access_flags = wr->access; 665 wqe_hdr->wr.fast_reg.rkey = wr->key; 666 667 return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages, 668 mr->npages); 669} 670 671/** 672 * pvrdma_post_send - post send work request entries on a QP 673 * @ibqp: the QP 674 * @wr: work request list to post 675 * @bad_wr: the first bad WR returned 676 * 677 * @return: 0 on success, otherwise errno returned. 678 */ 679int pvrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr, 680 const struct ib_send_wr **bad_wr) 681{ 682 struct pvrdma_qp *qp = to_vqp(ibqp); 683 struct pvrdma_dev *dev = to_vdev(ibqp->device); 684 unsigned long flags; 685 struct pvrdma_sq_wqe_hdr *wqe_hdr; 686 struct pvrdma_sge *sge; 687 int i, ret; 688 689 /* 690 * In states lower than RTS, we can fail immediately. In other states, 691 * just post and let the device figure it out. 692 */ 693 if (qp->state < IB_QPS_RTS) { 694 *bad_wr = wr; 695 return -EINVAL; 696 } 697 698 spin_lock_irqsave(&qp->sq.lock, flags); 699 700 while (wr) { 701 unsigned int tail = 0; 702 703 if (unlikely(!pvrdma_idx_ring_has_space( 704 qp->sq.ring, qp->sq.wqe_cnt, &tail))) { 705 dev_warn_ratelimited(&dev->pdev->dev, 706 "send queue is full\n"); 707 *bad_wr = wr; 708 ret = -ENOMEM; 709 goto out; 710 } 711 712 if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) { 713 dev_warn_ratelimited(&dev->pdev->dev, 714 "send SGE overflow\n"); 715 *bad_wr = wr; 716 ret = -EINVAL; 717 goto out; 718 } 719 720 if (unlikely(wr->opcode < 0)) { 721 dev_warn_ratelimited(&dev->pdev->dev, 722 "invalid send opcode\n"); 723 *bad_wr = wr; 724 ret = -EINVAL; 725 goto out; 726 } 727 728 /* 729 * Only support UD, RC. 730 * Need to check opcode table for thorough checking. 731 * opcode _UD _UC _RC 732 * _SEND x x x 733 * _SEND_WITH_IMM x x x 734 * _RDMA_WRITE x x 735 * _RDMA_WRITE_WITH_IMM x x 736 * _LOCAL_INV x x 737 * _SEND_WITH_INV x x 738 * _RDMA_READ x 739 * _ATOMIC_CMP_AND_SWP x 740 * _ATOMIC_FETCH_AND_ADD x 741 * _MASK_ATOMIC_CMP_AND_SWP x 742 * _MASK_ATOMIC_FETCH_AND_ADD x 743 * _REG_MR x 744 * 745 */ 746 if (qp->ibqp.qp_type != IB_QPT_UD && 747 qp->ibqp.qp_type != IB_QPT_RC && 748 wr->opcode != IB_WR_SEND) { 749 dev_warn_ratelimited(&dev->pdev->dev, 750 "unsupported queuepair type\n"); 751 *bad_wr = wr; 752 ret = -EINVAL; 753 goto out; 754 } else if (qp->ibqp.qp_type == IB_QPT_UD || 755 qp->ibqp.qp_type == IB_QPT_GSI) { 756 if (wr->opcode != IB_WR_SEND && 757 wr->opcode != IB_WR_SEND_WITH_IMM) { 758 dev_warn_ratelimited(&dev->pdev->dev, 759 "invalid send opcode\n"); 760 *bad_wr = wr; 761 ret = -EINVAL; 762 goto out; 763 } 764 } 765 766 wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail); 767 memset(wqe_hdr, 0, sizeof(*wqe_hdr)); 768 wqe_hdr->wr_id = wr->wr_id; 769 wqe_hdr->num_sge = wr->num_sge; 770 wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode); 771 wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags); 772 if (wr->opcode == IB_WR_SEND_WITH_IMM || 773 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) 774 wqe_hdr->ex.imm_data = wr->ex.imm_data; 775 776 if (unlikely(wqe_hdr->opcode == PVRDMA_WR_ERROR)) { 777 *bad_wr = wr; 778 ret = -EINVAL; 779 goto out; 780 } 781 782 switch (qp->ibqp.qp_type) { 783 case IB_QPT_GSI: 784 case IB_QPT_UD: 785 if (unlikely(!ud_wr(wr)->ah)) { 786 dev_warn_ratelimited(&dev->pdev->dev, 787 "invalid address handle\n"); 788 *bad_wr = wr; 789 ret = -EINVAL; 790 goto out; 791 } 792 793 /* 794 * Use qkey from qp context if high order bit set, 795 * otherwise from work request. 796 */ 797 wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn; 798 wqe_hdr->wr.ud.remote_qkey = 799 ud_wr(wr)->remote_qkey & 0x80000000 ? 800 qp->qkey : ud_wr(wr)->remote_qkey; 801 wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av; 802 803 break; 804 case IB_QPT_RC: 805 switch (wr->opcode) { 806 case IB_WR_RDMA_READ: 807 case IB_WR_RDMA_WRITE: 808 case IB_WR_RDMA_WRITE_WITH_IMM: 809 wqe_hdr->wr.rdma.remote_addr = 810 rdma_wr(wr)->remote_addr; 811 wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey; 812 break; 813 case IB_WR_LOCAL_INV: 814 case IB_WR_SEND_WITH_INV: 815 wqe_hdr->ex.invalidate_rkey = 816 wr->ex.invalidate_rkey; 817 break; 818 case IB_WR_ATOMIC_CMP_AND_SWP: 819 case IB_WR_ATOMIC_FETCH_AND_ADD: 820 wqe_hdr->wr.atomic.remote_addr = 821 atomic_wr(wr)->remote_addr; 822 wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey; 823 wqe_hdr->wr.atomic.compare_add = 824 atomic_wr(wr)->compare_add; 825 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) 826 wqe_hdr->wr.atomic.swap = 827 atomic_wr(wr)->swap; 828 break; 829 case IB_WR_REG_MR: 830 ret = set_reg_seg(wqe_hdr, reg_wr(wr)); 831 if (ret < 0) { 832 dev_warn_ratelimited(&dev->pdev->dev, 833 "Failed to set fast register work request\n"); 834 *bad_wr = wr; 835 goto out; 836 } 837 break; 838 default: 839 break; 840 } 841 842 break; 843 default: 844 dev_warn_ratelimited(&dev->pdev->dev, 845 "invalid queuepair type\n"); 846 ret = -EINVAL; 847 *bad_wr = wr; 848 goto out; 849 } 850 851 sge = (struct pvrdma_sge *)(wqe_hdr + 1); 852 for (i = 0; i < wr->num_sge; i++) { 853 /* Need to check wqe_size 0 or max size */ 854 sge->addr = wr->sg_list[i].addr; 855 sge->length = wr->sg_list[i].length; 856 sge->lkey = wr->sg_list[i].lkey; 857 sge++; 858 } 859 860 /* Make sure wqe is written before index update */ 861 smp_wmb(); 862 863 /* Update shared sq ring */ 864 pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail, 865 qp->sq.wqe_cnt); 866 867 wr = wr->next; 868 } 869 870 ret = 0; 871 872out: 873 spin_unlock_irqrestore(&qp->sq.lock, flags); 874 875 if (!ret) 876 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle); 877 878 return ret; 879} 880 881/** 882 * pvrdma_post_receive - post receive work request entries on a QP 883 * @ibqp: the QP 884 * @wr: the work request list to post 885 * @bad_wr: the first bad WR returned 886 * 887 * @return: 0 on success, otherwise errno returned. 888 */ 889int pvrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr, 890 const struct ib_recv_wr **bad_wr) 891{ 892 struct pvrdma_dev *dev = to_vdev(ibqp->device); 893 unsigned long flags; 894 struct pvrdma_qp *qp = to_vqp(ibqp); 895 struct pvrdma_rq_wqe_hdr *wqe_hdr; 896 struct pvrdma_sge *sge; 897 int ret = 0; 898 int i; 899 900 /* 901 * In the RESET state, we can fail immediately. For other states, 902 * just post and let the device figure it out. 903 */ 904 if (qp->state == IB_QPS_RESET) { 905 *bad_wr = wr; 906 return -EINVAL; 907 } 908 909 if (qp->srq) { 910 dev_warn(&dev->pdev->dev, "QP associated with SRQ\n"); 911 *bad_wr = wr; 912 return -EINVAL; 913 } 914 915 spin_lock_irqsave(&qp->rq.lock, flags); 916 917 while (wr) { 918 unsigned int tail = 0; 919 920 if (unlikely(wr->num_sge > qp->rq.max_sg || 921 wr->num_sge < 0)) { 922 ret = -EINVAL; 923 *bad_wr = wr; 924 dev_warn_ratelimited(&dev->pdev->dev, 925 "recv SGE overflow\n"); 926 goto out; 927 } 928 929 if (unlikely(!pvrdma_idx_ring_has_space( 930 qp->rq.ring, qp->rq.wqe_cnt, &tail))) { 931 ret = -ENOMEM; 932 *bad_wr = wr; 933 dev_warn_ratelimited(&dev->pdev->dev, 934 "recv queue full\n"); 935 goto out; 936 } 937 938 wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail); 939 wqe_hdr->wr_id = wr->wr_id; 940 wqe_hdr->num_sge = wr->num_sge; 941 wqe_hdr->total_len = 0; 942 943 sge = (struct pvrdma_sge *)(wqe_hdr + 1); 944 for (i = 0; i < wr->num_sge; i++) { 945 sge->addr = wr->sg_list[i].addr; 946 sge->length = wr->sg_list[i].length; 947 sge->lkey = wr->sg_list[i].lkey; 948 sge++; 949 } 950 951 /* Make sure wqe is written before index update */ 952 smp_wmb(); 953 954 /* Update shared rq ring */ 955 pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail, 956 qp->rq.wqe_cnt); 957 958 wr = wr->next; 959 } 960 961 spin_unlock_irqrestore(&qp->rq.lock, flags); 962 963 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle); 964 965 return ret; 966 967out: 968 spin_unlock_irqrestore(&qp->rq.lock, flags); 969 970 return ret; 971} 972 973/** 974 * pvrdma_query_qp - query a queue pair's attributes 975 * @ibqp: the queue pair to query 976 * @attr: the queue pair's attributes 977 * @attr_mask: attributes mask 978 * @init_attr: initial queue pair attributes 979 * 980 * @returns 0 on success, otherwise returns an errno. 981 */ 982int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 983 int attr_mask, struct ib_qp_init_attr *init_attr) 984{ 985 struct pvrdma_dev *dev = to_vdev(ibqp->device); 986 struct pvrdma_qp *qp = to_vqp(ibqp); 987 union pvrdma_cmd_req req; 988 union pvrdma_cmd_resp rsp; 989 struct pvrdma_cmd_query_qp *cmd = &req.query_qp; 990 struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp; 991 int ret = 0; 992 993 mutex_lock(&qp->mutex); 994 995 if (qp->state == IB_QPS_RESET) { 996 attr->qp_state = IB_QPS_RESET; 997 goto out; 998 } 999 1000 memset(cmd, 0, sizeof(*cmd)); 1001 cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP; 1002 cmd->qp_handle = qp->qp_handle; 1003 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask); 1004 1005 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP); 1006 if (ret < 0) { 1007 dev_warn(&dev->pdev->dev, 1008 "could not query queuepair, error: %d\n", ret); 1009 goto out; 1010 } 1011 1012 attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state); 1013 attr->cur_qp_state = 1014 pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state); 1015 attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu); 1016 attr->path_mig_state = 1017 pvrdma_mig_state_to_ib(resp->attrs.path_mig_state); 1018 attr->qkey = resp->attrs.qkey; 1019 attr->rq_psn = resp->attrs.rq_psn; 1020 attr->sq_psn = resp->attrs.sq_psn; 1021 attr->dest_qp_num = resp->attrs.dest_qp_num; 1022 attr->qp_access_flags = 1023 pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags); 1024 attr->pkey_index = resp->attrs.pkey_index; 1025 attr->alt_pkey_index = resp->attrs.alt_pkey_index; 1026 attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify; 1027 attr->sq_draining = resp->attrs.sq_draining; 1028 attr->max_rd_atomic = resp->attrs.max_rd_atomic; 1029 attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic; 1030 attr->min_rnr_timer = resp->attrs.min_rnr_timer; 1031 attr->port_num = resp->attrs.port_num; 1032 attr->timeout = resp->attrs.timeout; 1033 attr->retry_cnt = resp->attrs.retry_cnt; 1034 attr->rnr_retry = resp->attrs.rnr_retry; 1035 attr->alt_port_num = resp->attrs.alt_port_num; 1036 attr->alt_timeout = resp->attrs.alt_timeout; 1037 pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap); 1038 pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr); 1039 pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr); 1040 1041 qp->state = attr->qp_state; 1042 1043 ret = 0; 1044 1045out: 1046 attr->cur_qp_state = attr->qp_state; 1047 1048 init_attr->event_handler = qp->ibqp.event_handler; 1049 init_attr->qp_context = qp->ibqp.qp_context; 1050 init_attr->send_cq = qp->ibqp.send_cq; 1051 init_attr->recv_cq = qp->ibqp.recv_cq; 1052 init_attr->srq = qp->ibqp.srq; 1053 init_attr->xrcd = NULL; 1054 init_attr->cap = attr->cap; 1055 init_attr->sq_sig_type = 0; 1056 init_attr->qp_type = qp->ibqp.qp_type; 1057 init_attr->create_flags = 0; 1058 init_attr->port_num = qp->port; 1059 1060 mutex_unlock(&qp->mutex); 1061 return ret; 1062} 1063