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 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 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 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 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 } 126 127 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 } 152 153 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 */ 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 } 384 385 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 pvrdma_page_dir_cleanup(dev, &qp->pdir); 410 411 kfree(qp); 412 413 atomic_dec(&dev->num_qps); 414 } 415 416 /** 417 * pvrdma_destroy_qp - destroy a queue pair 418 * @qp: the queue pair to destroy 419 * 420 * @return: 0 on success. 421 */ 422 int pvrdma_destroy_qp(struct ib_qp *qp) 423 { 424 struct pvrdma_qp *vqp = to_vqp(qp); 425 union pvrdma_cmd_req req; 426 struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp; 427 int ret; 428 429 memset(cmd, 0, sizeof(*cmd)); 430 cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP; 431 cmd->qp_handle = vqp->qp_handle; 432 433 ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0); 434 if (ret < 0) 435 dev_warn(&to_vdev(qp->device)->pdev->dev, 436 "destroy queuepair failed, error: %d\n", ret); 437 438 pvrdma_free_qp(vqp); 439 440 return 0; 441 } 442 443 /** 444 * pvrdma_modify_qp - modify queue pair attributes 445 * @ibqp: the queue pair 446 * @attr: the new queue pair's attributes 447 * @attr_mask: attributes mask 448 * @udata: user data 449 * 450 * @returns 0 on success, otherwise returns an errno. 451 */ 452 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 453 int attr_mask, struct ib_udata *udata) 454 { 455 struct pvrdma_dev *dev = to_vdev(ibqp->device); 456 struct pvrdma_qp *qp = to_vqp(ibqp); 457 union pvrdma_cmd_req req; 458 union pvrdma_cmd_resp rsp; 459 struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp; 460 int cur_state, next_state; 461 int ret; 462 463 /* Sanity checking. Should need lock here */ 464 mutex_lock(&qp->mutex); 465 cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state : 466 qp->state; 467 next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state; 468 469 if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type, 470 attr_mask, IB_LINK_LAYER_ETHERNET)) { 471 ret = -EINVAL; 472 goto out; 473 } 474 475 if (attr_mask & IB_QP_PORT) { 476 if (attr->port_num == 0 || 477 attr->port_num > ibqp->device->phys_port_cnt) { 478 ret = -EINVAL; 479 goto out; 480 } 481 } 482 483 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 484 if (attr->min_rnr_timer > 31) { 485 ret = -EINVAL; 486 goto out; 487 } 488 } 489 490 if (attr_mask & IB_QP_PKEY_INDEX) { 491 if (attr->pkey_index >= dev->dsr->caps.max_pkeys) { 492 ret = -EINVAL; 493 goto out; 494 } 495 } 496 497 if (attr_mask & IB_QP_QKEY) 498 qp->qkey = attr->qkey; 499 500 if (cur_state == next_state && cur_state == IB_QPS_RESET) { 501 ret = 0; 502 goto out; 503 } 504 505 qp->state = next_state; 506 memset(cmd, 0, sizeof(*cmd)); 507 cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP; 508 cmd->qp_handle = qp->qp_handle; 509 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask); 510 cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state); 511 cmd->attrs.cur_qp_state = 512 ib_qp_state_to_pvrdma(attr->cur_qp_state); 513 cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu); 514 cmd->attrs.path_mig_state = 515 ib_mig_state_to_pvrdma(attr->path_mig_state); 516 cmd->attrs.qkey = attr->qkey; 517 cmd->attrs.rq_psn = attr->rq_psn; 518 cmd->attrs.sq_psn = attr->sq_psn; 519 cmd->attrs.dest_qp_num = attr->dest_qp_num; 520 cmd->attrs.qp_access_flags = 521 ib_access_flags_to_pvrdma(attr->qp_access_flags); 522 cmd->attrs.pkey_index = attr->pkey_index; 523 cmd->attrs.alt_pkey_index = attr->alt_pkey_index; 524 cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify; 525 cmd->attrs.sq_draining = attr->sq_draining; 526 cmd->attrs.max_rd_atomic = attr->max_rd_atomic; 527 cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic; 528 cmd->attrs.min_rnr_timer = attr->min_rnr_timer; 529 cmd->attrs.port_num = attr->port_num; 530 cmd->attrs.timeout = attr->timeout; 531 cmd->attrs.retry_cnt = attr->retry_cnt; 532 cmd->attrs.rnr_retry = attr->rnr_retry; 533 cmd->attrs.alt_port_num = attr->alt_port_num; 534 cmd->attrs.alt_timeout = attr->alt_timeout; 535 ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap); 536 rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr); 537 rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr); 538 539 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP); 540 if (ret < 0) { 541 dev_warn(&dev->pdev->dev, 542 "could not modify queuepair, error: %d\n", ret); 543 } else if (rsp.hdr.err > 0) { 544 dev_warn(&dev->pdev->dev, 545 "cannot modify queuepair, error: %d\n", rsp.hdr.err); 546 ret = -EINVAL; 547 } 548 549 if (ret == 0 && next_state == IB_QPS_RESET) 550 pvrdma_reset_qp(qp); 551 552 out: 553 mutex_unlock(&qp->mutex); 554 555 return ret; 556 } 557 558 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n) 559 { 560 return pvrdma_page_dir_get_ptr(&qp->pdir, 561 qp->sq.offset + n * qp->sq.wqe_size); 562 } 563 564 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n) 565 { 566 return pvrdma_page_dir_get_ptr(&qp->pdir, 567 qp->rq.offset + n * qp->rq.wqe_size); 568 } 569 570 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_wr *wr) 571 { 572 struct pvrdma_user_mr *mr = to_vmr(wr->mr); 573 574 wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova; 575 wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma; 576 wqe_hdr->wr.fast_reg.page_shift = mr->page_shift; 577 wqe_hdr->wr.fast_reg.page_list_len = mr->npages; 578 wqe_hdr->wr.fast_reg.length = mr->ibmr.length; 579 wqe_hdr->wr.fast_reg.access_flags = wr->access; 580 wqe_hdr->wr.fast_reg.rkey = wr->key; 581 582 return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages, 583 mr->npages); 584 } 585 586 /** 587 * pvrdma_post_send - post send work request entries on a QP 588 * @ibqp: the QP 589 * @wr: work request list to post 590 * @bad_wr: the first bad WR returned 591 * 592 * @return: 0 on success, otherwise errno returned. 593 */ 594 int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 595 struct ib_send_wr **bad_wr) 596 { 597 struct pvrdma_qp *qp = to_vqp(ibqp); 598 struct pvrdma_dev *dev = to_vdev(ibqp->device); 599 unsigned long flags; 600 struct pvrdma_sq_wqe_hdr *wqe_hdr; 601 struct pvrdma_sge *sge; 602 int i, ret; 603 604 /* 605 * In states lower than RTS, we can fail immediately. In other states, 606 * just post and let the device figure it out. 607 */ 608 if (qp->state < IB_QPS_RTS) { 609 *bad_wr = wr; 610 return -EINVAL; 611 } 612 613 spin_lock_irqsave(&qp->sq.lock, flags); 614 615 while (wr) { 616 unsigned int tail = 0; 617 618 if (unlikely(!pvrdma_idx_ring_has_space( 619 qp->sq.ring, qp->sq.wqe_cnt, &tail))) { 620 dev_warn_ratelimited(&dev->pdev->dev, 621 "send queue is full\n"); 622 *bad_wr = wr; 623 ret = -ENOMEM; 624 goto out; 625 } 626 627 if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) { 628 dev_warn_ratelimited(&dev->pdev->dev, 629 "send SGE overflow\n"); 630 *bad_wr = wr; 631 ret = -EINVAL; 632 goto out; 633 } 634 635 if (unlikely(wr->opcode < 0)) { 636 dev_warn_ratelimited(&dev->pdev->dev, 637 "invalid send opcode\n"); 638 *bad_wr = wr; 639 ret = -EINVAL; 640 goto out; 641 } 642 643 /* 644 * Only support UD, RC. 645 * Need to check opcode table for thorough checking. 646 * opcode _UD _UC _RC 647 * _SEND x x x 648 * _SEND_WITH_IMM x x x 649 * _RDMA_WRITE x x 650 * _RDMA_WRITE_WITH_IMM x x 651 * _LOCAL_INV x x 652 * _SEND_WITH_INV x x 653 * _RDMA_READ x 654 * _ATOMIC_CMP_AND_SWP x 655 * _ATOMIC_FETCH_AND_ADD x 656 * _MASK_ATOMIC_CMP_AND_SWP x 657 * _MASK_ATOMIC_FETCH_AND_ADD x 658 * _REG_MR x 659 * 660 */ 661 if (qp->ibqp.qp_type != IB_QPT_UD && 662 qp->ibqp.qp_type != IB_QPT_RC && 663 wr->opcode != IB_WR_SEND) { 664 dev_warn_ratelimited(&dev->pdev->dev, 665 "unsupported queuepair type\n"); 666 *bad_wr = wr; 667 ret = -EINVAL; 668 goto out; 669 } else if (qp->ibqp.qp_type == IB_QPT_UD || 670 qp->ibqp.qp_type == IB_QPT_GSI) { 671 if (wr->opcode != IB_WR_SEND && 672 wr->opcode != IB_WR_SEND_WITH_IMM) { 673 dev_warn_ratelimited(&dev->pdev->dev, 674 "invalid send opcode\n"); 675 *bad_wr = wr; 676 ret = -EINVAL; 677 goto out; 678 } 679 } 680 681 wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail); 682 memset(wqe_hdr, 0, sizeof(*wqe_hdr)); 683 wqe_hdr->wr_id = wr->wr_id; 684 wqe_hdr->num_sge = wr->num_sge; 685 wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode); 686 wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags); 687 if (wr->opcode == IB_WR_SEND_WITH_IMM || 688 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) 689 wqe_hdr->ex.imm_data = wr->ex.imm_data; 690 691 switch (qp->ibqp.qp_type) { 692 case IB_QPT_GSI: 693 case IB_QPT_UD: 694 if (unlikely(!ud_wr(wr)->ah)) { 695 dev_warn_ratelimited(&dev->pdev->dev, 696 "invalid address handle\n"); 697 *bad_wr = wr; 698 ret = -EINVAL; 699 goto out; 700 } 701 702 /* 703 * Use qkey from qp context if high order bit set, 704 * otherwise from work request. 705 */ 706 wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn; 707 wqe_hdr->wr.ud.remote_qkey = 708 ud_wr(wr)->remote_qkey & 0x80000000 ? 709 qp->qkey : ud_wr(wr)->remote_qkey; 710 wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av; 711 712 break; 713 case IB_QPT_RC: 714 switch (wr->opcode) { 715 case IB_WR_RDMA_READ: 716 case IB_WR_RDMA_WRITE: 717 case IB_WR_RDMA_WRITE_WITH_IMM: 718 wqe_hdr->wr.rdma.remote_addr = 719 rdma_wr(wr)->remote_addr; 720 wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey; 721 break; 722 case IB_WR_LOCAL_INV: 723 case IB_WR_SEND_WITH_INV: 724 wqe_hdr->ex.invalidate_rkey = 725 wr->ex.invalidate_rkey; 726 break; 727 case IB_WR_ATOMIC_CMP_AND_SWP: 728 case IB_WR_ATOMIC_FETCH_AND_ADD: 729 wqe_hdr->wr.atomic.remote_addr = 730 atomic_wr(wr)->remote_addr; 731 wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey; 732 wqe_hdr->wr.atomic.compare_add = 733 atomic_wr(wr)->compare_add; 734 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) 735 wqe_hdr->wr.atomic.swap = 736 atomic_wr(wr)->swap; 737 break; 738 case IB_WR_REG_MR: 739 ret = set_reg_seg(wqe_hdr, reg_wr(wr)); 740 if (ret < 0) { 741 dev_warn_ratelimited(&dev->pdev->dev, 742 "Failed to set fast register work request\n"); 743 *bad_wr = wr; 744 goto out; 745 } 746 break; 747 default: 748 break; 749 } 750 751 break; 752 default: 753 dev_warn_ratelimited(&dev->pdev->dev, 754 "invalid queuepair type\n"); 755 ret = -EINVAL; 756 *bad_wr = wr; 757 goto out; 758 } 759 760 sge = (struct pvrdma_sge *)(wqe_hdr + 1); 761 for (i = 0; i < wr->num_sge; i++) { 762 /* Need to check wqe_size 0 or max size */ 763 sge->addr = wr->sg_list[i].addr; 764 sge->length = wr->sg_list[i].length; 765 sge->lkey = wr->sg_list[i].lkey; 766 sge++; 767 } 768 769 /* Make sure wqe is written before index update */ 770 smp_wmb(); 771 772 /* Update shared sq ring */ 773 pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail, 774 qp->sq.wqe_cnt); 775 776 wr = wr->next; 777 } 778 779 ret = 0; 780 781 out: 782 spin_unlock_irqrestore(&qp->sq.lock, flags); 783 784 if (!ret) 785 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle); 786 787 return ret; 788 } 789 790 /** 791 * pvrdma_post_receive - post receive work request entries on a QP 792 * @ibqp: the QP 793 * @wr: the work request list to post 794 * @bad_wr: the first bad WR returned 795 * 796 * @return: 0 on success, otherwise errno returned. 797 */ 798 int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 799 struct ib_recv_wr **bad_wr) 800 { 801 struct pvrdma_dev *dev = to_vdev(ibqp->device); 802 unsigned long flags; 803 struct pvrdma_qp *qp = to_vqp(ibqp); 804 struct pvrdma_rq_wqe_hdr *wqe_hdr; 805 struct pvrdma_sge *sge; 806 int ret = 0; 807 int i; 808 809 /* 810 * In the RESET state, we can fail immediately. For other states, 811 * just post and let the device figure it out. 812 */ 813 if (qp->state == IB_QPS_RESET) { 814 *bad_wr = wr; 815 return -EINVAL; 816 } 817 818 spin_lock_irqsave(&qp->rq.lock, flags); 819 820 while (wr) { 821 unsigned int tail = 0; 822 823 if (unlikely(wr->num_sge > qp->rq.max_sg || 824 wr->num_sge < 0)) { 825 ret = -EINVAL; 826 *bad_wr = wr; 827 dev_warn_ratelimited(&dev->pdev->dev, 828 "recv SGE overflow\n"); 829 goto out; 830 } 831 832 if (unlikely(!pvrdma_idx_ring_has_space( 833 qp->rq.ring, qp->rq.wqe_cnt, &tail))) { 834 ret = -ENOMEM; 835 *bad_wr = wr; 836 dev_warn_ratelimited(&dev->pdev->dev, 837 "recv queue full\n"); 838 goto out; 839 } 840 841 wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail); 842 wqe_hdr->wr_id = wr->wr_id; 843 wqe_hdr->num_sge = wr->num_sge; 844 wqe_hdr->total_len = 0; 845 846 sge = (struct pvrdma_sge *)(wqe_hdr + 1); 847 for (i = 0; i < wr->num_sge; i++) { 848 sge->addr = wr->sg_list[i].addr; 849 sge->length = wr->sg_list[i].length; 850 sge->lkey = wr->sg_list[i].lkey; 851 sge++; 852 } 853 854 /* Make sure wqe is written before index update */ 855 smp_wmb(); 856 857 /* Update shared rq ring */ 858 pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail, 859 qp->rq.wqe_cnt); 860 861 wr = wr->next; 862 } 863 864 spin_unlock_irqrestore(&qp->rq.lock, flags); 865 866 pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle); 867 868 return ret; 869 870 out: 871 spin_unlock_irqrestore(&qp->rq.lock, flags); 872 873 return ret; 874 } 875 876 /** 877 * pvrdma_query_qp - query a queue pair's attributes 878 * @ibqp: the queue pair to query 879 * @attr: the queue pair's attributes 880 * @attr_mask: attributes mask 881 * @init_attr: initial queue pair attributes 882 * 883 * @returns 0 on success, otherwise returns an errno. 884 */ 885 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 886 int attr_mask, struct ib_qp_init_attr *init_attr) 887 { 888 struct pvrdma_dev *dev = to_vdev(ibqp->device); 889 struct pvrdma_qp *qp = to_vqp(ibqp); 890 union pvrdma_cmd_req req; 891 union pvrdma_cmd_resp rsp; 892 struct pvrdma_cmd_query_qp *cmd = &req.query_qp; 893 struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp; 894 int ret = 0; 895 896 mutex_lock(&qp->mutex); 897 898 if (qp->state == IB_QPS_RESET) { 899 attr->qp_state = IB_QPS_RESET; 900 goto out; 901 } 902 903 memset(cmd, 0, sizeof(*cmd)); 904 cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP; 905 cmd->qp_handle = qp->qp_handle; 906 cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask); 907 908 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP); 909 if (ret < 0) { 910 dev_warn(&dev->pdev->dev, 911 "could not query queuepair, error: %d\n", ret); 912 goto out; 913 } 914 915 attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state); 916 attr->cur_qp_state = 917 pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state); 918 attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu); 919 attr->path_mig_state = 920 pvrdma_mig_state_to_ib(resp->attrs.path_mig_state); 921 attr->qkey = resp->attrs.qkey; 922 attr->rq_psn = resp->attrs.rq_psn; 923 attr->sq_psn = resp->attrs.sq_psn; 924 attr->dest_qp_num = resp->attrs.dest_qp_num; 925 attr->qp_access_flags = 926 pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags); 927 attr->pkey_index = resp->attrs.pkey_index; 928 attr->alt_pkey_index = resp->attrs.alt_pkey_index; 929 attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify; 930 attr->sq_draining = resp->attrs.sq_draining; 931 attr->max_rd_atomic = resp->attrs.max_rd_atomic; 932 attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic; 933 attr->min_rnr_timer = resp->attrs.min_rnr_timer; 934 attr->port_num = resp->attrs.port_num; 935 attr->timeout = resp->attrs.timeout; 936 attr->retry_cnt = resp->attrs.retry_cnt; 937 attr->rnr_retry = resp->attrs.rnr_retry; 938 attr->alt_port_num = resp->attrs.alt_port_num; 939 attr->alt_timeout = resp->attrs.alt_timeout; 940 pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap); 941 pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr); 942 pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr); 943 944 qp->state = attr->qp_state; 945 946 ret = 0; 947 948 out: 949 attr->cur_qp_state = attr->qp_state; 950 951 init_attr->event_handler = qp->ibqp.event_handler; 952 init_attr->qp_context = qp->ibqp.qp_context; 953 init_attr->send_cq = qp->ibqp.send_cq; 954 init_attr->recv_cq = qp->ibqp.recv_cq; 955 init_attr->srq = qp->ibqp.srq; 956 init_attr->xrcd = NULL; 957 init_attr->cap = attr->cap; 958 init_attr->sq_sig_type = 0; 959 init_attr->qp_type = qp->ibqp.qp_type; 960 init_attr->create_flags = 0; 961 init_attr->port_num = qp->port; 962 963 mutex_unlock(&qp->mutex); 964 return ret; 965 } 966