1 /* 2 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved. 3 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34 #include <linux/dma-mapping.h> 35 #include <net/addrconf.h> 36 #include "rxe.h" 37 #include "rxe_loc.h" 38 #include "rxe_queue.h" 39 #include "rxe_hw_counters.h" 40 41 static int rxe_query_device(struct ib_device *dev, 42 struct ib_device_attr *attr, 43 struct ib_udata *uhw) 44 { 45 struct rxe_dev *rxe = to_rdev(dev); 46 47 if (uhw->inlen || uhw->outlen) 48 return -EINVAL; 49 50 *attr = rxe->attr; 51 return 0; 52 } 53 54 static int rxe_query_port(struct ib_device *dev, 55 u8 port_num, struct ib_port_attr *attr) 56 { 57 struct rxe_dev *rxe = to_rdev(dev); 58 struct rxe_port *port; 59 int rc = -EINVAL; 60 61 if (unlikely(port_num != 1)) { 62 pr_warn("invalid port_number %d\n", port_num); 63 goto out; 64 } 65 66 port = &rxe->port; 67 68 /* *attr being zeroed by the caller, avoid zeroing it here */ 69 *attr = port->attr; 70 71 mutex_lock(&rxe->usdev_lock); 72 rc = ib_get_eth_speed(dev, port_num, &attr->active_speed, 73 &attr->active_width); 74 mutex_unlock(&rxe->usdev_lock); 75 76 out: 77 return rc; 78 } 79 80 static struct net_device *rxe_get_netdev(struct ib_device *device, 81 u8 port_num) 82 { 83 struct rxe_dev *rxe = to_rdev(device); 84 85 if (rxe->ndev) { 86 dev_hold(rxe->ndev); 87 return rxe->ndev; 88 } 89 90 return NULL; 91 } 92 93 static int rxe_query_pkey(struct ib_device *device, 94 u8 port_num, u16 index, u16 *pkey) 95 { 96 struct rxe_dev *rxe = to_rdev(device); 97 struct rxe_port *port; 98 99 if (unlikely(port_num != 1)) { 100 dev_warn(device->dev.parent, "invalid port_num = %d\n", 101 port_num); 102 goto err1; 103 } 104 105 port = &rxe->port; 106 107 if (unlikely(index >= port->attr.pkey_tbl_len)) { 108 dev_warn(device->dev.parent, "invalid index = %d\n", 109 index); 110 goto err1; 111 } 112 113 *pkey = port->pkey_tbl[index]; 114 return 0; 115 116 err1: 117 return -EINVAL; 118 } 119 120 static int rxe_modify_device(struct ib_device *dev, 121 int mask, struct ib_device_modify *attr) 122 { 123 struct rxe_dev *rxe = to_rdev(dev); 124 125 if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) 126 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid); 127 128 if (mask & IB_DEVICE_MODIFY_NODE_DESC) { 129 memcpy(rxe->ib_dev.node_desc, 130 attr->node_desc, sizeof(rxe->ib_dev.node_desc)); 131 } 132 133 return 0; 134 } 135 136 static int rxe_modify_port(struct ib_device *dev, 137 u8 port_num, int mask, struct ib_port_modify *attr) 138 { 139 struct rxe_dev *rxe = to_rdev(dev); 140 struct rxe_port *port; 141 142 if (unlikely(port_num != 1)) { 143 pr_warn("invalid port_num = %d\n", port_num); 144 goto err1; 145 } 146 147 port = &rxe->port; 148 149 port->attr.port_cap_flags |= attr->set_port_cap_mask; 150 port->attr.port_cap_flags &= ~attr->clr_port_cap_mask; 151 152 if (mask & IB_PORT_RESET_QKEY_CNTR) 153 port->attr.qkey_viol_cntr = 0; 154 155 return 0; 156 157 err1: 158 return -EINVAL; 159 } 160 161 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev, 162 u8 port_num) 163 { 164 struct rxe_dev *rxe = to_rdev(dev); 165 166 return rxe_link_layer(rxe, port_num); 167 } 168 169 static struct ib_ucontext *rxe_alloc_ucontext(struct ib_device *dev, 170 struct ib_udata *udata) 171 { 172 struct rxe_dev *rxe = to_rdev(dev); 173 struct rxe_ucontext *uc; 174 175 uc = rxe_alloc(&rxe->uc_pool); 176 return uc ? &uc->ibuc : ERR_PTR(-ENOMEM); 177 } 178 179 static int rxe_dealloc_ucontext(struct ib_ucontext *ibuc) 180 { 181 struct rxe_ucontext *uc = to_ruc(ibuc); 182 183 rxe_drop_ref(uc); 184 return 0; 185 } 186 187 static int rxe_port_immutable(struct ib_device *dev, u8 port_num, 188 struct ib_port_immutable *immutable) 189 { 190 int err; 191 struct ib_port_attr attr; 192 193 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 194 195 err = ib_query_port(dev, port_num, &attr); 196 if (err) 197 return err; 198 199 immutable->pkey_tbl_len = attr.pkey_tbl_len; 200 immutable->gid_tbl_len = attr.gid_tbl_len; 201 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 202 203 return 0; 204 } 205 206 static struct ib_pd *rxe_alloc_pd(struct ib_device *dev, 207 struct ib_ucontext *context, 208 struct ib_udata *udata) 209 { 210 struct rxe_dev *rxe = to_rdev(dev); 211 struct rxe_pd *pd; 212 213 pd = rxe_alloc(&rxe->pd_pool); 214 return pd ? &pd->ibpd : ERR_PTR(-ENOMEM); 215 } 216 217 static int rxe_dealloc_pd(struct ib_pd *ibpd) 218 { 219 struct rxe_pd *pd = to_rpd(ibpd); 220 221 rxe_drop_ref(pd); 222 return 0; 223 } 224 225 static int rxe_init_av(struct rxe_dev *rxe, struct rdma_ah_attr *attr, 226 struct rxe_av *av) 227 { 228 int err; 229 union ib_gid sgid; 230 struct ib_gid_attr sgid_attr; 231 232 err = ib_get_cached_gid(&rxe->ib_dev, rdma_ah_get_port_num(attr), 233 rdma_ah_read_grh(attr)->sgid_index, &sgid, 234 &sgid_attr); 235 if (err) { 236 pr_err("Failed to query sgid. err = %d\n", err); 237 return err; 238 } 239 240 rxe_av_from_attr(rdma_ah_get_port_num(attr), av, attr); 241 rxe_av_fill_ip_info(av, attr, &sgid_attr, &sgid); 242 dev_put(sgid_attr.ndev); 243 return 0; 244 } 245 246 static struct ib_ah *rxe_create_ah(struct ib_pd *ibpd, 247 struct rdma_ah_attr *attr, 248 struct ib_udata *udata) 249 250 { 251 int err; 252 struct rxe_dev *rxe = to_rdev(ibpd->device); 253 struct rxe_pd *pd = to_rpd(ibpd); 254 struct rxe_ah *ah; 255 256 err = rxe_av_chk_attr(rxe, attr); 257 if (err) 258 goto err1; 259 260 ah = rxe_alloc(&rxe->ah_pool); 261 if (!ah) { 262 err = -ENOMEM; 263 goto err1; 264 } 265 266 rxe_add_ref(pd); 267 ah->pd = pd; 268 269 err = rxe_init_av(rxe, attr, &ah->av); 270 if (err) 271 goto err2; 272 273 return &ah->ibah; 274 275 err2: 276 rxe_drop_ref(pd); 277 rxe_drop_ref(ah); 278 err1: 279 return ERR_PTR(err); 280 } 281 282 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr) 283 { 284 int err; 285 struct rxe_dev *rxe = to_rdev(ibah->device); 286 struct rxe_ah *ah = to_rah(ibah); 287 288 err = rxe_av_chk_attr(rxe, attr); 289 if (err) 290 return err; 291 292 err = rxe_init_av(rxe, attr, &ah->av); 293 if (err) 294 return err; 295 296 return 0; 297 } 298 299 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr) 300 { 301 struct rxe_ah *ah = to_rah(ibah); 302 303 memset(attr, 0, sizeof(*attr)); 304 attr->type = ibah->type; 305 rxe_av_to_attr(&ah->av, attr); 306 return 0; 307 } 308 309 static int rxe_destroy_ah(struct ib_ah *ibah) 310 { 311 struct rxe_ah *ah = to_rah(ibah); 312 313 rxe_drop_ref(ah->pd); 314 rxe_drop_ref(ah); 315 return 0; 316 } 317 318 static int post_one_recv(struct rxe_rq *rq, struct ib_recv_wr *ibwr) 319 { 320 int err; 321 int i; 322 u32 length; 323 struct rxe_recv_wqe *recv_wqe; 324 int num_sge = ibwr->num_sge; 325 326 if (unlikely(queue_full(rq->queue))) { 327 err = -ENOMEM; 328 goto err1; 329 } 330 331 if (unlikely(num_sge > rq->max_sge)) { 332 err = -EINVAL; 333 goto err1; 334 } 335 336 length = 0; 337 for (i = 0; i < num_sge; i++) 338 length += ibwr->sg_list[i].length; 339 340 recv_wqe = producer_addr(rq->queue); 341 recv_wqe->wr_id = ibwr->wr_id; 342 recv_wqe->num_sge = num_sge; 343 344 memcpy(recv_wqe->dma.sge, ibwr->sg_list, 345 num_sge * sizeof(struct ib_sge)); 346 347 recv_wqe->dma.length = length; 348 recv_wqe->dma.resid = length; 349 recv_wqe->dma.num_sge = num_sge; 350 recv_wqe->dma.cur_sge = 0; 351 recv_wqe->dma.sge_offset = 0; 352 353 /* make sure all changes to the work queue are written before we 354 * update the producer pointer 355 */ 356 smp_wmb(); 357 358 advance_producer(rq->queue); 359 return 0; 360 361 err1: 362 return err; 363 } 364 365 static struct ib_srq *rxe_create_srq(struct ib_pd *ibpd, 366 struct ib_srq_init_attr *init, 367 struct ib_udata *udata) 368 { 369 int err; 370 struct rxe_dev *rxe = to_rdev(ibpd->device); 371 struct rxe_pd *pd = to_rpd(ibpd); 372 struct rxe_srq *srq; 373 struct ib_ucontext *context = udata ? ibpd->uobject->context : NULL; 374 struct rxe_create_srq_resp __user *uresp = NULL; 375 376 if (udata) { 377 if (udata->outlen < sizeof(*uresp)) 378 return ERR_PTR(-EINVAL); 379 uresp = udata->outbuf; 380 } 381 382 err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK); 383 if (err) 384 goto err1; 385 386 srq = rxe_alloc(&rxe->srq_pool); 387 if (!srq) { 388 err = -ENOMEM; 389 goto err1; 390 } 391 392 rxe_add_index(srq); 393 rxe_add_ref(pd); 394 srq->pd = pd; 395 396 err = rxe_srq_from_init(rxe, srq, init, context, uresp); 397 if (err) 398 goto err2; 399 400 return &srq->ibsrq; 401 402 err2: 403 rxe_drop_ref(pd); 404 rxe_drop_index(srq); 405 rxe_drop_ref(srq); 406 err1: 407 return ERR_PTR(err); 408 } 409 410 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 411 enum ib_srq_attr_mask mask, 412 struct ib_udata *udata) 413 { 414 int err; 415 struct rxe_srq *srq = to_rsrq(ibsrq); 416 struct rxe_dev *rxe = to_rdev(ibsrq->device); 417 struct rxe_modify_srq_cmd ucmd = {}; 418 419 if (udata) { 420 if (udata->inlen < sizeof(ucmd)) 421 return -EINVAL; 422 423 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)); 424 if (err) 425 return err; 426 } 427 428 err = rxe_srq_chk_attr(rxe, srq, attr, mask); 429 if (err) 430 goto err1; 431 432 err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd); 433 if (err) 434 goto err1; 435 436 return 0; 437 438 err1: 439 return err; 440 } 441 442 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr) 443 { 444 struct rxe_srq *srq = to_rsrq(ibsrq); 445 446 if (srq->error) 447 return -EINVAL; 448 449 attr->max_wr = srq->rq.queue->buf->index_mask; 450 attr->max_sge = srq->rq.max_sge; 451 attr->srq_limit = srq->limit; 452 return 0; 453 } 454 455 static int rxe_destroy_srq(struct ib_srq *ibsrq) 456 { 457 struct rxe_srq *srq = to_rsrq(ibsrq); 458 459 if (srq->rq.queue) 460 rxe_queue_cleanup(srq->rq.queue); 461 462 rxe_drop_ref(srq->pd); 463 rxe_drop_index(srq); 464 rxe_drop_ref(srq); 465 466 return 0; 467 } 468 469 static int rxe_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 470 struct ib_recv_wr **bad_wr) 471 { 472 int err = 0; 473 unsigned long flags; 474 struct rxe_srq *srq = to_rsrq(ibsrq); 475 476 spin_lock_irqsave(&srq->rq.producer_lock, flags); 477 478 while (wr) { 479 err = post_one_recv(&srq->rq, wr); 480 if (unlikely(err)) 481 break; 482 wr = wr->next; 483 } 484 485 spin_unlock_irqrestore(&srq->rq.producer_lock, flags); 486 487 if (err) 488 *bad_wr = wr; 489 490 return err; 491 } 492 493 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd, 494 struct ib_qp_init_attr *init, 495 struct ib_udata *udata) 496 { 497 int err; 498 struct rxe_dev *rxe = to_rdev(ibpd->device); 499 struct rxe_pd *pd = to_rpd(ibpd); 500 struct rxe_qp *qp; 501 struct rxe_create_qp_resp __user *uresp = NULL; 502 503 if (udata) { 504 if (udata->outlen < sizeof(*uresp)) 505 return ERR_PTR(-EINVAL); 506 uresp = udata->outbuf; 507 } 508 509 err = rxe_qp_chk_init(rxe, init); 510 if (err) 511 goto err1; 512 513 qp = rxe_alloc(&rxe->qp_pool); 514 if (!qp) { 515 err = -ENOMEM; 516 goto err1; 517 } 518 519 if (udata) { 520 if (udata->inlen) { 521 err = -EINVAL; 522 goto err2; 523 } 524 qp->is_user = 1; 525 } 526 527 rxe_add_index(qp); 528 529 err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibpd); 530 if (err) 531 goto err3; 532 533 return &qp->ibqp; 534 535 err3: 536 rxe_drop_index(qp); 537 err2: 538 rxe_drop_ref(qp); 539 err1: 540 return ERR_PTR(err); 541 } 542 543 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 544 int mask, struct ib_udata *udata) 545 { 546 int err; 547 struct rxe_dev *rxe = to_rdev(ibqp->device); 548 struct rxe_qp *qp = to_rqp(ibqp); 549 550 err = rxe_qp_chk_attr(rxe, qp, attr, mask); 551 if (err) 552 goto err1; 553 554 err = rxe_qp_from_attr(qp, attr, mask, udata); 555 if (err) 556 goto err1; 557 558 return 0; 559 560 err1: 561 return err; 562 } 563 564 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 565 int mask, struct ib_qp_init_attr *init) 566 { 567 struct rxe_qp *qp = to_rqp(ibqp); 568 569 rxe_qp_to_init(qp, init); 570 rxe_qp_to_attr(qp, attr, mask); 571 572 return 0; 573 } 574 575 static int rxe_destroy_qp(struct ib_qp *ibqp) 576 { 577 struct rxe_qp *qp = to_rqp(ibqp); 578 579 rxe_qp_destroy(qp); 580 rxe_drop_index(qp); 581 rxe_drop_ref(qp); 582 return 0; 583 } 584 585 static int validate_send_wr(struct rxe_qp *qp, struct ib_send_wr *ibwr, 586 unsigned int mask, unsigned int length) 587 { 588 int num_sge = ibwr->num_sge; 589 struct rxe_sq *sq = &qp->sq; 590 591 if (unlikely(num_sge > sq->max_sge)) 592 goto err1; 593 594 if (unlikely(mask & WR_ATOMIC_MASK)) { 595 if (length < 8) 596 goto err1; 597 598 if (atomic_wr(ibwr)->remote_addr & 0x7) 599 goto err1; 600 } 601 602 if (unlikely((ibwr->send_flags & IB_SEND_INLINE) && 603 (length > sq->max_inline))) 604 goto err1; 605 606 return 0; 607 608 err1: 609 return -EINVAL; 610 } 611 612 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr, 613 struct ib_send_wr *ibwr) 614 { 615 wr->wr_id = ibwr->wr_id; 616 wr->num_sge = ibwr->num_sge; 617 wr->opcode = ibwr->opcode; 618 wr->send_flags = ibwr->send_flags; 619 620 if (qp_type(qp) == IB_QPT_UD || 621 qp_type(qp) == IB_QPT_SMI || 622 qp_type(qp) == IB_QPT_GSI) { 623 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn; 624 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey; 625 if (qp_type(qp) == IB_QPT_GSI) 626 wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index; 627 if (wr->opcode == IB_WR_SEND_WITH_IMM) 628 wr->ex.imm_data = ibwr->ex.imm_data; 629 } else { 630 switch (wr->opcode) { 631 case IB_WR_RDMA_WRITE_WITH_IMM: 632 wr->ex.imm_data = ibwr->ex.imm_data; 633 /* fall through */ 634 case IB_WR_RDMA_READ: 635 case IB_WR_RDMA_WRITE: 636 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr; 637 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey; 638 break; 639 case IB_WR_SEND_WITH_IMM: 640 wr->ex.imm_data = ibwr->ex.imm_data; 641 break; 642 case IB_WR_SEND_WITH_INV: 643 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 644 break; 645 case IB_WR_ATOMIC_CMP_AND_SWP: 646 case IB_WR_ATOMIC_FETCH_AND_ADD: 647 wr->wr.atomic.remote_addr = 648 atomic_wr(ibwr)->remote_addr; 649 wr->wr.atomic.compare_add = 650 atomic_wr(ibwr)->compare_add; 651 wr->wr.atomic.swap = atomic_wr(ibwr)->swap; 652 wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey; 653 break; 654 case IB_WR_LOCAL_INV: 655 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 656 break; 657 case IB_WR_REG_MR: 658 wr->wr.reg.mr = reg_wr(ibwr)->mr; 659 wr->wr.reg.key = reg_wr(ibwr)->key; 660 wr->wr.reg.access = reg_wr(ibwr)->access; 661 break; 662 default: 663 break; 664 } 665 } 666 } 667 668 static int init_send_wqe(struct rxe_qp *qp, struct ib_send_wr *ibwr, 669 unsigned int mask, unsigned int length, 670 struct rxe_send_wqe *wqe) 671 { 672 int num_sge = ibwr->num_sge; 673 struct ib_sge *sge; 674 int i; 675 u8 *p; 676 677 init_send_wr(qp, &wqe->wr, ibwr); 678 679 if (qp_type(qp) == IB_QPT_UD || 680 qp_type(qp) == IB_QPT_SMI || 681 qp_type(qp) == IB_QPT_GSI) 682 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av)); 683 684 if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) { 685 p = wqe->dma.inline_data; 686 687 sge = ibwr->sg_list; 688 for (i = 0; i < num_sge; i++, sge++) { 689 memcpy(p, (void *)(uintptr_t)sge->addr, 690 sge->length); 691 692 p += sge->length; 693 } 694 } else if (mask & WR_REG_MASK) { 695 wqe->mask = mask; 696 wqe->state = wqe_state_posted; 697 return 0; 698 } else 699 memcpy(wqe->dma.sge, ibwr->sg_list, 700 num_sge * sizeof(struct ib_sge)); 701 702 wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr : 703 mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0; 704 wqe->mask = mask; 705 wqe->dma.length = length; 706 wqe->dma.resid = length; 707 wqe->dma.num_sge = num_sge; 708 wqe->dma.cur_sge = 0; 709 wqe->dma.sge_offset = 0; 710 wqe->state = wqe_state_posted; 711 wqe->ssn = atomic_add_return(1, &qp->ssn); 712 713 return 0; 714 } 715 716 static int post_one_send(struct rxe_qp *qp, struct ib_send_wr *ibwr, 717 unsigned int mask, u32 length) 718 { 719 int err; 720 struct rxe_sq *sq = &qp->sq; 721 struct rxe_send_wqe *send_wqe; 722 unsigned long flags; 723 724 err = validate_send_wr(qp, ibwr, mask, length); 725 if (err) 726 return err; 727 728 spin_lock_irqsave(&qp->sq.sq_lock, flags); 729 730 if (unlikely(queue_full(sq->queue))) { 731 err = -ENOMEM; 732 goto err1; 733 } 734 735 send_wqe = producer_addr(sq->queue); 736 737 err = init_send_wqe(qp, ibwr, mask, length, send_wqe); 738 if (unlikely(err)) 739 goto err1; 740 741 /* 742 * make sure all changes to the work queue are 743 * written before we update the producer pointer 744 */ 745 smp_wmb(); 746 747 advance_producer(sq->queue); 748 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 749 750 return 0; 751 752 err1: 753 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 754 return err; 755 } 756 757 static int rxe_post_send_kernel(struct rxe_qp *qp, struct ib_send_wr *wr, 758 struct ib_send_wr **bad_wr) 759 { 760 int err = 0; 761 unsigned int mask; 762 unsigned int length = 0; 763 int i; 764 int must_sched; 765 766 while (wr) { 767 mask = wr_opcode_mask(wr->opcode, qp); 768 if (unlikely(!mask)) { 769 err = -EINVAL; 770 *bad_wr = wr; 771 break; 772 } 773 774 if (unlikely((wr->send_flags & IB_SEND_INLINE) && 775 !(mask & WR_INLINE_MASK))) { 776 err = -EINVAL; 777 *bad_wr = wr; 778 break; 779 } 780 781 length = 0; 782 for (i = 0; i < wr->num_sge; i++) 783 length += wr->sg_list[i].length; 784 785 err = post_one_send(qp, wr, mask, length); 786 787 if (err) { 788 *bad_wr = wr; 789 break; 790 } 791 wr = wr->next; 792 } 793 794 /* 795 * Must sched in case of GSI QP because ib_send_mad() hold irq lock, 796 * and the requester call ip_local_out_sk() that takes spin_lock_bh. 797 */ 798 must_sched = (qp_type(qp) == IB_QPT_GSI) || 799 (queue_count(qp->sq.queue) > 1); 800 801 rxe_run_task(&qp->req.task, must_sched); 802 if (unlikely(qp->req.state == QP_STATE_ERROR)) 803 rxe_run_task(&qp->comp.task, 1); 804 805 return err; 806 } 807 808 static int rxe_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 809 struct ib_send_wr **bad_wr) 810 { 811 struct rxe_qp *qp = to_rqp(ibqp); 812 813 if (unlikely(!qp->valid)) { 814 *bad_wr = wr; 815 return -EINVAL; 816 } 817 818 if (unlikely(qp->req.state < QP_STATE_READY)) { 819 *bad_wr = wr; 820 return -EINVAL; 821 } 822 823 if (qp->is_user) { 824 /* Utilize process context to do protocol processing */ 825 rxe_run_task(&qp->req.task, 0); 826 return 0; 827 } else 828 return rxe_post_send_kernel(qp, wr, bad_wr); 829 } 830 831 static int rxe_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 832 struct ib_recv_wr **bad_wr) 833 { 834 int err = 0; 835 struct rxe_qp *qp = to_rqp(ibqp); 836 struct rxe_rq *rq = &qp->rq; 837 unsigned long flags; 838 839 if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) { 840 *bad_wr = wr; 841 err = -EINVAL; 842 goto err1; 843 } 844 845 if (unlikely(qp->srq)) { 846 *bad_wr = wr; 847 err = -EINVAL; 848 goto err1; 849 } 850 851 spin_lock_irqsave(&rq->producer_lock, flags); 852 853 while (wr) { 854 err = post_one_recv(rq, wr); 855 if (unlikely(err)) { 856 *bad_wr = wr; 857 break; 858 } 859 wr = wr->next; 860 } 861 862 spin_unlock_irqrestore(&rq->producer_lock, flags); 863 864 if (qp->resp.state == QP_STATE_ERROR) 865 rxe_run_task(&qp->resp.task, 1); 866 867 err1: 868 return err; 869 } 870 871 static struct ib_cq *rxe_create_cq(struct ib_device *dev, 872 const struct ib_cq_init_attr *attr, 873 struct ib_ucontext *context, 874 struct ib_udata *udata) 875 { 876 int err; 877 struct rxe_dev *rxe = to_rdev(dev); 878 struct rxe_cq *cq; 879 struct rxe_create_cq_resp __user *uresp = NULL; 880 881 if (udata) { 882 if (udata->outlen < sizeof(*uresp)) 883 return ERR_PTR(-EINVAL); 884 uresp = udata->outbuf; 885 } 886 887 if (attr->flags) 888 return ERR_PTR(-EINVAL); 889 890 err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector); 891 if (err) 892 goto err1; 893 894 cq = rxe_alloc(&rxe->cq_pool); 895 if (!cq) { 896 err = -ENOMEM; 897 goto err1; 898 } 899 900 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, 901 context, uresp); 902 if (err) 903 goto err2; 904 905 return &cq->ibcq; 906 907 err2: 908 rxe_drop_ref(cq); 909 err1: 910 return ERR_PTR(err); 911 } 912 913 static int rxe_destroy_cq(struct ib_cq *ibcq) 914 { 915 struct rxe_cq *cq = to_rcq(ibcq); 916 917 rxe_cq_disable(cq); 918 919 rxe_drop_ref(cq); 920 return 0; 921 } 922 923 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 924 { 925 int err; 926 struct rxe_cq *cq = to_rcq(ibcq); 927 struct rxe_dev *rxe = to_rdev(ibcq->device); 928 struct rxe_resize_cq_resp __user *uresp = NULL; 929 930 if (udata) { 931 if (udata->outlen < sizeof(*uresp)) 932 return -EINVAL; 933 uresp = udata->outbuf; 934 } 935 936 err = rxe_cq_chk_attr(rxe, cq, cqe, 0); 937 if (err) 938 goto err1; 939 940 err = rxe_cq_resize_queue(cq, cqe, uresp); 941 if (err) 942 goto err1; 943 944 return 0; 945 946 err1: 947 return err; 948 } 949 950 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 951 { 952 int i; 953 struct rxe_cq *cq = to_rcq(ibcq); 954 struct rxe_cqe *cqe; 955 unsigned long flags; 956 957 spin_lock_irqsave(&cq->cq_lock, flags); 958 for (i = 0; i < num_entries; i++) { 959 cqe = queue_head(cq->queue); 960 if (!cqe) 961 break; 962 963 memcpy(wc++, &cqe->ibwc, sizeof(*wc)); 964 advance_consumer(cq->queue); 965 } 966 spin_unlock_irqrestore(&cq->cq_lock, flags); 967 968 return i; 969 } 970 971 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt) 972 { 973 struct rxe_cq *cq = to_rcq(ibcq); 974 int count = queue_count(cq->queue); 975 976 return (count > wc_cnt) ? wc_cnt : count; 977 } 978 979 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 980 { 981 struct rxe_cq *cq = to_rcq(ibcq); 982 unsigned long irq_flags; 983 int ret = 0; 984 985 spin_lock_irqsave(&cq->cq_lock, irq_flags); 986 if (cq->notify != IB_CQ_NEXT_COMP) 987 cq->notify = flags & IB_CQ_SOLICITED_MASK; 988 989 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue)) 990 ret = 1; 991 992 spin_unlock_irqrestore(&cq->cq_lock, irq_flags); 993 994 return ret; 995 } 996 997 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access) 998 { 999 struct rxe_dev *rxe = to_rdev(ibpd->device); 1000 struct rxe_pd *pd = to_rpd(ibpd); 1001 struct rxe_mem *mr; 1002 int err; 1003 1004 mr = rxe_alloc(&rxe->mr_pool); 1005 if (!mr) { 1006 err = -ENOMEM; 1007 goto err1; 1008 } 1009 1010 rxe_add_index(mr); 1011 1012 rxe_add_ref(pd); 1013 1014 err = rxe_mem_init_dma(rxe, pd, access, mr); 1015 if (err) 1016 goto err2; 1017 1018 return &mr->ibmr; 1019 1020 err2: 1021 rxe_drop_ref(pd); 1022 rxe_drop_index(mr); 1023 rxe_drop_ref(mr); 1024 err1: 1025 return ERR_PTR(err); 1026 } 1027 1028 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd, 1029 u64 start, 1030 u64 length, 1031 u64 iova, 1032 int access, struct ib_udata *udata) 1033 { 1034 int err; 1035 struct rxe_dev *rxe = to_rdev(ibpd->device); 1036 struct rxe_pd *pd = to_rpd(ibpd); 1037 struct rxe_mem *mr; 1038 1039 mr = rxe_alloc(&rxe->mr_pool); 1040 if (!mr) { 1041 err = -ENOMEM; 1042 goto err2; 1043 } 1044 1045 rxe_add_index(mr); 1046 1047 rxe_add_ref(pd); 1048 1049 err = rxe_mem_init_user(rxe, pd, start, length, iova, 1050 access, udata, mr); 1051 if (err) 1052 goto err3; 1053 1054 return &mr->ibmr; 1055 1056 err3: 1057 rxe_drop_ref(pd); 1058 rxe_drop_index(mr); 1059 rxe_drop_ref(mr); 1060 err2: 1061 return ERR_PTR(err); 1062 } 1063 1064 static int rxe_dereg_mr(struct ib_mr *ibmr) 1065 { 1066 struct rxe_mem *mr = to_rmr(ibmr); 1067 1068 mr->state = RXE_MEM_STATE_ZOMBIE; 1069 rxe_drop_ref(mr->pd); 1070 rxe_drop_index(mr); 1071 rxe_drop_ref(mr); 1072 return 0; 1073 } 1074 1075 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, 1076 enum ib_mr_type mr_type, 1077 u32 max_num_sg) 1078 { 1079 struct rxe_dev *rxe = to_rdev(ibpd->device); 1080 struct rxe_pd *pd = to_rpd(ibpd); 1081 struct rxe_mem *mr; 1082 int err; 1083 1084 if (mr_type != IB_MR_TYPE_MEM_REG) 1085 return ERR_PTR(-EINVAL); 1086 1087 mr = rxe_alloc(&rxe->mr_pool); 1088 if (!mr) { 1089 err = -ENOMEM; 1090 goto err1; 1091 } 1092 1093 rxe_add_index(mr); 1094 1095 rxe_add_ref(pd); 1096 1097 err = rxe_mem_init_fast(rxe, pd, max_num_sg, mr); 1098 if (err) 1099 goto err2; 1100 1101 return &mr->ibmr; 1102 1103 err2: 1104 rxe_drop_ref(pd); 1105 rxe_drop_index(mr); 1106 rxe_drop_ref(mr); 1107 err1: 1108 return ERR_PTR(err); 1109 } 1110 1111 static int rxe_set_page(struct ib_mr *ibmr, u64 addr) 1112 { 1113 struct rxe_mem *mr = to_rmr(ibmr); 1114 struct rxe_map *map; 1115 struct rxe_phys_buf *buf; 1116 1117 if (unlikely(mr->nbuf == mr->num_buf)) 1118 return -ENOMEM; 1119 1120 map = mr->map[mr->nbuf / RXE_BUF_PER_MAP]; 1121 buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP]; 1122 1123 buf->addr = addr; 1124 buf->size = ibmr->page_size; 1125 mr->nbuf++; 1126 1127 return 0; 1128 } 1129 1130 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, 1131 int sg_nents, unsigned int *sg_offset) 1132 { 1133 struct rxe_mem *mr = to_rmr(ibmr); 1134 int n; 1135 1136 mr->nbuf = 0; 1137 1138 n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page); 1139 1140 mr->va = ibmr->iova; 1141 mr->iova = ibmr->iova; 1142 mr->length = ibmr->length; 1143 mr->page_shift = ilog2(ibmr->page_size); 1144 mr->page_mask = ibmr->page_size - 1; 1145 mr->offset = mr->iova & mr->page_mask; 1146 1147 return n; 1148 } 1149 1150 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1151 { 1152 int err; 1153 struct rxe_dev *rxe = to_rdev(ibqp->device); 1154 struct rxe_qp *qp = to_rqp(ibqp); 1155 struct rxe_mc_grp *grp; 1156 1157 /* takes a ref on grp if successful */ 1158 err = rxe_mcast_get_grp(rxe, mgid, &grp); 1159 if (err) 1160 return err; 1161 1162 err = rxe_mcast_add_grp_elem(rxe, qp, grp); 1163 1164 rxe_drop_ref(grp); 1165 return err; 1166 } 1167 1168 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1169 { 1170 struct rxe_dev *rxe = to_rdev(ibqp->device); 1171 struct rxe_qp *qp = to_rqp(ibqp); 1172 1173 return rxe_mcast_drop_grp_elem(rxe, qp, mgid); 1174 } 1175 1176 static ssize_t parent_show(struct device *device, 1177 struct device_attribute *attr, char *buf) 1178 { 1179 struct rxe_dev *rxe = container_of(device, struct rxe_dev, 1180 ib_dev.dev); 1181 1182 return snprintf(buf, 16, "%s\n", rxe_parent_name(rxe, 1)); 1183 } 1184 1185 static DEVICE_ATTR_RO(parent); 1186 1187 static struct device_attribute *rxe_dev_attributes[] = { 1188 &dev_attr_parent, 1189 }; 1190 1191 int rxe_register_device(struct rxe_dev *rxe) 1192 { 1193 int err; 1194 int i; 1195 struct ib_device *dev = &rxe->ib_dev; 1196 struct crypto_shash *tfm; 1197 1198 strlcpy(dev->name, "rxe%d", IB_DEVICE_NAME_MAX); 1199 strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc)); 1200 1201 dev->owner = THIS_MODULE; 1202 dev->node_type = RDMA_NODE_IB_CA; 1203 dev->phys_port_cnt = 1; 1204 dev->num_comp_vectors = num_possible_cpus(); 1205 dev->dev.parent = rxe_dma_device(rxe); 1206 dev->local_dma_lkey = 0; 1207 addrconf_addr_eui48((unsigned char *)&dev->node_guid, 1208 rxe->ndev->dev_addr); 1209 dev->dev.dma_ops = &dma_virt_ops; 1210 dma_coerce_mask_and_coherent(&dev->dev, 1211 dma_get_required_mask(&dev->dev)); 1212 1213 dev->uverbs_abi_ver = RXE_UVERBS_ABI_VERSION; 1214 dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) 1215 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) 1216 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) 1217 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) 1218 | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) 1219 | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) 1220 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) 1221 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) 1222 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) 1223 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) 1224 | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV) 1225 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) 1226 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) 1227 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) 1228 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) 1229 | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) 1230 | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV) 1231 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) 1232 | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ) 1233 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) 1234 | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ) 1235 | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ) 1236 | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) 1237 | BIT_ULL(IB_USER_VERBS_CMD_REG_MR) 1238 | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) 1239 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) 1240 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH) 1241 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH) 1242 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) 1243 | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) 1244 | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) 1245 ; 1246 1247 dev->query_device = rxe_query_device; 1248 dev->modify_device = rxe_modify_device; 1249 dev->query_port = rxe_query_port; 1250 dev->modify_port = rxe_modify_port; 1251 dev->get_link_layer = rxe_get_link_layer; 1252 dev->get_netdev = rxe_get_netdev; 1253 dev->query_pkey = rxe_query_pkey; 1254 dev->alloc_ucontext = rxe_alloc_ucontext; 1255 dev->dealloc_ucontext = rxe_dealloc_ucontext; 1256 dev->mmap = rxe_mmap; 1257 dev->get_port_immutable = rxe_port_immutable; 1258 dev->alloc_pd = rxe_alloc_pd; 1259 dev->dealloc_pd = rxe_dealloc_pd; 1260 dev->create_ah = rxe_create_ah; 1261 dev->modify_ah = rxe_modify_ah; 1262 dev->query_ah = rxe_query_ah; 1263 dev->destroy_ah = rxe_destroy_ah; 1264 dev->create_srq = rxe_create_srq; 1265 dev->modify_srq = rxe_modify_srq; 1266 dev->query_srq = rxe_query_srq; 1267 dev->destroy_srq = rxe_destroy_srq; 1268 dev->post_srq_recv = rxe_post_srq_recv; 1269 dev->create_qp = rxe_create_qp; 1270 dev->modify_qp = rxe_modify_qp; 1271 dev->query_qp = rxe_query_qp; 1272 dev->destroy_qp = rxe_destroy_qp; 1273 dev->post_send = rxe_post_send; 1274 dev->post_recv = rxe_post_recv; 1275 dev->create_cq = rxe_create_cq; 1276 dev->destroy_cq = rxe_destroy_cq; 1277 dev->resize_cq = rxe_resize_cq; 1278 dev->poll_cq = rxe_poll_cq; 1279 dev->peek_cq = rxe_peek_cq; 1280 dev->req_notify_cq = rxe_req_notify_cq; 1281 dev->get_dma_mr = rxe_get_dma_mr; 1282 dev->reg_user_mr = rxe_reg_user_mr; 1283 dev->dereg_mr = rxe_dereg_mr; 1284 dev->alloc_mr = rxe_alloc_mr; 1285 dev->map_mr_sg = rxe_map_mr_sg; 1286 dev->attach_mcast = rxe_attach_mcast; 1287 dev->detach_mcast = rxe_detach_mcast; 1288 dev->get_hw_stats = rxe_ib_get_hw_stats; 1289 dev->alloc_hw_stats = rxe_ib_alloc_hw_stats; 1290 1291 tfm = crypto_alloc_shash("crc32", 0, 0); 1292 if (IS_ERR(tfm)) { 1293 pr_err("failed to allocate crc algorithm err:%ld\n", 1294 PTR_ERR(tfm)); 1295 return PTR_ERR(tfm); 1296 } 1297 rxe->tfm = tfm; 1298 1299 dev->driver_id = RDMA_DRIVER_RXE; 1300 err = ib_register_device(dev, NULL); 1301 if (err) { 1302 pr_warn("%s failed with error %d\n", __func__, err); 1303 goto err1; 1304 } 1305 1306 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) { 1307 err = device_create_file(&dev->dev, rxe_dev_attributes[i]); 1308 if (err) { 1309 pr_warn("%s failed with error %d for attr number %d\n", 1310 __func__, err, i); 1311 goto err2; 1312 } 1313 } 1314 1315 return 0; 1316 1317 err2: 1318 ib_unregister_device(dev); 1319 err1: 1320 crypto_free_shash(rxe->tfm); 1321 1322 return err; 1323 } 1324 1325 int rxe_unregister_device(struct rxe_dev *rxe) 1326 { 1327 int i; 1328 struct ib_device *dev = &rxe->ib_dev; 1329 1330 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) 1331 device_remove_file(&dev->dev, rxe_dev_attributes[i]); 1332 1333 ib_unregister_device(dev); 1334 1335 return 0; 1336 } 1337