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; 764 765 while (wr) { 766 mask = wr_opcode_mask(wr->opcode, qp); 767 if (unlikely(!mask)) { 768 err = -EINVAL; 769 *bad_wr = wr; 770 break; 771 } 772 773 if (unlikely((wr->send_flags & IB_SEND_INLINE) && 774 !(mask & WR_INLINE_MASK))) { 775 err = -EINVAL; 776 *bad_wr = wr; 777 break; 778 } 779 780 length = 0; 781 for (i = 0; i < wr->num_sge; i++) 782 length += wr->sg_list[i].length; 783 784 err = post_one_send(qp, wr, mask, length); 785 786 if (err) { 787 *bad_wr = wr; 788 break; 789 } 790 wr = wr->next; 791 } 792 793 rxe_run_task(&qp->req.task, 1); 794 if (unlikely(qp->req.state == QP_STATE_ERROR)) 795 rxe_run_task(&qp->comp.task, 1); 796 797 return err; 798 } 799 800 static int rxe_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 801 struct ib_send_wr **bad_wr) 802 { 803 struct rxe_qp *qp = to_rqp(ibqp); 804 805 if (unlikely(!qp->valid)) { 806 *bad_wr = wr; 807 return -EINVAL; 808 } 809 810 if (unlikely(qp->req.state < QP_STATE_READY)) { 811 *bad_wr = wr; 812 return -EINVAL; 813 } 814 815 if (qp->is_user) { 816 /* Utilize process context to do protocol processing */ 817 rxe_run_task(&qp->req.task, 0); 818 return 0; 819 } else 820 return rxe_post_send_kernel(qp, wr, bad_wr); 821 } 822 823 static int rxe_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 824 struct ib_recv_wr **bad_wr) 825 { 826 int err = 0; 827 struct rxe_qp *qp = to_rqp(ibqp); 828 struct rxe_rq *rq = &qp->rq; 829 unsigned long flags; 830 831 if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) { 832 *bad_wr = wr; 833 err = -EINVAL; 834 goto err1; 835 } 836 837 if (unlikely(qp->srq)) { 838 *bad_wr = wr; 839 err = -EINVAL; 840 goto err1; 841 } 842 843 spin_lock_irqsave(&rq->producer_lock, flags); 844 845 while (wr) { 846 err = post_one_recv(rq, wr); 847 if (unlikely(err)) { 848 *bad_wr = wr; 849 break; 850 } 851 wr = wr->next; 852 } 853 854 spin_unlock_irqrestore(&rq->producer_lock, flags); 855 856 if (qp->resp.state == QP_STATE_ERROR) 857 rxe_run_task(&qp->resp.task, 1); 858 859 err1: 860 return err; 861 } 862 863 static struct ib_cq *rxe_create_cq(struct ib_device *dev, 864 const struct ib_cq_init_attr *attr, 865 struct ib_ucontext *context, 866 struct ib_udata *udata) 867 { 868 int err; 869 struct rxe_dev *rxe = to_rdev(dev); 870 struct rxe_cq *cq; 871 struct rxe_create_cq_resp __user *uresp = NULL; 872 873 if (udata) { 874 if (udata->outlen < sizeof(*uresp)) 875 return ERR_PTR(-EINVAL); 876 uresp = udata->outbuf; 877 } 878 879 if (attr->flags) 880 return ERR_PTR(-EINVAL); 881 882 err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector); 883 if (err) 884 goto err1; 885 886 cq = rxe_alloc(&rxe->cq_pool); 887 if (!cq) { 888 err = -ENOMEM; 889 goto err1; 890 } 891 892 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, 893 context, uresp); 894 if (err) 895 goto err2; 896 897 return &cq->ibcq; 898 899 err2: 900 rxe_drop_ref(cq); 901 err1: 902 return ERR_PTR(err); 903 } 904 905 static int rxe_destroy_cq(struct ib_cq *ibcq) 906 { 907 struct rxe_cq *cq = to_rcq(ibcq); 908 909 rxe_cq_disable(cq); 910 911 rxe_drop_ref(cq); 912 return 0; 913 } 914 915 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 916 { 917 int err; 918 struct rxe_cq *cq = to_rcq(ibcq); 919 struct rxe_dev *rxe = to_rdev(ibcq->device); 920 struct rxe_resize_cq_resp __user *uresp = NULL; 921 922 if (udata) { 923 if (udata->outlen < sizeof(*uresp)) 924 return -EINVAL; 925 uresp = udata->outbuf; 926 } 927 928 err = rxe_cq_chk_attr(rxe, cq, cqe, 0); 929 if (err) 930 goto err1; 931 932 err = rxe_cq_resize_queue(cq, cqe, uresp); 933 if (err) 934 goto err1; 935 936 return 0; 937 938 err1: 939 return err; 940 } 941 942 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 943 { 944 int i; 945 struct rxe_cq *cq = to_rcq(ibcq); 946 struct rxe_cqe *cqe; 947 unsigned long flags; 948 949 spin_lock_irqsave(&cq->cq_lock, flags); 950 for (i = 0; i < num_entries; i++) { 951 cqe = queue_head(cq->queue); 952 if (!cqe) 953 break; 954 955 memcpy(wc++, &cqe->ibwc, sizeof(*wc)); 956 advance_consumer(cq->queue); 957 } 958 spin_unlock_irqrestore(&cq->cq_lock, flags); 959 960 return i; 961 } 962 963 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt) 964 { 965 struct rxe_cq *cq = to_rcq(ibcq); 966 int count = queue_count(cq->queue); 967 968 return (count > wc_cnt) ? wc_cnt : count; 969 } 970 971 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 972 { 973 struct rxe_cq *cq = to_rcq(ibcq); 974 unsigned long irq_flags; 975 int ret = 0; 976 977 spin_lock_irqsave(&cq->cq_lock, irq_flags); 978 if (cq->notify != IB_CQ_NEXT_COMP) 979 cq->notify = flags & IB_CQ_SOLICITED_MASK; 980 981 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue)) 982 ret = 1; 983 984 spin_unlock_irqrestore(&cq->cq_lock, irq_flags); 985 986 return ret; 987 } 988 989 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access) 990 { 991 struct rxe_dev *rxe = to_rdev(ibpd->device); 992 struct rxe_pd *pd = to_rpd(ibpd); 993 struct rxe_mem *mr; 994 int err; 995 996 mr = rxe_alloc(&rxe->mr_pool); 997 if (!mr) { 998 err = -ENOMEM; 999 goto err1; 1000 } 1001 1002 rxe_add_index(mr); 1003 1004 rxe_add_ref(pd); 1005 1006 err = rxe_mem_init_dma(pd, access, mr); 1007 if (err) 1008 goto err2; 1009 1010 return &mr->ibmr; 1011 1012 err2: 1013 rxe_drop_ref(pd); 1014 rxe_drop_index(mr); 1015 rxe_drop_ref(mr); 1016 err1: 1017 return ERR_PTR(err); 1018 } 1019 1020 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd, 1021 u64 start, 1022 u64 length, 1023 u64 iova, 1024 int access, struct ib_udata *udata) 1025 { 1026 int err; 1027 struct rxe_dev *rxe = to_rdev(ibpd->device); 1028 struct rxe_pd *pd = to_rpd(ibpd); 1029 struct rxe_mem *mr; 1030 1031 mr = rxe_alloc(&rxe->mr_pool); 1032 if (!mr) { 1033 err = -ENOMEM; 1034 goto err2; 1035 } 1036 1037 rxe_add_index(mr); 1038 1039 rxe_add_ref(pd); 1040 1041 err = rxe_mem_init_user(pd, start, length, iova, 1042 access, udata, mr); 1043 if (err) 1044 goto err3; 1045 1046 return &mr->ibmr; 1047 1048 err3: 1049 rxe_drop_ref(pd); 1050 rxe_drop_index(mr); 1051 rxe_drop_ref(mr); 1052 err2: 1053 return ERR_PTR(err); 1054 } 1055 1056 static int rxe_dereg_mr(struct ib_mr *ibmr) 1057 { 1058 struct rxe_mem *mr = to_rmr(ibmr); 1059 1060 mr->state = RXE_MEM_STATE_ZOMBIE; 1061 rxe_drop_ref(mr->pd); 1062 rxe_drop_index(mr); 1063 rxe_drop_ref(mr); 1064 return 0; 1065 } 1066 1067 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, 1068 enum ib_mr_type mr_type, 1069 u32 max_num_sg) 1070 { 1071 struct rxe_dev *rxe = to_rdev(ibpd->device); 1072 struct rxe_pd *pd = to_rpd(ibpd); 1073 struct rxe_mem *mr; 1074 int err; 1075 1076 if (mr_type != IB_MR_TYPE_MEM_REG) 1077 return ERR_PTR(-EINVAL); 1078 1079 mr = rxe_alloc(&rxe->mr_pool); 1080 if (!mr) { 1081 err = -ENOMEM; 1082 goto err1; 1083 } 1084 1085 rxe_add_index(mr); 1086 1087 rxe_add_ref(pd); 1088 1089 err = rxe_mem_init_fast(pd, max_num_sg, mr); 1090 if (err) 1091 goto err2; 1092 1093 return &mr->ibmr; 1094 1095 err2: 1096 rxe_drop_ref(pd); 1097 rxe_drop_index(mr); 1098 rxe_drop_ref(mr); 1099 err1: 1100 return ERR_PTR(err); 1101 } 1102 1103 static int rxe_set_page(struct ib_mr *ibmr, u64 addr) 1104 { 1105 struct rxe_mem *mr = to_rmr(ibmr); 1106 struct rxe_map *map; 1107 struct rxe_phys_buf *buf; 1108 1109 if (unlikely(mr->nbuf == mr->num_buf)) 1110 return -ENOMEM; 1111 1112 map = mr->map[mr->nbuf / RXE_BUF_PER_MAP]; 1113 buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP]; 1114 1115 buf->addr = addr; 1116 buf->size = ibmr->page_size; 1117 mr->nbuf++; 1118 1119 return 0; 1120 } 1121 1122 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, 1123 int sg_nents, unsigned int *sg_offset) 1124 { 1125 struct rxe_mem *mr = to_rmr(ibmr); 1126 int n; 1127 1128 mr->nbuf = 0; 1129 1130 n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page); 1131 1132 mr->va = ibmr->iova; 1133 mr->iova = ibmr->iova; 1134 mr->length = ibmr->length; 1135 mr->page_shift = ilog2(ibmr->page_size); 1136 mr->page_mask = ibmr->page_size - 1; 1137 mr->offset = mr->iova & mr->page_mask; 1138 1139 return n; 1140 } 1141 1142 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1143 { 1144 int err; 1145 struct rxe_dev *rxe = to_rdev(ibqp->device); 1146 struct rxe_qp *qp = to_rqp(ibqp); 1147 struct rxe_mc_grp *grp; 1148 1149 /* takes a ref on grp if successful */ 1150 err = rxe_mcast_get_grp(rxe, mgid, &grp); 1151 if (err) 1152 return err; 1153 1154 err = rxe_mcast_add_grp_elem(rxe, qp, grp); 1155 1156 rxe_drop_ref(grp); 1157 return err; 1158 } 1159 1160 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1161 { 1162 struct rxe_dev *rxe = to_rdev(ibqp->device); 1163 struct rxe_qp *qp = to_rqp(ibqp); 1164 1165 return rxe_mcast_drop_grp_elem(rxe, qp, mgid); 1166 } 1167 1168 static ssize_t parent_show(struct device *device, 1169 struct device_attribute *attr, char *buf) 1170 { 1171 struct rxe_dev *rxe = container_of(device, struct rxe_dev, 1172 ib_dev.dev); 1173 1174 return snprintf(buf, 16, "%s\n", rxe_parent_name(rxe, 1)); 1175 } 1176 1177 static DEVICE_ATTR_RO(parent); 1178 1179 static struct device_attribute *rxe_dev_attributes[] = { 1180 &dev_attr_parent, 1181 }; 1182 1183 int rxe_register_device(struct rxe_dev *rxe) 1184 { 1185 int err; 1186 int i; 1187 struct ib_device *dev = &rxe->ib_dev; 1188 struct crypto_shash *tfm; 1189 1190 strlcpy(dev->name, "rxe%d", IB_DEVICE_NAME_MAX); 1191 strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc)); 1192 1193 dev->owner = THIS_MODULE; 1194 dev->node_type = RDMA_NODE_IB_CA; 1195 dev->phys_port_cnt = 1; 1196 dev->num_comp_vectors = num_possible_cpus(); 1197 dev->dev.parent = rxe_dma_device(rxe); 1198 dev->local_dma_lkey = 0; 1199 addrconf_addr_eui48((unsigned char *)&dev->node_guid, 1200 rxe->ndev->dev_addr); 1201 dev->dev.dma_ops = &dma_virt_ops; 1202 dma_coerce_mask_and_coherent(&dev->dev, 1203 dma_get_required_mask(&dev->dev)); 1204 1205 dev->uverbs_abi_ver = RXE_UVERBS_ABI_VERSION; 1206 dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) 1207 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) 1208 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) 1209 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) 1210 | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) 1211 | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) 1212 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) 1213 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) 1214 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) 1215 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) 1216 | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV) 1217 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) 1218 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) 1219 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) 1220 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) 1221 | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) 1222 | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV) 1223 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) 1224 | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ) 1225 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) 1226 | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ) 1227 | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ) 1228 | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) 1229 | BIT_ULL(IB_USER_VERBS_CMD_REG_MR) 1230 | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) 1231 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) 1232 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH) 1233 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH) 1234 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) 1235 | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) 1236 | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) 1237 ; 1238 1239 dev->query_device = rxe_query_device; 1240 dev->modify_device = rxe_modify_device; 1241 dev->query_port = rxe_query_port; 1242 dev->modify_port = rxe_modify_port; 1243 dev->get_link_layer = rxe_get_link_layer; 1244 dev->get_netdev = rxe_get_netdev; 1245 dev->query_pkey = rxe_query_pkey; 1246 dev->alloc_ucontext = rxe_alloc_ucontext; 1247 dev->dealloc_ucontext = rxe_dealloc_ucontext; 1248 dev->mmap = rxe_mmap; 1249 dev->get_port_immutable = rxe_port_immutable; 1250 dev->alloc_pd = rxe_alloc_pd; 1251 dev->dealloc_pd = rxe_dealloc_pd; 1252 dev->create_ah = rxe_create_ah; 1253 dev->modify_ah = rxe_modify_ah; 1254 dev->query_ah = rxe_query_ah; 1255 dev->destroy_ah = rxe_destroy_ah; 1256 dev->create_srq = rxe_create_srq; 1257 dev->modify_srq = rxe_modify_srq; 1258 dev->query_srq = rxe_query_srq; 1259 dev->destroy_srq = rxe_destroy_srq; 1260 dev->post_srq_recv = rxe_post_srq_recv; 1261 dev->create_qp = rxe_create_qp; 1262 dev->modify_qp = rxe_modify_qp; 1263 dev->query_qp = rxe_query_qp; 1264 dev->destroy_qp = rxe_destroy_qp; 1265 dev->post_send = rxe_post_send; 1266 dev->post_recv = rxe_post_recv; 1267 dev->create_cq = rxe_create_cq; 1268 dev->destroy_cq = rxe_destroy_cq; 1269 dev->resize_cq = rxe_resize_cq; 1270 dev->poll_cq = rxe_poll_cq; 1271 dev->peek_cq = rxe_peek_cq; 1272 dev->req_notify_cq = rxe_req_notify_cq; 1273 dev->get_dma_mr = rxe_get_dma_mr; 1274 dev->reg_user_mr = rxe_reg_user_mr; 1275 dev->dereg_mr = rxe_dereg_mr; 1276 dev->alloc_mr = rxe_alloc_mr; 1277 dev->map_mr_sg = rxe_map_mr_sg; 1278 dev->attach_mcast = rxe_attach_mcast; 1279 dev->detach_mcast = rxe_detach_mcast; 1280 dev->get_hw_stats = rxe_ib_get_hw_stats; 1281 dev->alloc_hw_stats = rxe_ib_alloc_hw_stats; 1282 1283 tfm = crypto_alloc_shash("crc32", 0, 0); 1284 if (IS_ERR(tfm)) { 1285 pr_err("failed to allocate crc algorithm err:%ld\n", 1286 PTR_ERR(tfm)); 1287 return PTR_ERR(tfm); 1288 } 1289 rxe->tfm = tfm; 1290 1291 dev->driver_id = RDMA_DRIVER_RXE; 1292 err = ib_register_device(dev, NULL); 1293 if (err) { 1294 pr_warn("%s failed with error %d\n", __func__, err); 1295 goto err1; 1296 } 1297 1298 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) { 1299 err = device_create_file(&dev->dev, rxe_dev_attributes[i]); 1300 if (err) { 1301 pr_warn("%s failed with error %d for attr number %d\n", 1302 __func__, err, i); 1303 goto err2; 1304 } 1305 } 1306 1307 return 0; 1308 1309 err2: 1310 ib_unregister_device(dev); 1311 err1: 1312 crypto_free_shash(rxe->tfm); 1313 1314 return err; 1315 } 1316 1317 int rxe_unregister_device(struct rxe_dev *rxe) 1318 { 1319 int i; 1320 struct ib_device *dev = &rxe->ib_dev; 1321 1322 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) 1323 device_remove_file(&dev->dev, rxe_dev_attributes[i]); 1324 1325 ib_unregister_device(dev); 1326 1327 return 0; 1328 } 1329