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 "rxe.h" 35 #include "rxe_loc.h" 36 #include "rxe_queue.h" 37 38 static int rxe_query_device(struct ib_device *dev, 39 struct ib_device_attr *attr, 40 struct ib_udata *uhw) 41 { 42 struct rxe_dev *rxe = to_rdev(dev); 43 44 if (uhw->inlen || uhw->outlen) 45 return -EINVAL; 46 47 *attr = rxe->attr; 48 return 0; 49 } 50 51 static void rxe_eth_speed_to_ib_speed(int speed, u8 *active_speed, 52 u8 *active_width) 53 { 54 if (speed <= 1000) { 55 *active_width = IB_WIDTH_1X; 56 *active_speed = IB_SPEED_SDR; 57 } else if (speed <= 10000) { 58 *active_width = IB_WIDTH_1X; 59 *active_speed = IB_SPEED_FDR10; 60 } else if (speed <= 20000) { 61 *active_width = IB_WIDTH_4X; 62 *active_speed = IB_SPEED_DDR; 63 } else if (speed <= 30000) { 64 *active_width = IB_WIDTH_4X; 65 *active_speed = IB_SPEED_QDR; 66 } else if (speed <= 40000) { 67 *active_width = IB_WIDTH_4X; 68 *active_speed = IB_SPEED_FDR10; 69 } else { 70 *active_width = IB_WIDTH_4X; 71 *active_speed = IB_SPEED_EDR; 72 } 73 } 74 75 static int rxe_query_port(struct ib_device *dev, 76 u8 port_num, struct ib_port_attr *attr) 77 { 78 struct rxe_dev *rxe = to_rdev(dev); 79 struct rxe_port *port; 80 u32 speed; 81 82 if (unlikely(port_num != 1)) { 83 pr_warn("invalid port_number %d\n", port_num); 84 goto err1; 85 } 86 87 port = &rxe->port; 88 89 *attr = port->attr; 90 91 mutex_lock(&rxe->usdev_lock); 92 if (rxe->ndev->ethtool_ops->get_link_ksettings) { 93 struct ethtool_link_ksettings ks; 94 95 rxe->ndev->ethtool_ops->get_link_ksettings(rxe->ndev, &ks); 96 speed = ks.base.speed; 97 } else if (rxe->ndev->ethtool_ops->get_settings) { 98 struct ethtool_cmd cmd; 99 100 rxe->ndev->ethtool_ops->get_settings(rxe->ndev, &cmd); 101 speed = cmd.speed; 102 } else { 103 pr_warn("%s speed is unknown, defaulting to 1000\n", 104 rxe->ndev->name); 105 speed = 1000; 106 } 107 rxe_eth_speed_to_ib_speed(speed, &attr->active_speed, 108 &attr->active_width); 109 mutex_unlock(&rxe->usdev_lock); 110 111 return 0; 112 113 err1: 114 return -EINVAL; 115 } 116 117 static int rxe_query_gid(struct ib_device *device, 118 u8 port_num, int index, union ib_gid *gid) 119 { 120 int ret; 121 122 if (index > RXE_PORT_GID_TBL_LEN) 123 return -EINVAL; 124 125 ret = ib_get_cached_gid(device, port_num, index, gid, NULL); 126 if (ret == -EAGAIN) { 127 memcpy(gid, &zgid, sizeof(*gid)); 128 return 0; 129 } 130 131 return ret; 132 } 133 134 static int rxe_add_gid(struct ib_device *device, u8 port_num, unsigned int 135 index, const union ib_gid *gid, 136 const struct ib_gid_attr *attr, void **context) 137 { 138 if (index >= RXE_PORT_GID_TBL_LEN) 139 return -EINVAL; 140 return 0; 141 } 142 143 static int rxe_del_gid(struct ib_device *device, u8 port_num, unsigned int 144 index, void **context) 145 { 146 if (index >= RXE_PORT_GID_TBL_LEN) 147 return -EINVAL; 148 return 0; 149 } 150 151 static struct net_device *rxe_get_netdev(struct ib_device *device, 152 u8 port_num) 153 { 154 struct rxe_dev *rxe = to_rdev(device); 155 156 if (rxe->ndev) { 157 dev_hold(rxe->ndev); 158 return rxe->ndev; 159 } 160 161 return NULL; 162 } 163 164 static int rxe_query_pkey(struct ib_device *device, 165 u8 port_num, u16 index, u16 *pkey) 166 { 167 struct rxe_dev *rxe = to_rdev(device); 168 struct rxe_port *port; 169 170 if (unlikely(port_num != 1)) { 171 dev_warn(device->dma_device, "invalid port_num = %d\n", 172 port_num); 173 goto err1; 174 } 175 176 port = &rxe->port; 177 178 if (unlikely(index >= port->attr.pkey_tbl_len)) { 179 dev_warn(device->dma_device, "invalid index = %d\n", 180 index); 181 goto err1; 182 } 183 184 *pkey = port->pkey_tbl[index]; 185 return 0; 186 187 err1: 188 return -EINVAL; 189 } 190 191 static int rxe_modify_device(struct ib_device *dev, 192 int mask, struct ib_device_modify *attr) 193 { 194 struct rxe_dev *rxe = to_rdev(dev); 195 196 if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) 197 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid); 198 199 if (mask & IB_DEVICE_MODIFY_NODE_DESC) { 200 memcpy(rxe->ib_dev.node_desc, 201 attr->node_desc, sizeof(rxe->ib_dev.node_desc)); 202 } 203 204 return 0; 205 } 206 207 static int rxe_modify_port(struct ib_device *dev, 208 u8 port_num, int mask, struct ib_port_modify *attr) 209 { 210 struct rxe_dev *rxe = to_rdev(dev); 211 struct rxe_port *port; 212 213 if (unlikely(port_num != 1)) { 214 pr_warn("invalid port_num = %d\n", port_num); 215 goto err1; 216 } 217 218 port = &rxe->port; 219 220 port->attr.port_cap_flags |= attr->set_port_cap_mask; 221 port->attr.port_cap_flags &= ~attr->clr_port_cap_mask; 222 223 if (mask & IB_PORT_RESET_QKEY_CNTR) 224 port->attr.qkey_viol_cntr = 0; 225 226 return 0; 227 228 err1: 229 return -EINVAL; 230 } 231 232 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev, 233 u8 port_num) 234 { 235 struct rxe_dev *rxe = to_rdev(dev); 236 237 return rxe->ifc_ops->link_layer(rxe, port_num); 238 } 239 240 static struct ib_ucontext *rxe_alloc_ucontext(struct ib_device *dev, 241 struct ib_udata *udata) 242 { 243 struct rxe_dev *rxe = to_rdev(dev); 244 struct rxe_ucontext *uc; 245 246 uc = rxe_alloc(&rxe->uc_pool); 247 return uc ? &uc->ibuc : ERR_PTR(-ENOMEM); 248 } 249 250 static int rxe_dealloc_ucontext(struct ib_ucontext *ibuc) 251 { 252 struct rxe_ucontext *uc = to_ruc(ibuc); 253 254 rxe_drop_ref(uc); 255 return 0; 256 } 257 258 static int rxe_port_immutable(struct ib_device *dev, u8 port_num, 259 struct ib_port_immutable *immutable) 260 { 261 int err; 262 struct ib_port_attr attr; 263 264 err = rxe_query_port(dev, port_num, &attr); 265 if (err) 266 return err; 267 268 immutable->pkey_tbl_len = attr.pkey_tbl_len; 269 immutable->gid_tbl_len = attr.gid_tbl_len; 270 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 271 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 272 273 return 0; 274 } 275 276 static struct ib_pd *rxe_alloc_pd(struct ib_device *dev, 277 struct ib_ucontext *context, 278 struct ib_udata *udata) 279 { 280 struct rxe_dev *rxe = to_rdev(dev); 281 struct rxe_pd *pd; 282 283 pd = rxe_alloc(&rxe->pd_pool); 284 return pd ? &pd->ibpd : ERR_PTR(-ENOMEM); 285 } 286 287 static int rxe_dealloc_pd(struct ib_pd *ibpd) 288 { 289 struct rxe_pd *pd = to_rpd(ibpd); 290 291 rxe_drop_ref(pd); 292 return 0; 293 } 294 295 static int rxe_init_av(struct rxe_dev *rxe, struct ib_ah_attr *attr, 296 struct rxe_av *av) 297 { 298 int err; 299 union ib_gid sgid; 300 struct ib_gid_attr sgid_attr; 301 302 err = ib_get_cached_gid(&rxe->ib_dev, attr->port_num, 303 attr->grh.sgid_index, &sgid, 304 &sgid_attr); 305 if (err) { 306 pr_err("Failed to query sgid. err = %d\n", err); 307 return err; 308 } 309 310 err = rxe_av_from_attr(rxe, attr->port_num, av, attr); 311 if (!err) 312 err = rxe_av_fill_ip_info(rxe, av, attr, &sgid_attr, &sgid); 313 314 if (sgid_attr.ndev) 315 dev_put(sgid_attr.ndev); 316 return err; 317 } 318 319 static struct ib_ah *rxe_create_ah(struct ib_pd *ibpd, struct ib_ah_attr *attr) 320 { 321 int err; 322 struct rxe_dev *rxe = to_rdev(ibpd->device); 323 struct rxe_pd *pd = to_rpd(ibpd); 324 struct rxe_ah *ah; 325 326 err = rxe_av_chk_attr(rxe, attr); 327 if (err) 328 goto err1; 329 330 ah = rxe_alloc(&rxe->ah_pool); 331 if (!ah) { 332 err = -ENOMEM; 333 goto err1; 334 } 335 336 rxe_add_ref(pd); 337 ah->pd = pd; 338 339 err = rxe_init_av(rxe, attr, &ah->av); 340 if (err) 341 goto err2; 342 343 return &ah->ibah; 344 345 err2: 346 rxe_drop_ref(pd); 347 rxe_drop_ref(ah); 348 err1: 349 return ERR_PTR(err); 350 } 351 352 static int rxe_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *attr) 353 { 354 int err; 355 struct rxe_dev *rxe = to_rdev(ibah->device); 356 struct rxe_ah *ah = to_rah(ibah); 357 358 err = rxe_av_chk_attr(rxe, attr); 359 if (err) 360 return err; 361 362 err = rxe_init_av(rxe, attr, &ah->av); 363 if (err) 364 return err; 365 366 return 0; 367 } 368 369 static int rxe_query_ah(struct ib_ah *ibah, struct ib_ah_attr *attr) 370 { 371 struct rxe_dev *rxe = to_rdev(ibah->device); 372 struct rxe_ah *ah = to_rah(ibah); 373 374 rxe_av_to_attr(rxe, &ah->av, attr); 375 return 0; 376 } 377 378 static int rxe_destroy_ah(struct ib_ah *ibah) 379 { 380 struct rxe_ah *ah = to_rah(ibah); 381 382 rxe_drop_ref(ah->pd); 383 rxe_drop_ref(ah); 384 return 0; 385 } 386 387 static int post_one_recv(struct rxe_rq *rq, struct ib_recv_wr *ibwr) 388 { 389 int err; 390 int i; 391 u32 length; 392 struct rxe_recv_wqe *recv_wqe; 393 int num_sge = ibwr->num_sge; 394 395 if (unlikely(queue_full(rq->queue))) { 396 err = -ENOMEM; 397 goto err1; 398 } 399 400 if (unlikely(num_sge > rq->max_sge)) { 401 err = -EINVAL; 402 goto err1; 403 } 404 405 length = 0; 406 for (i = 0; i < num_sge; i++) 407 length += ibwr->sg_list[i].length; 408 409 recv_wqe = producer_addr(rq->queue); 410 recv_wqe->wr_id = ibwr->wr_id; 411 recv_wqe->num_sge = num_sge; 412 413 memcpy(recv_wqe->dma.sge, ibwr->sg_list, 414 num_sge * sizeof(struct ib_sge)); 415 416 recv_wqe->dma.length = length; 417 recv_wqe->dma.resid = length; 418 recv_wqe->dma.num_sge = num_sge; 419 recv_wqe->dma.cur_sge = 0; 420 recv_wqe->dma.sge_offset = 0; 421 422 /* make sure all changes to the work queue are written before we 423 * update the producer pointer 424 */ 425 smp_wmb(); 426 427 advance_producer(rq->queue); 428 return 0; 429 430 err1: 431 return err; 432 } 433 434 static struct ib_srq *rxe_create_srq(struct ib_pd *ibpd, 435 struct ib_srq_init_attr *init, 436 struct ib_udata *udata) 437 { 438 int err; 439 struct rxe_dev *rxe = to_rdev(ibpd->device); 440 struct rxe_pd *pd = to_rpd(ibpd); 441 struct rxe_srq *srq; 442 struct ib_ucontext *context = udata ? ibpd->uobject->context : NULL; 443 444 err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK); 445 if (err) 446 goto err1; 447 448 srq = rxe_alloc(&rxe->srq_pool); 449 if (!srq) { 450 err = -ENOMEM; 451 goto err1; 452 } 453 454 rxe_add_index(srq); 455 rxe_add_ref(pd); 456 srq->pd = pd; 457 458 err = rxe_srq_from_init(rxe, srq, init, context, udata); 459 if (err) 460 goto err2; 461 462 return &srq->ibsrq; 463 464 err2: 465 rxe_drop_ref(pd); 466 rxe_drop_index(srq); 467 rxe_drop_ref(srq); 468 err1: 469 return ERR_PTR(err); 470 } 471 472 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 473 enum ib_srq_attr_mask mask, 474 struct ib_udata *udata) 475 { 476 int err; 477 struct rxe_srq *srq = to_rsrq(ibsrq); 478 struct rxe_dev *rxe = to_rdev(ibsrq->device); 479 480 err = rxe_srq_chk_attr(rxe, srq, attr, mask); 481 if (err) 482 goto err1; 483 484 err = rxe_srq_from_attr(rxe, srq, attr, mask, udata); 485 if (err) 486 goto err1; 487 488 return 0; 489 490 err1: 491 return err; 492 } 493 494 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr) 495 { 496 struct rxe_srq *srq = to_rsrq(ibsrq); 497 498 if (srq->error) 499 return -EINVAL; 500 501 attr->max_wr = srq->rq.queue->buf->index_mask; 502 attr->max_sge = srq->rq.max_sge; 503 attr->srq_limit = srq->limit; 504 return 0; 505 } 506 507 static int rxe_destroy_srq(struct ib_srq *ibsrq) 508 { 509 struct rxe_srq *srq = to_rsrq(ibsrq); 510 511 if (srq->rq.queue) 512 rxe_queue_cleanup(srq->rq.queue); 513 514 rxe_drop_ref(srq->pd); 515 rxe_drop_index(srq); 516 rxe_drop_ref(srq); 517 518 return 0; 519 } 520 521 static int rxe_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 522 struct ib_recv_wr **bad_wr) 523 { 524 int err = 0; 525 unsigned long flags; 526 struct rxe_srq *srq = to_rsrq(ibsrq); 527 528 spin_lock_irqsave(&srq->rq.producer_lock, flags); 529 530 while (wr) { 531 err = post_one_recv(&srq->rq, wr); 532 if (unlikely(err)) 533 break; 534 wr = wr->next; 535 } 536 537 spin_unlock_irqrestore(&srq->rq.producer_lock, flags); 538 539 if (err) 540 *bad_wr = wr; 541 542 return err; 543 } 544 545 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd, 546 struct ib_qp_init_attr *init, 547 struct ib_udata *udata) 548 { 549 int err; 550 struct rxe_dev *rxe = to_rdev(ibpd->device); 551 struct rxe_pd *pd = to_rpd(ibpd); 552 struct rxe_qp *qp; 553 554 err = rxe_qp_chk_init(rxe, init); 555 if (err) 556 goto err1; 557 558 qp = rxe_alloc(&rxe->qp_pool); 559 if (!qp) { 560 err = -ENOMEM; 561 goto err1; 562 } 563 564 if (udata) { 565 if (udata->inlen) { 566 err = -EINVAL; 567 goto err1; 568 } 569 qp->is_user = 1; 570 } 571 572 rxe_add_index(qp); 573 574 err = rxe_qp_from_init(rxe, qp, pd, init, udata, ibpd); 575 if (err) 576 goto err2; 577 578 return &qp->ibqp; 579 580 err2: 581 rxe_drop_index(qp); 582 rxe_drop_ref(qp); 583 err1: 584 return ERR_PTR(err); 585 } 586 587 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 588 int mask, struct ib_udata *udata) 589 { 590 int err; 591 struct rxe_dev *rxe = to_rdev(ibqp->device); 592 struct rxe_qp *qp = to_rqp(ibqp); 593 594 err = rxe_qp_chk_attr(rxe, qp, attr, mask); 595 if (err) 596 goto err1; 597 598 err = rxe_qp_from_attr(qp, attr, mask, udata); 599 if (err) 600 goto err1; 601 602 return 0; 603 604 err1: 605 return err; 606 } 607 608 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 609 int mask, struct ib_qp_init_attr *init) 610 { 611 struct rxe_qp *qp = to_rqp(ibqp); 612 613 rxe_qp_to_init(qp, init); 614 rxe_qp_to_attr(qp, attr, mask); 615 616 return 0; 617 } 618 619 static int rxe_destroy_qp(struct ib_qp *ibqp) 620 { 621 struct rxe_qp *qp = to_rqp(ibqp); 622 623 rxe_qp_destroy(qp); 624 rxe_drop_index(qp); 625 rxe_drop_ref(qp); 626 return 0; 627 } 628 629 static int validate_send_wr(struct rxe_qp *qp, struct ib_send_wr *ibwr, 630 unsigned int mask, unsigned int length) 631 { 632 int num_sge = ibwr->num_sge; 633 struct rxe_sq *sq = &qp->sq; 634 635 if (unlikely(num_sge > sq->max_sge)) 636 goto err1; 637 638 if (unlikely(mask & WR_ATOMIC_MASK)) { 639 if (length < 8) 640 goto err1; 641 642 if (atomic_wr(ibwr)->remote_addr & 0x7) 643 goto err1; 644 } 645 646 if (unlikely((ibwr->send_flags & IB_SEND_INLINE) && 647 (length > sq->max_inline))) 648 goto err1; 649 650 return 0; 651 652 err1: 653 return -EINVAL; 654 } 655 656 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr, 657 struct ib_send_wr *ibwr) 658 { 659 wr->wr_id = ibwr->wr_id; 660 wr->num_sge = ibwr->num_sge; 661 wr->opcode = ibwr->opcode; 662 wr->send_flags = ibwr->send_flags; 663 664 if (qp_type(qp) == IB_QPT_UD || 665 qp_type(qp) == IB_QPT_SMI || 666 qp_type(qp) == IB_QPT_GSI) { 667 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn; 668 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey; 669 if (qp_type(qp) == IB_QPT_GSI) 670 wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index; 671 if (wr->opcode == IB_WR_SEND_WITH_IMM) 672 wr->ex.imm_data = ibwr->ex.imm_data; 673 } else { 674 switch (wr->opcode) { 675 case IB_WR_RDMA_WRITE_WITH_IMM: 676 wr->ex.imm_data = ibwr->ex.imm_data; 677 case IB_WR_RDMA_READ: 678 case IB_WR_RDMA_WRITE: 679 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr; 680 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey; 681 break; 682 case IB_WR_SEND_WITH_IMM: 683 wr->ex.imm_data = ibwr->ex.imm_data; 684 break; 685 case IB_WR_SEND_WITH_INV: 686 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 687 break; 688 case IB_WR_ATOMIC_CMP_AND_SWP: 689 case IB_WR_ATOMIC_FETCH_AND_ADD: 690 wr->wr.atomic.remote_addr = 691 atomic_wr(ibwr)->remote_addr; 692 wr->wr.atomic.compare_add = 693 atomic_wr(ibwr)->compare_add; 694 wr->wr.atomic.swap = atomic_wr(ibwr)->swap; 695 wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey; 696 break; 697 case IB_WR_LOCAL_INV: 698 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 699 break; 700 case IB_WR_REG_MR: 701 wr->wr.reg.mr = reg_wr(ibwr)->mr; 702 wr->wr.reg.key = reg_wr(ibwr)->key; 703 wr->wr.reg.access = reg_wr(ibwr)->access; 704 break; 705 default: 706 break; 707 } 708 } 709 } 710 711 static int init_send_wqe(struct rxe_qp *qp, struct ib_send_wr *ibwr, 712 unsigned int mask, unsigned int length, 713 struct rxe_send_wqe *wqe) 714 { 715 int num_sge = ibwr->num_sge; 716 struct ib_sge *sge; 717 int i; 718 u8 *p; 719 720 init_send_wr(qp, &wqe->wr, ibwr); 721 722 if (qp_type(qp) == IB_QPT_UD || 723 qp_type(qp) == IB_QPT_SMI || 724 qp_type(qp) == IB_QPT_GSI) 725 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av)); 726 727 if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) { 728 p = wqe->dma.inline_data; 729 730 sge = ibwr->sg_list; 731 for (i = 0; i < num_sge; i++, sge++) { 732 if (qp->is_user && copy_from_user(p, (__user void *) 733 (uintptr_t)sge->addr, sge->length)) 734 return -EFAULT; 735 736 else if (!qp->is_user) 737 memcpy(p, (void *)(uintptr_t)sge->addr, 738 sge->length); 739 740 p += sge->length; 741 } 742 } else if (mask & WR_REG_MASK) { 743 wqe->mask = mask; 744 wqe->state = wqe_state_posted; 745 return 0; 746 } else 747 memcpy(wqe->dma.sge, ibwr->sg_list, 748 num_sge * sizeof(struct ib_sge)); 749 750 wqe->iova = (mask & WR_ATOMIC_MASK) ? 751 atomic_wr(ibwr)->remote_addr : 752 rdma_wr(ibwr)->remote_addr; 753 wqe->mask = mask; 754 wqe->dma.length = length; 755 wqe->dma.resid = length; 756 wqe->dma.num_sge = num_sge; 757 wqe->dma.cur_sge = 0; 758 wqe->dma.sge_offset = 0; 759 wqe->state = wqe_state_posted; 760 wqe->ssn = atomic_add_return(1, &qp->ssn); 761 762 return 0; 763 } 764 765 static int post_one_send(struct rxe_qp *qp, struct ib_send_wr *ibwr, 766 unsigned int mask, u32 length) 767 { 768 int err; 769 struct rxe_sq *sq = &qp->sq; 770 struct rxe_send_wqe *send_wqe; 771 unsigned long flags; 772 773 err = validate_send_wr(qp, ibwr, mask, length); 774 if (err) 775 return err; 776 777 spin_lock_irqsave(&qp->sq.sq_lock, flags); 778 779 if (unlikely(queue_full(sq->queue))) { 780 err = -ENOMEM; 781 goto err1; 782 } 783 784 send_wqe = producer_addr(sq->queue); 785 786 err = init_send_wqe(qp, ibwr, mask, length, send_wqe); 787 if (unlikely(err)) 788 goto err1; 789 790 /* 791 * make sure all changes to the work queue are 792 * written before we update the producer pointer 793 */ 794 smp_wmb(); 795 796 advance_producer(sq->queue); 797 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 798 799 return 0; 800 801 err1: 802 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 803 return err; 804 } 805 806 static int rxe_post_send_kernel(struct rxe_qp *qp, struct ib_send_wr *wr, 807 struct ib_send_wr **bad_wr) 808 { 809 int err = 0; 810 unsigned int mask; 811 unsigned int length = 0; 812 int i; 813 int must_sched; 814 815 while (wr) { 816 mask = wr_opcode_mask(wr->opcode, qp); 817 if (unlikely(!mask)) { 818 err = -EINVAL; 819 *bad_wr = wr; 820 break; 821 } 822 823 if (unlikely((wr->send_flags & IB_SEND_INLINE) && 824 !(mask & WR_INLINE_MASK))) { 825 err = -EINVAL; 826 *bad_wr = wr; 827 break; 828 } 829 830 length = 0; 831 for (i = 0; i < wr->num_sge; i++) 832 length += wr->sg_list[i].length; 833 834 err = post_one_send(qp, wr, mask, length); 835 836 if (err) { 837 *bad_wr = wr; 838 break; 839 } 840 wr = wr->next; 841 } 842 843 /* 844 * Must sched in case of GSI QP because ib_send_mad() hold irq lock, 845 * and the requester call ip_local_out_sk() that takes spin_lock_bh. 846 */ 847 must_sched = (qp_type(qp) == IB_QPT_GSI) || 848 (queue_count(qp->sq.queue) > 1); 849 850 rxe_run_task(&qp->req.task, must_sched); 851 852 return err; 853 } 854 855 static int rxe_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 856 struct ib_send_wr **bad_wr) 857 { 858 struct rxe_qp *qp = to_rqp(ibqp); 859 860 if (unlikely(!qp->valid)) { 861 *bad_wr = wr; 862 return -EINVAL; 863 } 864 865 if (unlikely(qp->req.state < QP_STATE_READY)) { 866 *bad_wr = wr; 867 return -EINVAL; 868 } 869 870 if (qp->is_user) { 871 /* Utilize process context to do protocol processing */ 872 rxe_run_task(&qp->req.task, 0); 873 return 0; 874 } else 875 return rxe_post_send_kernel(qp, wr, bad_wr); 876 } 877 878 static int rxe_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 879 struct ib_recv_wr **bad_wr) 880 { 881 int err = 0; 882 struct rxe_qp *qp = to_rqp(ibqp); 883 struct rxe_rq *rq = &qp->rq; 884 unsigned long flags; 885 886 if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) { 887 *bad_wr = wr; 888 err = -EINVAL; 889 goto err1; 890 } 891 892 if (unlikely(qp->srq)) { 893 *bad_wr = wr; 894 err = -EINVAL; 895 goto err1; 896 } 897 898 spin_lock_irqsave(&rq->producer_lock, flags); 899 900 while (wr) { 901 err = post_one_recv(rq, wr); 902 if (unlikely(err)) { 903 *bad_wr = wr; 904 break; 905 } 906 wr = wr->next; 907 } 908 909 spin_unlock_irqrestore(&rq->producer_lock, flags); 910 911 err1: 912 return err; 913 } 914 915 static struct ib_cq *rxe_create_cq(struct ib_device *dev, 916 const struct ib_cq_init_attr *attr, 917 struct ib_ucontext *context, 918 struct ib_udata *udata) 919 { 920 int err; 921 struct rxe_dev *rxe = to_rdev(dev); 922 struct rxe_cq *cq; 923 924 if (attr->flags) 925 return ERR_PTR(-EINVAL); 926 927 err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector, udata); 928 if (err) 929 goto err1; 930 931 cq = rxe_alloc(&rxe->cq_pool); 932 if (!cq) { 933 err = -ENOMEM; 934 goto err1; 935 } 936 937 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, 938 context, udata); 939 if (err) 940 goto err2; 941 942 return &cq->ibcq; 943 944 err2: 945 rxe_drop_ref(cq); 946 err1: 947 return ERR_PTR(err); 948 } 949 950 static int rxe_destroy_cq(struct ib_cq *ibcq) 951 { 952 struct rxe_cq *cq = to_rcq(ibcq); 953 954 rxe_drop_ref(cq); 955 return 0; 956 } 957 958 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 959 { 960 int err; 961 struct rxe_cq *cq = to_rcq(ibcq); 962 struct rxe_dev *rxe = to_rdev(ibcq->device); 963 964 err = rxe_cq_chk_attr(rxe, cq, cqe, 0, udata); 965 if (err) 966 goto err1; 967 968 err = rxe_cq_resize_queue(cq, cqe, udata); 969 if (err) 970 goto err1; 971 972 return 0; 973 974 err1: 975 return err; 976 } 977 978 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 979 { 980 int i; 981 struct rxe_cq *cq = to_rcq(ibcq); 982 struct rxe_cqe *cqe; 983 unsigned long flags; 984 985 spin_lock_irqsave(&cq->cq_lock, flags); 986 for (i = 0; i < num_entries; i++) { 987 cqe = queue_head(cq->queue); 988 if (!cqe) 989 break; 990 991 memcpy(wc++, &cqe->ibwc, sizeof(*wc)); 992 advance_consumer(cq->queue); 993 } 994 spin_unlock_irqrestore(&cq->cq_lock, flags); 995 996 return i; 997 } 998 999 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt) 1000 { 1001 struct rxe_cq *cq = to_rcq(ibcq); 1002 int count = queue_count(cq->queue); 1003 1004 return (count > wc_cnt) ? wc_cnt : count; 1005 } 1006 1007 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 1008 { 1009 struct rxe_cq *cq = to_rcq(ibcq); 1010 1011 if (cq->notify != IB_CQ_NEXT_COMP) 1012 cq->notify = flags & IB_CQ_SOLICITED_MASK; 1013 1014 return 0; 1015 } 1016 1017 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access) 1018 { 1019 struct rxe_dev *rxe = to_rdev(ibpd->device); 1020 struct rxe_pd *pd = to_rpd(ibpd); 1021 struct rxe_mem *mr; 1022 int err; 1023 1024 mr = rxe_alloc(&rxe->mr_pool); 1025 if (!mr) { 1026 err = -ENOMEM; 1027 goto err1; 1028 } 1029 1030 rxe_add_index(mr); 1031 1032 rxe_add_ref(pd); 1033 1034 err = rxe_mem_init_dma(rxe, pd, access, mr); 1035 if (err) 1036 goto err2; 1037 1038 return &mr->ibmr; 1039 1040 err2: 1041 rxe_drop_ref(pd); 1042 rxe_drop_index(mr); 1043 rxe_drop_ref(mr); 1044 err1: 1045 return ERR_PTR(err); 1046 } 1047 1048 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd, 1049 u64 start, 1050 u64 length, 1051 u64 iova, 1052 int access, struct ib_udata *udata) 1053 { 1054 int err; 1055 struct rxe_dev *rxe = to_rdev(ibpd->device); 1056 struct rxe_pd *pd = to_rpd(ibpd); 1057 struct rxe_mem *mr; 1058 1059 mr = rxe_alloc(&rxe->mr_pool); 1060 if (!mr) { 1061 err = -ENOMEM; 1062 goto err2; 1063 } 1064 1065 rxe_add_index(mr); 1066 1067 rxe_add_ref(pd); 1068 1069 err = rxe_mem_init_user(rxe, pd, start, length, iova, 1070 access, udata, mr); 1071 if (err) 1072 goto err3; 1073 1074 return &mr->ibmr; 1075 1076 err3: 1077 rxe_drop_ref(pd); 1078 rxe_drop_index(mr); 1079 rxe_drop_ref(mr); 1080 err2: 1081 return ERR_PTR(err); 1082 } 1083 1084 static int rxe_dereg_mr(struct ib_mr *ibmr) 1085 { 1086 struct rxe_mem *mr = to_rmr(ibmr); 1087 1088 mr->state = RXE_MEM_STATE_ZOMBIE; 1089 rxe_drop_ref(mr->pd); 1090 rxe_drop_index(mr); 1091 rxe_drop_ref(mr); 1092 return 0; 1093 } 1094 1095 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, 1096 enum ib_mr_type mr_type, 1097 u32 max_num_sg) 1098 { 1099 struct rxe_dev *rxe = to_rdev(ibpd->device); 1100 struct rxe_pd *pd = to_rpd(ibpd); 1101 struct rxe_mem *mr; 1102 int err; 1103 1104 if (mr_type != IB_MR_TYPE_MEM_REG) 1105 return ERR_PTR(-EINVAL); 1106 1107 mr = rxe_alloc(&rxe->mr_pool); 1108 if (!mr) { 1109 err = -ENOMEM; 1110 goto err1; 1111 } 1112 1113 rxe_add_index(mr); 1114 1115 rxe_add_ref(pd); 1116 1117 err = rxe_mem_init_fast(rxe, pd, max_num_sg, mr); 1118 if (err) 1119 goto err2; 1120 1121 return &mr->ibmr; 1122 1123 err2: 1124 rxe_drop_ref(pd); 1125 rxe_drop_index(mr); 1126 rxe_drop_ref(mr); 1127 err1: 1128 return ERR_PTR(err); 1129 } 1130 1131 static int rxe_set_page(struct ib_mr *ibmr, u64 addr) 1132 { 1133 struct rxe_mem *mr = to_rmr(ibmr); 1134 struct rxe_map *map; 1135 struct rxe_phys_buf *buf; 1136 1137 if (unlikely(mr->nbuf == mr->num_buf)) 1138 return -ENOMEM; 1139 1140 map = mr->map[mr->nbuf / RXE_BUF_PER_MAP]; 1141 buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP]; 1142 1143 buf->addr = addr; 1144 buf->size = ibmr->page_size; 1145 mr->nbuf++; 1146 1147 return 0; 1148 } 1149 1150 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, 1151 int sg_nents, unsigned int *sg_offset) 1152 { 1153 struct rxe_mem *mr = to_rmr(ibmr); 1154 int n; 1155 1156 mr->nbuf = 0; 1157 1158 n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page); 1159 1160 mr->va = ibmr->iova; 1161 mr->iova = ibmr->iova; 1162 mr->length = ibmr->length; 1163 mr->page_shift = ilog2(ibmr->page_size); 1164 mr->page_mask = ibmr->page_size - 1; 1165 mr->offset = mr->iova & mr->page_mask; 1166 1167 return n; 1168 } 1169 1170 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1171 { 1172 int err; 1173 struct rxe_dev *rxe = to_rdev(ibqp->device); 1174 struct rxe_qp *qp = to_rqp(ibqp); 1175 struct rxe_mc_grp *grp; 1176 1177 /* takes a ref on grp if successful */ 1178 err = rxe_mcast_get_grp(rxe, mgid, &grp); 1179 if (err) 1180 return err; 1181 1182 err = rxe_mcast_add_grp_elem(rxe, qp, grp); 1183 1184 rxe_drop_ref(grp); 1185 return err; 1186 } 1187 1188 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1189 { 1190 struct rxe_dev *rxe = to_rdev(ibqp->device); 1191 struct rxe_qp *qp = to_rqp(ibqp); 1192 1193 return rxe_mcast_drop_grp_elem(rxe, qp, mgid); 1194 } 1195 1196 static ssize_t rxe_show_parent(struct device *device, 1197 struct device_attribute *attr, char *buf) 1198 { 1199 struct rxe_dev *rxe = container_of(device, struct rxe_dev, 1200 ib_dev.dev); 1201 char *name; 1202 1203 name = rxe->ifc_ops->parent_name(rxe, 1); 1204 return snprintf(buf, 16, "%s\n", name); 1205 } 1206 1207 static DEVICE_ATTR(parent, S_IRUGO, rxe_show_parent, NULL); 1208 1209 static struct device_attribute *rxe_dev_attributes[] = { 1210 &dev_attr_parent, 1211 }; 1212 1213 int rxe_register_device(struct rxe_dev *rxe) 1214 { 1215 int err; 1216 int i; 1217 struct ib_device *dev = &rxe->ib_dev; 1218 1219 strlcpy(dev->name, "rxe%d", IB_DEVICE_NAME_MAX); 1220 strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc)); 1221 1222 dev->owner = THIS_MODULE; 1223 dev->node_type = RDMA_NODE_IB_CA; 1224 dev->phys_port_cnt = 1; 1225 dev->num_comp_vectors = RXE_NUM_COMP_VECTORS; 1226 dev->dma_device = rxe->ifc_ops->dma_device(rxe); 1227 dev->local_dma_lkey = 0; 1228 dev->node_guid = rxe->ifc_ops->node_guid(rxe); 1229 dev->dma_ops = &rxe_dma_mapping_ops; 1230 1231 dev->uverbs_abi_ver = RXE_UVERBS_ABI_VERSION; 1232 dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) 1233 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) 1234 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) 1235 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) 1236 | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) 1237 | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) 1238 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) 1239 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) 1240 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) 1241 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) 1242 | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV) 1243 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) 1244 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) 1245 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) 1246 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) 1247 | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) 1248 | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV) 1249 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) 1250 | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ) 1251 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) 1252 | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ) 1253 | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ) 1254 | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) 1255 | BIT_ULL(IB_USER_VERBS_CMD_REG_MR) 1256 | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) 1257 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) 1258 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH) 1259 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH) 1260 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) 1261 | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) 1262 | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) 1263 ; 1264 1265 dev->query_device = rxe_query_device; 1266 dev->modify_device = rxe_modify_device; 1267 dev->query_port = rxe_query_port; 1268 dev->modify_port = rxe_modify_port; 1269 dev->get_link_layer = rxe_get_link_layer; 1270 dev->query_gid = rxe_query_gid; 1271 dev->get_netdev = rxe_get_netdev; 1272 dev->add_gid = rxe_add_gid; 1273 dev->del_gid = rxe_del_gid; 1274 dev->query_pkey = rxe_query_pkey; 1275 dev->alloc_ucontext = rxe_alloc_ucontext; 1276 dev->dealloc_ucontext = rxe_dealloc_ucontext; 1277 dev->mmap = rxe_mmap; 1278 dev->get_port_immutable = rxe_port_immutable; 1279 dev->alloc_pd = rxe_alloc_pd; 1280 dev->dealloc_pd = rxe_dealloc_pd; 1281 dev->create_ah = rxe_create_ah; 1282 dev->modify_ah = rxe_modify_ah; 1283 dev->query_ah = rxe_query_ah; 1284 dev->destroy_ah = rxe_destroy_ah; 1285 dev->create_srq = rxe_create_srq; 1286 dev->modify_srq = rxe_modify_srq; 1287 dev->query_srq = rxe_query_srq; 1288 dev->destroy_srq = rxe_destroy_srq; 1289 dev->post_srq_recv = rxe_post_srq_recv; 1290 dev->create_qp = rxe_create_qp; 1291 dev->modify_qp = rxe_modify_qp; 1292 dev->query_qp = rxe_query_qp; 1293 dev->destroy_qp = rxe_destroy_qp; 1294 dev->post_send = rxe_post_send; 1295 dev->post_recv = rxe_post_recv; 1296 dev->create_cq = rxe_create_cq; 1297 dev->destroy_cq = rxe_destroy_cq; 1298 dev->resize_cq = rxe_resize_cq; 1299 dev->poll_cq = rxe_poll_cq; 1300 dev->peek_cq = rxe_peek_cq; 1301 dev->req_notify_cq = rxe_req_notify_cq; 1302 dev->get_dma_mr = rxe_get_dma_mr; 1303 dev->reg_user_mr = rxe_reg_user_mr; 1304 dev->dereg_mr = rxe_dereg_mr; 1305 dev->alloc_mr = rxe_alloc_mr; 1306 dev->map_mr_sg = rxe_map_mr_sg; 1307 dev->attach_mcast = rxe_attach_mcast; 1308 dev->detach_mcast = rxe_detach_mcast; 1309 1310 err = ib_register_device(dev, NULL); 1311 if (err) { 1312 pr_warn("rxe_register_device failed, err = %d\n", err); 1313 goto err1; 1314 } 1315 1316 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) { 1317 err = device_create_file(&dev->dev, rxe_dev_attributes[i]); 1318 if (err) { 1319 pr_warn("device_create_file failed, i = %d, err = %d\n", 1320 i, err); 1321 goto err2; 1322 } 1323 } 1324 1325 return 0; 1326 1327 err2: 1328 ib_unregister_device(dev); 1329 err1: 1330 return err; 1331 } 1332 1333 int rxe_unregister_device(struct rxe_dev *rxe) 1334 { 1335 int i; 1336 struct ib_device *dev = &rxe->ib_dev; 1337 1338 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) 1339 device_remove_file(&dev->dev, rxe_dev_attributes[i]); 1340 1341 ib_unregister_device(dev); 1342 1343 return 0; 1344 } 1345