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 "rxe.h" 36 #include "rxe_loc.h" 37 #include "rxe_queue.h" 38 39 static int rxe_query_device(struct ib_device *dev, 40 struct ib_device_attr *attr, 41 struct ib_udata *uhw) 42 { 43 struct rxe_dev *rxe = to_rdev(dev); 44 45 if (uhw->inlen || uhw->outlen) 46 return -EINVAL; 47 48 *attr = rxe->attr; 49 return 0; 50 } 51 52 static void rxe_eth_speed_to_ib_speed(int speed, u8 *active_speed, 53 u8 *active_width) 54 { 55 if (speed <= 1000) { 56 *active_width = IB_WIDTH_1X; 57 *active_speed = IB_SPEED_SDR; 58 } else if (speed <= 10000) { 59 *active_width = IB_WIDTH_1X; 60 *active_speed = IB_SPEED_FDR10; 61 } else if (speed <= 20000) { 62 *active_width = IB_WIDTH_4X; 63 *active_speed = IB_SPEED_DDR; 64 } else if (speed <= 30000) { 65 *active_width = IB_WIDTH_4X; 66 *active_speed = IB_SPEED_QDR; 67 } else if (speed <= 40000) { 68 *active_width = IB_WIDTH_4X; 69 *active_speed = IB_SPEED_FDR10; 70 } else { 71 *active_width = IB_WIDTH_4X; 72 *active_speed = IB_SPEED_EDR; 73 } 74 } 75 76 static int rxe_query_port(struct ib_device *dev, 77 u8 port_num, struct ib_port_attr *attr) 78 { 79 struct rxe_dev *rxe = to_rdev(dev); 80 struct rxe_port *port; 81 u32 speed; 82 83 if (unlikely(port_num != 1)) { 84 pr_warn("invalid port_number %d\n", port_num); 85 goto err1; 86 } 87 88 port = &rxe->port; 89 90 /* *attr being zeroed by the caller, avoid zeroing it here */ 91 *attr = port->attr; 92 93 mutex_lock(&rxe->usdev_lock); 94 if (rxe->ndev->ethtool_ops->get_link_ksettings) { 95 struct ethtool_link_ksettings ks; 96 97 rxe->ndev->ethtool_ops->get_link_ksettings(rxe->ndev, &ks); 98 speed = ks.base.speed; 99 } else if (rxe->ndev->ethtool_ops->get_settings) { 100 struct ethtool_cmd cmd; 101 102 rxe->ndev->ethtool_ops->get_settings(rxe->ndev, &cmd); 103 speed = cmd.speed; 104 } else { 105 pr_warn("%s speed is unknown, defaulting to 1000\n", 106 rxe->ndev->name); 107 speed = 1000; 108 } 109 rxe_eth_speed_to_ib_speed(speed, &attr->active_speed, 110 &attr->active_width); 111 mutex_unlock(&rxe->usdev_lock); 112 113 return 0; 114 115 err1: 116 return -EINVAL; 117 } 118 119 static int rxe_query_gid(struct ib_device *device, 120 u8 port_num, int index, union ib_gid *gid) 121 { 122 int ret; 123 124 if (index > RXE_PORT_GID_TBL_LEN) 125 return -EINVAL; 126 127 ret = ib_get_cached_gid(device, port_num, index, gid, NULL); 128 if (ret == -EAGAIN) { 129 memcpy(gid, &zgid, sizeof(*gid)); 130 return 0; 131 } 132 133 return ret; 134 } 135 136 static int rxe_add_gid(struct ib_device *device, u8 port_num, unsigned int 137 index, const union ib_gid *gid, 138 const struct ib_gid_attr *attr, void **context) 139 { 140 if (index >= RXE_PORT_GID_TBL_LEN) 141 return -EINVAL; 142 return 0; 143 } 144 145 static int rxe_del_gid(struct ib_device *device, u8 port_num, unsigned int 146 index, void **context) 147 { 148 if (index >= RXE_PORT_GID_TBL_LEN) 149 return -EINVAL; 150 return 0; 151 } 152 153 static struct net_device *rxe_get_netdev(struct ib_device *device, 154 u8 port_num) 155 { 156 struct rxe_dev *rxe = to_rdev(device); 157 158 if (rxe->ndev) { 159 dev_hold(rxe->ndev); 160 return rxe->ndev; 161 } 162 163 return NULL; 164 } 165 166 static int rxe_query_pkey(struct ib_device *device, 167 u8 port_num, u16 index, u16 *pkey) 168 { 169 struct rxe_dev *rxe = to_rdev(device); 170 struct rxe_port *port; 171 172 if (unlikely(port_num != 1)) { 173 dev_warn(device->dev.parent, "invalid port_num = %d\n", 174 port_num); 175 goto err1; 176 } 177 178 port = &rxe->port; 179 180 if (unlikely(index >= port->attr.pkey_tbl_len)) { 181 dev_warn(device->dev.parent, "invalid index = %d\n", 182 index); 183 goto err1; 184 } 185 186 *pkey = port->pkey_tbl[index]; 187 return 0; 188 189 err1: 190 return -EINVAL; 191 } 192 193 static int rxe_modify_device(struct ib_device *dev, 194 int mask, struct ib_device_modify *attr) 195 { 196 struct rxe_dev *rxe = to_rdev(dev); 197 198 if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) 199 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid); 200 201 if (mask & IB_DEVICE_MODIFY_NODE_DESC) { 202 memcpy(rxe->ib_dev.node_desc, 203 attr->node_desc, sizeof(rxe->ib_dev.node_desc)); 204 } 205 206 return 0; 207 } 208 209 static int rxe_modify_port(struct ib_device *dev, 210 u8 port_num, int mask, struct ib_port_modify *attr) 211 { 212 struct rxe_dev *rxe = to_rdev(dev); 213 struct rxe_port *port; 214 215 if (unlikely(port_num != 1)) { 216 pr_warn("invalid port_num = %d\n", port_num); 217 goto err1; 218 } 219 220 port = &rxe->port; 221 222 port->attr.port_cap_flags |= attr->set_port_cap_mask; 223 port->attr.port_cap_flags &= ~attr->clr_port_cap_mask; 224 225 if (mask & IB_PORT_RESET_QKEY_CNTR) 226 port->attr.qkey_viol_cntr = 0; 227 228 return 0; 229 230 err1: 231 return -EINVAL; 232 } 233 234 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev, 235 u8 port_num) 236 { 237 struct rxe_dev *rxe = to_rdev(dev); 238 239 return rxe_link_layer(rxe, port_num); 240 } 241 242 static struct ib_ucontext *rxe_alloc_ucontext(struct ib_device *dev, 243 struct ib_udata *udata) 244 { 245 struct rxe_dev *rxe = to_rdev(dev); 246 struct rxe_ucontext *uc; 247 248 uc = rxe_alloc(&rxe->uc_pool); 249 return uc ? &uc->ibuc : ERR_PTR(-ENOMEM); 250 } 251 252 static int rxe_dealloc_ucontext(struct ib_ucontext *ibuc) 253 { 254 struct rxe_ucontext *uc = to_ruc(ibuc); 255 256 rxe_drop_ref(uc); 257 return 0; 258 } 259 260 static int rxe_port_immutable(struct ib_device *dev, u8 port_num, 261 struct ib_port_immutable *immutable) 262 { 263 int err; 264 struct ib_port_attr attr; 265 266 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 267 268 err = ib_query_port(dev, port_num, &attr); 269 if (err) 270 return err; 271 272 immutable->pkey_tbl_len = attr.pkey_tbl_len; 273 immutable->gid_tbl_len = attr.gid_tbl_len; 274 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 275 276 return 0; 277 } 278 279 static struct ib_pd *rxe_alloc_pd(struct ib_device *dev, 280 struct ib_ucontext *context, 281 struct ib_udata *udata) 282 { 283 struct rxe_dev *rxe = to_rdev(dev); 284 struct rxe_pd *pd; 285 286 pd = rxe_alloc(&rxe->pd_pool); 287 return pd ? &pd->ibpd : ERR_PTR(-ENOMEM); 288 } 289 290 static int rxe_dealloc_pd(struct ib_pd *ibpd) 291 { 292 struct rxe_pd *pd = to_rpd(ibpd); 293 294 rxe_drop_ref(pd); 295 return 0; 296 } 297 298 static int rxe_init_av(struct rxe_dev *rxe, struct ib_ah_attr *attr, 299 struct rxe_av *av) 300 { 301 int err; 302 union ib_gid sgid; 303 struct ib_gid_attr sgid_attr; 304 305 err = ib_get_cached_gid(&rxe->ib_dev, attr->port_num, 306 attr->grh.sgid_index, &sgid, 307 &sgid_attr); 308 if (err) { 309 pr_err("Failed to query sgid. err = %d\n", err); 310 return err; 311 } 312 313 err = rxe_av_from_attr(rxe, attr->port_num, av, attr); 314 if (!err) 315 err = rxe_av_fill_ip_info(rxe, av, attr, &sgid_attr, &sgid); 316 317 if (sgid_attr.ndev) 318 dev_put(sgid_attr.ndev); 319 return err; 320 } 321 322 static struct ib_ah *rxe_create_ah(struct ib_pd *ibpd, struct ib_ah_attr *attr, 323 struct ib_udata *udata) 324 325 { 326 int err; 327 struct rxe_dev *rxe = to_rdev(ibpd->device); 328 struct rxe_pd *pd = to_rpd(ibpd); 329 struct rxe_ah *ah; 330 331 err = rxe_av_chk_attr(rxe, attr); 332 if (err) 333 goto err1; 334 335 ah = rxe_alloc(&rxe->ah_pool); 336 if (!ah) { 337 err = -ENOMEM; 338 goto err1; 339 } 340 341 rxe_add_ref(pd); 342 ah->pd = pd; 343 344 err = rxe_init_av(rxe, attr, &ah->av); 345 if (err) 346 goto err2; 347 348 return &ah->ibah; 349 350 err2: 351 rxe_drop_ref(pd); 352 rxe_drop_ref(ah); 353 err1: 354 return ERR_PTR(err); 355 } 356 357 static int rxe_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *attr) 358 { 359 int err; 360 struct rxe_dev *rxe = to_rdev(ibah->device); 361 struct rxe_ah *ah = to_rah(ibah); 362 363 err = rxe_av_chk_attr(rxe, attr); 364 if (err) 365 return err; 366 367 err = rxe_init_av(rxe, attr, &ah->av); 368 if (err) 369 return err; 370 371 return 0; 372 } 373 374 static int rxe_query_ah(struct ib_ah *ibah, struct ib_ah_attr *attr) 375 { 376 struct rxe_dev *rxe = to_rdev(ibah->device); 377 struct rxe_ah *ah = to_rah(ibah); 378 379 rxe_av_to_attr(rxe, &ah->av, attr); 380 return 0; 381 } 382 383 static int rxe_destroy_ah(struct ib_ah *ibah) 384 { 385 struct rxe_ah *ah = to_rah(ibah); 386 387 rxe_drop_ref(ah->pd); 388 rxe_drop_ref(ah); 389 return 0; 390 } 391 392 static int post_one_recv(struct rxe_rq *rq, struct ib_recv_wr *ibwr) 393 { 394 int err; 395 int i; 396 u32 length; 397 struct rxe_recv_wqe *recv_wqe; 398 int num_sge = ibwr->num_sge; 399 400 if (unlikely(queue_full(rq->queue))) { 401 err = -ENOMEM; 402 goto err1; 403 } 404 405 if (unlikely(num_sge > rq->max_sge)) { 406 err = -EINVAL; 407 goto err1; 408 } 409 410 length = 0; 411 for (i = 0; i < num_sge; i++) 412 length += ibwr->sg_list[i].length; 413 414 recv_wqe = producer_addr(rq->queue); 415 recv_wqe->wr_id = ibwr->wr_id; 416 recv_wqe->num_sge = num_sge; 417 418 memcpy(recv_wqe->dma.sge, ibwr->sg_list, 419 num_sge * sizeof(struct ib_sge)); 420 421 recv_wqe->dma.length = length; 422 recv_wqe->dma.resid = length; 423 recv_wqe->dma.num_sge = num_sge; 424 recv_wqe->dma.cur_sge = 0; 425 recv_wqe->dma.sge_offset = 0; 426 427 /* make sure all changes to the work queue are written before we 428 * update the producer pointer 429 */ 430 smp_wmb(); 431 432 advance_producer(rq->queue); 433 return 0; 434 435 err1: 436 return err; 437 } 438 439 static struct ib_srq *rxe_create_srq(struct ib_pd *ibpd, 440 struct ib_srq_init_attr *init, 441 struct ib_udata *udata) 442 { 443 int err; 444 struct rxe_dev *rxe = to_rdev(ibpd->device); 445 struct rxe_pd *pd = to_rpd(ibpd); 446 struct rxe_srq *srq; 447 struct ib_ucontext *context = udata ? ibpd->uobject->context : NULL; 448 449 err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK); 450 if (err) 451 goto err1; 452 453 srq = rxe_alloc(&rxe->srq_pool); 454 if (!srq) { 455 err = -ENOMEM; 456 goto err1; 457 } 458 459 rxe_add_index(srq); 460 rxe_add_ref(pd); 461 srq->pd = pd; 462 463 err = rxe_srq_from_init(rxe, srq, init, context, udata); 464 if (err) 465 goto err2; 466 467 return &srq->ibsrq; 468 469 err2: 470 rxe_drop_ref(pd); 471 rxe_drop_index(srq); 472 rxe_drop_ref(srq); 473 err1: 474 return ERR_PTR(err); 475 } 476 477 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 478 enum ib_srq_attr_mask mask, 479 struct ib_udata *udata) 480 { 481 int err; 482 struct rxe_srq *srq = to_rsrq(ibsrq); 483 struct rxe_dev *rxe = to_rdev(ibsrq->device); 484 485 err = rxe_srq_chk_attr(rxe, srq, attr, mask); 486 if (err) 487 goto err1; 488 489 err = rxe_srq_from_attr(rxe, srq, attr, mask, udata); 490 if (err) 491 goto err1; 492 493 return 0; 494 495 err1: 496 return err; 497 } 498 499 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr) 500 { 501 struct rxe_srq *srq = to_rsrq(ibsrq); 502 503 if (srq->error) 504 return -EINVAL; 505 506 attr->max_wr = srq->rq.queue->buf->index_mask; 507 attr->max_sge = srq->rq.max_sge; 508 attr->srq_limit = srq->limit; 509 return 0; 510 } 511 512 static int rxe_destroy_srq(struct ib_srq *ibsrq) 513 { 514 struct rxe_srq *srq = to_rsrq(ibsrq); 515 516 if (srq->rq.queue) 517 rxe_queue_cleanup(srq->rq.queue); 518 519 rxe_drop_ref(srq->pd); 520 rxe_drop_index(srq); 521 rxe_drop_ref(srq); 522 523 return 0; 524 } 525 526 static int rxe_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 527 struct ib_recv_wr **bad_wr) 528 { 529 int err = 0; 530 unsigned long flags; 531 struct rxe_srq *srq = to_rsrq(ibsrq); 532 533 spin_lock_irqsave(&srq->rq.producer_lock, flags); 534 535 while (wr) { 536 err = post_one_recv(&srq->rq, wr); 537 if (unlikely(err)) 538 break; 539 wr = wr->next; 540 } 541 542 spin_unlock_irqrestore(&srq->rq.producer_lock, flags); 543 544 if (err) 545 *bad_wr = wr; 546 547 return err; 548 } 549 550 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd, 551 struct ib_qp_init_attr *init, 552 struct ib_udata *udata) 553 { 554 int err; 555 struct rxe_dev *rxe = to_rdev(ibpd->device); 556 struct rxe_pd *pd = to_rpd(ibpd); 557 struct rxe_qp *qp; 558 559 err = rxe_qp_chk_init(rxe, init); 560 if (err) 561 goto err1; 562 563 qp = rxe_alloc(&rxe->qp_pool); 564 if (!qp) { 565 err = -ENOMEM; 566 goto err1; 567 } 568 569 if (udata) { 570 if (udata->inlen) { 571 err = -EINVAL; 572 goto err2; 573 } 574 qp->is_user = 1; 575 } 576 577 rxe_add_index(qp); 578 579 err = rxe_qp_from_init(rxe, qp, pd, init, udata, ibpd); 580 if (err) 581 goto err3; 582 583 return &qp->ibqp; 584 585 err3: 586 rxe_drop_index(qp); 587 err2: 588 rxe_drop_ref(qp); 589 err1: 590 return ERR_PTR(err); 591 } 592 593 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 594 int mask, struct ib_udata *udata) 595 { 596 int err; 597 struct rxe_dev *rxe = to_rdev(ibqp->device); 598 struct rxe_qp *qp = to_rqp(ibqp); 599 600 err = rxe_qp_chk_attr(rxe, qp, attr, mask); 601 if (err) 602 goto err1; 603 604 err = rxe_qp_from_attr(qp, attr, mask, udata); 605 if (err) 606 goto err1; 607 608 return 0; 609 610 err1: 611 return err; 612 } 613 614 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 615 int mask, struct ib_qp_init_attr *init) 616 { 617 struct rxe_qp *qp = to_rqp(ibqp); 618 619 rxe_qp_to_init(qp, init); 620 rxe_qp_to_attr(qp, attr, mask); 621 622 return 0; 623 } 624 625 static int rxe_destroy_qp(struct ib_qp *ibqp) 626 { 627 struct rxe_qp *qp = to_rqp(ibqp); 628 629 rxe_qp_destroy(qp); 630 rxe_drop_index(qp); 631 rxe_drop_ref(qp); 632 return 0; 633 } 634 635 static int validate_send_wr(struct rxe_qp *qp, struct ib_send_wr *ibwr, 636 unsigned int mask, unsigned int length) 637 { 638 int num_sge = ibwr->num_sge; 639 struct rxe_sq *sq = &qp->sq; 640 641 if (unlikely(num_sge > sq->max_sge)) 642 goto err1; 643 644 if (unlikely(mask & WR_ATOMIC_MASK)) { 645 if (length < 8) 646 goto err1; 647 648 if (atomic_wr(ibwr)->remote_addr & 0x7) 649 goto err1; 650 } 651 652 if (unlikely((ibwr->send_flags & IB_SEND_INLINE) && 653 (length > sq->max_inline))) 654 goto err1; 655 656 return 0; 657 658 err1: 659 return -EINVAL; 660 } 661 662 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr, 663 struct ib_send_wr *ibwr) 664 { 665 wr->wr_id = ibwr->wr_id; 666 wr->num_sge = ibwr->num_sge; 667 wr->opcode = ibwr->opcode; 668 wr->send_flags = ibwr->send_flags; 669 670 if (qp_type(qp) == IB_QPT_UD || 671 qp_type(qp) == IB_QPT_SMI || 672 qp_type(qp) == IB_QPT_GSI) { 673 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn; 674 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey; 675 if (qp_type(qp) == IB_QPT_GSI) 676 wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index; 677 if (wr->opcode == IB_WR_SEND_WITH_IMM) 678 wr->ex.imm_data = ibwr->ex.imm_data; 679 } else { 680 switch (wr->opcode) { 681 case IB_WR_RDMA_WRITE_WITH_IMM: 682 wr->ex.imm_data = ibwr->ex.imm_data; 683 case IB_WR_RDMA_READ: 684 case IB_WR_RDMA_WRITE: 685 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr; 686 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey; 687 break; 688 case IB_WR_SEND_WITH_IMM: 689 wr->ex.imm_data = ibwr->ex.imm_data; 690 break; 691 case IB_WR_SEND_WITH_INV: 692 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 693 break; 694 case IB_WR_ATOMIC_CMP_AND_SWP: 695 case IB_WR_ATOMIC_FETCH_AND_ADD: 696 wr->wr.atomic.remote_addr = 697 atomic_wr(ibwr)->remote_addr; 698 wr->wr.atomic.compare_add = 699 atomic_wr(ibwr)->compare_add; 700 wr->wr.atomic.swap = atomic_wr(ibwr)->swap; 701 wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey; 702 break; 703 case IB_WR_LOCAL_INV: 704 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 705 break; 706 case IB_WR_REG_MR: 707 wr->wr.reg.mr = reg_wr(ibwr)->mr; 708 wr->wr.reg.key = reg_wr(ibwr)->key; 709 wr->wr.reg.access = reg_wr(ibwr)->access; 710 break; 711 default: 712 break; 713 } 714 } 715 } 716 717 static int init_send_wqe(struct rxe_qp *qp, struct ib_send_wr *ibwr, 718 unsigned int mask, unsigned int length, 719 struct rxe_send_wqe *wqe) 720 { 721 int num_sge = ibwr->num_sge; 722 struct ib_sge *sge; 723 int i; 724 u8 *p; 725 726 init_send_wr(qp, &wqe->wr, ibwr); 727 728 if (qp_type(qp) == IB_QPT_UD || 729 qp_type(qp) == IB_QPT_SMI || 730 qp_type(qp) == IB_QPT_GSI) 731 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av)); 732 733 if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) { 734 p = wqe->dma.inline_data; 735 736 sge = ibwr->sg_list; 737 for (i = 0; i < num_sge; i++, sge++) { 738 if (qp->is_user && copy_from_user(p, (__user void *) 739 (uintptr_t)sge->addr, sge->length)) 740 return -EFAULT; 741 742 else if (!qp->is_user) 743 memcpy(p, (void *)(uintptr_t)sge->addr, 744 sge->length); 745 746 p += sge->length; 747 } 748 } else if (mask & WR_REG_MASK) { 749 wqe->mask = mask; 750 wqe->state = wqe_state_posted; 751 return 0; 752 } else 753 memcpy(wqe->dma.sge, ibwr->sg_list, 754 num_sge * sizeof(struct ib_sge)); 755 756 wqe->iova = (mask & WR_ATOMIC_MASK) ? 757 atomic_wr(ibwr)->remote_addr : 758 rdma_wr(ibwr)->remote_addr; 759 wqe->mask = mask; 760 wqe->dma.length = length; 761 wqe->dma.resid = length; 762 wqe->dma.num_sge = num_sge; 763 wqe->dma.cur_sge = 0; 764 wqe->dma.sge_offset = 0; 765 wqe->state = wqe_state_posted; 766 wqe->ssn = atomic_add_return(1, &qp->ssn); 767 768 return 0; 769 } 770 771 static int post_one_send(struct rxe_qp *qp, struct ib_send_wr *ibwr, 772 unsigned int mask, u32 length) 773 { 774 int err; 775 struct rxe_sq *sq = &qp->sq; 776 struct rxe_send_wqe *send_wqe; 777 unsigned long flags; 778 779 err = validate_send_wr(qp, ibwr, mask, length); 780 if (err) 781 return err; 782 783 spin_lock_irqsave(&qp->sq.sq_lock, flags); 784 785 if (unlikely(queue_full(sq->queue))) { 786 err = -ENOMEM; 787 goto err1; 788 } 789 790 send_wqe = producer_addr(sq->queue); 791 792 err = init_send_wqe(qp, ibwr, mask, length, send_wqe); 793 if (unlikely(err)) 794 goto err1; 795 796 /* 797 * make sure all changes to the work queue are 798 * written before we update the producer pointer 799 */ 800 smp_wmb(); 801 802 advance_producer(sq->queue); 803 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 804 805 return 0; 806 807 err1: 808 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 809 return err; 810 } 811 812 static int rxe_post_send_kernel(struct rxe_qp *qp, struct ib_send_wr *wr, 813 struct ib_send_wr **bad_wr) 814 { 815 int err = 0; 816 unsigned int mask; 817 unsigned int length = 0; 818 int i; 819 int must_sched; 820 821 while (wr) { 822 mask = wr_opcode_mask(wr->opcode, qp); 823 if (unlikely(!mask)) { 824 err = -EINVAL; 825 *bad_wr = wr; 826 break; 827 } 828 829 if (unlikely((wr->send_flags & IB_SEND_INLINE) && 830 !(mask & WR_INLINE_MASK))) { 831 err = -EINVAL; 832 *bad_wr = wr; 833 break; 834 } 835 836 length = 0; 837 for (i = 0; i < wr->num_sge; i++) 838 length += wr->sg_list[i].length; 839 840 err = post_one_send(qp, wr, mask, length); 841 842 if (err) { 843 *bad_wr = wr; 844 break; 845 } 846 wr = wr->next; 847 } 848 849 /* 850 * Must sched in case of GSI QP because ib_send_mad() hold irq lock, 851 * and the requester call ip_local_out_sk() that takes spin_lock_bh. 852 */ 853 must_sched = (qp_type(qp) == IB_QPT_GSI) || 854 (queue_count(qp->sq.queue) > 1); 855 856 rxe_run_task(&qp->req.task, must_sched); 857 858 return err; 859 } 860 861 static int rxe_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 862 struct ib_send_wr **bad_wr) 863 { 864 struct rxe_qp *qp = to_rqp(ibqp); 865 866 if (unlikely(!qp->valid)) { 867 *bad_wr = wr; 868 return -EINVAL; 869 } 870 871 if (unlikely(qp->req.state < QP_STATE_READY)) { 872 *bad_wr = wr; 873 return -EINVAL; 874 } 875 876 if (qp->is_user) { 877 /* Utilize process context to do protocol processing */ 878 rxe_run_task(&qp->req.task, 0); 879 return 0; 880 } else 881 return rxe_post_send_kernel(qp, wr, bad_wr); 882 } 883 884 static int rxe_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 885 struct ib_recv_wr **bad_wr) 886 { 887 int err = 0; 888 struct rxe_qp *qp = to_rqp(ibqp); 889 struct rxe_rq *rq = &qp->rq; 890 unsigned long flags; 891 892 if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) { 893 *bad_wr = wr; 894 err = -EINVAL; 895 goto err1; 896 } 897 898 if (unlikely(qp->srq)) { 899 *bad_wr = wr; 900 err = -EINVAL; 901 goto err1; 902 } 903 904 spin_lock_irqsave(&rq->producer_lock, flags); 905 906 while (wr) { 907 err = post_one_recv(rq, wr); 908 if (unlikely(err)) { 909 *bad_wr = wr; 910 break; 911 } 912 wr = wr->next; 913 } 914 915 spin_unlock_irqrestore(&rq->producer_lock, flags); 916 917 err1: 918 return err; 919 } 920 921 static struct ib_cq *rxe_create_cq(struct ib_device *dev, 922 const struct ib_cq_init_attr *attr, 923 struct ib_ucontext *context, 924 struct ib_udata *udata) 925 { 926 int err; 927 struct rxe_dev *rxe = to_rdev(dev); 928 struct rxe_cq *cq; 929 930 if (attr->flags) 931 return ERR_PTR(-EINVAL); 932 933 err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector, udata); 934 if (err) 935 goto err1; 936 937 cq = rxe_alloc(&rxe->cq_pool); 938 if (!cq) { 939 err = -ENOMEM; 940 goto err1; 941 } 942 943 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, 944 context, udata); 945 if (err) 946 goto err2; 947 948 return &cq->ibcq; 949 950 err2: 951 rxe_drop_ref(cq); 952 err1: 953 return ERR_PTR(err); 954 } 955 956 static int rxe_destroy_cq(struct ib_cq *ibcq) 957 { 958 struct rxe_cq *cq = to_rcq(ibcq); 959 960 rxe_drop_ref(cq); 961 return 0; 962 } 963 964 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 965 { 966 int err; 967 struct rxe_cq *cq = to_rcq(ibcq); 968 struct rxe_dev *rxe = to_rdev(ibcq->device); 969 970 err = rxe_cq_chk_attr(rxe, cq, cqe, 0, udata); 971 if (err) 972 goto err1; 973 974 err = rxe_cq_resize_queue(cq, cqe, udata); 975 if (err) 976 goto err1; 977 978 return 0; 979 980 err1: 981 return err; 982 } 983 984 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 985 { 986 int i; 987 struct rxe_cq *cq = to_rcq(ibcq); 988 struct rxe_cqe *cqe; 989 unsigned long flags; 990 991 spin_lock_irqsave(&cq->cq_lock, flags); 992 for (i = 0; i < num_entries; i++) { 993 cqe = queue_head(cq->queue); 994 if (!cqe) 995 break; 996 997 memcpy(wc++, &cqe->ibwc, sizeof(*wc)); 998 advance_consumer(cq->queue); 999 } 1000 spin_unlock_irqrestore(&cq->cq_lock, flags); 1001 1002 return i; 1003 } 1004 1005 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt) 1006 { 1007 struct rxe_cq *cq = to_rcq(ibcq); 1008 int count = queue_count(cq->queue); 1009 1010 return (count > wc_cnt) ? wc_cnt : count; 1011 } 1012 1013 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 1014 { 1015 struct rxe_cq *cq = to_rcq(ibcq); 1016 unsigned long irq_flags; 1017 int ret = 0; 1018 1019 spin_lock_irqsave(&cq->cq_lock, irq_flags); 1020 if (cq->notify != IB_CQ_NEXT_COMP) 1021 cq->notify = flags & IB_CQ_SOLICITED_MASK; 1022 1023 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue)) 1024 ret = 1; 1025 1026 spin_unlock_irqrestore(&cq->cq_lock, irq_flags); 1027 1028 return ret; 1029 } 1030 1031 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access) 1032 { 1033 struct rxe_dev *rxe = to_rdev(ibpd->device); 1034 struct rxe_pd *pd = to_rpd(ibpd); 1035 struct rxe_mem *mr; 1036 int err; 1037 1038 mr = rxe_alloc(&rxe->mr_pool); 1039 if (!mr) { 1040 err = -ENOMEM; 1041 goto err1; 1042 } 1043 1044 rxe_add_index(mr); 1045 1046 rxe_add_ref(pd); 1047 1048 err = rxe_mem_init_dma(rxe, pd, access, mr); 1049 if (err) 1050 goto err2; 1051 1052 return &mr->ibmr; 1053 1054 err2: 1055 rxe_drop_ref(pd); 1056 rxe_drop_index(mr); 1057 rxe_drop_ref(mr); 1058 err1: 1059 return ERR_PTR(err); 1060 } 1061 1062 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd, 1063 u64 start, 1064 u64 length, 1065 u64 iova, 1066 int access, struct ib_udata *udata) 1067 { 1068 int err; 1069 struct rxe_dev *rxe = to_rdev(ibpd->device); 1070 struct rxe_pd *pd = to_rpd(ibpd); 1071 struct rxe_mem *mr; 1072 1073 mr = rxe_alloc(&rxe->mr_pool); 1074 if (!mr) { 1075 err = -ENOMEM; 1076 goto err2; 1077 } 1078 1079 rxe_add_index(mr); 1080 1081 rxe_add_ref(pd); 1082 1083 err = rxe_mem_init_user(rxe, pd, start, length, iova, 1084 access, udata, mr); 1085 if (err) 1086 goto err3; 1087 1088 return &mr->ibmr; 1089 1090 err3: 1091 rxe_drop_ref(pd); 1092 rxe_drop_index(mr); 1093 rxe_drop_ref(mr); 1094 err2: 1095 return ERR_PTR(err); 1096 } 1097 1098 static int rxe_dereg_mr(struct ib_mr *ibmr) 1099 { 1100 struct rxe_mem *mr = to_rmr(ibmr); 1101 1102 mr->state = RXE_MEM_STATE_ZOMBIE; 1103 rxe_drop_ref(mr->pd); 1104 rxe_drop_index(mr); 1105 rxe_drop_ref(mr); 1106 return 0; 1107 } 1108 1109 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, 1110 enum ib_mr_type mr_type, 1111 u32 max_num_sg) 1112 { 1113 struct rxe_dev *rxe = to_rdev(ibpd->device); 1114 struct rxe_pd *pd = to_rpd(ibpd); 1115 struct rxe_mem *mr; 1116 int err; 1117 1118 if (mr_type != IB_MR_TYPE_MEM_REG) 1119 return ERR_PTR(-EINVAL); 1120 1121 mr = rxe_alloc(&rxe->mr_pool); 1122 if (!mr) { 1123 err = -ENOMEM; 1124 goto err1; 1125 } 1126 1127 rxe_add_index(mr); 1128 1129 rxe_add_ref(pd); 1130 1131 err = rxe_mem_init_fast(rxe, pd, max_num_sg, mr); 1132 if (err) 1133 goto err2; 1134 1135 return &mr->ibmr; 1136 1137 err2: 1138 rxe_drop_ref(pd); 1139 rxe_drop_index(mr); 1140 rxe_drop_ref(mr); 1141 err1: 1142 return ERR_PTR(err); 1143 } 1144 1145 static int rxe_set_page(struct ib_mr *ibmr, u64 addr) 1146 { 1147 struct rxe_mem *mr = to_rmr(ibmr); 1148 struct rxe_map *map; 1149 struct rxe_phys_buf *buf; 1150 1151 if (unlikely(mr->nbuf == mr->num_buf)) 1152 return -ENOMEM; 1153 1154 map = mr->map[mr->nbuf / RXE_BUF_PER_MAP]; 1155 buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP]; 1156 1157 buf->addr = addr; 1158 buf->size = ibmr->page_size; 1159 mr->nbuf++; 1160 1161 return 0; 1162 } 1163 1164 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, 1165 int sg_nents, unsigned int *sg_offset) 1166 { 1167 struct rxe_mem *mr = to_rmr(ibmr); 1168 int n; 1169 1170 mr->nbuf = 0; 1171 1172 n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page); 1173 1174 mr->va = ibmr->iova; 1175 mr->iova = ibmr->iova; 1176 mr->length = ibmr->length; 1177 mr->page_shift = ilog2(ibmr->page_size); 1178 mr->page_mask = ibmr->page_size - 1; 1179 mr->offset = mr->iova & mr->page_mask; 1180 1181 return n; 1182 } 1183 1184 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1185 { 1186 int err; 1187 struct rxe_dev *rxe = to_rdev(ibqp->device); 1188 struct rxe_qp *qp = to_rqp(ibqp); 1189 struct rxe_mc_grp *grp; 1190 1191 /* takes a ref on grp if successful */ 1192 err = rxe_mcast_get_grp(rxe, mgid, &grp); 1193 if (err) 1194 return err; 1195 1196 err = rxe_mcast_add_grp_elem(rxe, qp, grp); 1197 1198 rxe_drop_ref(grp); 1199 return err; 1200 } 1201 1202 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1203 { 1204 struct rxe_dev *rxe = to_rdev(ibqp->device); 1205 struct rxe_qp *qp = to_rqp(ibqp); 1206 1207 return rxe_mcast_drop_grp_elem(rxe, qp, mgid); 1208 } 1209 1210 static ssize_t rxe_show_parent(struct device *device, 1211 struct device_attribute *attr, char *buf) 1212 { 1213 struct rxe_dev *rxe = container_of(device, struct rxe_dev, 1214 ib_dev.dev); 1215 1216 return snprintf(buf, 16, "%s\n", rxe_parent_name(rxe, 1)); 1217 } 1218 1219 static DEVICE_ATTR(parent, S_IRUGO, rxe_show_parent, NULL); 1220 1221 static struct device_attribute *rxe_dev_attributes[] = { 1222 &dev_attr_parent, 1223 }; 1224 1225 int rxe_register_device(struct rxe_dev *rxe) 1226 { 1227 int err; 1228 int i; 1229 struct ib_device *dev = &rxe->ib_dev; 1230 1231 strlcpy(dev->name, "rxe%d", IB_DEVICE_NAME_MAX); 1232 strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc)); 1233 1234 dev->owner = THIS_MODULE; 1235 dev->node_type = RDMA_NODE_IB_CA; 1236 dev->phys_port_cnt = 1; 1237 dev->num_comp_vectors = RXE_NUM_COMP_VECTORS; 1238 dev->dev.parent = rxe_dma_device(rxe); 1239 dev->local_dma_lkey = 0; 1240 dev->node_guid = rxe_node_guid(rxe); 1241 dev->dev.dma_ops = &dma_virt_ops; 1242 1243 dev->uverbs_abi_ver = RXE_UVERBS_ABI_VERSION; 1244 dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) 1245 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) 1246 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) 1247 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) 1248 | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) 1249 | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) 1250 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) 1251 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) 1252 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) 1253 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) 1254 | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV) 1255 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) 1256 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) 1257 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) 1258 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) 1259 | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) 1260 | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV) 1261 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) 1262 | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ) 1263 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) 1264 | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ) 1265 | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ) 1266 | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) 1267 | BIT_ULL(IB_USER_VERBS_CMD_REG_MR) 1268 | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) 1269 | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) 1270 | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH) 1271 | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH) 1272 | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) 1273 | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) 1274 | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) 1275 ; 1276 1277 dev->query_device = rxe_query_device; 1278 dev->modify_device = rxe_modify_device; 1279 dev->query_port = rxe_query_port; 1280 dev->modify_port = rxe_modify_port; 1281 dev->get_link_layer = rxe_get_link_layer; 1282 dev->query_gid = rxe_query_gid; 1283 dev->get_netdev = rxe_get_netdev; 1284 dev->add_gid = rxe_add_gid; 1285 dev->del_gid = rxe_del_gid; 1286 dev->query_pkey = rxe_query_pkey; 1287 dev->alloc_ucontext = rxe_alloc_ucontext; 1288 dev->dealloc_ucontext = rxe_dealloc_ucontext; 1289 dev->mmap = rxe_mmap; 1290 dev->get_port_immutable = rxe_port_immutable; 1291 dev->alloc_pd = rxe_alloc_pd; 1292 dev->dealloc_pd = rxe_dealloc_pd; 1293 dev->create_ah = rxe_create_ah; 1294 dev->modify_ah = rxe_modify_ah; 1295 dev->query_ah = rxe_query_ah; 1296 dev->destroy_ah = rxe_destroy_ah; 1297 dev->create_srq = rxe_create_srq; 1298 dev->modify_srq = rxe_modify_srq; 1299 dev->query_srq = rxe_query_srq; 1300 dev->destroy_srq = rxe_destroy_srq; 1301 dev->post_srq_recv = rxe_post_srq_recv; 1302 dev->create_qp = rxe_create_qp; 1303 dev->modify_qp = rxe_modify_qp; 1304 dev->query_qp = rxe_query_qp; 1305 dev->destroy_qp = rxe_destroy_qp; 1306 dev->post_send = rxe_post_send; 1307 dev->post_recv = rxe_post_recv; 1308 dev->create_cq = rxe_create_cq; 1309 dev->destroy_cq = rxe_destroy_cq; 1310 dev->resize_cq = rxe_resize_cq; 1311 dev->poll_cq = rxe_poll_cq; 1312 dev->peek_cq = rxe_peek_cq; 1313 dev->req_notify_cq = rxe_req_notify_cq; 1314 dev->get_dma_mr = rxe_get_dma_mr; 1315 dev->reg_user_mr = rxe_reg_user_mr; 1316 dev->dereg_mr = rxe_dereg_mr; 1317 dev->alloc_mr = rxe_alloc_mr; 1318 dev->map_mr_sg = rxe_map_mr_sg; 1319 dev->attach_mcast = rxe_attach_mcast; 1320 dev->detach_mcast = rxe_detach_mcast; 1321 1322 err = ib_register_device(dev, NULL); 1323 if (err) { 1324 pr_warn("rxe_register_device failed, err = %d\n", err); 1325 goto err1; 1326 } 1327 1328 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) { 1329 err = device_create_file(&dev->dev, rxe_dev_attributes[i]); 1330 if (err) { 1331 pr_warn("device_create_file failed, i = %d, err = %d\n", 1332 i, err); 1333 goto err2; 1334 } 1335 } 1336 1337 return 0; 1338 1339 err2: 1340 ib_unregister_device(dev); 1341 err1: 1342 return err; 1343 } 1344 1345 int rxe_unregister_device(struct rxe_dev *rxe) 1346 { 1347 int i; 1348 struct ib_device *dev = &rxe->ib_dev; 1349 1350 for (i = 0; i < ARRAY_SIZE(rxe_dev_attributes); ++i) 1351 device_remove_file(&dev->dev, rxe_dev_attributes[i]); 1352 1353 ib_unregister_device(dev); 1354 1355 return 0; 1356 } 1357