1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* 3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved. 4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved. 5 */ 6 7 #include <linux/dma-mapping.h> 8 #include <net/addrconf.h> 9 #include <rdma/uverbs_ioctl.h> 10 #include "rxe.h" 11 #include "rxe_loc.h" 12 #include "rxe_queue.h" 13 #include "rxe_hw_counters.h" 14 15 static int rxe_query_device(struct ib_device *dev, 16 struct ib_device_attr *attr, 17 struct ib_udata *uhw) 18 { 19 struct rxe_dev *rxe = to_rdev(dev); 20 21 if (uhw->inlen || uhw->outlen) 22 return -EINVAL; 23 24 *attr = rxe->attr; 25 return 0; 26 } 27 28 static int rxe_query_port(struct ib_device *dev, 29 u32 port_num, struct ib_port_attr *attr) 30 { 31 struct rxe_dev *rxe = to_rdev(dev); 32 struct rxe_port *port; 33 int rc; 34 35 port = &rxe->port; 36 37 /* *attr being zeroed by the caller, avoid zeroing it here */ 38 *attr = port->attr; 39 40 mutex_lock(&rxe->usdev_lock); 41 rc = ib_get_eth_speed(dev, port_num, &attr->active_speed, 42 &attr->active_width); 43 44 if (attr->state == IB_PORT_ACTIVE) 45 attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 46 else if (dev_get_flags(rxe->ndev) & IFF_UP) 47 attr->phys_state = IB_PORT_PHYS_STATE_POLLING; 48 else 49 attr->phys_state = IB_PORT_PHYS_STATE_DISABLED; 50 51 mutex_unlock(&rxe->usdev_lock); 52 53 return rc; 54 } 55 56 static int rxe_query_pkey(struct ib_device *device, 57 u32 port_num, u16 index, u16 *pkey) 58 { 59 if (index > 0) 60 return -EINVAL; 61 62 *pkey = IB_DEFAULT_PKEY_FULL; 63 return 0; 64 } 65 66 static int rxe_modify_device(struct ib_device *dev, 67 int mask, struct ib_device_modify *attr) 68 { 69 struct rxe_dev *rxe = to_rdev(dev); 70 71 if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID | 72 IB_DEVICE_MODIFY_NODE_DESC)) 73 return -EOPNOTSUPP; 74 75 if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) 76 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid); 77 78 if (mask & IB_DEVICE_MODIFY_NODE_DESC) { 79 memcpy(rxe->ib_dev.node_desc, 80 attr->node_desc, sizeof(rxe->ib_dev.node_desc)); 81 } 82 83 return 0; 84 } 85 86 static int rxe_modify_port(struct ib_device *dev, 87 u32 port_num, int mask, struct ib_port_modify *attr) 88 { 89 struct rxe_dev *rxe = to_rdev(dev); 90 struct rxe_port *port; 91 92 port = &rxe->port; 93 94 port->attr.port_cap_flags |= attr->set_port_cap_mask; 95 port->attr.port_cap_flags &= ~attr->clr_port_cap_mask; 96 97 if (mask & IB_PORT_RESET_QKEY_CNTR) 98 port->attr.qkey_viol_cntr = 0; 99 100 return 0; 101 } 102 103 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev, 104 u32 port_num) 105 { 106 return IB_LINK_LAYER_ETHERNET; 107 } 108 109 static int rxe_alloc_ucontext(struct ib_ucontext *ibuc, struct ib_udata *udata) 110 { 111 struct rxe_dev *rxe = to_rdev(ibuc->device); 112 struct rxe_ucontext *uc = to_ruc(ibuc); 113 114 return rxe_add_to_pool(&rxe->uc_pool, uc); 115 } 116 117 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc) 118 { 119 struct rxe_ucontext *uc = to_ruc(ibuc); 120 121 rxe_drop_ref(uc); 122 } 123 124 static int rxe_port_immutable(struct ib_device *dev, u32 port_num, 125 struct ib_port_immutable *immutable) 126 { 127 int err; 128 struct ib_port_attr attr; 129 130 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 131 132 err = ib_query_port(dev, port_num, &attr); 133 if (err) 134 return err; 135 136 immutable->pkey_tbl_len = attr.pkey_tbl_len; 137 immutable->gid_tbl_len = attr.gid_tbl_len; 138 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 139 140 return 0; 141 } 142 143 static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 144 { 145 struct rxe_dev *rxe = to_rdev(ibpd->device); 146 struct rxe_pd *pd = to_rpd(ibpd); 147 148 return rxe_add_to_pool(&rxe->pd_pool, pd); 149 } 150 151 static int rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 152 { 153 struct rxe_pd *pd = to_rpd(ibpd); 154 155 rxe_drop_ref(pd); 156 return 0; 157 } 158 159 static int rxe_create_ah(struct ib_ah *ibah, 160 struct rdma_ah_init_attr *init_attr, 161 struct ib_udata *udata) 162 163 { 164 int err; 165 struct rxe_dev *rxe = to_rdev(ibah->device); 166 struct rxe_ah *ah = to_rah(ibah); 167 168 err = rxe_av_chk_attr(rxe, init_attr->ah_attr); 169 if (err) 170 return err; 171 172 err = rxe_add_to_pool(&rxe->ah_pool, ah); 173 if (err) 174 return err; 175 176 rxe_init_av(init_attr->ah_attr, &ah->av); 177 return 0; 178 } 179 180 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr) 181 { 182 int err; 183 struct rxe_dev *rxe = to_rdev(ibah->device); 184 struct rxe_ah *ah = to_rah(ibah); 185 186 err = rxe_av_chk_attr(rxe, attr); 187 if (err) 188 return err; 189 190 rxe_init_av(attr, &ah->av); 191 return 0; 192 } 193 194 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr) 195 { 196 struct rxe_ah *ah = to_rah(ibah); 197 198 memset(attr, 0, sizeof(*attr)); 199 attr->type = ibah->type; 200 rxe_av_to_attr(&ah->av, attr); 201 return 0; 202 } 203 204 static int rxe_destroy_ah(struct ib_ah *ibah, u32 flags) 205 { 206 struct rxe_ah *ah = to_rah(ibah); 207 208 rxe_drop_ref(ah); 209 return 0; 210 } 211 212 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr) 213 { 214 int err; 215 int i; 216 u32 length; 217 struct rxe_recv_wqe *recv_wqe; 218 int num_sge = ibwr->num_sge; 219 int full; 220 221 if (rq->is_user) 222 full = queue_full(rq->queue, QUEUE_TYPE_FROM_USER); 223 else 224 full = queue_full(rq->queue, QUEUE_TYPE_KERNEL); 225 226 if (unlikely(full)) { 227 err = -ENOMEM; 228 goto err1; 229 } 230 231 if (unlikely(num_sge > rq->max_sge)) { 232 err = -EINVAL; 233 goto err1; 234 } 235 236 length = 0; 237 for (i = 0; i < num_sge; i++) 238 length += ibwr->sg_list[i].length; 239 240 if (rq->is_user) 241 recv_wqe = producer_addr(rq->queue, QUEUE_TYPE_FROM_USER); 242 else 243 recv_wqe = producer_addr(rq->queue, QUEUE_TYPE_KERNEL); 244 245 recv_wqe->wr_id = ibwr->wr_id; 246 recv_wqe->num_sge = num_sge; 247 248 memcpy(recv_wqe->dma.sge, ibwr->sg_list, 249 num_sge * sizeof(struct ib_sge)); 250 251 recv_wqe->dma.length = length; 252 recv_wqe->dma.resid = length; 253 recv_wqe->dma.num_sge = num_sge; 254 recv_wqe->dma.cur_sge = 0; 255 recv_wqe->dma.sge_offset = 0; 256 257 if (rq->is_user) 258 advance_producer(rq->queue, QUEUE_TYPE_FROM_USER); 259 else 260 advance_producer(rq->queue, QUEUE_TYPE_KERNEL); 261 262 return 0; 263 264 err1: 265 return err; 266 } 267 268 static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init, 269 struct ib_udata *udata) 270 { 271 int err; 272 struct rxe_dev *rxe = to_rdev(ibsrq->device); 273 struct rxe_pd *pd = to_rpd(ibsrq->pd); 274 struct rxe_srq *srq = to_rsrq(ibsrq); 275 struct rxe_create_srq_resp __user *uresp = NULL; 276 277 if (init->srq_type != IB_SRQT_BASIC) 278 return -EOPNOTSUPP; 279 280 if (udata) { 281 if (udata->outlen < sizeof(*uresp)) 282 return -EINVAL; 283 uresp = udata->outbuf; 284 srq->is_user = true; 285 } else { 286 srq->is_user = false; 287 } 288 289 err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK); 290 if (err) 291 goto err1; 292 293 err = rxe_add_to_pool(&rxe->srq_pool, srq); 294 if (err) 295 goto err1; 296 297 rxe_add_ref(pd); 298 srq->pd = pd; 299 300 err = rxe_srq_from_init(rxe, srq, init, udata, uresp); 301 if (err) 302 goto err2; 303 304 return 0; 305 306 err2: 307 rxe_drop_ref(pd); 308 rxe_drop_ref(srq); 309 err1: 310 return err; 311 } 312 313 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 314 enum ib_srq_attr_mask mask, 315 struct ib_udata *udata) 316 { 317 int err; 318 struct rxe_srq *srq = to_rsrq(ibsrq); 319 struct rxe_dev *rxe = to_rdev(ibsrq->device); 320 struct rxe_modify_srq_cmd ucmd = {}; 321 322 if (udata) { 323 if (udata->inlen < sizeof(ucmd)) 324 return -EINVAL; 325 326 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)); 327 if (err) 328 return err; 329 } 330 331 err = rxe_srq_chk_attr(rxe, srq, attr, mask); 332 if (err) 333 goto err1; 334 335 err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd, udata); 336 if (err) 337 goto err1; 338 339 return 0; 340 341 err1: 342 return err; 343 } 344 345 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr) 346 { 347 struct rxe_srq *srq = to_rsrq(ibsrq); 348 349 if (srq->error) 350 return -EINVAL; 351 352 attr->max_wr = srq->rq.queue->buf->index_mask; 353 attr->max_sge = srq->rq.max_sge; 354 attr->srq_limit = srq->limit; 355 return 0; 356 } 357 358 static int rxe_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata) 359 { 360 struct rxe_srq *srq = to_rsrq(ibsrq); 361 362 if (srq->rq.queue) 363 rxe_queue_cleanup(srq->rq.queue); 364 365 rxe_drop_ref(srq->pd); 366 rxe_drop_ref(srq); 367 return 0; 368 } 369 370 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr, 371 const struct ib_recv_wr **bad_wr) 372 { 373 int err = 0; 374 unsigned long flags; 375 struct rxe_srq *srq = to_rsrq(ibsrq); 376 377 spin_lock_irqsave(&srq->rq.producer_lock, flags); 378 379 while (wr) { 380 err = post_one_recv(&srq->rq, wr); 381 if (unlikely(err)) 382 break; 383 wr = wr->next; 384 } 385 386 spin_unlock_irqrestore(&srq->rq.producer_lock, flags); 387 388 if (err) 389 *bad_wr = wr; 390 391 return err; 392 } 393 394 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd, 395 struct ib_qp_init_attr *init, 396 struct ib_udata *udata) 397 { 398 int err; 399 struct rxe_dev *rxe = to_rdev(ibpd->device); 400 struct rxe_pd *pd = to_rpd(ibpd); 401 struct rxe_qp *qp; 402 struct rxe_create_qp_resp __user *uresp = NULL; 403 404 if (udata) { 405 if (udata->outlen < sizeof(*uresp)) 406 return ERR_PTR(-EINVAL); 407 uresp = udata->outbuf; 408 } 409 410 if (init->create_flags) 411 return ERR_PTR(-EOPNOTSUPP); 412 413 err = rxe_qp_chk_init(rxe, init); 414 if (err) 415 goto err1; 416 417 qp = rxe_alloc(&rxe->qp_pool); 418 if (!qp) { 419 err = -ENOMEM; 420 goto err1; 421 } 422 423 if (udata) { 424 if (udata->inlen) { 425 err = -EINVAL; 426 goto err2; 427 } 428 qp->is_user = true; 429 } else { 430 qp->is_user = false; 431 } 432 433 rxe_add_index(qp); 434 435 err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibpd, udata); 436 if (err) 437 goto err3; 438 439 return &qp->ibqp; 440 441 err3: 442 rxe_drop_index(qp); 443 err2: 444 rxe_drop_ref(qp); 445 err1: 446 return ERR_PTR(err); 447 } 448 449 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 450 int mask, struct ib_udata *udata) 451 { 452 int err; 453 struct rxe_dev *rxe = to_rdev(ibqp->device); 454 struct rxe_qp *qp = to_rqp(ibqp); 455 456 if (mask & ~IB_QP_ATTR_STANDARD_BITS) 457 return -EOPNOTSUPP; 458 459 err = rxe_qp_chk_attr(rxe, qp, attr, mask); 460 if (err) 461 goto err1; 462 463 err = rxe_qp_from_attr(qp, attr, mask, udata); 464 if (err) 465 goto err1; 466 467 return 0; 468 469 err1: 470 return err; 471 } 472 473 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 474 int mask, struct ib_qp_init_attr *init) 475 { 476 struct rxe_qp *qp = to_rqp(ibqp); 477 478 rxe_qp_to_init(qp, init); 479 rxe_qp_to_attr(qp, attr, mask); 480 481 return 0; 482 } 483 484 static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) 485 { 486 struct rxe_qp *qp = to_rqp(ibqp); 487 488 rxe_qp_destroy(qp); 489 rxe_drop_index(qp); 490 rxe_drop_ref(qp); 491 return 0; 492 } 493 494 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr, 495 unsigned int mask, unsigned int length) 496 { 497 int num_sge = ibwr->num_sge; 498 struct rxe_sq *sq = &qp->sq; 499 500 if (unlikely(num_sge > sq->max_sge)) 501 goto err1; 502 503 if (unlikely(mask & WR_ATOMIC_MASK)) { 504 if (length < 8) 505 goto err1; 506 507 if (atomic_wr(ibwr)->remote_addr & 0x7) 508 goto err1; 509 } 510 511 if (unlikely((ibwr->send_flags & IB_SEND_INLINE) && 512 (length > sq->max_inline))) 513 goto err1; 514 515 return 0; 516 517 err1: 518 return -EINVAL; 519 } 520 521 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr, 522 const struct ib_send_wr *ibwr) 523 { 524 wr->wr_id = ibwr->wr_id; 525 wr->num_sge = ibwr->num_sge; 526 wr->opcode = ibwr->opcode; 527 wr->send_flags = ibwr->send_flags; 528 529 if (qp_type(qp) == IB_QPT_UD || 530 qp_type(qp) == IB_QPT_SMI || 531 qp_type(qp) == IB_QPT_GSI) { 532 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn; 533 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey; 534 if (qp_type(qp) == IB_QPT_GSI) 535 wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index; 536 if (wr->opcode == IB_WR_SEND_WITH_IMM) 537 wr->ex.imm_data = ibwr->ex.imm_data; 538 } else { 539 switch (wr->opcode) { 540 case IB_WR_RDMA_WRITE_WITH_IMM: 541 wr->ex.imm_data = ibwr->ex.imm_data; 542 fallthrough; 543 case IB_WR_RDMA_READ: 544 case IB_WR_RDMA_WRITE: 545 wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr; 546 wr->wr.rdma.rkey = rdma_wr(ibwr)->rkey; 547 break; 548 case IB_WR_SEND_WITH_IMM: 549 wr->ex.imm_data = ibwr->ex.imm_data; 550 break; 551 case IB_WR_SEND_WITH_INV: 552 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 553 break; 554 case IB_WR_ATOMIC_CMP_AND_SWP: 555 case IB_WR_ATOMIC_FETCH_AND_ADD: 556 wr->wr.atomic.remote_addr = 557 atomic_wr(ibwr)->remote_addr; 558 wr->wr.atomic.compare_add = 559 atomic_wr(ibwr)->compare_add; 560 wr->wr.atomic.swap = atomic_wr(ibwr)->swap; 561 wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey; 562 break; 563 case IB_WR_LOCAL_INV: 564 wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey; 565 break; 566 case IB_WR_REG_MR: 567 wr->wr.reg.mr = reg_wr(ibwr)->mr; 568 wr->wr.reg.key = reg_wr(ibwr)->key; 569 wr->wr.reg.access = reg_wr(ibwr)->access; 570 break; 571 default: 572 break; 573 } 574 } 575 } 576 577 static void copy_inline_data_to_wqe(struct rxe_send_wqe *wqe, 578 const struct ib_send_wr *ibwr) 579 { 580 struct ib_sge *sge = ibwr->sg_list; 581 u8 *p = wqe->dma.inline_data; 582 int i; 583 584 for (i = 0; i < ibwr->num_sge; i++, sge++) { 585 memcpy(p, (void *)(uintptr_t)sge->addr, sge->length); 586 p += sge->length; 587 } 588 } 589 590 static void init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr, 591 unsigned int mask, unsigned int length, 592 struct rxe_send_wqe *wqe) 593 { 594 int num_sge = ibwr->num_sge; 595 596 init_send_wr(qp, &wqe->wr, ibwr); 597 598 /* local operation */ 599 if (unlikely(mask & WR_LOCAL_OP_MASK)) { 600 wqe->mask = mask; 601 wqe->state = wqe_state_posted; 602 return; 603 } 604 605 if (qp_type(qp) == IB_QPT_UD || 606 qp_type(qp) == IB_QPT_SMI || 607 qp_type(qp) == IB_QPT_GSI) 608 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av)); 609 610 if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) 611 copy_inline_data_to_wqe(wqe, ibwr); 612 else 613 memcpy(wqe->dma.sge, ibwr->sg_list, 614 num_sge * sizeof(struct ib_sge)); 615 616 wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr : 617 mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0; 618 wqe->mask = mask; 619 wqe->dma.length = length; 620 wqe->dma.resid = length; 621 wqe->dma.num_sge = num_sge; 622 wqe->dma.cur_sge = 0; 623 wqe->dma.sge_offset = 0; 624 wqe->state = wqe_state_posted; 625 wqe->ssn = atomic_add_return(1, &qp->ssn); 626 } 627 628 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr, 629 unsigned int mask, u32 length) 630 { 631 int err; 632 struct rxe_sq *sq = &qp->sq; 633 struct rxe_send_wqe *send_wqe; 634 unsigned long flags; 635 int full; 636 637 err = validate_send_wr(qp, ibwr, mask, length); 638 if (err) 639 return err; 640 641 spin_lock_irqsave(&qp->sq.sq_lock, flags); 642 643 if (qp->is_user) 644 full = queue_full(sq->queue, QUEUE_TYPE_FROM_USER); 645 else 646 full = queue_full(sq->queue, QUEUE_TYPE_KERNEL); 647 648 if (unlikely(full)) { 649 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 650 return -ENOMEM; 651 } 652 653 if (qp->is_user) 654 send_wqe = producer_addr(sq->queue, QUEUE_TYPE_FROM_USER); 655 else 656 send_wqe = producer_addr(sq->queue, QUEUE_TYPE_KERNEL); 657 658 init_send_wqe(qp, ibwr, mask, length, send_wqe); 659 660 if (qp->is_user) 661 advance_producer(sq->queue, QUEUE_TYPE_FROM_USER); 662 else 663 advance_producer(sq->queue, QUEUE_TYPE_KERNEL); 664 665 spin_unlock_irqrestore(&qp->sq.sq_lock, flags); 666 667 return 0; 668 } 669 670 static int rxe_post_send_kernel(struct rxe_qp *qp, const struct ib_send_wr *wr, 671 const struct ib_send_wr **bad_wr) 672 { 673 int err = 0; 674 unsigned int mask; 675 unsigned int length = 0; 676 int i; 677 struct ib_send_wr *next; 678 679 while (wr) { 680 mask = wr_opcode_mask(wr->opcode, qp); 681 if (unlikely(!mask)) { 682 err = -EINVAL; 683 *bad_wr = wr; 684 break; 685 } 686 687 if (unlikely((wr->send_flags & IB_SEND_INLINE) && 688 !(mask & WR_INLINE_MASK))) { 689 err = -EINVAL; 690 *bad_wr = wr; 691 break; 692 } 693 694 next = wr->next; 695 696 length = 0; 697 for (i = 0; i < wr->num_sge; i++) 698 length += wr->sg_list[i].length; 699 700 err = post_one_send(qp, wr, mask, length); 701 702 if (err) { 703 *bad_wr = wr; 704 break; 705 } 706 wr = next; 707 } 708 709 rxe_run_task(&qp->req.task, 1); 710 if (unlikely(qp->req.state == QP_STATE_ERROR)) 711 rxe_run_task(&qp->comp.task, 1); 712 713 return err; 714 } 715 716 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr, 717 const struct ib_send_wr **bad_wr) 718 { 719 struct rxe_qp *qp = to_rqp(ibqp); 720 721 if (unlikely(!qp->valid)) { 722 *bad_wr = wr; 723 return -EINVAL; 724 } 725 726 if (unlikely(qp->req.state < QP_STATE_READY)) { 727 *bad_wr = wr; 728 return -EINVAL; 729 } 730 731 if (qp->is_user) { 732 /* Utilize process context to do protocol processing */ 733 rxe_run_task(&qp->req.task, 0); 734 return 0; 735 } else 736 return rxe_post_send_kernel(qp, wr, bad_wr); 737 } 738 739 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr, 740 const struct ib_recv_wr **bad_wr) 741 { 742 int err = 0; 743 struct rxe_qp *qp = to_rqp(ibqp); 744 struct rxe_rq *rq = &qp->rq; 745 unsigned long flags; 746 747 if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) { 748 *bad_wr = wr; 749 err = -EINVAL; 750 goto err1; 751 } 752 753 if (unlikely(qp->srq)) { 754 *bad_wr = wr; 755 err = -EINVAL; 756 goto err1; 757 } 758 759 spin_lock_irqsave(&rq->producer_lock, flags); 760 761 while (wr) { 762 err = post_one_recv(rq, wr); 763 if (unlikely(err)) { 764 *bad_wr = wr; 765 break; 766 } 767 wr = wr->next; 768 } 769 770 spin_unlock_irqrestore(&rq->producer_lock, flags); 771 772 if (qp->resp.state == QP_STATE_ERROR) 773 rxe_run_task(&qp->resp.task, 1); 774 775 err1: 776 return err; 777 } 778 779 static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 780 struct ib_udata *udata) 781 { 782 int err; 783 struct ib_device *dev = ibcq->device; 784 struct rxe_dev *rxe = to_rdev(dev); 785 struct rxe_cq *cq = to_rcq(ibcq); 786 struct rxe_create_cq_resp __user *uresp = NULL; 787 788 if (udata) { 789 if (udata->outlen < sizeof(*uresp)) 790 return -EINVAL; 791 uresp = udata->outbuf; 792 } 793 794 if (attr->flags) 795 return -EOPNOTSUPP; 796 797 err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector); 798 if (err) 799 return err; 800 801 err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, udata, 802 uresp); 803 if (err) 804 return err; 805 806 return rxe_add_to_pool(&rxe->cq_pool, cq); 807 } 808 809 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata) 810 { 811 struct rxe_cq *cq = to_rcq(ibcq); 812 813 rxe_cq_disable(cq); 814 815 rxe_drop_ref(cq); 816 return 0; 817 } 818 819 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 820 { 821 int err; 822 struct rxe_cq *cq = to_rcq(ibcq); 823 struct rxe_dev *rxe = to_rdev(ibcq->device); 824 struct rxe_resize_cq_resp __user *uresp = NULL; 825 826 if (udata) { 827 if (udata->outlen < sizeof(*uresp)) 828 return -EINVAL; 829 uresp = udata->outbuf; 830 } 831 832 err = rxe_cq_chk_attr(rxe, cq, cqe, 0); 833 if (err) 834 goto err1; 835 836 err = rxe_cq_resize_queue(cq, cqe, uresp, udata); 837 if (err) 838 goto err1; 839 840 return 0; 841 842 err1: 843 return err; 844 } 845 846 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 847 { 848 int i; 849 struct rxe_cq *cq = to_rcq(ibcq); 850 struct rxe_cqe *cqe; 851 unsigned long flags; 852 853 spin_lock_irqsave(&cq->cq_lock, flags); 854 for (i = 0; i < num_entries; i++) { 855 if (cq->is_user) 856 cqe = queue_head(cq->queue, QUEUE_TYPE_TO_USER); 857 else 858 cqe = queue_head(cq->queue, QUEUE_TYPE_KERNEL); 859 if (!cqe) 860 break; 861 862 memcpy(wc++, &cqe->ibwc, sizeof(*wc)); 863 if (cq->is_user) 864 advance_consumer(cq->queue, QUEUE_TYPE_TO_USER); 865 else 866 advance_consumer(cq->queue, QUEUE_TYPE_KERNEL); 867 } 868 spin_unlock_irqrestore(&cq->cq_lock, flags); 869 870 return i; 871 } 872 873 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt) 874 { 875 struct rxe_cq *cq = to_rcq(ibcq); 876 int count; 877 878 if (cq->is_user) 879 count = queue_count(cq->queue, QUEUE_TYPE_TO_USER); 880 else 881 count = queue_count(cq->queue, QUEUE_TYPE_KERNEL); 882 883 return (count > wc_cnt) ? wc_cnt : count; 884 } 885 886 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 887 { 888 struct rxe_cq *cq = to_rcq(ibcq); 889 unsigned long irq_flags; 890 int ret = 0; 891 int empty; 892 893 spin_lock_irqsave(&cq->cq_lock, irq_flags); 894 if (cq->notify != IB_CQ_NEXT_COMP) 895 cq->notify = flags & IB_CQ_SOLICITED_MASK; 896 897 if (cq->is_user) 898 empty = queue_empty(cq->queue, QUEUE_TYPE_TO_USER); 899 else 900 empty = queue_empty(cq->queue, QUEUE_TYPE_KERNEL); 901 902 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !empty) 903 ret = 1; 904 905 spin_unlock_irqrestore(&cq->cq_lock, irq_flags); 906 907 return ret; 908 } 909 910 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access) 911 { 912 struct rxe_dev *rxe = to_rdev(ibpd->device); 913 struct rxe_pd *pd = to_rpd(ibpd); 914 struct rxe_mr *mr; 915 916 mr = rxe_alloc(&rxe->mr_pool); 917 if (!mr) 918 return ERR_PTR(-ENOMEM); 919 920 rxe_add_index(mr); 921 rxe_add_ref(pd); 922 rxe_mr_init_dma(pd, access, mr); 923 924 return &mr->ibmr; 925 } 926 927 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd, 928 u64 start, 929 u64 length, 930 u64 iova, 931 int access, struct ib_udata *udata) 932 { 933 int err; 934 struct rxe_dev *rxe = to_rdev(ibpd->device); 935 struct rxe_pd *pd = to_rpd(ibpd); 936 struct rxe_mr *mr; 937 938 mr = rxe_alloc(&rxe->mr_pool); 939 if (!mr) { 940 err = -ENOMEM; 941 goto err2; 942 } 943 944 rxe_add_index(mr); 945 946 rxe_add_ref(pd); 947 948 err = rxe_mr_init_user(pd, start, length, iova, access, mr); 949 if (err) 950 goto err3; 951 952 return &mr->ibmr; 953 954 err3: 955 rxe_drop_ref(pd); 956 rxe_drop_index(mr); 957 rxe_drop_ref(mr); 958 err2: 959 return ERR_PTR(err); 960 } 961 962 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type, 963 u32 max_num_sg) 964 { 965 struct rxe_dev *rxe = to_rdev(ibpd->device); 966 struct rxe_pd *pd = to_rpd(ibpd); 967 struct rxe_mr *mr; 968 int err; 969 970 if (mr_type != IB_MR_TYPE_MEM_REG) 971 return ERR_PTR(-EINVAL); 972 973 mr = rxe_alloc(&rxe->mr_pool); 974 if (!mr) { 975 err = -ENOMEM; 976 goto err1; 977 } 978 979 rxe_add_index(mr); 980 981 rxe_add_ref(pd); 982 983 err = rxe_mr_init_fast(pd, max_num_sg, mr); 984 if (err) 985 goto err2; 986 987 return &mr->ibmr; 988 989 err2: 990 rxe_drop_ref(pd); 991 rxe_drop_index(mr); 992 rxe_drop_ref(mr); 993 err1: 994 return ERR_PTR(err); 995 } 996 997 static int rxe_set_page(struct ib_mr *ibmr, u64 addr) 998 { 999 struct rxe_mr *mr = to_rmr(ibmr); 1000 struct rxe_map *map; 1001 struct rxe_phys_buf *buf; 1002 1003 if (unlikely(mr->nbuf == mr->num_buf)) 1004 return -ENOMEM; 1005 1006 map = mr->map[mr->nbuf / RXE_BUF_PER_MAP]; 1007 buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP]; 1008 1009 buf->addr = addr; 1010 buf->size = ibmr->page_size; 1011 mr->nbuf++; 1012 1013 return 0; 1014 } 1015 1016 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, 1017 int sg_nents, unsigned int *sg_offset) 1018 { 1019 struct rxe_mr *mr = to_rmr(ibmr); 1020 int n; 1021 1022 mr->nbuf = 0; 1023 1024 n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page); 1025 1026 mr->va = ibmr->iova; 1027 mr->iova = ibmr->iova; 1028 mr->length = ibmr->length; 1029 mr->page_shift = ilog2(ibmr->page_size); 1030 mr->page_mask = ibmr->page_size - 1; 1031 mr->offset = mr->iova & mr->page_mask; 1032 1033 return n; 1034 } 1035 1036 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1037 { 1038 int err; 1039 struct rxe_dev *rxe = to_rdev(ibqp->device); 1040 struct rxe_qp *qp = to_rqp(ibqp); 1041 struct rxe_mc_grp *grp; 1042 1043 /* takes a ref on grp if successful */ 1044 err = rxe_mcast_get_grp(rxe, mgid, &grp); 1045 if (err) 1046 return err; 1047 1048 err = rxe_mcast_add_grp_elem(rxe, qp, grp); 1049 1050 rxe_drop_ref(grp); 1051 return err; 1052 } 1053 1054 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid) 1055 { 1056 struct rxe_dev *rxe = to_rdev(ibqp->device); 1057 struct rxe_qp *qp = to_rqp(ibqp); 1058 1059 return rxe_mcast_drop_grp_elem(rxe, qp, mgid); 1060 } 1061 1062 static ssize_t parent_show(struct device *device, 1063 struct device_attribute *attr, char *buf) 1064 { 1065 struct rxe_dev *rxe = 1066 rdma_device_to_drv_device(device, struct rxe_dev, ib_dev); 1067 1068 return sysfs_emit(buf, "%s\n", rxe_parent_name(rxe, 1)); 1069 } 1070 1071 static DEVICE_ATTR_RO(parent); 1072 1073 static struct attribute *rxe_dev_attributes[] = { 1074 &dev_attr_parent.attr, 1075 NULL 1076 }; 1077 1078 static const struct attribute_group rxe_attr_group = { 1079 .attrs = rxe_dev_attributes, 1080 }; 1081 1082 static int rxe_enable_driver(struct ib_device *ib_dev) 1083 { 1084 struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev); 1085 1086 rxe_set_port_state(rxe); 1087 dev_info(&rxe->ib_dev.dev, "added %s\n", netdev_name(rxe->ndev)); 1088 return 0; 1089 } 1090 1091 static const struct ib_device_ops rxe_dev_ops = { 1092 .owner = THIS_MODULE, 1093 .driver_id = RDMA_DRIVER_RXE, 1094 .uverbs_abi_ver = RXE_UVERBS_ABI_VERSION, 1095 1096 .alloc_hw_port_stats = rxe_ib_alloc_hw_port_stats, 1097 .alloc_mr = rxe_alloc_mr, 1098 .alloc_mw = rxe_alloc_mw, 1099 .alloc_pd = rxe_alloc_pd, 1100 .alloc_ucontext = rxe_alloc_ucontext, 1101 .attach_mcast = rxe_attach_mcast, 1102 .create_ah = rxe_create_ah, 1103 .create_cq = rxe_create_cq, 1104 .create_qp = rxe_create_qp, 1105 .create_srq = rxe_create_srq, 1106 .create_user_ah = rxe_create_ah, 1107 .dealloc_driver = rxe_dealloc, 1108 .dealloc_mw = rxe_dealloc_mw, 1109 .dealloc_pd = rxe_dealloc_pd, 1110 .dealloc_ucontext = rxe_dealloc_ucontext, 1111 .dereg_mr = rxe_dereg_mr, 1112 .destroy_ah = rxe_destroy_ah, 1113 .destroy_cq = rxe_destroy_cq, 1114 .destroy_qp = rxe_destroy_qp, 1115 .destroy_srq = rxe_destroy_srq, 1116 .detach_mcast = rxe_detach_mcast, 1117 .device_group = &rxe_attr_group, 1118 .enable_driver = rxe_enable_driver, 1119 .get_dma_mr = rxe_get_dma_mr, 1120 .get_hw_stats = rxe_ib_get_hw_stats, 1121 .get_link_layer = rxe_get_link_layer, 1122 .get_port_immutable = rxe_port_immutable, 1123 .map_mr_sg = rxe_map_mr_sg, 1124 .mmap = rxe_mmap, 1125 .modify_ah = rxe_modify_ah, 1126 .modify_device = rxe_modify_device, 1127 .modify_port = rxe_modify_port, 1128 .modify_qp = rxe_modify_qp, 1129 .modify_srq = rxe_modify_srq, 1130 .peek_cq = rxe_peek_cq, 1131 .poll_cq = rxe_poll_cq, 1132 .post_recv = rxe_post_recv, 1133 .post_send = rxe_post_send, 1134 .post_srq_recv = rxe_post_srq_recv, 1135 .query_ah = rxe_query_ah, 1136 .query_device = rxe_query_device, 1137 .query_pkey = rxe_query_pkey, 1138 .query_port = rxe_query_port, 1139 .query_qp = rxe_query_qp, 1140 .query_srq = rxe_query_srq, 1141 .reg_user_mr = rxe_reg_user_mr, 1142 .req_notify_cq = rxe_req_notify_cq, 1143 .resize_cq = rxe_resize_cq, 1144 1145 INIT_RDMA_OBJ_SIZE(ib_ah, rxe_ah, ibah), 1146 INIT_RDMA_OBJ_SIZE(ib_cq, rxe_cq, ibcq), 1147 INIT_RDMA_OBJ_SIZE(ib_pd, rxe_pd, ibpd), 1148 INIT_RDMA_OBJ_SIZE(ib_srq, rxe_srq, ibsrq), 1149 INIT_RDMA_OBJ_SIZE(ib_ucontext, rxe_ucontext, ibuc), 1150 INIT_RDMA_OBJ_SIZE(ib_mw, rxe_mw, ibmw), 1151 }; 1152 1153 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name) 1154 { 1155 int err; 1156 struct ib_device *dev = &rxe->ib_dev; 1157 struct crypto_shash *tfm; 1158 1159 strscpy(dev->node_desc, "rxe", sizeof(dev->node_desc)); 1160 1161 dev->node_type = RDMA_NODE_IB_CA; 1162 dev->phys_port_cnt = 1; 1163 dev->num_comp_vectors = num_possible_cpus(); 1164 dev->local_dma_lkey = 0; 1165 addrconf_addr_eui48((unsigned char *)&dev->node_guid, 1166 rxe->ndev->dev_addr); 1167 1168 dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) | 1169 BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ); 1170 1171 ib_set_device_ops(dev, &rxe_dev_ops); 1172 err = ib_device_set_netdev(&rxe->ib_dev, rxe->ndev, 1); 1173 if (err) 1174 return err; 1175 1176 tfm = crypto_alloc_shash("crc32", 0, 0); 1177 if (IS_ERR(tfm)) { 1178 pr_err("failed to allocate crc algorithm err:%ld\n", 1179 PTR_ERR(tfm)); 1180 return PTR_ERR(tfm); 1181 } 1182 rxe->tfm = tfm; 1183 1184 err = ib_register_device(dev, ibdev_name, NULL); 1185 if (err) 1186 pr_warn("%s failed with error %d\n", __func__, err); 1187 1188 /* 1189 * Note that rxe may be invalid at this point if another thread 1190 * unregistered it. 1191 */ 1192 return err; 1193 } 1194