1 /* 2 * NVMe over Fabrics RDMA host code. 3 * Copyright (c) 2015-2016 HGST, a Western Digital Company. 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms and conditions of the GNU General Public License, 7 * version 2, as published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12 * more details. 13 */ 14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 15 #include <linux/module.h> 16 #include <linux/init.h> 17 #include <linux/slab.h> 18 #include <rdma/mr_pool.h> 19 #include <linux/err.h> 20 #include <linux/string.h> 21 #include <linux/atomic.h> 22 #include <linux/blk-mq.h> 23 #include <linux/blk-mq-rdma.h> 24 #include <linux/types.h> 25 #include <linux/list.h> 26 #include <linux/mutex.h> 27 #include <linux/scatterlist.h> 28 #include <linux/nvme.h> 29 #include <asm/unaligned.h> 30 31 #include <rdma/ib_verbs.h> 32 #include <rdma/rdma_cm.h> 33 #include <linux/nvme-rdma.h> 34 35 #include "nvme.h" 36 #include "fabrics.h" 37 38 39 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */ 40 41 #define NVME_RDMA_MAX_SEGMENTS 256 42 43 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1 44 45 struct nvme_rdma_device { 46 struct ib_device *dev; 47 struct ib_pd *pd; 48 struct kref ref; 49 struct list_head entry; 50 }; 51 52 struct nvme_rdma_qe { 53 struct ib_cqe cqe; 54 void *data; 55 u64 dma; 56 }; 57 58 struct nvme_rdma_queue; 59 struct nvme_rdma_request { 60 struct nvme_request req; 61 struct ib_mr *mr; 62 struct nvme_rdma_qe sqe; 63 union nvme_result result; 64 __le16 status; 65 refcount_t ref; 66 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS]; 67 u32 num_sge; 68 int nents; 69 struct ib_reg_wr reg_wr; 70 struct ib_cqe reg_cqe; 71 struct nvme_rdma_queue *queue; 72 struct sg_table sg_table; 73 struct scatterlist first_sgl[]; 74 }; 75 76 enum nvme_rdma_queue_flags { 77 NVME_RDMA_Q_ALLOCATED = 0, 78 NVME_RDMA_Q_LIVE = 1, 79 NVME_RDMA_Q_TR_READY = 2, 80 }; 81 82 struct nvme_rdma_queue { 83 struct nvme_rdma_qe *rsp_ring; 84 int queue_size; 85 size_t cmnd_capsule_len; 86 struct nvme_rdma_ctrl *ctrl; 87 struct nvme_rdma_device *device; 88 struct ib_cq *ib_cq; 89 struct ib_qp *qp; 90 91 unsigned long flags; 92 struct rdma_cm_id *cm_id; 93 int cm_error; 94 struct completion cm_done; 95 }; 96 97 struct nvme_rdma_ctrl { 98 /* read only in the hot path */ 99 struct nvme_rdma_queue *queues; 100 101 /* other member variables */ 102 struct blk_mq_tag_set tag_set; 103 struct work_struct err_work; 104 105 struct nvme_rdma_qe async_event_sqe; 106 107 struct delayed_work reconnect_work; 108 109 struct list_head list; 110 111 struct blk_mq_tag_set admin_tag_set; 112 struct nvme_rdma_device *device; 113 114 u32 max_fr_pages; 115 116 struct sockaddr_storage addr; 117 struct sockaddr_storage src_addr; 118 119 struct nvme_ctrl ctrl; 120 }; 121 122 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl) 123 { 124 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl); 125 } 126 127 static LIST_HEAD(device_list); 128 static DEFINE_MUTEX(device_list_mutex); 129 130 static LIST_HEAD(nvme_rdma_ctrl_list); 131 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex); 132 133 /* 134 * Disabling this option makes small I/O goes faster, but is fundamentally 135 * unsafe. With it turned off we will have to register a global rkey that 136 * allows read and write access to all physical memory. 137 */ 138 static bool register_always = true; 139 module_param(register_always, bool, 0444); 140 MODULE_PARM_DESC(register_always, 141 "Use memory registration even for contiguous memory regions"); 142 143 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, 144 struct rdma_cm_event *event); 145 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc); 146 147 static const struct blk_mq_ops nvme_rdma_mq_ops; 148 static const struct blk_mq_ops nvme_rdma_admin_mq_ops; 149 150 /* XXX: really should move to a generic header sooner or later.. */ 151 static inline void put_unaligned_le24(u32 val, u8 *p) 152 { 153 *p++ = val; 154 *p++ = val >> 8; 155 *p++ = val >> 16; 156 } 157 158 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue) 159 { 160 return queue - queue->ctrl->queues; 161 } 162 163 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue) 164 { 165 return queue->cmnd_capsule_len - sizeof(struct nvme_command); 166 } 167 168 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, 169 size_t capsule_size, enum dma_data_direction dir) 170 { 171 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir); 172 kfree(qe->data); 173 } 174 175 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, 176 size_t capsule_size, enum dma_data_direction dir) 177 { 178 qe->data = kzalloc(capsule_size, GFP_KERNEL); 179 if (!qe->data) 180 return -ENOMEM; 181 182 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir); 183 if (ib_dma_mapping_error(ibdev, qe->dma)) { 184 kfree(qe->data); 185 return -ENOMEM; 186 } 187 188 return 0; 189 } 190 191 static void nvme_rdma_free_ring(struct ib_device *ibdev, 192 struct nvme_rdma_qe *ring, size_t ib_queue_size, 193 size_t capsule_size, enum dma_data_direction dir) 194 { 195 int i; 196 197 for (i = 0; i < ib_queue_size; i++) 198 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir); 199 kfree(ring); 200 } 201 202 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev, 203 size_t ib_queue_size, size_t capsule_size, 204 enum dma_data_direction dir) 205 { 206 struct nvme_rdma_qe *ring; 207 int i; 208 209 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL); 210 if (!ring) 211 return NULL; 212 213 for (i = 0; i < ib_queue_size; i++) { 214 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir)) 215 goto out_free_ring; 216 } 217 218 return ring; 219 220 out_free_ring: 221 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir); 222 return NULL; 223 } 224 225 static void nvme_rdma_qp_event(struct ib_event *event, void *context) 226 { 227 pr_debug("QP event %s (%d)\n", 228 ib_event_msg(event->event), event->event); 229 230 } 231 232 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue) 233 { 234 wait_for_completion_interruptible_timeout(&queue->cm_done, 235 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1); 236 return queue->cm_error; 237 } 238 239 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor) 240 { 241 struct nvme_rdma_device *dev = queue->device; 242 struct ib_qp_init_attr init_attr; 243 int ret; 244 245 memset(&init_attr, 0, sizeof(init_attr)); 246 init_attr.event_handler = nvme_rdma_qp_event; 247 /* +1 for drain */ 248 init_attr.cap.max_send_wr = factor * queue->queue_size + 1; 249 /* +1 for drain */ 250 init_attr.cap.max_recv_wr = queue->queue_size + 1; 251 init_attr.cap.max_recv_sge = 1; 252 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS; 253 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 254 init_attr.qp_type = IB_QPT_RC; 255 init_attr.send_cq = queue->ib_cq; 256 init_attr.recv_cq = queue->ib_cq; 257 258 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr); 259 260 queue->qp = queue->cm_id->qp; 261 return ret; 262 } 263 264 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set, 265 struct request *rq, unsigned int hctx_idx) 266 { 267 struct nvme_rdma_ctrl *ctrl = set->driver_data; 268 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 269 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0; 270 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx]; 271 struct nvme_rdma_device *dev = queue->device; 272 273 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command), 274 DMA_TO_DEVICE); 275 } 276 277 static int nvme_rdma_init_request(struct blk_mq_tag_set *set, 278 struct request *rq, unsigned int hctx_idx, 279 unsigned int numa_node) 280 { 281 struct nvme_rdma_ctrl *ctrl = set->driver_data; 282 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 283 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0; 284 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx]; 285 struct nvme_rdma_device *dev = queue->device; 286 struct ib_device *ibdev = dev->dev; 287 int ret; 288 289 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command), 290 DMA_TO_DEVICE); 291 if (ret) 292 return ret; 293 294 req->queue = queue; 295 296 return 0; 297 } 298 299 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 300 unsigned int hctx_idx) 301 { 302 struct nvme_rdma_ctrl *ctrl = data; 303 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1]; 304 305 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count); 306 307 hctx->driver_data = queue; 308 return 0; 309 } 310 311 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, 312 unsigned int hctx_idx) 313 { 314 struct nvme_rdma_ctrl *ctrl = data; 315 struct nvme_rdma_queue *queue = &ctrl->queues[0]; 316 317 BUG_ON(hctx_idx != 0); 318 319 hctx->driver_data = queue; 320 return 0; 321 } 322 323 static void nvme_rdma_free_dev(struct kref *ref) 324 { 325 struct nvme_rdma_device *ndev = 326 container_of(ref, struct nvme_rdma_device, ref); 327 328 mutex_lock(&device_list_mutex); 329 list_del(&ndev->entry); 330 mutex_unlock(&device_list_mutex); 331 332 ib_dealloc_pd(ndev->pd); 333 kfree(ndev); 334 } 335 336 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev) 337 { 338 kref_put(&dev->ref, nvme_rdma_free_dev); 339 } 340 341 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev) 342 { 343 return kref_get_unless_zero(&dev->ref); 344 } 345 346 static struct nvme_rdma_device * 347 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id) 348 { 349 struct nvme_rdma_device *ndev; 350 351 mutex_lock(&device_list_mutex); 352 list_for_each_entry(ndev, &device_list, entry) { 353 if (ndev->dev->node_guid == cm_id->device->node_guid && 354 nvme_rdma_dev_get(ndev)) 355 goto out_unlock; 356 } 357 358 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL); 359 if (!ndev) 360 goto out_err; 361 362 ndev->dev = cm_id->device; 363 kref_init(&ndev->ref); 364 365 ndev->pd = ib_alloc_pd(ndev->dev, 366 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY); 367 if (IS_ERR(ndev->pd)) 368 goto out_free_dev; 369 370 if (!(ndev->dev->attrs.device_cap_flags & 371 IB_DEVICE_MEM_MGT_EXTENSIONS)) { 372 dev_err(&ndev->dev->dev, 373 "Memory registrations not supported.\n"); 374 goto out_free_pd; 375 } 376 377 list_add(&ndev->entry, &device_list); 378 out_unlock: 379 mutex_unlock(&device_list_mutex); 380 return ndev; 381 382 out_free_pd: 383 ib_dealloc_pd(ndev->pd); 384 out_free_dev: 385 kfree(ndev); 386 out_err: 387 mutex_unlock(&device_list_mutex); 388 return NULL; 389 } 390 391 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue) 392 { 393 struct nvme_rdma_device *dev; 394 struct ib_device *ibdev; 395 396 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags)) 397 return; 398 399 dev = queue->device; 400 ibdev = dev->dev; 401 402 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs); 403 404 /* 405 * The cm_id object might have been destroyed during RDMA connection 406 * establishment error flow to avoid getting other cma events, thus 407 * the destruction of the QP shouldn't use rdma_cm API. 408 */ 409 ib_destroy_qp(queue->qp); 410 ib_free_cq(queue->ib_cq); 411 412 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, 413 sizeof(struct nvme_completion), DMA_FROM_DEVICE); 414 415 nvme_rdma_dev_put(dev); 416 } 417 418 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev) 419 { 420 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, 421 ibdev->attrs.max_fast_reg_page_list_len); 422 } 423 424 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue) 425 { 426 struct ib_device *ibdev; 427 const int send_wr_factor = 3; /* MR, SEND, INV */ 428 const int cq_factor = send_wr_factor + 1; /* + RECV */ 429 int comp_vector, idx = nvme_rdma_queue_idx(queue); 430 int ret; 431 432 queue->device = nvme_rdma_find_get_device(queue->cm_id); 433 if (!queue->device) { 434 dev_err(queue->cm_id->device->dev.parent, 435 "no client data found!\n"); 436 return -ECONNREFUSED; 437 } 438 ibdev = queue->device->dev; 439 440 /* 441 * Spread I/O queues completion vectors according their queue index. 442 * Admin queues can always go on completion vector 0. 443 */ 444 comp_vector = idx == 0 ? idx : idx - 1; 445 446 /* +1 for ib_stop_cq */ 447 queue->ib_cq = ib_alloc_cq(ibdev, queue, 448 cq_factor * queue->queue_size + 1, 449 comp_vector, IB_POLL_SOFTIRQ); 450 if (IS_ERR(queue->ib_cq)) { 451 ret = PTR_ERR(queue->ib_cq); 452 goto out_put_dev; 453 } 454 455 ret = nvme_rdma_create_qp(queue, send_wr_factor); 456 if (ret) 457 goto out_destroy_ib_cq; 458 459 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size, 460 sizeof(struct nvme_completion), DMA_FROM_DEVICE); 461 if (!queue->rsp_ring) { 462 ret = -ENOMEM; 463 goto out_destroy_qp; 464 } 465 466 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs, 467 queue->queue_size, 468 IB_MR_TYPE_MEM_REG, 469 nvme_rdma_get_max_fr_pages(ibdev)); 470 if (ret) { 471 dev_err(queue->ctrl->ctrl.device, 472 "failed to initialize MR pool sized %d for QID %d\n", 473 queue->queue_size, idx); 474 goto out_destroy_ring; 475 } 476 477 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags); 478 479 return 0; 480 481 out_destroy_ring: 482 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, 483 sizeof(struct nvme_completion), DMA_FROM_DEVICE); 484 out_destroy_qp: 485 rdma_destroy_qp(queue->cm_id); 486 out_destroy_ib_cq: 487 ib_free_cq(queue->ib_cq); 488 out_put_dev: 489 nvme_rdma_dev_put(queue->device); 490 return ret; 491 } 492 493 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl, 494 int idx, size_t queue_size) 495 { 496 struct nvme_rdma_queue *queue; 497 struct sockaddr *src_addr = NULL; 498 int ret; 499 500 queue = &ctrl->queues[idx]; 501 queue->ctrl = ctrl; 502 init_completion(&queue->cm_done); 503 504 if (idx > 0) 505 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16; 506 else 507 queue->cmnd_capsule_len = sizeof(struct nvme_command); 508 509 queue->queue_size = queue_size; 510 511 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue, 512 RDMA_PS_TCP, IB_QPT_RC); 513 if (IS_ERR(queue->cm_id)) { 514 dev_info(ctrl->ctrl.device, 515 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id)); 516 return PTR_ERR(queue->cm_id); 517 } 518 519 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) 520 src_addr = (struct sockaddr *)&ctrl->src_addr; 521 522 queue->cm_error = -ETIMEDOUT; 523 ret = rdma_resolve_addr(queue->cm_id, src_addr, 524 (struct sockaddr *)&ctrl->addr, 525 NVME_RDMA_CONNECT_TIMEOUT_MS); 526 if (ret) { 527 dev_info(ctrl->ctrl.device, 528 "rdma_resolve_addr failed (%d).\n", ret); 529 goto out_destroy_cm_id; 530 } 531 532 ret = nvme_rdma_wait_for_cm(queue); 533 if (ret) { 534 dev_info(ctrl->ctrl.device, 535 "rdma connection establishment failed (%d)\n", ret); 536 goto out_destroy_cm_id; 537 } 538 539 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags); 540 541 return 0; 542 543 out_destroy_cm_id: 544 rdma_destroy_id(queue->cm_id); 545 nvme_rdma_destroy_queue_ib(queue); 546 return ret; 547 } 548 549 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue) 550 { 551 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags)) 552 return; 553 554 rdma_disconnect(queue->cm_id); 555 ib_drain_qp(queue->qp); 556 } 557 558 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue) 559 { 560 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags)) 561 return; 562 563 nvme_rdma_destroy_queue_ib(queue); 564 rdma_destroy_id(queue->cm_id); 565 } 566 567 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl) 568 { 569 int i; 570 571 for (i = 1; i < ctrl->ctrl.queue_count; i++) 572 nvme_rdma_free_queue(&ctrl->queues[i]); 573 } 574 575 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl) 576 { 577 int i; 578 579 for (i = 1; i < ctrl->ctrl.queue_count; i++) 580 nvme_rdma_stop_queue(&ctrl->queues[i]); 581 } 582 583 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx) 584 { 585 int ret; 586 587 if (idx) 588 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx); 589 else 590 ret = nvmf_connect_admin_queue(&ctrl->ctrl); 591 592 if (!ret) 593 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags); 594 else 595 dev_info(ctrl->ctrl.device, 596 "failed to connect queue: %d ret=%d\n", idx, ret); 597 return ret; 598 } 599 600 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl) 601 { 602 int i, ret = 0; 603 604 for (i = 1; i < ctrl->ctrl.queue_count; i++) { 605 ret = nvme_rdma_start_queue(ctrl, i); 606 if (ret) 607 goto out_stop_queues; 608 } 609 610 return 0; 611 612 out_stop_queues: 613 for (i--; i >= 1; i--) 614 nvme_rdma_stop_queue(&ctrl->queues[i]); 615 return ret; 616 } 617 618 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl) 619 { 620 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; 621 struct ib_device *ibdev = ctrl->device->dev; 622 unsigned int nr_io_queues; 623 int i, ret; 624 625 nr_io_queues = min(opts->nr_io_queues, num_online_cpus()); 626 627 /* 628 * we map queues according to the device irq vectors for 629 * optimal locality so we don't need more queues than 630 * completion vectors. 631 */ 632 nr_io_queues = min_t(unsigned int, nr_io_queues, 633 ibdev->num_comp_vectors); 634 635 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues); 636 if (ret) 637 return ret; 638 639 ctrl->ctrl.queue_count = nr_io_queues + 1; 640 if (ctrl->ctrl.queue_count < 2) 641 return 0; 642 643 dev_info(ctrl->ctrl.device, 644 "creating %d I/O queues.\n", nr_io_queues); 645 646 for (i = 1; i < ctrl->ctrl.queue_count; i++) { 647 ret = nvme_rdma_alloc_queue(ctrl, i, 648 ctrl->ctrl.sqsize + 1); 649 if (ret) 650 goto out_free_queues; 651 } 652 653 return 0; 654 655 out_free_queues: 656 for (i--; i >= 1; i--) 657 nvme_rdma_free_queue(&ctrl->queues[i]); 658 659 return ret; 660 } 661 662 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl, 663 struct blk_mq_tag_set *set) 664 { 665 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); 666 667 blk_mq_free_tag_set(set); 668 nvme_rdma_dev_put(ctrl->device); 669 } 670 671 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl, 672 bool admin) 673 { 674 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); 675 struct blk_mq_tag_set *set; 676 int ret; 677 678 if (admin) { 679 set = &ctrl->admin_tag_set; 680 memset(set, 0, sizeof(*set)); 681 set->ops = &nvme_rdma_admin_mq_ops; 682 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH; 683 set->reserved_tags = 2; /* connect + keep-alive */ 684 set->numa_node = NUMA_NO_NODE; 685 set->cmd_size = sizeof(struct nvme_rdma_request) + 686 SG_CHUNK_SIZE * sizeof(struct scatterlist); 687 set->driver_data = ctrl; 688 set->nr_hw_queues = 1; 689 set->timeout = ADMIN_TIMEOUT; 690 set->flags = BLK_MQ_F_NO_SCHED; 691 } else { 692 set = &ctrl->tag_set; 693 memset(set, 0, sizeof(*set)); 694 set->ops = &nvme_rdma_mq_ops; 695 set->queue_depth = nctrl->sqsize + 1; 696 set->reserved_tags = 1; /* fabric connect */ 697 set->numa_node = NUMA_NO_NODE; 698 set->flags = BLK_MQ_F_SHOULD_MERGE; 699 set->cmd_size = sizeof(struct nvme_rdma_request) + 700 SG_CHUNK_SIZE * sizeof(struct scatterlist); 701 set->driver_data = ctrl; 702 set->nr_hw_queues = nctrl->queue_count - 1; 703 set->timeout = NVME_IO_TIMEOUT; 704 } 705 706 ret = blk_mq_alloc_tag_set(set); 707 if (ret) 708 goto out; 709 710 /* 711 * We need a reference on the device as long as the tag_set is alive, 712 * as the MRs in the request structures need a valid ib_device. 713 */ 714 ret = nvme_rdma_dev_get(ctrl->device); 715 if (!ret) { 716 ret = -EINVAL; 717 goto out_free_tagset; 718 } 719 720 return set; 721 722 out_free_tagset: 723 blk_mq_free_tag_set(set); 724 out: 725 return ERR_PTR(ret); 726 } 727 728 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl, 729 bool remove) 730 { 731 if (remove) { 732 blk_cleanup_queue(ctrl->ctrl.admin_q); 733 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset); 734 } 735 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe, 736 sizeof(struct nvme_command), DMA_TO_DEVICE); 737 nvme_rdma_free_queue(&ctrl->queues[0]); 738 } 739 740 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl, 741 bool new) 742 { 743 int error; 744 745 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH); 746 if (error) 747 return error; 748 749 ctrl->device = ctrl->queues[0].device; 750 751 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev); 752 753 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe, 754 sizeof(struct nvme_command), DMA_TO_DEVICE); 755 if (error) 756 goto out_free_queue; 757 758 if (new) { 759 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true); 760 if (IS_ERR(ctrl->ctrl.admin_tagset)) { 761 error = PTR_ERR(ctrl->ctrl.admin_tagset); 762 goto out_free_async_qe; 763 } 764 765 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set); 766 if (IS_ERR(ctrl->ctrl.admin_q)) { 767 error = PTR_ERR(ctrl->ctrl.admin_q); 768 goto out_free_tagset; 769 } 770 } 771 772 error = nvme_rdma_start_queue(ctrl, 0); 773 if (error) 774 goto out_cleanup_queue; 775 776 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP, 777 &ctrl->ctrl.cap); 778 if (error) { 779 dev_err(ctrl->ctrl.device, 780 "prop_get NVME_REG_CAP failed\n"); 781 goto out_stop_queue; 782 } 783 784 ctrl->ctrl.sqsize = 785 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize); 786 787 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap); 788 if (error) 789 goto out_stop_queue; 790 791 ctrl->ctrl.max_hw_sectors = 792 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9); 793 794 error = nvme_init_identify(&ctrl->ctrl); 795 if (error) 796 goto out_stop_queue; 797 798 return 0; 799 800 out_stop_queue: 801 nvme_rdma_stop_queue(&ctrl->queues[0]); 802 out_cleanup_queue: 803 if (new) 804 blk_cleanup_queue(ctrl->ctrl.admin_q); 805 out_free_tagset: 806 if (new) 807 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset); 808 out_free_async_qe: 809 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe, 810 sizeof(struct nvme_command), DMA_TO_DEVICE); 811 out_free_queue: 812 nvme_rdma_free_queue(&ctrl->queues[0]); 813 return error; 814 } 815 816 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl, 817 bool remove) 818 { 819 if (remove) { 820 blk_cleanup_queue(ctrl->ctrl.connect_q); 821 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset); 822 } 823 nvme_rdma_free_io_queues(ctrl); 824 } 825 826 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new) 827 { 828 int ret; 829 830 ret = nvme_rdma_alloc_io_queues(ctrl); 831 if (ret) 832 return ret; 833 834 if (new) { 835 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false); 836 if (IS_ERR(ctrl->ctrl.tagset)) { 837 ret = PTR_ERR(ctrl->ctrl.tagset); 838 goto out_free_io_queues; 839 } 840 841 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set); 842 if (IS_ERR(ctrl->ctrl.connect_q)) { 843 ret = PTR_ERR(ctrl->ctrl.connect_q); 844 goto out_free_tag_set; 845 } 846 } else { 847 blk_mq_update_nr_hw_queues(&ctrl->tag_set, 848 ctrl->ctrl.queue_count - 1); 849 } 850 851 ret = nvme_rdma_start_io_queues(ctrl); 852 if (ret) 853 goto out_cleanup_connect_q; 854 855 return 0; 856 857 out_cleanup_connect_q: 858 if (new) 859 blk_cleanup_queue(ctrl->ctrl.connect_q); 860 out_free_tag_set: 861 if (new) 862 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset); 863 out_free_io_queues: 864 nvme_rdma_free_io_queues(ctrl); 865 return ret; 866 } 867 868 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl) 869 { 870 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); 871 872 cancel_work_sync(&ctrl->err_work); 873 cancel_delayed_work_sync(&ctrl->reconnect_work); 874 } 875 876 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl) 877 { 878 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); 879 880 if (list_empty(&ctrl->list)) 881 goto free_ctrl; 882 883 mutex_lock(&nvme_rdma_ctrl_mutex); 884 list_del(&ctrl->list); 885 mutex_unlock(&nvme_rdma_ctrl_mutex); 886 887 nvmf_free_options(nctrl->opts); 888 free_ctrl: 889 kfree(ctrl->queues); 890 kfree(ctrl); 891 } 892 893 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl) 894 { 895 /* If we are resetting/deleting then do nothing */ 896 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) { 897 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW || 898 ctrl->ctrl.state == NVME_CTRL_LIVE); 899 return; 900 } 901 902 if (nvmf_should_reconnect(&ctrl->ctrl)) { 903 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n", 904 ctrl->ctrl.opts->reconnect_delay); 905 queue_delayed_work(nvme_wq, &ctrl->reconnect_work, 906 ctrl->ctrl.opts->reconnect_delay * HZ); 907 } else { 908 nvme_delete_ctrl(&ctrl->ctrl); 909 } 910 } 911 912 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work) 913 { 914 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work), 915 struct nvme_rdma_ctrl, reconnect_work); 916 bool changed; 917 int ret; 918 919 ++ctrl->ctrl.nr_reconnects; 920 921 ret = nvme_rdma_configure_admin_queue(ctrl, false); 922 if (ret) 923 goto requeue; 924 925 if (ctrl->ctrl.queue_count > 1) { 926 ret = nvme_rdma_configure_io_queues(ctrl, false); 927 if (ret) 928 goto destroy_admin; 929 } 930 931 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); 932 if (!changed) { 933 /* state change failure is ok if we're in DELETING state */ 934 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING); 935 return; 936 } 937 938 nvme_start_ctrl(&ctrl->ctrl); 939 940 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n", 941 ctrl->ctrl.nr_reconnects); 942 943 ctrl->ctrl.nr_reconnects = 0; 944 945 return; 946 947 destroy_admin: 948 nvme_rdma_stop_queue(&ctrl->queues[0]); 949 nvme_rdma_destroy_admin_queue(ctrl, false); 950 requeue: 951 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n", 952 ctrl->ctrl.nr_reconnects); 953 nvme_rdma_reconnect_or_remove(ctrl); 954 } 955 956 static void nvme_rdma_error_recovery_work(struct work_struct *work) 957 { 958 struct nvme_rdma_ctrl *ctrl = container_of(work, 959 struct nvme_rdma_ctrl, err_work); 960 961 nvme_stop_keep_alive(&ctrl->ctrl); 962 963 if (ctrl->ctrl.queue_count > 1) { 964 nvme_stop_queues(&ctrl->ctrl); 965 nvme_rdma_stop_io_queues(ctrl); 966 blk_mq_tagset_busy_iter(&ctrl->tag_set, 967 nvme_cancel_request, &ctrl->ctrl); 968 nvme_rdma_destroy_io_queues(ctrl, false); 969 } 970 971 blk_mq_quiesce_queue(ctrl->ctrl.admin_q); 972 nvme_rdma_stop_queue(&ctrl->queues[0]); 973 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, 974 nvme_cancel_request, &ctrl->ctrl); 975 nvme_rdma_destroy_admin_queue(ctrl, false); 976 977 /* 978 * queues are not a live anymore, so restart the queues to fail fast 979 * new IO 980 */ 981 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); 982 nvme_start_queues(&ctrl->ctrl); 983 984 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { 985 /* state change failure is ok if we're in DELETING state */ 986 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING); 987 return; 988 } 989 990 nvme_rdma_reconnect_or_remove(ctrl); 991 } 992 993 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl) 994 { 995 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING)) 996 return; 997 998 queue_work(nvme_wq, &ctrl->err_work); 999 } 1000 1001 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc, 1002 const char *op) 1003 { 1004 struct nvme_rdma_queue *queue = cq->cq_context; 1005 struct nvme_rdma_ctrl *ctrl = queue->ctrl; 1006 1007 if (ctrl->ctrl.state == NVME_CTRL_LIVE) 1008 dev_info(ctrl->ctrl.device, 1009 "%s for CQE 0x%p failed with status %s (%d)\n", 1010 op, wc->wr_cqe, 1011 ib_wc_status_msg(wc->status), wc->status); 1012 nvme_rdma_error_recovery(ctrl); 1013 } 1014 1015 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc) 1016 { 1017 if (unlikely(wc->status != IB_WC_SUCCESS)) 1018 nvme_rdma_wr_error(cq, wc, "MEMREG"); 1019 } 1020 1021 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc) 1022 { 1023 struct nvme_rdma_request *req = 1024 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe); 1025 struct request *rq = blk_mq_rq_from_pdu(req); 1026 1027 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1028 nvme_rdma_wr_error(cq, wc, "LOCAL_INV"); 1029 return; 1030 } 1031 1032 if (refcount_dec_and_test(&req->ref)) 1033 nvme_end_request(rq, req->status, req->result); 1034 1035 } 1036 1037 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue, 1038 struct nvme_rdma_request *req) 1039 { 1040 struct ib_send_wr *bad_wr; 1041 struct ib_send_wr wr = { 1042 .opcode = IB_WR_LOCAL_INV, 1043 .next = NULL, 1044 .num_sge = 0, 1045 .send_flags = IB_SEND_SIGNALED, 1046 .ex.invalidate_rkey = req->mr->rkey, 1047 }; 1048 1049 req->reg_cqe.done = nvme_rdma_inv_rkey_done; 1050 wr.wr_cqe = &req->reg_cqe; 1051 1052 return ib_post_send(queue->qp, &wr, &bad_wr); 1053 } 1054 1055 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue, 1056 struct request *rq) 1057 { 1058 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1059 struct nvme_rdma_device *dev = queue->device; 1060 struct ib_device *ibdev = dev->dev; 1061 1062 if (!blk_rq_payload_bytes(rq)) 1063 return; 1064 1065 if (req->mr) { 1066 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr); 1067 req->mr = NULL; 1068 } 1069 1070 ib_dma_unmap_sg(ibdev, req->sg_table.sgl, 1071 req->nents, rq_data_dir(rq) == 1072 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE); 1073 1074 nvme_cleanup_cmd(rq); 1075 sg_free_table_chained(&req->sg_table, true); 1076 } 1077 1078 static int nvme_rdma_set_sg_null(struct nvme_command *c) 1079 { 1080 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1081 1082 sg->addr = 0; 1083 put_unaligned_le24(0, sg->length); 1084 put_unaligned_le32(0, sg->key); 1085 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; 1086 return 0; 1087 } 1088 1089 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue, 1090 struct nvme_rdma_request *req, struct nvme_command *c) 1091 { 1092 struct nvme_sgl_desc *sg = &c->common.dptr.sgl; 1093 1094 req->sge[1].addr = sg_dma_address(req->sg_table.sgl); 1095 req->sge[1].length = sg_dma_len(req->sg_table.sgl); 1096 req->sge[1].lkey = queue->device->pd->local_dma_lkey; 1097 1098 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff); 1099 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl)); 1100 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET; 1101 1102 req->num_sge++; 1103 return 0; 1104 } 1105 1106 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue, 1107 struct nvme_rdma_request *req, struct nvme_command *c) 1108 { 1109 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1110 1111 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl)); 1112 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length); 1113 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key); 1114 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; 1115 return 0; 1116 } 1117 1118 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue, 1119 struct nvme_rdma_request *req, struct nvme_command *c, 1120 int count) 1121 { 1122 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1123 int nr; 1124 1125 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs); 1126 if (WARN_ON_ONCE(!req->mr)) 1127 return -EAGAIN; 1128 1129 /* 1130 * Align the MR to a 4K page size to match the ctrl page size and 1131 * the block virtual boundary. 1132 */ 1133 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K); 1134 if (unlikely(nr < count)) { 1135 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr); 1136 req->mr = NULL; 1137 if (nr < 0) 1138 return nr; 1139 return -EINVAL; 1140 } 1141 1142 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey)); 1143 1144 req->reg_cqe.done = nvme_rdma_memreg_done; 1145 memset(&req->reg_wr, 0, sizeof(req->reg_wr)); 1146 req->reg_wr.wr.opcode = IB_WR_REG_MR; 1147 req->reg_wr.wr.wr_cqe = &req->reg_cqe; 1148 req->reg_wr.wr.num_sge = 0; 1149 req->reg_wr.mr = req->mr; 1150 req->reg_wr.key = req->mr->rkey; 1151 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE | 1152 IB_ACCESS_REMOTE_READ | 1153 IB_ACCESS_REMOTE_WRITE; 1154 1155 sg->addr = cpu_to_le64(req->mr->iova); 1156 put_unaligned_le24(req->mr->length, sg->length); 1157 put_unaligned_le32(req->mr->rkey, sg->key); 1158 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) | 1159 NVME_SGL_FMT_INVALIDATE; 1160 1161 return 0; 1162 } 1163 1164 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue, 1165 struct request *rq, struct nvme_command *c) 1166 { 1167 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1168 struct nvme_rdma_device *dev = queue->device; 1169 struct ib_device *ibdev = dev->dev; 1170 int count, ret; 1171 1172 req->num_sge = 1; 1173 refcount_set(&req->ref, 2); /* send and recv completions */ 1174 1175 c->common.flags |= NVME_CMD_SGL_METABUF; 1176 1177 if (!blk_rq_payload_bytes(rq)) 1178 return nvme_rdma_set_sg_null(c); 1179 1180 req->sg_table.sgl = req->first_sgl; 1181 ret = sg_alloc_table_chained(&req->sg_table, 1182 blk_rq_nr_phys_segments(rq), req->sg_table.sgl); 1183 if (ret) 1184 return -ENOMEM; 1185 1186 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl); 1187 1188 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents, 1189 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE); 1190 if (unlikely(count <= 0)) { 1191 ret = -EIO; 1192 goto out_free_table; 1193 } 1194 1195 if (count == 1) { 1196 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) && 1197 blk_rq_payload_bytes(rq) <= 1198 nvme_rdma_inline_data_size(queue)) { 1199 ret = nvme_rdma_map_sg_inline(queue, req, c); 1200 goto out; 1201 } 1202 1203 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) { 1204 ret = nvme_rdma_map_sg_single(queue, req, c); 1205 goto out; 1206 } 1207 } 1208 1209 ret = nvme_rdma_map_sg_fr(queue, req, c, count); 1210 out: 1211 if (unlikely(ret)) 1212 goto out_unmap_sg; 1213 1214 return 0; 1215 1216 out_unmap_sg: 1217 ib_dma_unmap_sg(ibdev, req->sg_table.sgl, 1218 req->nents, rq_data_dir(rq) == 1219 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE); 1220 out_free_table: 1221 sg_free_table_chained(&req->sg_table, true); 1222 return ret; 1223 } 1224 1225 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc) 1226 { 1227 struct nvme_rdma_qe *qe = 1228 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); 1229 struct nvme_rdma_request *req = 1230 container_of(qe, struct nvme_rdma_request, sqe); 1231 struct request *rq = blk_mq_rq_from_pdu(req); 1232 1233 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1234 nvme_rdma_wr_error(cq, wc, "SEND"); 1235 return; 1236 } 1237 1238 if (refcount_dec_and_test(&req->ref)) 1239 nvme_end_request(rq, req->status, req->result); 1240 } 1241 1242 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue, 1243 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge, 1244 struct ib_send_wr *first) 1245 { 1246 struct ib_send_wr wr, *bad_wr; 1247 int ret; 1248 1249 sge->addr = qe->dma; 1250 sge->length = sizeof(struct nvme_command), 1251 sge->lkey = queue->device->pd->local_dma_lkey; 1252 1253 wr.next = NULL; 1254 wr.wr_cqe = &qe->cqe; 1255 wr.sg_list = sge; 1256 wr.num_sge = num_sge; 1257 wr.opcode = IB_WR_SEND; 1258 wr.send_flags = IB_SEND_SIGNALED; 1259 1260 if (first) 1261 first->next = ≀ 1262 else 1263 first = ≀ 1264 1265 ret = ib_post_send(queue->qp, first, &bad_wr); 1266 if (unlikely(ret)) { 1267 dev_err(queue->ctrl->ctrl.device, 1268 "%s failed with error code %d\n", __func__, ret); 1269 } 1270 return ret; 1271 } 1272 1273 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue, 1274 struct nvme_rdma_qe *qe) 1275 { 1276 struct ib_recv_wr wr, *bad_wr; 1277 struct ib_sge list; 1278 int ret; 1279 1280 list.addr = qe->dma; 1281 list.length = sizeof(struct nvme_completion); 1282 list.lkey = queue->device->pd->local_dma_lkey; 1283 1284 qe->cqe.done = nvme_rdma_recv_done; 1285 1286 wr.next = NULL; 1287 wr.wr_cqe = &qe->cqe; 1288 wr.sg_list = &list; 1289 wr.num_sge = 1; 1290 1291 ret = ib_post_recv(queue->qp, &wr, &bad_wr); 1292 if (unlikely(ret)) { 1293 dev_err(queue->ctrl->ctrl.device, 1294 "%s failed with error code %d\n", __func__, ret); 1295 } 1296 return ret; 1297 } 1298 1299 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue) 1300 { 1301 u32 queue_idx = nvme_rdma_queue_idx(queue); 1302 1303 if (queue_idx == 0) 1304 return queue->ctrl->admin_tag_set.tags[queue_idx]; 1305 return queue->ctrl->tag_set.tags[queue_idx - 1]; 1306 } 1307 1308 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc) 1309 { 1310 if (unlikely(wc->status != IB_WC_SUCCESS)) 1311 nvme_rdma_wr_error(cq, wc, "ASYNC"); 1312 } 1313 1314 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg) 1315 { 1316 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg); 1317 struct nvme_rdma_queue *queue = &ctrl->queues[0]; 1318 struct ib_device *dev = queue->device->dev; 1319 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe; 1320 struct nvme_command *cmd = sqe->data; 1321 struct ib_sge sge; 1322 int ret; 1323 1324 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE); 1325 1326 memset(cmd, 0, sizeof(*cmd)); 1327 cmd->common.opcode = nvme_admin_async_event; 1328 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH; 1329 cmd->common.flags |= NVME_CMD_SGL_METABUF; 1330 nvme_rdma_set_sg_null(cmd); 1331 1332 sqe->cqe.done = nvme_rdma_async_done; 1333 1334 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd), 1335 DMA_TO_DEVICE); 1336 1337 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL); 1338 WARN_ON_ONCE(ret); 1339 } 1340 1341 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue, 1342 struct nvme_completion *cqe, struct ib_wc *wc, int tag) 1343 { 1344 struct request *rq; 1345 struct nvme_rdma_request *req; 1346 int ret = 0; 1347 1348 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id); 1349 if (!rq) { 1350 dev_err(queue->ctrl->ctrl.device, 1351 "tag 0x%x on QP %#x not found\n", 1352 cqe->command_id, queue->qp->qp_num); 1353 nvme_rdma_error_recovery(queue->ctrl); 1354 return ret; 1355 } 1356 req = blk_mq_rq_to_pdu(rq); 1357 1358 req->status = cqe->status; 1359 req->result = cqe->result; 1360 1361 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) { 1362 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) { 1363 dev_err(queue->ctrl->ctrl.device, 1364 "Bogus remote invalidation for rkey %#x\n", 1365 req->mr->rkey); 1366 nvme_rdma_error_recovery(queue->ctrl); 1367 } 1368 } else if (req->mr) { 1369 ret = nvme_rdma_inv_rkey(queue, req); 1370 if (unlikely(ret < 0)) { 1371 dev_err(queue->ctrl->ctrl.device, 1372 "Queueing INV WR for rkey %#x failed (%d)\n", 1373 req->mr->rkey, ret); 1374 nvme_rdma_error_recovery(queue->ctrl); 1375 } 1376 /* the local invalidation completion will end the request */ 1377 return 0; 1378 } 1379 1380 if (refcount_dec_and_test(&req->ref)) { 1381 if (rq->tag == tag) 1382 ret = 1; 1383 nvme_end_request(rq, req->status, req->result); 1384 } 1385 1386 return ret; 1387 } 1388 1389 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag) 1390 { 1391 struct nvme_rdma_qe *qe = 1392 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); 1393 struct nvme_rdma_queue *queue = cq->cq_context; 1394 struct ib_device *ibdev = queue->device->dev; 1395 struct nvme_completion *cqe = qe->data; 1396 const size_t len = sizeof(struct nvme_completion); 1397 int ret = 0; 1398 1399 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1400 nvme_rdma_wr_error(cq, wc, "RECV"); 1401 return 0; 1402 } 1403 1404 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE); 1405 /* 1406 * AEN requests are special as they don't time out and can 1407 * survive any kind of queue freeze and often don't respond to 1408 * aborts. We don't even bother to allocate a struct request 1409 * for them but rather special case them here. 1410 */ 1411 if (unlikely(nvme_rdma_queue_idx(queue) == 0 && 1412 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) 1413 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status, 1414 &cqe->result); 1415 else 1416 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag); 1417 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE); 1418 1419 nvme_rdma_post_recv(queue, qe); 1420 return ret; 1421 } 1422 1423 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc) 1424 { 1425 __nvme_rdma_recv_done(cq, wc, -1); 1426 } 1427 1428 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue) 1429 { 1430 int ret, i; 1431 1432 for (i = 0; i < queue->queue_size; i++) { 1433 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]); 1434 if (ret) 1435 goto out_destroy_queue_ib; 1436 } 1437 1438 return 0; 1439 1440 out_destroy_queue_ib: 1441 nvme_rdma_destroy_queue_ib(queue); 1442 return ret; 1443 } 1444 1445 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue, 1446 struct rdma_cm_event *ev) 1447 { 1448 struct rdma_cm_id *cm_id = queue->cm_id; 1449 int status = ev->status; 1450 const char *rej_msg; 1451 const struct nvme_rdma_cm_rej *rej_data; 1452 u8 rej_data_len; 1453 1454 rej_msg = rdma_reject_msg(cm_id, status); 1455 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len); 1456 1457 if (rej_data && rej_data_len >= sizeof(u16)) { 1458 u16 sts = le16_to_cpu(rej_data->sts); 1459 1460 dev_err(queue->ctrl->ctrl.device, 1461 "Connect rejected: status %d (%s) nvme status %d (%s).\n", 1462 status, rej_msg, sts, nvme_rdma_cm_msg(sts)); 1463 } else { 1464 dev_err(queue->ctrl->ctrl.device, 1465 "Connect rejected: status %d (%s).\n", status, rej_msg); 1466 } 1467 1468 return -ECONNRESET; 1469 } 1470 1471 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue) 1472 { 1473 int ret; 1474 1475 ret = nvme_rdma_create_queue_ib(queue); 1476 if (ret) 1477 return ret; 1478 1479 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS); 1480 if (ret) { 1481 dev_err(queue->ctrl->ctrl.device, 1482 "rdma_resolve_route failed (%d).\n", 1483 queue->cm_error); 1484 goto out_destroy_queue; 1485 } 1486 1487 return 0; 1488 1489 out_destroy_queue: 1490 nvme_rdma_destroy_queue_ib(queue); 1491 return ret; 1492 } 1493 1494 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue) 1495 { 1496 struct nvme_rdma_ctrl *ctrl = queue->ctrl; 1497 struct rdma_conn_param param = { }; 1498 struct nvme_rdma_cm_req priv = { }; 1499 int ret; 1500 1501 param.qp_num = queue->qp->qp_num; 1502 param.flow_control = 1; 1503 1504 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom; 1505 /* maximum retry count */ 1506 param.retry_count = 7; 1507 param.rnr_retry_count = 7; 1508 param.private_data = &priv; 1509 param.private_data_len = sizeof(priv); 1510 1511 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0); 1512 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue)); 1513 /* 1514 * set the admin queue depth to the minimum size 1515 * specified by the Fabrics standard. 1516 */ 1517 if (priv.qid == 0) { 1518 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH); 1519 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1); 1520 } else { 1521 /* 1522 * current interpretation of the fabrics spec 1523 * is at minimum you make hrqsize sqsize+1, or a 1524 * 1's based representation of sqsize. 1525 */ 1526 priv.hrqsize = cpu_to_le16(queue->queue_size); 1527 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize); 1528 } 1529 1530 ret = rdma_connect(queue->cm_id, ¶m); 1531 if (ret) { 1532 dev_err(ctrl->ctrl.device, 1533 "rdma_connect failed (%d).\n", ret); 1534 goto out_destroy_queue_ib; 1535 } 1536 1537 return 0; 1538 1539 out_destroy_queue_ib: 1540 nvme_rdma_destroy_queue_ib(queue); 1541 return ret; 1542 } 1543 1544 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, 1545 struct rdma_cm_event *ev) 1546 { 1547 struct nvme_rdma_queue *queue = cm_id->context; 1548 int cm_error = 0; 1549 1550 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n", 1551 rdma_event_msg(ev->event), ev->event, 1552 ev->status, cm_id); 1553 1554 switch (ev->event) { 1555 case RDMA_CM_EVENT_ADDR_RESOLVED: 1556 cm_error = nvme_rdma_addr_resolved(queue); 1557 break; 1558 case RDMA_CM_EVENT_ROUTE_RESOLVED: 1559 cm_error = nvme_rdma_route_resolved(queue); 1560 break; 1561 case RDMA_CM_EVENT_ESTABLISHED: 1562 queue->cm_error = nvme_rdma_conn_established(queue); 1563 /* complete cm_done regardless of success/failure */ 1564 complete(&queue->cm_done); 1565 return 0; 1566 case RDMA_CM_EVENT_REJECTED: 1567 nvme_rdma_destroy_queue_ib(queue); 1568 cm_error = nvme_rdma_conn_rejected(queue, ev); 1569 break; 1570 case RDMA_CM_EVENT_ROUTE_ERROR: 1571 case RDMA_CM_EVENT_CONNECT_ERROR: 1572 case RDMA_CM_EVENT_UNREACHABLE: 1573 nvme_rdma_destroy_queue_ib(queue); 1574 case RDMA_CM_EVENT_ADDR_ERROR: 1575 dev_dbg(queue->ctrl->ctrl.device, 1576 "CM error event %d\n", ev->event); 1577 cm_error = -ECONNRESET; 1578 break; 1579 case RDMA_CM_EVENT_DISCONNECTED: 1580 case RDMA_CM_EVENT_ADDR_CHANGE: 1581 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 1582 dev_dbg(queue->ctrl->ctrl.device, 1583 "disconnect received - connection closed\n"); 1584 nvme_rdma_error_recovery(queue->ctrl); 1585 break; 1586 case RDMA_CM_EVENT_DEVICE_REMOVAL: 1587 /* device removal is handled via the ib_client API */ 1588 break; 1589 default: 1590 dev_err(queue->ctrl->ctrl.device, 1591 "Unexpected RDMA CM event (%d)\n", ev->event); 1592 nvme_rdma_error_recovery(queue->ctrl); 1593 break; 1594 } 1595 1596 if (cm_error) { 1597 queue->cm_error = cm_error; 1598 complete(&queue->cm_done); 1599 } 1600 1601 return 0; 1602 } 1603 1604 static enum blk_eh_timer_return 1605 nvme_rdma_timeout(struct request *rq, bool reserved) 1606 { 1607 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1608 1609 dev_warn(req->queue->ctrl->ctrl.device, 1610 "I/O %d QID %d timeout, reset controller\n", 1611 rq->tag, nvme_rdma_queue_idx(req->queue)); 1612 1613 /* queue error recovery */ 1614 nvme_rdma_error_recovery(req->queue->ctrl); 1615 1616 /* fail with DNR on cmd timeout */ 1617 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR; 1618 1619 return BLK_EH_DONE; 1620 } 1621 1622 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx, 1623 const struct blk_mq_queue_data *bd) 1624 { 1625 struct nvme_ns *ns = hctx->queue->queuedata; 1626 struct nvme_rdma_queue *queue = hctx->driver_data; 1627 struct request *rq = bd->rq; 1628 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1629 struct nvme_rdma_qe *sqe = &req->sqe; 1630 struct nvme_command *c = sqe->data; 1631 struct ib_device *dev; 1632 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags); 1633 blk_status_t ret; 1634 int err; 1635 1636 WARN_ON_ONCE(rq->tag < 0); 1637 1638 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready)) 1639 return nvmf_fail_nonready_command(rq); 1640 1641 dev = queue->device->dev; 1642 ib_dma_sync_single_for_cpu(dev, sqe->dma, 1643 sizeof(struct nvme_command), DMA_TO_DEVICE); 1644 1645 ret = nvme_setup_cmd(ns, rq, c); 1646 if (ret) 1647 return ret; 1648 1649 blk_mq_start_request(rq); 1650 1651 err = nvme_rdma_map_data(queue, rq, c); 1652 if (unlikely(err < 0)) { 1653 dev_err(queue->ctrl->ctrl.device, 1654 "Failed to map data (%d)\n", err); 1655 nvme_cleanup_cmd(rq); 1656 goto err; 1657 } 1658 1659 sqe->cqe.done = nvme_rdma_send_done; 1660 1661 ib_dma_sync_single_for_device(dev, sqe->dma, 1662 sizeof(struct nvme_command), DMA_TO_DEVICE); 1663 1664 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge, 1665 req->mr ? &req->reg_wr.wr : NULL); 1666 if (unlikely(err)) { 1667 nvme_rdma_unmap_data(queue, rq); 1668 goto err; 1669 } 1670 1671 return BLK_STS_OK; 1672 err: 1673 if (err == -ENOMEM || err == -EAGAIN) 1674 return BLK_STS_RESOURCE; 1675 return BLK_STS_IOERR; 1676 } 1677 1678 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag) 1679 { 1680 struct nvme_rdma_queue *queue = hctx->driver_data; 1681 struct ib_cq *cq = queue->ib_cq; 1682 struct ib_wc wc; 1683 int found = 0; 1684 1685 while (ib_poll_cq(cq, 1, &wc) > 0) { 1686 struct ib_cqe *cqe = wc.wr_cqe; 1687 1688 if (cqe) { 1689 if (cqe->done == nvme_rdma_recv_done) 1690 found |= __nvme_rdma_recv_done(cq, &wc, tag); 1691 else 1692 cqe->done(cq, &wc); 1693 } 1694 } 1695 1696 return found; 1697 } 1698 1699 static void nvme_rdma_complete_rq(struct request *rq) 1700 { 1701 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1702 1703 nvme_rdma_unmap_data(req->queue, rq); 1704 nvme_complete_rq(rq); 1705 } 1706 1707 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set) 1708 { 1709 struct nvme_rdma_ctrl *ctrl = set->driver_data; 1710 1711 return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0); 1712 } 1713 1714 static const struct blk_mq_ops nvme_rdma_mq_ops = { 1715 .queue_rq = nvme_rdma_queue_rq, 1716 .complete = nvme_rdma_complete_rq, 1717 .init_request = nvme_rdma_init_request, 1718 .exit_request = nvme_rdma_exit_request, 1719 .init_hctx = nvme_rdma_init_hctx, 1720 .poll = nvme_rdma_poll, 1721 .timeout = nvme_rdma_timeout, 1722 .map_queues = nvme_rdma_map_queues, 1723 }; 1724 1725 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = { 1726 .queue_rq = nvme_rdma_queue_rq, 1727 .complete = nvme_rdma_complete_rq, 1728 .init_request = nvme_rdma_init_request, 1729 .exit_request = nvme_rdma_exit_request, 1730 .init_hctx = nvme_rdma_init_admin_hctx, 1731 .timeout = nvme_rdma_timeout, 1732 }; 1733 1734 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown) 1735 { 1736 if (ctrl->ctrl.queue_count > 1) { 1737 nvme_stop_queues(&ctrl->ctrl); 1738 nvme_rdma_stop_io_queues(ctrl); 1739 blk_mq_tagset_busy_iter(&ctrl->tag_set, 1740 nvme_cancel_request, &ctrl->ctrl); 1741 nvme_rdma_destroy_io_queues(ctrl, shutdown); 1742 } 1743 1744 if (shutdown) 1745 nvme_shutdown_ctrl(&ctrl->ctrl); 1746 else 1747 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap); 1748 1749 blk_mq_quiesce_queue(ctrl->ctrl.admin_q); 1750 nvme_rdma_stop_queue(&ctrl->queues[0]); 1751 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, 1752 nvme_cancel_request, &ctrl->ctrl); 1753 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); 1754 nvme_rdma_destroy_admin_queue(ctrl, shutdown); 1755 } 1756 1757 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl) 1758 { 1759 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true); 1760 } 1761 1762 static void nvme_rdma_reset_ctrl_work(struct work_struct *work) 1763 { 1764 struct nvme_rdma_ctrl *ctrl = 1765 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work); 1766 int ret; 1767 bool changed; 1768 1769 nvme_stop_ctrl(&ctrl->ctrl); 1770 nvme_rdma_shutdown_ctrl(ctrl, false); 1771 1772 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { 1773 /* state change failure should never happen */ 1774 WARN_ON_ONCE(1); 1775 return; 1776 } 1777 1778 ret = nvme_rdma_configure_admin_queue(ctrl, false); 1779 if (ret) 1780 goto out_fail; 1781 1782 if (ctrl->ctrl.queue_count > 1) { 1783 ret = nvme_rdma_configure_io_queues(ctrl, false); 1784 if (ret) 1785 goto out_fail; 1786 } 1787 1788 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); 1789 if (!changed) { 1790 /* state change failure is ok if we're in DELETING state */ 1791 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING); 1792 return; 1793 } 1794 1795 nvme_start_ctrl(&ctrl->ctrl); 1796 1797 return; 1798 1799 out_fail: 1800 ++ctrl->ctrl.nr_reconnects; 1801 nvme_rdma_reconnect_or_remove(ctrl); 1802 } 1803 1804 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = { 1805 .name = "rdma", 1806 .module = THIS_MODULE, 1807 .flags = NVME_F_FABRICS, 1808 .reg_read32 = nvmf_reg_read32, 1809 .reg_read64 = nvmf_reg_read64, 1810 .reg_write32 = nvmf_reg_write32, 1811 .free_ctrl = nvme_rdma_free_ctrl, 1812 .submit_async_event = nvme_rdma_submit_async_event, 1813 .delete_ctrl = nvme_rdma_delete_ctrl, 1814 .get_address = nvmf_get_address, 1815 .stop_ctrl = nvme_rdma_stop_ctrl, 1816 }; 1817 1818 static inline bool 1819 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl, 1820 struct nvmf_ctrl_options *opts) 1821 { 1822 char *stdport = __stringify(NVME_RDMA_IP_PORT); 1823 1824 1825 if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) || 1826 strcmp(opts->traddr, ctrl->ctrl.opts->traddr)) 1827 return false; 1828 1829 if (opts->mask & NVMF_OPT_TRSVCID && 1830 ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) { 1831 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid)) 1832 return false; 1833 } else if (opts->mask & NVMF_OPT_TRSVCID) { 1834 if (strcmp(opts->trsvcid, stdport)) 1835 return false; 1836 } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) { 1837 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid)) 1838 return false; 1839 } 1840 /* else, it's a match as both have stdport. Fall to next checks */ 1841 1842 /* 1843 * checking the local address is rough. In most cases, one 1844 * is not specified and the host port is selected by the stack. 1845 * 1846 * Assume no match if: 1847 * local address is specified and address is not the same 1848 * local address is not specified but remote is, or vice versa 1849 * (admin using specific host_traddr when it matters). 1850 */ 1851 if (opts->mask & NVMF_OPT_HOST_TRADDR && 1852 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) { 1853 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr)) 1854 return false; 1855 } else if (opts->mask & NVMF_OPT_HOST_TRADDR || 1856 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) 1857 return false; 1858 /* 1859 * if neither controller had an host port specified, assume it's 1860 * a match as everything else matched. 1861 */ 1862 1863 return true; 1864 } 1865 1866 /* 1867 * Fails a connection request if it matches an existing controller 1868 * (association) with the same tuple: 1869 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN> 1870 * 1871 * if local address is not specified in the request, it will match an 1872 * existing controller with all the other parameters the same and no 1873 * local port address specified as well. 1874 * 1875 * The ports don't need to be compared as they are intrinsically 1876 * already matched by the port pointers supplied. 1877 */ 1878 static bool 1879 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts) 1880 { 1881 struct nvme_rdma_ctrl *ctrl; 1882 bool found = false; 1883 1884 mutex_lock(&nvme_rdma_ctrl_mutex); 1885 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { 1886 found = __nvme_rdma_options_match(ctrl, opts); 1887 if (found) 1888 break; 1889 } 1890 mutex_unlock(&nvme_rdma_ctrl_mutex); 1891 1892 return found; 1893 } 1894 1895 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev, 1896 struct nvmf_ctrl_options *opts) 1897 { 1898 struct nvme_rdma_ctrl *ctrl; 1899 int ret; 1900 bool changed; 1901 char *port; 1902 1903 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL); 1904 if (!ctrl) 1905 return ERR_PTR(-ENOMEM); 1906 ctrl->ctrl.opts = opts; 1907 INIT_LIST_HEAD(&ctrl->list); 1908 1909 if (opts->mask & NVMF_OPT_TRSVCID) 1910 port = opts->trsvcid; 1911 else 1912 port = __stringify(NVME_RDMA_IP_PORT); 1913 1914 ret = inet_pton_with_scope(&init_net, AF_UNSPEC, 1915 opts->traddr, port, &ctrl->addr); 1916 if (ret) { 1917 pr_err("malformed address passed: %s:%s\n", opts->traddr, port); 1918 goto out_free_ctrl; 1919 } 1920 1921 if (opts->mask & NVMF_OPT_HOST_TRADDR) { 1922 ret = inet_pton_with_scope(&init_net, AF_UNSPEC, 1923 opts->host_traddr, NULL, &ctrl->src_addr); 1924 if (ret) { 1925 pr_err("malformed src address passed: %s\n", 1926 opts->host_traddr); 1927 goto out_free_ctrl; 1928 } 1929 } 1930 1931 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) { 1932 ret = -EALREADY; 1933 goto out_free_ctrl; 1934 } 1935 1936 INIT_DELAYED_WORK(&ctrl->reconnect_work, 1937 nvme_rdma_reconnect_ctrl_work); 1938 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work); 1939 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work); 1940 1941 ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */ 1942 ctrl->ctrl.sqsize = opts->queue_size - 1; 1943 ctrl->ctrl.kato = opts->kato; 1944 1945 ret = -ENOMEM; 1946 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues), 1947 GFP_KERNEL); 1948 if (!ctrl->queues) 1949 goto out_free_ctrl; 1950 1951 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops, 1952 0 /* no quirks, we're perfect! */); 1953 if (ret) 1954 goto out_kfree_queues; 1955 1956 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING); 1957 WARN_ON_ONCE(!changed); 1958 1959 ret = nvme_rdma_configure_admin_queue(ctrl, true); 1960 if (ret) 1961 goto out_uninit_ctrl; 1962 1963 /* sanity check icdoff */ 1964 if (ctrl->ctrl.icdoff) { 1965 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n"); 1966 ret = -EINVAL; 1967 goto out_remove_admin_queue; 1968 } 1969 1970 /* sanity check keyed sgls */ 1971 if (!(ctrl->ctrl.sgls & (1 << 2))) { 1972 dev_err(ctrl->ctrl.device, 1973 "Mandatory keyed sgls are not supported!\n"); 1974 ret = -EINVAL; 1975 goto out_remove_admin_queue; 1976 } 1977 1978 /* only warn if argument is too large here, will clamp later */ 1979 if (opts->queue_size > ctrl->ctrl.sqsize + 1) { 1980 dev_warn(ctrl->ctrl.device, 1981 "queue_size %zu > ctrl sqsize %u, clamping down\n", 1982 opts->queue_size, ctrl->ctrl.sqsize + 1); 1983 } 1984 1985 /* warn if maxcmd is lower than sqsize+1 */ 1986 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) { 1987 dev_warn(ctrl->ctrl.device, 1988 "sqsize %u > ctrl maxcmd %u, clamping down\n", 1989 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd); 1990 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1; 1991 } 1992 1993 if (opts->nr_io_queues) { 1994 ret = nvme_rdma_configure_io_queues(ctrl, true); 1995 if (ret) 1996 goto out_remove_admin_queue; 1997 } 1998 1999 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); 2000 WARN_ON_ONCE(!changed); 2001 2002 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n", 2003 ctrl->ctrl.opts->subsysnqn, &ctrl->addr); 2004 2005 nvme_get_ctrl(&ctrl->ctrl); 2006 2007 mutex_lock(&nvme_rdma_ctrl_mutex); 2008 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list); 2009 mutex_unlock(&nvme_rdma_ctrl_mutex); 2010 2011 nvme_start_ctrl(&ctrl->ctrl); 2012 2013 return &ctrl->ctrl; 2014 2015 out_remove_admin_queue: 2016 nvme_rdma_stop_queue(&ctrl->queues[0]); 2017 nvme_rdma_destroy_admin_queue(ctrl, true); 2018 out_uninit_ctrl: 2019 nvme_uninit_ctrl(&ctrl->ctrl); 2020 nvme_put_ctrl(&ctrl->ctrl); 2021 if (ret > 0) 2022 ret = -EIO; 2023 return ERR_PTR(ret); 2024 out_kfree_queues: 2025 kfree(ctrl->queues); 2026 out_free_ctrl: 2027 kfree(ctrl); 2028 return ERR_PTR(ret); 2029 } 2030 2031 static struct nvmf_transport_ops nvme_rdma_transport = { 2032 .name = "rdma", 2033 .module = THIS_MODULE, 2034 .required_opts = NVMF_OPT_TRADDR, 2035 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY | 2036 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO, 2037 .create_ctrl = nvme_rdma_create_ctrl, 2038 }; 2039 2040 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data) 2041 { 2042 struct nvme_rdma_ctrl *ctrl; 2043 struct nvme_rdma_device *ndev; 2044 bool found = false; 2045 2046 mutex_lock(&device_list_mutex); 2047 list_for_each_entry(ndev, &device_list, entry) { 2048 if (ndev->dev == ib_device) { 2049 found = true; 2050 break; 2051 } 2052 } 2053 mutex_unlock(&device_list_mutex); 2054 2055 if (!found) 2056 return; 2057 2058 /* Delete all controllers using this device */ 2059 mutex_lock(&nvme_rdma_ctrl_mutex); 2060 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { 2061 if (ctrl->device->dev != ib_device) 2062 continue; 2063 nvme_delete_ctrl(&ctrl->ctrl); 2064 } 2065 mutex_unlock(&nvme_rdma_ctrl_mutex); 2066 2067 flush_workqueue(nvme_delete_wq); 2068 } 2069 2070 static struct ib_client nvme_rdma_ib_client = { 2071 .name = "nvme_rdma", 2072 .remove = nvme_rdma_remove_one 2073 }; 2074 2075 static int __init nvme_rdma_init_module(void) 2076 { 2077 int ret; 2078 2079 ret = ib_register_client(&nvme_rdma_ib_client); 2080 if (ret) 2081 return ret; 2082 2083 ret = nvmf_register_transport(&nvme_rdma_transport); 2084 if (ret) 2085 goto err_unreg_client; 2086 2087 return 0; 2088 2089 err_unreg_client: 2090 ib_unregister_client(&nvme_rdma_ib_client); 2091 return ret; 2092 } 2093 2094 static void __exit nvme_rdma_cleanup_module(void) 2095 { 2096 nvmf_unregister_transport(&nvme_rdma_transport); 2097 ib_unregister_client(&nvme_rdma_ib_client); 2098 } 2099 2100 module_init(nvme_rdma_init_module); 2101 module_exit(nvme_rdma_cleanup_module); 2102 2103 MODULE_LICENSE("GPL v2"); 2104