1// SPDX-License-Identifier: GPL-2.0 2/* 3 * NVMe over Fabrics RDMA host code. 4 * Copyright (c) 2015-2016 HGST, a Western Digital Company. 5 */ 6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 7#include <linux/module.h> 8#include <linux/init.h> 9#include <linux/slab.h> 10#include <rdma/mr_pool.h> 11#include <linux/err.h> 12#include <linux/string.h> 13#include <linux/atomic.h> 14#include <linux/blk-mq.h> 15#include <linux/blk-integrity.h> 16#include <linux/types.h> 17#include <linux/list.h> 18#include <linux/mutex.h> 19#include <linux/scatterlist.h> 20#include <linux/nvme.h> 21#include <asm/unaligned.h> 22 23#include <rdma/ib_verbs.h> 24#include <rdma/rdma_cm.h> 25#include <linux/nvme-rdma.h> 26 27#include "nvme.h" 28#include "fabrics.h" 29 30 31#define NVME_RDMA_CM_TIMEOUT_MS 3000 /* 3 second */ 32 33#define NVME_RDMA_MAX_SEGMENTS 256 34 35#define NVME_RDMA_MAX_INLINE_SEGMENTS 4 36 37#define NVME_RDMA_DATA_SGL_SIZE \ 38 (sizeof(struct scatterlist) * NVME_INLINE_SG_CNT) 39#define NVME_RDMA_METADATA_SGL_SIZE \ 40 (sizeof(struct scatterlist) * NVME_INLINE_METADATA_SG_CNT) 41 42struct nvme_rdma_device { 43 struct ib_device *dev; 44 struct ib_pd *pd; 45 struct kref ref; 46 struct list_head entry; 47 unsigned int num_inline_segments; 48}; 49 50struct nvme_rdma_qe { 51 struct ib_cqe cqe; 52 void *data; 53 u64 dma; 54}; 55 56struct nvme_rdma_sgl { 57 int nents; 58 struct sg_table sg_table; 59}; 60 61struct nvme_rdma_queue; 62struct nvme_rdma_request { 63 struct nvme_request req; 64 struct ib_mr *mr; 65 struct nvme_rdma_qe sqe; 66 union nvme_result result; 67 __le16 status; 68 refcount_t ref; 69 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS]; 70 u32 num_sge; 71 struct ib_reg_wr reg_wr; 72 struct ib_cqe reg_cqe; 73 struct nvme_rdma_queue *queue; 74 struct nvme_rdma_sgl data_sgl; 75 struct nvme_rdma_sgl *metadata_sgl; 76 bool use_sig_mr; 77}; 78 79enum nvme_rdma_queue_flags { 80 NVME_RDMA_Q_ALLOCATED = 0, 81 NVME_RDMA_Q_LIVE = 1, 82 NVME_RDMA_Q_TR_READY = 2, 83}; 84 85struct nvme_rdma_queue { 86 struct nvme_rdma_qe *rsp_ring; 87 int queue_size; 88 size_t cmnd_capsule_len; 89 struct nvme_rdma_ctrl *ctrl; 90 struct nvme_rdma_device *device; 91 struct ib_cq *ib_cq; 92 struct ib_qp *qp; 93 94 unsigned long flags; 95 struct rdma_cm_id *cm_id; 96 int cm_error; 97 struct completion cm_done; 98 bool pi_support; 99 int cq_size; 100 struct mutex queue_lock; 101}; 102 103struct nvme_rdma_ctrl { 104 /* read only in the hot path */ 105 struct nvme_rdma_queue *queues; 106 107 /* other member variables */ 108 struct blk_mq_tag_set tag_set; 109 struct work_struct err_work; 110 111 struct nvme_rdma_qe async_event_sqe; 112 113 struct delayed_work reconnect_work; 114 115 struct list_head list; 116 117 struct blk_mq_tag_set admin_tag_set; 118 struct nvme_rdma_device *device; 119 120 u32 max_fr_pages; 121 122 struct sockaddr_storage addr; 123 struct sockaddr_storage src_addr; 124 125 struct nvme_ctrl ctrl; 126 bool use_inline_data; 127 u32 io_queues[HCTX_MAX_TYPES]; 128}; 129 130static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl) 131{ 132 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl); 133} 134 135static LIST_HEAD(device_list); 136static DEFINE_MUTEX(device_list_mutex); 137 138static LIST_HEAD(nvme_rdma_ctrl_list); 139static DEFINE_MUTEX(nvme_rdma_ctrl_mutex); 140 141/* 142 * Disabling this option makes small I/O goes faster, but is fundamentally 143 * unsafe. With it turned off we will have to register a global rkey that 144 * allows read and write access to all physical memory. 145 */ 146static bool register_always = true; 147module_param(register_always, bool, 0444); 148MODULE_PARM_DESC(register_always, 149 "Use memory registration even for contiguous memory regions"); 150 151static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, 152 struct rdma_cm_event *event); 153static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc); 154static void nvme_rdma_complete_rq(struct request *rq); 155 156static const struct blk_mq_ops nvme_rdma_mq_ops; 157static const struct blk_mq_ops nvme_rdma_admin_mq_ops; 158 159static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue) 160{ 161 return queue - queue->ctrl->queues; 162} 163 164static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue) 165{ 166 return nvme_rdma_queue_idx(queue) > 167 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] + 168 queue->ctrl->io_queues[HCTX_TYPE_READ]; 169} 170 171static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue) 172{ 173 return queue->cmnd_capsule_len - sizeof(struct nvme_command); 174} 175 176static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, 177 size_t capsule_size, enum dma_data_direction dir) 178{ 179 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir); 180 kfree(qe->data); 181} 182 183static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, 184 size_t capsule_size, enum dma_data_direction dir) 185{ 186 qe->data = kzalloc(capsule_size, GFP_KERNEL); 187 if (!qe->data) 188 return -ENOMEM; 189 190 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir); 191 if (ib_dma_mapping_error(ibdev, qe->dma)) { 192 kfree(qe->data); 193 qe->data = NULL; 194 return -ENOMEM; 195 } 196 197 return 0; 198} 199 200static void nvme_rdma_free_ring(struct ib_device *ibdev, 201 struct nvme_rdma_qe *ring, size_t ib_queue_size, 202 size_t capsule_size, enum dma_data_direction dir) 203{ 204 int i; 205 206 for (i = 0; i < ib_queue_size; i++) 207 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir); 208 kfree(ring); 209} 210 211static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev, 212 size_t ib_queue_size, size_t capsule_size, 213 enum dma_data_direction dir) 214{ 215 struct nvme_rdma_qe *ring; 216 int i; 217 218 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL); 219 if (!ring) 220 return NULL; 221 222 /* 223 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue 224 * lifetime. It's safe, since any chage in the underlying RDMA device 225 * will issue error recovery and queue re-creation. 226 */ 227 for (i = 0; i < ib_queue_size; i++) { 228 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir)) 229 goto out_free_ring; 230 } 231 232 return ring; 233 234out_free_ring: 235 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir); 236 return NULL; 237} 238 239static void nvme_rdma_qp_event(struct ib_event *event, void *context) 240{ 241 pr_debug("QP event %s (%d)\n", 242 ib_event_msg(event->event), event->event); 243 244} 245 246static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue) 247{ 248 int ret; 249 250 ret = wait_for_completion_interruptible(&queue->cm_done); 251 if (ret) 252 return ret; 253 WARN_ON_ONCE(queue->cm_error > 0); 254 return queue->cm_error; 255} 256 257static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor) 258{ 259 struct nvme_rdma_device *dev = queue->device; 260 struct ib_qp_init_attr init_attr; 261 int ret; 262 263 memset(&init_attr, 0, sizeof(init_attr)); 264 init_attr.event_handler = nvme_rdma_qp_event; 265 /* +1 for drain */ 266 init_attr.cap.max_send_wr = factor * queue->queue_size + 1; 267 /* +1 for drain */ 268 init_attr.cap.max_recv_wr = queue->queue_size + 1; 269 init_attr.cap.max_recv_sge = 1; 270 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments; 271 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 272 init_attr.qp_type = IB_QPT_RC; 273 init_attr.send_cq = queue->ib_cq; 274 init_attr.recv_cq = queue->ib_cq; 275 if (queue->pi_support) 276 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN; 277 init_attr.qp_context = queue; 278 279 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr); 280 281 queue->qp = queue->cm_id->qp; 282 return ret; 283} 284 285static void nvme_rdma_exit_request(struct blk_mq_tag_set *set, 286 struct request *rq, unsigned int hctx_idx) 287{ 288 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 289 290 kfree(req->sqe.data); 291} 292 293static int nvme_rdma_init_request(struct blk_mq_tag_set *set, 294 struct request *rq, unsigned int hctx_idx, 295 unsigned int numa_node) 296{ 297 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(set->driver_data); 298 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 299 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0; 300 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx]; 301 302 nvme_req(rq)->ctrl = &ctrl->ctrl; 303 req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL); 304 if (!req->sqe.data) 305 return -ENOMEM; 306 307 /* metadata nvme_rdma_sgl struct is located after command's data SGL */ 308 if (queue->pi_support) 309 req->metadata_sgl = (void *)nvme_req(rq) + 310 sizeof(struct nvme_rdma_request) + 311 NVME_RDMA_DATA_SGL_SIZE; 312 313 req->queue = queue; 314 nvme_req(rq)->cmd = req->sqe.data; 315 316 return 0; 317} 318 319static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 320 unsigned int hctx_idx) 321{ 322 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(data); 323 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1]; 324 325 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count); 326 327 hctx->driver_data = queue; 328 return 0; 329} 330 331static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, 332 unsigned int hctx_idx) 333{ 334 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(data); 335 struct nvme_rdma_queue *queue = &ctrl->queues[0]; 336 337 BUG_ON(hctx_idx != 0); 338 339 hctx->driver_data = queue; 340 return 0; 341} 342 343static void nvme_rdma_free_dev(struct kref *ref) 344{ 345 struct nvme_rdma_device *ndev = 346 container_of(ref, struct nvme_rdma_device, ref); 347 348 mutex_lock(&device_list_mutex); 349 list_del(&ndev->entry); 350 mutex_unlock(&device_list_mutex); 351 352 ib_dealloc_pd(ndev->pd); 353 kfree(ndev); 354} 355 356static void nvme_rdma_dev_put(struct nvme_rdma_device *dev) 357{ 358 kref_put(&dev->ref, nvme_rdma_free_dev); 359} 360 361static int nvme_rdma_dev_get(struct nvme_rdma_device *dev) 362{ 363 return kref_get_unless_zero(&dev->ref); 364} 365 366static struct nvme_rdma_device * 367nvme_rdma_find_get_device(struct rdma_cm_id *cm_id) 368{ 369 struct nvme_rdma_device *ndev; 370 371 mutex_lock(&device_list_mutex); 372 list_for_each_entry(ndev, &device_list, entry) { 373 if (ndev->dev->node_guid == cm_id->device->node_guid && 374 nvme_rdma_dev_get(ndev)) 375 goto out_unlock; 376 } 377 378 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL); 379 if (!ndev) 380 goto out_err; 381 382 ndev->dev = cm_id->device; 383 kref_init(&ndev->ref); 384 385 ndev->pd = ib_alloc_pd(ndev->dev, 386 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY); 387 if (IS_ERR(ndev->pd)) 388 goto out_free_dev; 389 390 if (!(ndev->dev->attrs.device_cap_flags & 391 IB_DEVICE_MEM_MGT_EXTENSIONS)) { 392 dev_err(&ndev->dev->dev, 393 "Memory registrations not supported.\n"); 394 goto out_free_pd; 395 } 396 397 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS, 398 ndev->dev->attrs.max_send_sge - 1); 399 list_add(&ndev->entry, &device_list); 400out_unlock: 401 mutex_unlock(&device_list_mutex); 402 return ndev; 403 404out_free_pd: 405 ib_dealloc_pd(ndev->pd); 406out_free_dev: 407 kfree(ndev); 408out_err: 409 mutex_unlock(&device_list_mutex); 410 return NULL; 411} 412 413static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue) 414{ 415 if (nvme_rdma_poll_queue(queue)) 416 ib_free_cq(queue->ib_cq); 417 else 418 ib_cq_pool_put(queue->ib_cq, queue->cq_size); 419} 420 421static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue) 422{ 423 struct nvme_rdma_device *dev; 424 struct ib_device *ibdev; 425 426 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags)) 427 return; 428 429 dev = queue->device; 430 ibdev = dev->dev; 431 432 if (queue->pi_support) 433 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs); 434 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs); 435 436 /* 437 * The cm_id object might have been destroyed during RDMA connection 438 * establishment error flow to avoid getting other cma events, thus 439 * the destruction of the QP shouldn't use rdma_cm API. 440 */ 441 ib_destroy_qp(queue->qp); 442 nvme_rdma_free_cq(queue); 443 444 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, 445 sizeof(struct nvme_completion), DMA_FROM_DEVICE); 446 447 nvme_rdma_dev_put(dev); 448} 449 450static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support) 451{ 452 u32 max_page_list_len; 453 454 if (pi_support) 455 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len; 456 else 457 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len; 458 459 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1); 460} 461 462static int nvme_rdma_create_cq(struct ib_device *ibdev, 463 struct nvme_rdma_queue *queue) 464{ 465 int ret, comp_vector, idx = nvme_rdma_queue_idx(queue); 466 467 /* 468 * Spread I/O queues completion vectors according their queue index. 469 * Admin queues can always go on completion vector 0. 470 */ 471 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors; 472 473 /* Polling queues need direct cq polling context */ 474 if (nvme_rdma_poll_queue(queue)) 475 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size, 476 comp_vector, IB_POLL_DIRECT); 477 else 478 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size, 479 comp_vector, IB_POLL_SOFTIRQ); 480 481 if (IS_ERR(queue->ib_cq)) { 482 ret = PTR_ERR(queue->ib_cq); 483 return ret; 484 } 485 486 return 0; 487} 488 489static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue) 490{ 491 struct ib_device *ibdev; 492 const int send_wr_factor = 3; /* MR, SEND, INV */ 493 const int cq_factor = send_wr_factor + 1; /* + RECV */ 494 int ret, pages_per_mr; 495 496 queue->device = nvme_rdma_find_get_device(queue->cm_id); 497 if (!queue->device) { 498 dev_err(queue->cm_id->device->dev.parent, 499 "no client data found!\n"); 500 return -ECONNREFUSED; 501 } 502 ibdev = queue->device->dev; 503 504 /* +1 for ib_drain_qp */ 505 queue->cq_size = cq_factor * queue->queue_size + 1; 506 507 ret = nvme_rdma_create_cq(ibdev, queue); 508 if (ret) 509 goto out_put_dev; 510 511 ret = nvme_rdma_create_qp(queue, send_wr_factor); 512 if (ret) 513 goto out_destroy_ib_cq; 514 515 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size, 516 sizeof(struct nvme_completion), DMA_FROM_DEVICE); 517 if (!queue->rsp_ring) { 518 ret = -ENOMEM; 519 goto out_destroy_qp; 520 } 521 522 /* 523 * Currently we don't use SG_GAPS MR's so if the first entry is 524 * misaligned we'll end up using two entries for a single data page, 525 * so one additional entry is required. 526 */ 527 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1; 528 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs, 529 queue->queue_size, 530 IB_MR_TYPE_MEM_REG, 531 pages_per_mr, 0); 532 if (ret) { 533 dev_err(queue->ctrl->ctrl.device, 534 "failed to initialize MR pool sized %d for QID %d\n", 535 queue->queue_size, nvme_rdma_queue_idx(queue)); 536 goto out_destroy_ring; 537 } 538 539 if (queue->pi_support) { 540 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs, 541 queue->queue_size, IB_MR_TYPE_INTEGRITY, 542 pages_per_mr, pages_per_mr); 543 if (ret) { 544 dev_err(queue->ctrl->ctrl.device, 545 "failed to initialize PI MR pool sized %d for QID %d\n", 546 queue->queue_size, nvme_rdma_queue_idx(queue)); 547 goto out_destroy_mr_pool; 548 } 549 } 550 551 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags); 552 553 return 0; 554 555out_destroy_mr_pool: 556 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs); 557out_destroy_ring: 558 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, 559 sizeof(struct nvme_completion), DMA_FROM_DEVICE); 560out_destroy_qp: 561 rdma_destroy_qp(queue->cm_id); 562out_destroy_ib_cq: 563 nvme_rdma_free_cq(queue); 564out_put_dev: 565 nvme_rdma_dev_put(queue->device); 566 return ret; 567} 568 569static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl, 570 int idx, size_t queue_size) 571{ 572 struct nvme_rdma_queue *queue; 573 struct sockaddr *src_addr = NULL; 574 int ret; 575 576 queue = &ctrl->queues[idx]; 577 mutex_init(&queue->queue_lock); 578 queue->ctrl = ctrl; 579 if (idx && ctrl->ctrl.max_integrity_segments) 580 queue->pi_support = true; 581 else 582 queue->pi_support = false; 583 init_completion(&queue->cm_done); 584 585 if (idx > 0) 586 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16; 587 else 588 queue->cmnd_capsule_len = sizeof(struct nvme_command); 589 590 queue->queue_size = queue_size; 591 592 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue, 593 RDMA_PS_TCP, IB_QPT_RC); 594 if (IS_ERR(queue->cm_id)) { 595 dev_info(ctrl->ctrl.device, 596 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id)); 597 ret = PTR_ERR(queue->cm_id); 598 goto out_destroy_mutex; 599 } 600 601 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) 602 src_addr = (struct sockaddr *)&ctrl->src_addr; 603 604 queue->cm_error = -ETIMEDOUT; 605 ret = rdma_resolve_addr(queue->cm_id, src_addr, 606 (struct sockaddr *)&ctrl->addr, 607 NVME_RDMA_CM_TIMEOUT_MS); 608 if (ret) { 609 dev_info(ctrl->ctrl.device, 610 "rdma_resolve_addr failed (%d).\n", ret); 611 goto out_destroy_cm_id; 612 } 613 614 ret = nvme_rdma_wait_for_cm(queue); 615 if (ret) { 616 dev_info(ctrl->ctrl.device, 617 "rdma connection establishment failed (%d)\n", ret); 618 goto out_destroy_cm_id; 619 } 620 621 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags); 622 623 return 0; 624 625out_destroy_cm_id: 626 rdma_destroy_id(queue->cm_id); 627 nvme_rdma_destroy_queue_ib(queue); 628out_destroy_mutex: 629 mutex_destroy(&queue->queue_lock); 630 return ret; 631} 632 633static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue) 634{ 635 rdma_disconnect(queue->cm_id); 636 ib_drain_qp(queue->qp); 637} 638 639static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue) 640{ 641 if (!test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags)) 642 return; 643 644 mutex_lock(&queue->queue_lock); 645 if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags)) 646 __nvme_rdma_stop_queue(queue); 647 mutex_unlock(&queue->queue_lock); 648} 649 650static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue) 651{ 652 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags)) 653 return; 654 655 rdma_destroy_id(queue->cm_id); 656 nvme_rdma_destroy_queue_ib(queue); 657 mutex_destroy(&queue->queue_lock); 658} 659 660static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl) 661{ 662 int i; 663 664 for (i = 1; i < ctrl->ctrl.queue_count; i++) 665 nvme_rdma_free_queue(&ctrl->queues[i]); 666} 667 668static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl) 669{ 670 int i; 671 672 for (i = 1; i < ctrl->ctrl.queue_count; i++) 673 nvme_rdma_stop_queue(&ctrl->queues[i]); 674} 675 676static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx) 677{ 678 struct nvme_rdma_queue *queue = &ctrl->queues[idx]; 679 int ret; 680 681 if (idx) 682 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx); 683 else 684 ret = nvmf_connect_admin_queue(&ctrl->ctrl); 685 686 if (!ret) { 687 set_bit(NVME_RDMA_Q_LIVE, &queue->flags); 688 } else { 689 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags)) 690 __nvme_rdma_stop_queue(queue); 691 dev_info(ctrl->ctrl.device, 692 "failed to connect queue: %d ret=%d\n", idx, ret); 693 } 694 return ret; 695} 696 697static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl, 698 int first, int last) 699{ 700 int i, ret = 0; 701 702 for (i = first; i < last; i++) { 703 ret = nvme_rdma_start_queue(ctrl, i); 704 if (ret) 705 goto out_stop_queues; 706 } 707 708 return 0; 709 710out_stop_queues: 711 for (i--; i >= first; i--) 712 nvme_rdma_stop_queue(&ctrl->queues[i]); 713 return ret; 714} 715 716static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl) 717{ 718 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; 719 unsigned int nr_io_queues; 720 int i, ret; 721 722 nr_io_queues = nvmf_nr_io_queues(opts); 723 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues); 724 if (ret) 725 return ret; 726 727 if (nr_io_queues == 0) { 728 dev_err(ctrl->ctrl.device, 729 "unable to set any I/O queues\n"); 730 return -ENOMEM; 731 } 732 733 ctrl->ctrl.queue_count = nr_io_queues + 1; 734 dev_info(ctrl->ctrl.device, 735 "creating %d I/O queues.\n", nr_io_queues); 736 737 nvmf_set_io_queues(opts, nr_io_queues, ctrl->io_queues); 738 for (i = 1; i < ctrl->ctrl.queue_count; i++) { 739 ret = nvme_rdma_alloc_queue(ctrl, i, 740 ctrl->ctrl.sqsize + 1); 741 if (ret) 742 goto out_free_queues; 743 } 744 745 return 0; 746 747out_free_queues: 748 for (i--; i >= 1; i--) 749 nvme_rdma_free_queue(&ctrl->queues[i]); 750 751 return ret; 752} 753 754static int nvme_rdma_alloc_tag_set(struct nvme_ctrl *ctrl) 755{ 756 unsigned int cmd_size = sizeof(struct nvme_rdma_request) + 757 NVME_RDMA_DATA_SGL_SIZE; 758 759 if (ctrl->max_integrity_segments) 760 cmd_size += sizeof(struct nvme_rdma_sgl) + 761 NVME_RDMA_METADATA_SGL_SIZE; 762 763 return nvme_alloc_io_tag_set(ctrl, &to_rdma_ctrl(ctrl)->tag_set, 764 &nvme_rdma_mq_ops, 765 ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2, 766 cmd_size); 767} 768 769static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl) 770{ 771 if (ctrl->async_event_sqe.data) { 772 cancel_work_sync(&ctrl->ctrl.async_event_work); 773 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe, 774 sizeof(struct nvme_command), DMA_TO_DEVICE); 775 ctrl->async_event_sqe.data = NULL; 776 } 777 nvme_rdma_free_queue(&ctrl->queues[0]); 778} 779 780static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl, 781 bool new) 782{ 783 bool pi_capable = false; 784 int error; 785 786 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH); 787 if (error) 788 return error; 789 790 ctrl->device = ctrl->queues[0].device; 791 ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev); 792 793 /* T10-PI support */ 794 if (ctrl->device->dev->attrs.kernel_cap_flags & 795 IBK_INTEGRITY_HANDOVER) 796 pi_capable = true; 797 798 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev, 799 pi_capable); 800 801 /* 802 * Bind the async event SQE DMA mapping to the admin queue lifetime. 803 * It's safe, since any chage in the underlying RDMA device will issue 804 * error recovery and queue re-creation. 805 */ 806 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe, 807 sizeof(struct nvme_command), DMA_TO_DEVICE); 808 if (error) 809 goto out_free_queue; 810 811 if (new) { 812 error = nvme_alloc_admin_tag_set(&ctrl->ctrl, 813 &ctrl->admin_tag_set, &nvme_rdma_admin_mq_ops, 814 sizeof(struct nvme_rdma_request) + 815 NVME_RDMA_DATA_SGL_SIZE); 816 if (error) 817 goto out_free_async_qe; 818 819 } 820 821 error = nvme_rdma_start_queue(ctrl, 0); 822 if (error) 823 goto out_remove_admin_tag_set; 824 825 error = nvme_enable_ctrl(&ctrl->ctrl); 826 if (error) 827 goto out_stop_queue; 828 829 ctrl->ctrl.max_segments = ctrl->max_fr_pages; 830 ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9); 831 if (pi_capable) 832 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages; 833 else 834 ctrl->ctrl.max_integrity_segments = 0; 835 836 nvme_unquiesce_admin_queue(&ctrl->ctrl); 837 838 error = nvme_init_ctrl_finish(&ctrl->ctrl, false); 839 if (error) 840 goto out_quiesce_queue; 841 842 return 0; 843 844out_quiesce_queue: 845 nvme_quiesce_admin_queue(&ctrl->ctrl); 846 blk_sync_queue(ctrl->ctrl.admin_q); 847out_stop_queue: 848 nvme_rdma_stop_queue(&ctrl->queues[0]); 849 nvme_cancel_admin_tagset(&ctrl->ctrl); 850out_remove_admin_tag_set: 851 if (new) 852 nvme_remove_admin_tag_set(&ctrl->ctrl); 853out_free_async_qe: 854 if (ctrl->async_event_sqe.data) { 855 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe, 856 sizeof(struct nvme_command), DMA_TO_DEVICE); 857 ctrl->async_event_sqe.data = NULL; 858 } 859out_free_queue: 860 nvme_rdma_free_queue(&ctrl->queues[0]); 861 return error; 862} 863 864static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new) 865{ 866 int ret, nr_queues; 867 868 ret = nvme_rdma_alloc_io_queues(ctrl); 869 if (ret) 870 return ret; 871 872 if (new) { 873 ret = nvme_rdma_alloc_tag_set(&ctrl->ctrl); 874 if (ret) 875 goto out_free_io_queues; 876 } 877 878 /* 879 * Only start IO queues for which we have allocated the tagset 880 * and limitted it to the available queues. On reconnects, the 881 * queue number might have changed. 882 */ 883 nr_queues = min(ctrl->tag_set.nr_hw_queues + 1, ctrl->ctrl.queue_count); 884 ret = nvme_rdma_start_io_queues(ctrl, 1, nr_queues); 885 if (ret) 886 goto out_cleanup_tagset; 887 888 if (!new) { 889 nvme_start_freeze(&ctrl->ctrl); 890 nvme_unquiesce_io_queues(&ctrl->ctrl); 891 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) { 892 /* 893 * If we timed out waiting for freeze we are likely to 894 * be stuck. Fail the controller initialization just 895 * to be safe. 896 */ 897 ret = -ENODEV; 898 nvme_unfreeze(&ctrl->ctrl); 899 goto out_wait_freeze_timed_out; 900 } 901 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset, 902 ctrl->ctrl.queue_count - 1); 903 nvme_unfreeze(&ctrl->ctrl); 904 } 905 906 /* 907 * If the number of queues has increased (reconnect case) 908 * start all new queues now. 909 */ 910 ret = nvme_rdma_start_io_queues(ctrl, nr_queues, 911 ctrl->tag_set.nr_hw_queues + 1); 912 if (ret) 913 goto out_wait_freeze_timed_out; 914 915 return 0; 916 917out_wait_freeze_timed_out: 918 nvme_quiesce_io_queues(&ctrl->ctrl); 919 nvme_sync_io_queues(&ctrl->ctrl); 920 nvme_rdma_stop_io_queues(ctrl); 921out_cleanup_tagset: 922 nvme_cancel_tagset(&ctrl->ctrl); 923 if (new) 924 nvme_remove_io_tag_set(&ctrl->ctrl); 925out_free_io_queues: 926 nvme_rdma_free_io_queues(ctrl); 927 return ret; 928} 929 930static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl, 931 bool remove) 932{ 933 nvme_quiesce_admin_queue(&ctrl->ctrl); 934 blk_sync_queue(ctrl->ctrl.admin_q); 935 nvme_rdma_stop_queue(&ctrl->queues[0]); 936 nvme_cancel_admin_tagset(&ctrl->ctrl); 937 if (remove) { 938 nvme_unquiesce_admin_queue(&ctrl->ctrl); 939 nvme_remove_admin_tag_set(&ctrl->ctrl); 940 } 941 nvme_rdma_destroy_admin_queue(ctrl); 942} 943 944static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl, 945 bool remove) 946{ 947 if (ctrl->ctrl.queue_count > 1) { 948 nvme_quiesce_io_queues(&ctrl->ctrl); 949 nvme_sync_io_queues(&ctrl->ctrl); 950 nvme_rdma_stop_io_queues(ctrl); 951 nvme_cancel_tagset(&ctrl->ctrl); 952 if (remove) { 953 nvme_unquiesce_io_queues(&ctrl->ctrl); 954 nvme_remove_io_tag_set(&ctrl->ctrl); 955 } 956 nvme_rdma_free_io_queues(ctrl); 957 } 958} 959 960static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl) 961{ 962 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); 963 964 flush_work(&ctrl->err_work); 965 cancel_delayed_work_sync(&ctrl->reconnect_work); 966} 967 968static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl) 969{ 970 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); 971 972 if (list_empty(&ctrl->list)) 973 goto free_ctrl; 974 975 mutex_lock(&nvme_rdma_ctrl_mutex); 976 list_del(&ctrl->list); 977 mutex_unlock(&nvme_rdma_ctrl_mutex); 978 979 nvmf_free_options(nctrl->opts); 980free_ctrl: 981 kfree(ctrl->queues); 982 kfree(ctrl); 983} 984 985static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl) 986{ 987 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl); 988 989 /* If we are resetting/deleting then do nothing */ 990 if (state != NVME_CTRL_CONNECTING) { 991 WARN_ON_ONCE(state == NVME_CTRL_NEW || state == NVME_CTRL_LIVE); 992 return; 993 } 994 995 if (nvmf_should_reconnect(&ctrl->ctrl)) { 996 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n", 997 ctrl->ctrl.opts->reconnect_delay); 998 queue_delayed_work(nvme_wq, &ctrl->reconnect_work, 999 ctrl->ctrl.opts->reconnect_delay * HZ); 1000 } else { 1001 nvme_delete_ctrl(&ctrl->ctrl); 1002 } 1003} 1004 1005static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new) 1006{ 1007 int ret; 1008 bool changed; 1009 1010 ret = nvme_rdma_configure_admin_queue(ctrl, new); 1011 if (ret) 1012 return ret; 1013 1014 if (ctrl->ctrl.icdoff) { 1015 ret = -EOPNOTSUPP; 1016 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n"); 1017 goto destroy_admin; 1018 } 1019 1020 if (!(ctrl->ctrl.sgls & (1 << 2))) { 1021 ret = -EOPNOTSUPP; 1022 dev_err(ctrl->ctrl.device, 1023 "Mandatory keyed sgls are not supported!\n"); 1024 goto destroy_admin; 1025 } 1026 1027 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) { 1028 dev_warn(ctrl->ctrl.device, 1029 "queue_size %zu > ctrl sqsize %u, clamping down\n", 1030 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1); 1031 } 1032 1033 if (ctrl->ctrl.sqsize + 1 > NVME_RDMA_MAX_QUEUE_SIZE) { 1034 dev_warn(ctrl->ctrl.device, 1035 "ctrl sqsize %u > max queue size %u, clamping down\n", 1036 ctrl->ctrl.sqsize + 1, NVME_RDMA_MAX_QUEUE_SIZE); 1037 ctrl->ctrl.sqsize = NVME_RDMA_MAX_QUEUE_SIZE - 1; 1038 } 1039 1040 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) { 1041 dev_warn(ctrl->ctrl.device, 1042 "sqsize %u > ctrl maxcmd %u, clamping down\n", 1043 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd); 1044 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1; 1045 } 1046 1047 if (ctrl->ctrl.sgls & (1 << 20)) 1048 ctrl->use_inline_data = true; 1049 1050 if (ctrl->ctrl.queue_count > 1) { 1051 ret = nvme_rdma_configure_io_queues(ctrl, new); 1052 if (ret) 1053 goto destroy_admin; 1054 } 1055 1056 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); 1057 if (!changed) { 1058 /* 1059 * state change failure is ok if we started ctrl delete, 1060 * unless we're during creation of a new controller to 1061 * avoid races with teardown flow. 1062 */ 1063 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl); 1064 1065 WARN_ON_ONCE(state != NVME_CTRL_DELETING && 1066 state != NVME_CTRL_DELETING_NOIO); 1067 WARN_ON_ONCE(new); 1068 ret = -EINVAL; 1069 goto destroy_io; 1070 } 1071 1072 nvme_start_ctrl(&ctrl->ctrl); 1073 return 0; 1074 1075destroy_io: 1076 if (ctrl->ctrl.queue_count > 1) { 1077 nvme_quiesce_io_queues(&ctrl->ctrl); 1078 nvme_sync_io_queues(&ctrl->ctrl); 1079 nvme_rdma_stop_io_queues(ctrl); 1080 nvme_cancel_tagset(&ctrl->ctrl); 1081 if (new) 1082 nvme_remove_io_tag_set(&ctrl->ctrl); 1083 nvme_rdma_free_io_queues(ctrl); 1084 } 1085destroy_admin: 1086 nvme_quiesce_admin_queue(&ctrl->ctrl); 1087 blk_sync_queue(ctrl->ctrl.admin_q); 1088 nvme_rdma_stop_queue(&ctrl->queues[0]); 1089 nvme_cancel_admin_tagset(&ctrl->ctrl); 1090 if (new) 1091 nvme_remove_admin_tag_set(&ctrl->ctrl); 1092 nvme_rdma_destroy_admin_queue(ctrl); 1093 return ret; 1094} 1095 1096static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work) 1097{ 1098 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work), 1099 struct nvme_rdma_ctrl, reconnect_work); 1100 1101 ++ctrl->ctrl.nr_reconnects; 1102 1103 if (nvme_rdma_setup_ctrl(ctrl, false)) 1104 goto requeue; 1105 1106 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n", 1107 ctrl->ctrl.nr_reconnects); 1108 1109 ctrl->ctrl.nr_reconnects = 0; 1110 1111 return; 1112 1113requeue: 1114 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n", 1115 ctrl->ctrl.nr_reconnects); 1116 nvme_rdma_reconnect_or_remove(ctrl); 1117} 1118 1119static void nvme_rdma_error_recovery_work(struct work_struct *work) 1120{ 1121 struct nvme_rdma_ctrl *ctrl = container_of(work, 1122 struct nvme_rdma_ctrl, err_work); 1123 1124 nvme_stop_keep_alive(&ctrl->ctrl); 1125 flush_work(&ctrl->ctrl.async_event_work); 1126 nvme_rdma_teardown_io_queues(ctrl, false); 1127 nvme_unquiesce_io_queues(&ctrl->ctrl); 1128 nvme_rdma_teardown_admin_queue(ctrl, false); 1129 nvme_unquiesce_admin_queue(&ctrl->ctrl); 1130 nvme_auth_stop(&ctrl->ctrl); 1131 1132 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { 1133 /* state change failure is ok if we started ctrl delete */ 1134 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl); 1135 1136 WARN_ON_ONCE(state != NVME_CTRL_DELETING && 1137 state != NVME_CTRL_DELETING_NOIO); 1138 return; 1139 } 1140 1141 nvme_rdma_reconnect_or_remove(ctrl); 1142} 1143 1144static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl) 1145{ 1146 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING)) 1147 return; 1148 1149 dev_warn(ctrl->ctrl.device, "starting error recovery\n"); 1150 queue_work(nvme_reset_wq, &ctrl->err_work); 1151} 1152 1153static void nvme_rdma_end_request(struct nvme_rdma_request *req) 1154{ 1155 struct request *rq = blk_mq_rq_from_pdu(req); 1156 1157 if (!refcount_dec_and_test(&req->ref)) 1158 return; 1159 if (!nvme_try_complete_req(rq, req->status, req->result)) 1160 nvme_rdma_complete_rq(rq); 1161} 1162 1163static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc, 1164 const char *op) 1165{ 1166 struct nvme_rdma_queue *queue = wc->qp->qp_context; 1167 struct nvme_rdma_ctrl *ctrl = queue->ctrl; 1168 1169 if (nvme_ctrl_state(&ctrl->ctrl) == NVME_CTRL_LIVE) 1170 dev_info(ctrl->ctrl.device, 1171 "%s for CQE 0x%p failed with status %s (%d)\n", 1172 op, wc->wr_cqe, 1173 ib_wc_status_msg(wc->status), wc->status); 1174 nvme_rdma_error_recovery(ctrl); 1175} 1176 1177static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc) 1178{ 1179 if (unlikely(wc->status != IB_WC_SUCCESS)) 1180 nvme_rdma_wr_error(cq, wc, "MEMREG"); 1181} 1182 1183static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc) 1184{ 1185 struct nvme_rdma_request *req = 1186 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe); 1187 1188 if (unlikely(wc->status != IB_WC_SUCCESS)) 1189 nvme_rdma_wr_error(cq, wc, "LOCAL_INV"); 1190 else 1191 nvme_rdma_end_request(req); 1192} 1193 1194static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue, 1195 struct nvme_rdma_request *req) 1196{ 1197 struct ib_send_wr wr = { 1198 .opcode = IB_WR_LOCAL_INV, 1199 .next = NULL, 1200 .num_sge = 0, 1201 .send_flags = IB_SEND_SIGNALED, 1202 .ex.invalidate_rkey = req->mr->rkey, 1203 }; 1204 1205 req->reg_cqe.done = nvme_rdma_inv_rkey_done; 1206 wr.wr_cqe = &req->reg_cqe; 1207 1208 return ib_post_send(queue->qp, &wr, NULL); 1209} 1210 1211static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq) 1212{ 1213 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1214 1215 if (blk_integrity_rq(rq)) { 1216 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl, 1217 req->metadata_sgl->nents, rq_dma_dir(rq)); 1218 sg_free_table_chained(&req->metadata_sgl->sg_table, 1219 NVME_INLINE_METADATA_SG_CNT); 1220 } 1221 1222 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents, 1223 rq_dma_dir(rq)); 1224 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT); 1225} 1226 1227static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue, 1228 struct request *rq) 1229{ 1230 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1231 struct nvme_rdma_device *dev = queue->device; 1232 struct ib_device *ibdev = dev->dev; 1233 struct list_head *pool = &queue->qp->rdma_mrs; 1234 1235 if (!blk_rq_nr_phys_segments(rq)) 1236 return; 1237 1238 if (req->use_sig_mr) 1239 pool = &queue->qp->sig_mrs; 1240 1241 if (req->mr) { 1242 ib_mr_pool_put(queue->qp, pool, req->mr); 1243 req->mr = NULL; 1244 } 1245 1246 nvme_rdma_dma_unmap_req(ibdev, rq); 1247} 1248 1249static int nvme_rdma_set_sg_null(struct nvme_command *c) 1250{ 1251 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1252 1253 sg->addr = 0; 1254 put_unaligned_le24(0, sg->length); 1255 put_unaligned_le32(0, sg->key); 1256 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; 1257 return 0; 1258} 1259 1260static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue, 1261 struct nvme_rdma_request *req, struct nvme_command *c, 1262 int count) 1263{ 1264 struct nvme_sgl_desc *sg = &c->common.dptr.sgl; 1265 struct ib_sge *sge = &req->sge[1]; 1266 struct scatterlist *sgl; 1267 u32 len = 0; 1268 int i; 1269 1270 for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) { 1271 sge->addr = sg_dma_address(sgl); 1272 sge->length = sg_dma_len(sgl); 1273 sge->lkey = queue->device->pd->local_dma_lkey; 1274 len += sge->length; 1275 sge++; 1276 } 1277 1278 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff); 1279 sg->length = cpu_to_le32(len); 1280 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET; 1281 1282 req->num_sge += count; 1283 return 0; 1284} 1285 1286static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue, 1287 struct nvme_rdma_request *req, struct nvme_command *c) 1288{ 1289 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1290 1291 sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl)); 1292 put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length); 1293 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key); 1294 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; 1295 return 0; 1296} 1297 1298static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue, 1299 struct nvme_rdma_request *req, struct nvme_command *c, 1300 int count) 1301{ 1302 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1303 int nr; 1304 1305 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs); 1306 if (WARN_ON_ONCE(!req->mr)) 1307 return -EAGAIN; 1308 1309 /* 1310 * Align the MR to a 4K page size to match the ctrl page size and 1311 * the block virtual boundary. 1312 */ 1313 nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL, 1314 SZ_4K); 1315 if (unlikely(nr < count)) { 1316 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr); 1317 req->mr = NULL; 1318 if (nr < 0) 1319 return nr; 1320 return -EINVAL; 1321 } 1322 1323 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey)); 1324 1325 req->reg_cqe.done = nvme_rdma_memreg_done; 1326 memset(&req->reg_wr, 0, sizeof(req->reg_wr)); 1327 req->reg_wr.wr.opcode = IB_WR_REG_MR; 1328 req->reg_wr.wr.wr_cqe = &req->reg_cqe; 1329 req->reg_wr.wr.num_sge = 0; 1330 req->reg_wr.mr = req->mr; 1331 req->reg_wr.key = req->mr->rkey; 1332 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE | 1333 IB_ACCESS_REMOTE_READ | 1334 IB_ACCESS_REMOTE_WRITE; 1335 1336 sg->addr = cpu_to_le64(req->mr->iova); 1337 put_unaligned_le24(req->mr->length, sg->length); 1338 put_unaligned_le32(req->mr->rkey, sg->key); 1339 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) | 1340 NVME_SGL_FMT_INVALIDATE; 1341 1342 return 0; 1343} 1344 1345static void nvme_rdma_set_sig_domain(struct blk_integrity *bi, 1346 struct nvme_command *cmd, struct ib_sig_domain *domain, 1347 u16 control, u8 pi_type) 1348{ 1349 domain->sig_type = IB_SIG_TYPE_T10_DIF; 1350 domain->sig.dif.bg_type = IB_T10DIF_CRC; 1351 domain->sig.dif.pi_interval = 1 << bi->interval_exp; 1352 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag); 1353 if (control & NVME_RW_PRINFO_PRCHK_REF) 1354 domain->sig.dif.ref_remap = true; 1355 1356 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag); 1357 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask); 1358 domain->sig.dif.app_escape = true; 1359 if (pi_type == NVME_NS_DPS_PI_TYPE3) 1360 domain->sig.dif.ref_escape = true; 1361} 1362 1363static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi, 1364 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs, 1365 u8 pi_type) 1366{ 1367 u16 control = le16_to_cpu(cmd->rw.control); 1368 1369 memset(sig_attrs, 0, sizeof(*sig_attrs)); 1370 if (control & NVME_RW_PRINFO_PRACT) { 1371 /* for WRITE_INSERT/READ_STRIP no memory domain */ 1372 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE; 1373 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control, 1374 pi_type); 1375 /* Clear the PRACT bit since HCA will generate/verify the PI */ 1376 control &= ~NVME_RW_PRINFO_PRACT; 1377 cmd->rw.control = cpu_to_le16(control); 1378 } else { 1379 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */ 1380 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control, 1381 pi_type); 1382 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control, 1383 pi_type); 1384 } 1385} 1386 1387static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask) 1388{ 1389 *mask = 0; 1390 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF) 1391 *mask |= IB_SIG_CHECK_REFTAG; 1392 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD) 1393 *mask |= IB_SIG_CHECK_GUARD; 1394} 1395 1396static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc) 1397{ 1398 if (unlikely(wc->status != IB_WC_SUCCESS)) 1399 nvme_rdma_wr_error(cq, wc, "SIG"); 1400} 1401 1402static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue, 1403 struct nvme_rdma_request *req, struct nvme_command *c, 1404 int count, int pi_count) 1405{ 1406 struct nvme_rdma_sgl *sgl = &req->data_sgl; 1407 struct ib_reg_wr *wr = &req->reg_wr; 1408 struct request *rq = blk_mq_rq_from_pdu(req); 1409 struct nvme_ns *ns = rq->q->queuedata; 1410 struct bio *bio = rq->bio; 1411 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; 1412 int nr; 1413 1414 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs); 1415 if (WARN_ON_ONCE(!req->mr)) 1416 return -EAGAIN; 1417 1418 nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL, 1419 req->metadata_sgl->sg_table.sgl, pi_count, NULL, 1420 SZ_4K); 1421 if (unlikely(nr)) 1422 goto mr_put; 1423 1424 nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c, 1425 req->mr->sig_attrs, ns->pi_type); 1426 nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask); 1427 1428 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey)); 1429 1430 req->reg_cqe.done = nvme_rdma_sig_done; 1431 memset(wr, 0, sizeof(*wr)); 1432 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY; 1433 wr->wr.wr_cqe = &req->reg_cqe; 1434 wr->wr.num_sge = 0; 1435 wr->wr.send_flags = 0; 1436 wr->mr = req->mr; 1437 wr->key = req->mr->rkey; 1438 wr->access = IB_ACCESS_LOCAL_WRITE | 1439 IB_ACCESS_REMOTE_READ | 1440 IB_ACCESS_REMOTE_WRITE; 1441 1442 sg->addr = cpu_to_le64(req->mr->iova); 1443 put_unaligned_le24(req->mr->length, sg->length); 1444 put_unaligned_le32(req->mr->rkey, sg->key); 1445 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; 1446 1447 return 0; 1448 1449mr_put: 1450 ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr); 1451 req->mr = NULL; 1452 if (nr < 0) 1453 return nr; 1454 return -EINVAL; 1455} 1456 1457static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq, 1458 int *count, int *pi_count) 1459{ 1460 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1461 int ret; 1462 1463 req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1); 1464 ret = sg_alloc_table_chained(&req->data_sgl.sg_table, 1465 blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl, 1466 NVME_INLINE_SG_CNT); 1467 if (ret) 1468 return -ENOMEM; 1469 1470 req->data_sgl.nents = blk_rq_map_sg(rq->q, rq, 1471 req->data_sgl.sg_table.sgl); 1472 1473 *count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl, 1474 req->data_sgl.nents, rq_dma_dir(rq)); 1475 if (unlikely(*count <= 0)) { 1476 ret = -EIO; 1477 goto out_free_table; 1478 } 1479 1480 if (blk_integrity_rq(rq)) { 1481 req->metadata_sgl->sg_table.sgl = 1482 (struct scatterlist *)(req->metadata_sgl + 1); 1483 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table, 1484 blk_rq_count_integrity_sg(rq->q, rq->bio), 1485 req->metadata_sgl->sg_table.sgl, 1486 NVME_INLINE_METADATA_SG_CNT); 1487 if (unlikely(ret)) { 1488 ret = -ENOMEM; 1489 goto out_unmap_sg; 1490 } 1491 1492 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q, 1493 rq->bio, req->metadata_sgl->sg_table.sgl); 1494 *pi_count = ib_dma_map_sg(ibdev, 1495 req->metadata_sgl->sg_table.sgl, 1496 req->metadata_sgl->nents, 1497 rq_dma_dir(rq)); 1498 if (unlikely(*pi_count <= 0)) { 1499 ret = -EIO; 1500 goto out_free_pi_table; 1501 } 1502 } 1503 1504 return 0; 1505 1506out_free_pi_table: 1507 sg_free_table_chained(&req->metadata_sgl->sg_table, 1508 NVME_INLINE_METADATA_SG_CNT); 1509out_unmap_sg: 1510 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents, 1511 rq_dma_dir(rq)); 1512out_free_table: 1513 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT); 1514 return ret; 1515} 1516 1517static int nvme_rdma_map_data(struct nvme_rdma_queue *queue, 1518 struct request *rq, struct nvme_command *c) 1519{ 1520 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1521 struct nvme_rdma_device *dev = queue->device; 1522 struct ib_device *ibdev = dev->dev; 1523 int pi_count = 0; 1524 int count, ret; 1525 1526 req->num_sge = 1; 1527 refcount_set(&req->ref, 2); /* send and recv completions */ 1528 1529 c->common.flags |= NVME_CMD_SGL_METABUF; 1530 1531 if (!blk_rq_nr_phys_segments(rq)) 1532 return nvme_rdma_set_sg_null(c); 1533 1534 ret = nvme_rdma_dma_map_req(ibdev, rq, &count, &pi_count); 1535 if (unlikely(ret)) 1536 return ret; 1537 1538 if (req->use_sig_mr) { 1539 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count); 1540 goto out; 1541 } 1542 1543 if (count <= dev->num_inline_segments) { 1544 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) && 1545 queue->ctrl->use_inline_data && 1546 blk_rq_payload_bytes(rq) <= 1547 nvme_rdma_inline_data_size(queue)) { 1548 ret = nvme_rdma_map_sg_inline(queue, req, c, count); 1549 goto out; 1550 } 1551 1552 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) { 1553 ret = nvme_rdma_map_sg_single(queue, req, c); 1554 goto out; 1555 } 1556 } 1557 1558 ret = nvme_rdma_map_sg_fr(queue, req, c, count); 1559out: 1560 if (unlikely(ret)) 1561 goto out_dma_unmap_req; 1562 1563 return 0; 1564 1565out_dma_unmap_req: 1566 nvme_rdma_dma_unmap_req(ibdev, rq); 1567 return ret; 1568} 1569 1570static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc) 1571{ 1572 struct nvme_rdma_qe *qe = 1573 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); 1574 struct nvme_rdma_request *req = 1575 container_of(qe, struct nvme_rdma_request, sqe); 1576 1577 if (unlikely(wc->status != IB_WC_SUCCESS)) 1578 nvme_rdma_wr_error(cq, wc, "SEND"); 1579 else 1580 nvme_rdma_end_request(req); 1581} 1582 1583static int nvme_rdma_post_send(struct nvme_rdma_queue *queue, 1584 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge, 1585 struct ib_send_wr *first) 1586{ 1587 struct ib_send_wr wr; 1588 int ret; 1589 1590 sge->addr = qe->dma; 1591 sge->length = sizeof(struct nvme_command); 1592 sge->lkey = queue->device->pd->local_dma_lkey; 1593 1594 wr.next = NULL; 1595 wr.wr_cqe = &qe->cqe; 1596 wr.sg_list = sge; 1597 wr.num_sge = num_sge; 1598 wr.opcode = IB_WR_SEND; 1599 wr.send_flags = IB_SEND_SIGNALED; 1600 1601 if (first) 1602 first->next = ≀ 1603 else 1604 first = ≀ 1605 1606 ret = ib_post_send(queue->qp, first, NULL); 1607 if (unlikely(ret)) { 1608 dev_err(queue->ctrl->ctrl.device, 1609 "%s failed with error code %d\n", __func__, ret); 1610 } 1611 return ret; 1612} 1613 1614static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue, 1615 struct nvme_rdma_qe *qe) 1616{ 1617 struct ib_recv_wr wr; 1618 struct ib_sge list; 1619 int ret; 1620 1621 list.addr = qe->dma; 1622 list.length = sizeof(struct nvme_completion); 1623 list.lkey = queue->device->pd->local_dma_lkey; 1624 1625 qe->cqe.done = nvme_rdma_recv_done; 1626 1627 wr.next = NULL; 1628 wr.wr_cqe = &qe->cqe; 1629 wr.sg_list = &list; 1630 wr.num_sge = 1; 1631 1632 ret = ib_post_recv(queue->qp, &wr, NULL); 1633 if (unlikely(ret)) { 1634 dev_err(queue->ctrl->ctrl.device, 1635 "%s failed with error code %d\n", __func__, ret); 1636 } 1637 return ret; 1638} 1639 1640static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue) 1641{ 1642 u32 queue_idx = nvme_rdma_queue_idx(queue); 1643 1644 if (queue_idx == 0) 1645 return queue->ctrl->admin_tag_set.tags[queue_idx]; 1646 return queue->ctrl->tag_set.tags[queue_idx - 1]; 1647} 1648 1649static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc) 1650{ 1651 if (unlikely(wc->status != IB_WC_SUCCESS)) 1652 nvme_rdma_wr_error(cq, wc, "ASYNC"); 1653} 1654 1655static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg) 1656{ 1657 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg); 1658 struct nvme_rdma_queue *queue = &ctrl->queues[0]; 1659 struct ib_device *dev = queue->device->dev; 1660 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe; 1661 struct nvme_command *cmd = sqe->data; 1662 struct ib_sge sge; 1663 int ret; 1664 1665 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE); 1666 1667 memset(cmd, 0, sizeof(*cmd)); 1668 cmd->common.opcode = nvme_admin_async_event; 1669 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH; 1670 cmd->common.flags |= NVME_CMD_SGL_METABUF; 1671 nvme_rdma_set_sg_null(cmd); 1672 1673 sqe->cqe.done = nvme_rdma_async_done; 1674 1675 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd), 1676 DMA_TO_DEVICE); 1677 1678 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL); 1679 WARN_ON_ONCE(ret); 1680} 1681 1682static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue, 1683 struct nvme_completion *cqe, struct ib_wc *wc) 1684{ 1685 struct request *rq; 1686 struct nvme_rdma_request *req; 1687 1688 rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id); 1689 if (!rq) { 1690 dev_err(queue->ctrl->ctrl.device, 1691 "got bad command_id %#x on QP %#x\n", 1692 cqe->command_id, queue->qp->qp_num); 1693 nvme_rdma_error_recovery(queue->ctrl); 1694 return; 1695 } 1696 req = blk_mq_rq_to_pdu(rq); 1697 1698 req->status = cqe->status; 1699 req->result = cqe->result; 1700 1701 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) { 1702 if (unlikely(!req->mr || 1703 wc->ex.invalidate_rkey != req->mr->rkey)) { 1704 dev_err(queue->ctrl->ctrl.device, 1705 "Bogus remote invalidation for rkey %#x\n", 1706 req->mr ? req->mr->rkey : 0); 1707 nvme_rdma_error_recovery(queue->ctrl); 1708 } 1709 } else if (req->mr) { 1710 int ret; 1711 1712 ret = nvme_rdma_inv_rkey(queue, req); 1713 if (unlikely(ret < 0)) { 1714 dev_err(queue->ctrl->ctrl.device, 1715 "Queueing INV WR for rkey %#x failed (%d)\n", 1716 req->mr->rkey, ret); 1717 nvme_rdma_error_recovery(queue->ctrl); 1718 } 1719 /* the local invalidation completion will end the request */ 1720 return; 1721 } 1722 1723 nvme_rdma_end_request(req); 1724} 1725 1726static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc) 1727{ 1728 struct nvme_rdma_qe *qe = 1729 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); 1730 struct nvme_rdma_queue *queue = wc->qp->qp_context; 1731 struct ib_device *ibdev = queue->device->dev; 1732 struct nvme_completion *cqe = qe->data; 1733 const size_t len = sizeof(struct nvme_completion); 1734 1735 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1736 nvme_rdma_wr_error(cq, wc, "RECV"); 1737 return; 1738 } 1739 1740 /* sanity checking for received data length */ 1741 if (unlikely(wc->byte_len < len)) { 1742 dev_err(queue->ctrl->ctrl.device, 1743 "Unexpected nvme completion length(%d)\n", wc->byte_len); 1744 nvme_rdma_error_recovery(queue->ctrl); 1745 return; 1746 } 1747 1748 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE); 1749 /* 1750 * AEN requests are special as they don't time out and can 1751 * survive any kind of queue freeze and often don't respond to 1752 * aborts. We don't even bother to allocate a struct request 1753 * for them but rather special case them here. 1754 */ 1755 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue), 1756 cqe->command_id))) 1757 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status, 1758 &cqe->result); 1759 else 1760 nvme_rdma_process_nvme_rsp(queue, cqe, wc); 1761 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE); 1762 1763 nvme_rdma_post_recv(queue, qe); 1764} 1765 1766static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue) 1767{ 1768 int ret, i; 1769 1770 for (i = 0; i < queue->queue_size; i++) { 1771 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]); 1772 if (ret) 1773 return ret; 1774 } 1775 1776 return 0; 1777} 1778 1779static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue, 1780 struct rdma_cm_event *ev) 1781{ 1782 struct rdma_cm_id *cm_id = queue->cm_id; 1783 int status = ev->status; 1784 const char *rej_msg; 1785 const struct nvme_rdma_cm_rej *rej_data; 1786 u8 rej_data_len; 1787 1788 rej_msg = rdma_reject_msg(cm_id, status); 1789 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len); 1790 1791 if (rej_data && rej_data_len >= sizeof(u16)) { 1792 u16 sts = le16_to_cpu(rej_data->sts); 1793 1794 dev_err(queue->ctrl->ctrl.device, 1795 "Connect rejected: status %d (%s) nvme status %d (%s).\n", 1796 status, rej_msg, sts, nvme_rdma_cm_msg(sts)); 1797 } else { 1798 dev_err(queue->ctrl->ctrl.device, 1799 "Connect rejected: status %d (%s).\n", status, rej_msg); 1800 } 1801 1802 return -ECONNRESET; 1803} 1804 1805static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue) 1806{ 1807 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl; 1808 int ret; 1809 1810 ret = nvme_rdma_create_queue_ib(queue); 1811 if (ret) 1812 return ret; 1813 1814 if (ctrl->opts->tos >= 0) 1815 rdma_set_service_type(queue->cm_id, ctrl->opts->tos); 1816 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CM_TIMEOUT_MS); 1817 if (ret) { 1818 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n", 1819 queue->cm_error); 1820 goto out_destroy_queue; 1821 } 1822 1823 return 0; 1824 1825out_destroy_queue: 1826 nvme_rdma_destroy_queue_ib(queue); 1827 return ret; 1828} 1829 1830static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue) 1831{ 1832 struct nvme_rdma_ctrl *ctrl = queue->ctrl; 1833 struct rdma_conn_param param = { }; 1834 struct nvme_rdma_cm_req priv = { }; 1835 int ret; 1836 1837 param.qp_num = queue->qp->qp_num; 1838 param.flow_control = 1; 1839 1840 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom; 1841 /* maximum retry count */ 1842 param.retry_count = 7; 1843 param.rnr_retry_count = 7; 1844 param.private_data = &priv; 1845 param.private_data_len = sizeof(priv); 1846 1847 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0); 1848 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue)); 1849 /* 1850 * set the admin queue depth to the minimum size 1851 * specified by the Fabrics standard. 1852 */ 1853 if (priv.qid == 0) { 1854 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH); 1855 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1); 1856 } else { 1857 /* 1858 * current interpretation of the fabrics spec 1859 * is at minimum you make hrqsize sqsize+1, or a 1860 * 1's based representation of sqsize. 1861 */ 1862 priv.hrqsize = cpu_to_le16(queue->queue_size); 1863 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize); 1864 } 1865 1866 ret = rdma_connect_locked(queue->cm_id, ¶m); 1867 if (ret) { 1868 dev_err(ctrl->ctrl.device, 1869 "rdma_connect_locked failed (%d).\n", ret); 1870 return ret; 1871 } 1872 1873 return 0; 1874} 1875 1876static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, 1877 struct rdma_cm_event *ev) 1878{ 1879 struct nvme_rdma_queue *queue = cm_id->context; 1880 int cm_error = 0; 1881 1882 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n", 1883 rdma_event_msg(ev->event), ev->event, 1884 ev->status, cm_id); 1885 1886 switch (ev->event) { 1887 case RDMA_CM_EVENT_ADDR_RESOLVED: 1888 cm_error = nvme_rdma_addr_resolved(queue); 1889 break; 1890 case RDMA_CM_EVENT_ROUTE_RESOLVED: 1891 cm_error = nvme_rdma_route_resolved(queue); 1892 break; 1893 case RDMA_CM_EVENT_ESTABLISHED: 1894 queue->cm_error = nvme_rdma_conn_established(queue); 1895 /* complete cm_done regardless of success/failure */ 1896 complete(&queue->cm_done); 1897 return 0; 1898 case RDMA_CM_EVENT_REJECTED: 1899 cm_error = nvme_rdma_conn_rejected(queue, ev); 1900 break; 1901 case RDMA_CM_EVENT_ROUTE_ERROR: 1902 case RDMA_CM_EVENT_CONNECT_ERROR: 1903 case RDMA_CM_EVENT_UNREACHABLE: 1904 case RDMA_CM_EVENT_ADDR_ERROR: 1905 dev_dbg(queue->ctrl->ctrl.device, 1906 "CM error event %d\n", ev->event); 1907 cm_error = -ECONNRESET; 1908 break; 1909 case RDMA_CM_EVENT_DISCONNECTED: 1910 case RDMA_CM_EVENT_ADDR_CHANGE: 1911 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 1912 dev_dbg(queue->ctrl->ctrl.device, 1913 "disconnect received - connection closed\n"); 1914 nvme_rdma_error_recovery(queue->ctrl); 1915 break; 1916 case RDMA_CM_EVENT_DEVICE_REMOVAL: 1917 /* device removal is handled via the ib_client API */ 1918 break; 1919 default: 1920 dev_err(queue->ctrl->ctrl.device, 1921 "Unexpected RDMA CM event (%d)\n", ev->event); 1922 nvme_rdma_error_recovery(queue->ctrl); 1923 break; 1924 } 1925 1926 if (cm_error) { 1927 queue->cm_error = cm_error; 1928 complete(&queue->cm_done); 1929 } 1930 1931 return 0; 1932} 1933 1934static void nvme_rdma_complete_timed_out(struct request *rq) 1935{ 1936 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1937 struct nvme_rdma_queue *queue = req->queue; 1938 1939 nvme_rdma_stop_queue(queue); 1940 nvmf_complete_timed_out_request(rq); 1941} 1942 1943static enum blk_eh_timer_return nvme_rdma_timeout(struct request *rq) 1944{ 1945 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1946 struct nvme_rdma_queue *queue = req->queue; 1947 struct nvme_rdma_ctrl *ctrl = queue->ctrl; 1948 1949 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n", 1950 rq->tag, nvme_rdma_queue_idx(queue)); 1951 1952 if (nvme_ctrl_state(&ctrl->ctrl) != NVME_CTRL_LIVE) { 1953 /* 1954 * If we are resetting, connecting or deleting we should 1955 * complete immediately because we may block controller 1956 * teardown or setup sequence 1957 * - ctrl disable/shutdown fabrics requests 1958 * - connect requests 1959 * - initialization admin requests 1960 * - I/O requests that entered after unquiescing and 1961 * the controller stopped responding 1962 * 1963 * All other requests should be cancelled by the error 1964 * recovery work, so it's fine that we fail it here. 1965 */ 1966 nvme_rdma_complete_timed_out(rq); 1967 return BLK_EH_DONE; 1968 } 1969 1970 /* 1971 * LIVE state should trigger the normal error recovery which will 1972 * handle completing this request. 1973 */ 1974 nvme_rdma_error_recovery(ctrl); 1975 return BLK_EH_RESET_TIMER; 1976} 1977 1978static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx, 1979 const struct blk_mq_queue_data *bd) 1980{ 1981 struct nvme_ns *ns = hctx->queue->queuedata; 1982 struct nvme_rdma_queue *queue = hctx->driver_data; 1983 struct request *rq = bd->rq; 1984 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 1985 struct nvme_rdma_qe *sqe = &req->sqe; 1986 struct nvme_command *c = nvme_req(rq)->cmd; 1987 struct ib_device *dev; 1988 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags); 1989 blk_status_t ret; 1990 int err; 1991 1992 WARN_ON_ONCE(rq->tag < 0); 1993 1994 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready)) 1995 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq); 1996 1997 dev = queue->device->dev; 1998 1999 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data, 2000 sizeof(struct nvme_command), 2001 DMA_TO_DEVICE); 2002 err = ib_dma_mapping_error(dev, req->sqe.dma); 2003 if (unlikely(err)) 2004 return BLK_STS_RESOURCE; 2005 2006 ib_dma_sync_single_for_cpu(dev, sqe->dma, 2007 sizeof(struct nvme_command), DMA_TO_DEVICE); 2008 2009 ret = nvme_setup_cmd(ns, rq); 2010 if (ret) 2011 goto unmap_qe; 2012 2013 nvme_start_request(rq); 2014 2015 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) && 2016 queue->pi_support && 2017 (c->common.opcode == nvme_cmd_write || 2018 c->common.opcode == nvme_cmd_read) && 2019 nvme_ns_has_pi(ns)) 2020 req->use_sig_mr = true; 2021 else 2022 req->use_sig_mr = false; 2023 2024 err = nvme_rdma_map_data(queue, rq, c); 2025 if (unlikely(err < 0)) { 2026 dev_err(queue->ctrl->ctrl.device, 2027 "Failed to map data (%d)\n", err); 2028 goto err; 2029 } 2030 2031 sqe->cqe.done = nvme_rdma_send_done; 2032 2033 ib_dma_sync_single_for_device(dev, sqe->dma, 2034 sizeof(struct nvme_command), DMA_TO_DEVICE); 2035 2036 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge, 2037 req->mr ? &req->reg_wr.wr : NULL); 2038 if (unlikely(err)) 2039 goto err_unmap; 2040 2041 return BLK_STS_OK; 2042 2043err_unmap: 2044 nvme_rdma_unmap_data(queue, rq); 2045err: 2046 if (err == -EIO) 2047 ret = nvme_host_path_error(rq); 2048 else if (err == -ENOMEM || err == -EAGAIN) 2049 ret = BLK_STS_RESOURCE; 2050 else 2051 ret = BLK_STS_IOERR; 2052 nvme_cleanup_cmd(rq); 2053unmap_qe: 2054 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command), 2055 DMA_TO_DEVICE); 2056 return ret; 2057} 2058 2059static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) 2060{ 2061 struct nvme_rdma_queue *queue = hctx->driver_data; 2062 2063 return ib_process_cq_direct(queue->ib_cq, -1); 2064} 2065 2066static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req) 2067{ 2068 struct request *rq = blk_mq_rq_from_pdu(req); 2069 struct ib_mr_status mr_status; 2070 int ret; 2071 2072 ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status); 2073 if (ret) { 2074 pr_err("ib_check_mr_status failed, ret %d\n", ret); 2075 nvme_req(rq)->status = NVME_SC_INVALID_PI; 2076 return; 2077 } 2078 2079 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) { 2080 switch (mr_status.sig_err.err_type) { 2081 case IB_SIG_BAD_GUARD: 2082 nvme_req(rq)->status = NVME_SC_GUARD_CHECK; 2083 break; 2084 case IB_SIG_BAD_REFTAG: 2085 nvme_req(rq)->status = NVME_SC_REFTAG_CHECK; 2086 break; 2087 case IB_SIG_BAD_APPTAG: 2088 nvme_req(rq)->status = NVME_SC_APPTAG_CHECK; 2089 break; 2090 } 2091 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n", 2092 mr_status.sig_err.err_type, mr_status.sig_err.expected, 2093 mr_status.sig_err.actual); 2094 } 2095} 2096 2097static void nvme_rdma_complete_rq(struct request *rq) 2098{ 2099 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); 2100 struct nvme_rdma_queue *queue = req->queue; 2101 struct ib_device *ibdev = queue->device->dev; 2102 2103 if (req->use_sig_mr) 2104 nvme_rdma_check_pi_status(req); 2105 2106 nvme_rdma_unmap_data(queue, rq); 2107 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command), 2108 DMA_TO_DEVICE); 2109 nvme_complete_rq(rq); 2110} 2111 2112static void nvme_rdma_map_queues(struct blk_mq_tag_set *set) 2113{ 2114 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(set->driver_data); 2115 2116 nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues); 2117} 2118 2119static const struct blk_mq_ops nvme_rdma_mq_ops = { 2120 .queue_rq = nvme_rdma_queue_rq, 2121 .complete = nvme_rdma_complete_rq, 2122 .init_request = nvme_rdma_init_request, 2123 .exit_request = nvme_rdma_exit_request, 2124 .init_hctx = nvme_rdma_init_hctx, 2125 .timeout = nvme_rdma_timeout, 2126 .map_queues = nvme_rdma_map_queues, 2127 .poll = nvme_rdma_poll, 2128}; 2129 2130static const struct blk_mq_ops nvme_rdma_admin_mq_ops = { 2131 .queue_rq = nvme_rdma_queue_rq, 2132 .complete = nvme_rdma_complete_rq, 2133 .init_request = nvme_rdma_init_request, 2134 .exit_request = nvme_rdma_exit_request, 2135 .init_hctx = nvme_rdma_init_admin_hctx, 2136 .timeout = nvme_rdma_timeout, 2137}; 2138 2139static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown) 2140{ 2141 nvme_rdma_teardown_io_queues(ctrl, shutdown); 2142 nvme_quiesce_admin_queue(&ctrl->ctrl); 2143 nvme_disable_ctrl(&ctrl->ctrl, shutdown); 2144 nvme_rdma_teardown_admin_queue(ctrl, shutdown); 2145} 2146 2147static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl) 2148{ 2149 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true); 2150} 2151 2152static void nvme_rdma_reset_ctrl_work(struct work_struct *work) 2153{ 2154 struct nvme_rdma_ctrl *ctrl = 2155 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work); 2156 2157 nvme_stop_ctrl(&ctrl->ctrl); 2158 nvme_rdma_shutdown_ctrl(ctrl, false); 2159 2160 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { 2161 /* state change failure should never happen */ 2162 WARN_ON_ONCE(1); 2163 return; 2164 } 2165 2166 if (nvme_rdma_setup_ctrl(ctrl, false)) 2167 goto out_fail; 2168 2169 return; 2170 2171out_fail: 2172 ++ctrl->ctrl.nr_reconnects; 2173 nvme_rdma_reconnect_or_remove(ctrl); 2174} 2175 2176static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = { 2177 .name = "rdma", 2178 .module = THIS_MODULE, 2179 .flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED, 2180 .reg_read32 = nvmf_reg_read32, 2181 .reg_read64 = nvmf_reg_read64, 2182 .reg_write32 = nvmf_reg_write32, 2183 .free_ctrl = nvme_rdma_free_ctrl, 2184 .submit_async_event = nvme_rdma_submit_async_event, 2185 .delete_ctrl = nvme_rdma_delete_ctrl, 2186 .get_address = nvmf_get_address, 2187 .stop_ctrl = nvme_rdma_stop_ctrl, 2188}; 2189 2190/* 2191 * Fails a connection request if it matches an existing controller 2192 * (association) with the same tuple: 2193 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN> 2194 * 2195 * if local address is not specified in the request, it will match an 2196 * existing controller with all the other parameters the same and no 2197 * local port address specified as well. 2198 * 2199 * The ports don't need to be compared as they are intrinsically 2200 * already matched by the port pointers supplied. 2201 */ 2202static bool 2203nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts) 2204{ 2205 struct nvme_rdma_ctrl *ctrl; 2206 bool found = false; 2207 2208 mutex_lock(&nvme_rdma_ctrl_mutex); 2209 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { 2210 found = nvmf_ip_options_match(&ctrl->ctrl, opts); 2211 if (found) 2212 break; 2213 } 2214 mutex_unlock(&nvme_rdma_ctrl_mutex); 2215 2216 return found; 2217} 2218 2219static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev, 2220 struct nvmf_ctrl_options *opts) 2221{ 2222 struct nvme_rdma_ctrl *ctrl; 2223 int ret; 2224 bool changed; 2225 2226 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL); 2227 if (!ctrl) 2228 return ERR_PTR(-ENOMEM); 2229 ctrl->ctrl.opts = opts; 2230 INIT_LIST_HEAD(&ctrl->list); 2231 2232 if (!(opts->mask & NVMF_OPT_TRSVCID)) { 2233 opts->trsvcid = 2234 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL); 2235 if (!opts->trsvcid) { 2236 ret = -ENOMEM; 2237 goto out_free_ctrl; 2238 } 2239 opts->mask |= NVMF_OPT_TRSVCID; 2240 } 2241 2242 ret = inet_pton_with_scope(&init_net, AF_UNSPEC, 2243 opts->traddr, opts->trsvcid, &ctrl->addr); 2244 if (ret) { 2245 pr_err("malformed address passed: %s:%s\n", 2246 opts->traddr, opts->trsvcid); 2247 goto out_free_ctrl; 2248 } 2249 2250 if (opts->mask & NVMF_OPT_HOST_TRADDR) { 2251 ret = inet_pton_with_scope(&init_net, AF_UNSPEC, 2252 opts->host_traddr, NULL, &ctrl->src_addr); 2253 if (ret) { 2254 pr_err("malformed src address passed: %s\n", 2255 opts->host_traddr); 2256 goto out_free_ctrl; 2257 } 2258 } 2259 2260 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) { 2261 ret = -EALREADY; 2262 goto out_free_ctrl; 2263 } 2264 2265 INIT_DELAYED_WORK(&ctrl->reconnect_work, 2266 nvme_rdma_reconnect_ctrl_work); 2267 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work); 2268 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work); 2269 2270 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues + 2271 opts->nr_poll_queues + 1; 2272 ctrl->ctrl.sqsize = opts->queue_size - 1; 2273 ctrl->ctrl.kato = opts->kato; 2274 2275 ret = -ENOMEM; 2276 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues), 2277 GFP_KERNEL); 2278 if (!ctrl->queues) 2279 goto out_free_ctrl; 2280 2281 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops, 2282 0 /* no quirks, we're perfect! */); 2283 if (ret) 2284 goto out_kfree_queues; 2285 2286 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING); 2287 WARN_ON_ONCE(!changed); 2288 2289 ret = nvme_rdma_setup_ctrl(ctrl, true); 2290 if (ret) 2291 goto out_uninit_ctrl; 2292 2293 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n", 2294 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr); 2295 2296 mutex_lock(&nvme_rdma_ctrl_mutex); 2297 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list); 2298 mutex_unlock(&nvme_rdma_ctrl_mutex); 2299 2300 return &ctrl->ctrl; 2301 2302out_uninit_ctrl: 2303 nvme_uninit_ctrl(&ctrl->ctrl); 2304 nvme_put_ctrl(&ctrl->ctrl); 2305 if (ret > 0) 2306 ret = -EIO; 2307 return ERR_PTR(ret); 2308out_kfree_queues: 2309 kfree(ctrl->queues); 2310out_free_ctrl: 2311 kfree(ctrl); 2312 return ERR_PTR(ret); 2313} 2314 2315static struct nvmf_transport_ops nvme_rdma_transport = { 2316 .name = "rdma", 2317 .module = THIS_MODULE, 2318 .required_opts = NVMF_OPT_TRADDR, 2319 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY | 2320 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO | 2321 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES | 2322 NVMF_OPT_TOS, 2323 .create_ctrl = nvme_rdma_create_ctrl, 2324}; 2325 2326static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data) 2327{ 2328 struct nvme_rdma_ctrl *ctrl; 2329 struct nvme_rdma_device *ndev; 2330 bool found = false; 2331 2332 mutex_lock(&device_list_mutex); 2333 list_for_each_entry(ndev, &device_list, entry) { 2334 if (ndev->dev == ib_device) { 2335 found = true; 2336 break; 2337 } 2338 } 2339 mutex_unlock(&device_list_mutex); 2340 2341 if (!found) 2342 return; 2343 2344 /* Delete all controllers using this device */ 2345 mutex_lock(&nvme_rdma_ctrl_mutex); 2346 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { 2347 if (ctrl->device->dev != ib_device) 2348 continue; 2349 nvme_delete_ctrl(&ctrl->ctrl); 2350 } 2351 mutex_unlock(&nvme_rdma_ctrl_mutex); 2352 2353 flush_workqueue(nvme_delete_wq); 2354} 2355 2356static struct ib_client nvme_rdma_ib_client = { 2357 .name = "nvme_rdma", 2358 .remove = nvme_rdma_remove_one 2359}; 2360 2361static int __init nvme_rdma_init_module(void) 2362{ 2363 int ret; 2364 2365 ret = ib_register_client(&nvme_rdma_ib_client); 2366 if (ret) 2367 return ret; 2368 2369 ret = nvmf_register_transport(&nvme_rdma_transport); 2370 if (ret) 2371 goto err_unreg_client; 2372 2373 return 0; 2374 2375err_unreg_client: 2376 ib_unregister_client(&nvme_rdma_ib_client); 2377 return ret; 2378} 2379 2380static void __exit nvme_rdma_cleanup_module(void) 2381{ 2382 struct nvme_rdma_ctrl *ctrl; 2383 2384 nvmf_unregister_transport(&nvme_rdma_transport); 2385 ib_unregister_client(&nvme_rdma_ib_client); 2386 2387 mutex_lock(&nvme_rdma_ctrl_mutex); 2388 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) 2389 nvme_delete_ctrl(&ctrl->ctrl); 2390 mutex_unlock(&nvme_rdma_ctrl_mutex); 2391 flush_workqueue(nvme_delete_wq); 2392} 2393 2394module_init(nvme_rdma_init_module); 2395module_exit(nvme_rdma_cleanup_module); 2396 2397MODULE_LICENSE("GPL v2"); 2398