drivers/nvme/host/rdma.c

   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
  42 struct 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
  50 struct nvme_rdma_qe {
  51         struct ib_cqe           cqe;
  52         void                    *data;
  53         u64                     dma;
  54 };
  55
  56 struct nvme_rdma_sgl {
  57         int                     nents;
  58         struct sg_table         sg_table;
  59 };
  60
  61 struct nvme_rdma_queue;
  62 struct 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
  79 enum 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
  85 struct 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
 103 struct 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
 130 static 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
 135 static LIST_HEAD(device_list);
 136 static DEFINE_MUTEX(device_list_mutex);
 137
 138 static LIST_HEAD(nvme_rdma_ctrl_list);
 139 static 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  */
 146 static bool register_always = true;
 147 module_param(register_always, bool, 0444);
 148 MODULE_PARM_DESC(register_always,
 149          "Use memory registration even for contiguous memory regions");
 150
 151 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 152                 struct rdma_cm_event *event);
 153 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 154 static void nvme_rdma_complete_rq(struct request *rq);
 155
 156 static const struct blk_mq_ops nvme_rdma_mq_ops;
 157 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 158
 159 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 160 {
 161         return queue - queue->ctrl->queues;
 162 }
 163
 164 static 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
 171 static 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
 176 static 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
 183 static 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
 200 static 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
 211 static 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
 234 out_free_ring:
 235         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 236         return NULL;
 237 }
 238
 239 static 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
 246 static 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
 257 static 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
 285 static 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
 293 static 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
 319 static 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
 331 static 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
 343 static 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
 356 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 357 {
 358         kref_put(&dev->ref, nvme_rdma_free_dev);
 359 }
 360
 361 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 362 {
 363         return kref_get_unless_zero(&dev->ref);
 364 }
 365
 366 static struct nvme_rdma_device *
 367 nvme_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);
 400 out_unlock:
 401         mutex_unlock(&device_list_mutex);
 402         return ndev;
 403
 404 out_free_pd:
 405         ib_dealloc_pd(ndev->pd);
 406 out_free_dev:
 407         kfree(ndev);
 408 out_err:
 409         mutex_unlock(&device_list_mutex);
 410         return NULL;
 411 }
 412
 413 static 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
 421 static 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
 450 static 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
 462 static 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
 489 static 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
 555 out_destroy_mr_pool:
 556         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 557 out_destroy_ring:
 558         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 559                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 560 out_destroy_qp:
 561         rdma_destroy_qp(queue->cm_id);
 562 out_destroy_ib_cq:
 563         nvme_rdma_free_cq(queue);
 564 out_put_dev:
 565         nvme_rdma_dev_put(queue->device);
 566         return ret;
 567 }
 568
 569 static 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
 625 out_destroy_cm_id:
 626         rdma_destroy_id(queue->cm_id);
 627         nvme_rdma_destroy_queue_ib(queue);
 628 out_destroy_mutex:
 629         mutex_destroy(&queue->queue_lock);
 630         return ret;
 631 }
 632
 633 static 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
 639 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 640 {
 641         mutex_lock(&queue->queue_lock);
 642         if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 643                 __nvme_rdma_stop_queue(queue);
 644         mutex_unlock(&queue->queue_lock);
 645 }
 646
 647 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 648 {
 649         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 650                 return;
 651
 652         rdma_destroy_id(queue->cm_id);
 653         nvme_rdma_destroy_queue_ib(queue);
 654         mutex_destroy(&queue->queue_lock);
 655 }
 656
 657 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 658 {
 659         int i;
 660
 661         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 662                 nvme_rdma_free_queue(&ctrl->queues[i]);
 663 }
 664
 665 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 666 {
 667         int i;
 668
 669         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 670                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 671 }
 672
 673 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 674 {
 675         struct nvme_rdma_queue *queue = &ctrl->queues[idx];
 676         int ret;
 677
 678         if (idx)
 679                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 680         else
 681                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 682
 683         if (!ret) {
 684                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 685         } else {
 686                 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 687                         __nvme_rdma_stop_queue(queue);
 688                 dev_info(ctrl->ctrl.device,
 689                         "failed to connect queue: %d ret=%d\n", idx, ret);
 690         }
 691         return ret;
 692 }
 693
 694 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl,
 695                                      int first, int last)
 696 {
 697         int i, ret = 0;
 698
 699         for (i = first; i < last; i++) {
 700                 ret = nvme_rdma_start_queue(ctrl, i);
 701                 if (ret)
 702                         goto out_stop_queues;
 703         }
 704
 705         return 0;
 706
 707 out_stop_queues:
 708         for (i--; i >= first; i--)
 709                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 710         return ret;
 711 }
 712
 713 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 714 {
 715         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 716         unsigned int nr_io_queues;
 717         int i, ret;
 718
 719         nr_io_queues = nvmf_nr_io_queues(opts);
 720         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 721         if (ret)
 722                 return ret;
 723
 724         if (nr_io_queues == 0) {
 725                 dev_err(ctrl->ctrl.device,
 726                         "unable to set any I/O queues\n");
 727                 return -ENOMEM;
 728         }
 729
 730         ctrl->ctrl.queue_count = nr_io_queues + 1;
 731         dev_info(ctrl->ctrl.device,
 732                 "creating %d I/O queues.\n", nr_io_queues);
 733
 734         nvmf_set_io_queues(opts, nr_io_queues, ctrl->io_queues);
 735         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 736                 ret = nvme_rdma_alloc_queue(ctrl, i,
 737                                 ctrl->ctrl.sqsize + 1);
 738                 if (ret)
 739                         goto out_free_queues;
 740         }
 741
 742         return 0;
 743
 744 out_free_queues:
 745         for (i--; i >= 1; i--)
 746                 nvme_rdma_free_queue(&ctrl->queues[i]);
 747
 748         return ret;
 749 }
 750
 751 static int nvme_rdma_alloc_tag_set(struct nvme_ctrl *ctrl)
 752 {
 753         unsigned int cmd_size = sizeof(struct nvme_rdma_request) +
 754                                 NVME_RDMA_DATA_SGL_SIZE;
 755
 756         if (ctrl->max_integrity_segments)
 757                 cmd_size += sizeof(struct nvme_rdma_sgl) +
 758                             NVME_RDMA_METADATA_SGL_SIZE;
 759
 760         return nvme_alloc_io_tag_set(ctrl, &to_rdma_ctrl(ctrl)->tag_set,
 761                         &nvme_rdma_mq_ops,
 762                         ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
 763                         cmd_size);
 764 }
 765
 766 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
 767 {
 768         if (ctrl->async_event_sqe.data) {
 769                 cancel_work_sync(&ctrl->ctrl.async_event_work);
 770                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 771                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 772                 ctrl->async_event_sqe.data = NULL;
 773         }
 774         nvme_rdma_free_queue(&ctrl->queues[0]);
 775 }
 776
 777 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 778                 bool new)
 779 {
 780         bool pi_capable = false;
 781         int error;
 782
 783         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 784         if (error)
 785                 return error;
 786
 787         ctrl->device = ctrl->queues[0].device;
 788         ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
 789
 790         /* T10-PI support */
 791         if (ctrl->device->dev->attrs.kernel_cap_flags &
 792             IBK_INTEGRITY_HANDOVER)
 793                 pi_capable = true;
 794
 795         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
 796                                                         pi_capable);
 797
 798         /*
 799          * Bind the async event SQE DMA mapping to the admin queue lifetime.
 800          * It's safe, since any chage in the underlying RDMA device will issue
 801          * error recovery and queue re-creation.
 802          */
 803         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 804                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 805         if (error)
 806                 goto out_free_queue;
 807
 808         if (new) {
 809                 error = nvme_alloc_admin_tag_set(&ctrl->ctrl,
 810                                 &ctrl->admin_tag_set, &nvme_rdma_admin_mq_ops,
 811                                 sizeof(struct nvme_rdma_request) +
 812                                 NVME_RDMA_DATA_SGL_SIZE);
 813                 if (error)
 814                         goto out_free_async_qe;
 815
 816         }
 817
 818         error = nvme_rdma_start_queue(ctrl, 0);
 819         if (error)
 820                 goto out_remove_admin_tag_set;
 821
 822         error = nvme_enable_ctrl(&ctrl->ctrl);
 823         if (error)
 824                 goto out_stop_queue;
 825
 826         ctrl->ctrl.max_segments = ctrl->max_fr_pages;
 827         ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
 828         if (pi_capable)
 829                 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
 830         else
 831                 ctrl->ctrl.max_integrity_segments = 0;
 832
 833         nvme_unquiesce_admin_queue(&ctrl->ctrl);
 834
 835         error = nvme_init_ctrl_finish(&ctrl->ctrl, false);
 836         if (error)
 837                 goto out_quiesce_queue;
 838
 839         return 0;
 840
 841 out_quiesce_queue:
 842         nvme_quiesce_admin_queue(&ctrl->ctrl);
 843         blk_sync_queue(ctrl->ctrl.admin_q);
 844 out_stop_queue:
 845         nvme_rdma_stop_queue(&ctrl->queues[0]);
 846         nvme_cancel_admin_tagset(&ctrl->ctrl);
 847 out_remove_admin_tag_set:
 848         if (new)
 849                 nvme_remove_admin_tag_set(&ctrl->ctrl);
 850 out_free_async_qe:
 851         if (ctrl->async_event_sqe.data) {
 852                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 853                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 854                 ctrl->async_event_sqe.data = NULL;
 855         }
 856 out_free_queue:
 857         nvme_rdma_free_queue(&ctrl->queues[0]);
 858         return error;
 859 }
 860
 861 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 862 {
 863         int ret, nr_queues;
 864
 865         ret = nvme_rdma_alloc_io_queues(ctrl);
 866         if (ret)
 867                 return ret;
 868
 869         if (new) {
 870                 ret = nvme_rdma_alloc_tag_set(&ctrl->ctrl);
 871                 if (ret)
 872                         goto out_free_io_queues;
 873         }
 874
 875         /*
 876          * Only start IO queues for which we have allocated the tagset
 877          * and limitted it to the available queues. On reconnects, the
 878          * queue number might have changed.
 879          */
 880         nr_queues = min(ctrl->tag_set.nr_hw_queues + 1, ctrl->ctrl.queue_count);
 881         ret = nvme_rdma_start_io_queues(ctrl, 1, nr_queues);
 882         if (ret)
 883                 goto out_cleanup_tagset;
 884
 885         if (!new) {
 886                 nvme_start_freeze(&ctrl->ctrl);
 887                 nvme_unquiesce_io_queues(&ctrl->ctrl);
 888                 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
 889                         /*
 890                          * If we timed out waiting for freeze we are likely to
 891                          * be stuck.  Fail the controller initialization just
 892                          * to be safe.
 893                          */
 894                         ret = -ENODEV;
 895                         nvme_unfreeze(&ctrl->ctrl);
 896                         goto out_wait_freeze_timed_out;
 897                 }
 898                 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
 899                         ctrl->ctrl.queue_count - 1);
 900                 nvme_unfreeze(&ctrl->ctrl);
 901         }
 902
 903         /*
 904          * If the number of queues has increased (reconnect case)
 905          * start all new queues now.
 906          */
 907         ret = nvme_rdma_start_io_queues(ctrl, nr_queues,
 908                                         ctrl->tag_set.nr_hw_queues + 1);
 909         if (ret)
 910                 goto out_wait_freeze_timed_out;
 911
 912         return 0;
 913
 914 out_wait_freeze_timed_out:
 915         nvme_quiesce_io_queues(&ctrl->ctrl);
 916         nvme_sync_io_queues(&ctrl->ctrl);
 917         nvme_rdma_stop_io_queues(ctrl);
 918 out_cleanup_tagset:
 919         nvme_cancel_tagset(&ctrl->ctrl);
 920         if (new)
 921                 nvme_remove_io_tag_set(&ctrl->ctrl);
 922 out_free_io_queues:
 923         nvme_rdma_free_io_queues(ctrl);
 924         return ret;
 925 }
 926
 927 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
 928                 bool remove)
 929 {
 930         nvme_quiesce_admin_queue(&ctrl->ctrl);
 931         blk_sync_queue(ctrl->ctrl.admin_q);
 932         nvme_rdma_stop_queue(&ctrl->queues[0]);
 933         nvme_cancel_admin_tagset(&ctrl->ctrl);
 934         if (remove) {
 935                 nvme_unquiesce_admin_queue(&ctrl->ctrl);
 936                 nvme_remove_admin_tag_set(&ctrl->ctrl);
 937         }
 938         nvme_rdma_destroy_admin_queue(ctrl);
 939 }
 940
 941 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
 942                 bool remove)
 943 {
 944         if (ctrl->ctrl.queue_count > 1) {
 945                 nvme_quiesce_io_queues(&ctrl->ctrl);
 946                 nvme_sync_io_queues(&ctrl->ctrl);
 947                 nvme_rdma_stop_io_queues(ctrl);
 948                 nvme_cancel_tagset(&ctrl->ctrl);
 949                 if (remove) {
 950                         nvme_unquiesce_io_queues(&ctrl->ctrl);
 951                         nvme_remove_io_tag_set(&ctrl->ctrl);
 952                 }
 953                 nvme_rdma_free_io_queues(ctrl);
 954         }
 955 }
 956
 957 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
 958 {
 959         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 960
 961         flush_work(&ctrl->err_work);
 962         cancel_delayed_work_sync(&ctrl->reconnect_work);
 963 }
 964
 965 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 966 {
 967         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 968
 969         if (list_empty(&ctrl->list))
 970                 goto free_ctrl;
 971
 972         mutex_lock(&nvme_rdma_ctrl_mutex);
 973         list_del(&ctrl->list);
 974         mutex_unlock(&nvme_rdma_ctrl_mutex);
 975
 976         nvmf_free_options(nctrl->opts);
 977 free_ctrl:
 978         kfree(ctrl->queues);
 979         kfree(ctrl);
 980 }
 981
 982 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 983 {
 984         /* If we are resetting/deleting then do nothing */
 985         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 986                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 987                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 988                 return;
 989         }
 990
 991         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 992                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 993                         ctrl->ctrl.opts->reconnect_delay);
 994                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 995                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 996         } else {
 997                 nvme_delete_ctrl(&ctrl->ctrl);
 998         }
 999 }
1000
1001 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1002 {
1003         int ret;
1004         bool changed;
1005
1006         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1007         if (ret)
1008                 return ret;
1009
1010         if (ctrl->ctrl.icdoff) {
1011                 ret = -EOPNOTSUPP;
1012                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1013                 goto destroy_admin;
1014         }
1015
1016         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1017                 ret = -EOPNOTSUPP;
1018                 dev_err(ctrl->ctrl.device,
1019                         "Mandatory keyed sgls are not supported!\n");
1020                 goto destroy_admin;
1021         }
1022
1023         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1024                 dev_warn(ctrl->ctrl.device,
1025                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1026                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1027         }
1028
1029         if (ctrl->ctrl.sqsize + 1 > NVME_RDMA_MAX_QUEUE_SIZE) {
1030                 dev_warn(ctrl->ctrl.device,
1031                         "ctrl sqsize %u > max queue size %u, clamping down\n",
1032                         ctrl->ctrl.sqsize + 1, NVME_RDMA_MAX_QUEUE_SIZE);
1033                 ctrl->ctrl.sqsize = NVME_RDMA_MAX_QUEUE_SIZE - 1;
1034         }
1035
1036         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1037                 dev_warn(ctrl->ctrl.device,
1038                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1039                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1040                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1041         }
1042
1043         if (ctrl->ctrl.sgls & (1 << 20))
1044                 ctrl->use_inline_data = true;
1045
1046         if (ctrl->ctrl.queue_count > 1) {
1047                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1048                 if (ret)
1049                         goto destroy_admin;
1050         }
1051
1052         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1053         if (!changed) {
1054                 /*
1055                  * state change failure is ok if we started ctrl delete,
1056                  * unless we're during creation of a new controller to
1057                  * avoid races with teardown flow.
1058                  */
1059                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1060                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1061                 WARN_ON_ONCE(new);
1062                 ret = -EINVAL;
1063                 goto destroy_io;
1064         }
1065
1066         nvme_start_ctrl(&ctrl->ctrl);
1067         return 0;
1068
1069 destroy_io:
1070         if (ctrl->ctrl.queue_count > 1) {
1071                 nvme_quiesce_io_queues(&ctrl->ctrl);
1072                 nvme_sync_io_queues(&ctrl->ctrl);
1073                 nvme_rdma_stop_io_queues(ctrl);
1074                 nvme_cancel_tagset(&ctrl->ctrl);
1075                 if (new)
1076                         nvme_remove_io_tag_set(&ctrl->ctrl);
1077                 nvme_rdma_free_io_queues(ctrl);
1078         }
1079 destroy_admin:
1080         nvme_quiesce_admin_queue(&ctrl->ctrl);
1081         blk_sync_queue(ctrl->ctrl.admin_q);
1082         nvme_rdma_stop_queue(&ctrl->queues[0]);
1083         nvme_cancel_admin_tagset(&ctrl->ctrl);
1084         if (new)
1085                 nvme_remove_admin_tag_set(&ctrl->ctrl);
1086         nvme_rdma_destroy_admin_queue(ctrl);
1087         return ret;
1088 }
1089
1090 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1091 {
1092         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1093                         struct nvme_rdma_ctrl, reconnect_work);
1094
1095         ++ctrl->ctrl.nr_reconnects;
1096
1097         if (nvme_rdma_setup_ctrl(ctrl, false))
1098                 goto requeue;
1099
1100         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1101                         ctrl->ctrl.nr_reconnects);
1102
1103         ctrl->ctrl.nr_reconnects = 0;
1104
1105         return;
1106
1107 requeue:
1108         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1109                         ctrl->ctrl.nr_reconnects);
1110         nvme_rdma_reconnect_or_remove(ctrl);
1111 }
1112
1113 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1114 {
1115         struct nvme_rdma_ctrl *ctrl = container_of(work,
1116                         struct nvme_rdma_ctrl, err_work);
1117
1118         nvme_stop_keep_alive(&ctrl->ctrl);
1119         flush_work(&ctrl->ctrl.async_event_work);
1120         nvme_rdma_teardown_io_queues(ctrl, false);
1121         nvme_unquiesce_io_queues(&ctrl->ctrl);
1122         nvme_rdma_teardown_admin_queue(ctrl, false);
1123         nvme_unquiesce_admin_queue(&ctrl->ctrl);
1124         nvme_auth_stop(&ctrl->ctrl);
1125
1126         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1127                 /* state change failure is ok if we started ctrl delete */
1128                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1129                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1130                 return;
1131         }
1132
1133         nvme_rdma_reconnect_or_remove(ctrl);
1134 }
1135
1136 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1137 {
1138         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1139                 return;
1140
1141         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1142         queue_work(nvme_reset_wq, &ctrl->err_work);
1143 }
1144
1145 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1146 {
1147         struct request *rq = blk_mq_rq_from_pdu(req);
1148
1149         if (!refcount_dec_and_test(&req->ref))
1150                 return;
1151         if (!nvme_try_complete_req(rq, req->status, req->result))
1152                 nvme_rdma_complete_rq(rq);
1153 }
1154
1155 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1156                 const char *op)
1157 {
1158         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1159         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1160
1161         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1162                 dev_info(ctrl->ctrl.device,
1163                              "%s for CQE 0x%p failed with status %s (%d)\n",
1164                              op, wc->wr_cqe,
1165                              ib_wc_status_msg(wc->status), wc->status);
1166         nvme_rdma_error_recovery(ctrl);
1167 }
1168
1169 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1170 {
1171         if (unlikely(wc->status != IB_WC_SUCCESS))
1172                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1173 }
1174
1175 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1176 {
1177         struct nvme_rdma_request *req =
1178                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1179
1180         if (unlikely(wc->status != IB_WC_SUCCESS))
1181                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1182         else
1183                 nvme_rdma_end_request(req);
1184 }
1185
1186 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1187                 struct nvme_rdma_request *req)
1188 {
1189         struct ib_send_wr wr = {
1190                 .opcode             = IB_WR_LOCAL_INV,
1191                 .next               = NULL,
1192                 .num_sge            = 0,
1193                 .send_flags         = IB_SEND_SIGNALED,
1194                 .ex.invalidate_rkey = req->mr->rkey,
1195         };
1196
1197         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1198         wr.wr_cqe = &req->reg_cqe;
1199
1200         return ib_post_send(queue->qp, &wr, NULL);
1201 }
1202
1203 static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq)
1204 {
1205         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1206
1207         if (blk_integrity_rq(rq)) {
1208                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1209                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1210                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1211                                       NVME_INLINE_METADATA_SG_CNT);
1212         }
1213
1214         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1215                         rq_dma_dir(rq));
1216         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1217 }
1218
1219 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1220                 struct request *rq)
1221 {
1222         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1223         struct nvme_rdma_device *dev = queue->device;
1224         struct ib_device *ibdev = dev->dev;
1225         struct list_head *pool = &queue->qp->rdma_mrs;
1226
1227         if (!blk_rq_nr_phys_segments(rq))
1228                 return;
1229
1230         if (req->use_sig_mr)
1231                 pool = &queue->qp->sig_mrs;
1232
1233         if (req->mr) {
1234                 ib_mr_pool_put(queue->qp, pool, req->mr);
1235                 req->mr = NULL;
1236         }
1237
1238         nvme_rdma_dma_unmap_req(ibdev, rq);
1239 }
1240
1241 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1242 {
1243         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1244
1245         sg->addr = 0;
1246         put_unaligned_le24(0, sg->length);
1247         put_unaligned_le32(0, sg->key);
1248         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1249         return 0;
1250 }
1251
1252 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1253                 struct nvme_rdma_request *req, struct nvme_command *c,
1254                 int count)
1255 {
1256         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1257         struct ib_sge *sge = &req->sge[1];
1258         struct scatterlist *sgl;
1259         u32 len = 0;
1260         int i;
1261
1262         for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1263                 sge->addr = sg_dma_address(sgl);
1264                 sge->length = sg_dma_len(sgl);
1265                 sge->lkey = queue->device->pd->local_dma_lkey;
1266                 len += sge->length;
1267                 sge++;
1268         }
1269
1270         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1271         sg->length = cpu_to_le32(len);
1272         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1273
1274         req->num_sge += count;
1275         return 0;
1276 }
1277
1278 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1279                 struct nvme_rdma_request *req, struct nvme_command *c)
1280 {
1281         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1282
1283         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1284         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1285         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1286         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1287         return 0;
1288 }
1289
1290 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1291                 struct nvme_rdma_request *req, struct nvme_command *c,
1292                 int count)
1293 {
1294         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1295         int nr;
1296
1297         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1298         if (WARN_ON_ONCE(!req->mr))
1299                 return -EAGAIN;
1300
1301         /*
1302          * Align the MR to a 4K page size to match the ctrl page size and
1303          * the block virtual boundary.
1304          */
1305         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1306                           SZ_4K);
1307         if (unlikely(nr < count)) {
1308                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1309                 req->mr = NULL;
1310                 if (nr < 0)
1311                         return nr;
1312                 return -EINVAL;
1313         }
1314
1315         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1316
1317         req->reg_cqe.done = nvme_rdma_memreg_done;
1318         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1319         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1320         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1321         req->reg_wr.wr.num_sge = 0;
1322         req->reg_wr.mr = req->mr;
1323         req->reg_wr.key = req->mr->rkey;
1324         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1325                              IB_ACCESS_REMOTE_READ |
1326                              IB_ACCESS_REMOTE_WRITE;
1327
1328         sg->addr = cpu_to_le64(req->mr->iova);
1329         put_unaligned_le24(req->mr->length, sg->length);
1330         put_unaligned_le32(req->mr->rkey, sg->key);
1331         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1332                         NVME_SGL_FMT_INVALIDATE;
1333
1334         return 0;
1335 }
1336
1337 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1338                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1339                 u16 control, u8 pi_type)
1340 {
1341         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1342         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1343         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1344         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1345         if (control & NVME_RW_PRINFO_PRCHK_REF)
1346                 domain->sig.dif.ref_remap = true;
1347
1348         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1349         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1350         domain->sig.dif.app_escape = true;
1351         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1352                 domain->sig.dif.ref_escape = true;
1353 }
1354
1355 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1356                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1357                 u8 pi_type)
1358 {
1359         u16 control = le16_to_cpu(cmd->rw.control);
1360
1361         memset(sig_attrs, 0, sizeof(*sig_attrs));
1362         if (control & NVME_RW_PRINFO_PRACT) {
1363                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1364                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1365                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1366                                          pi_type);
1367                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1368                 control &= ~NVME_RW_PRINFO_PRACT;
1369                 cmd->rw.control = cpu_to_le16(control);
1370         } else {
1371                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1372                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1373                                          pi_type);
1374                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1375                                          pi_type);
1376         }
1377 }
1378
1379 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1380 {
1381         *mask = 0;
1382         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1383                 *mask |= IB_SIG_CHECK_REFTAG;
1384         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1385                 *mask |= IB_SIG_CHECK_GUARD;
1386 }
1387
1388 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1389 {
1390         if (unlikely(wc->status != IB_WC_SUCCESS))
1391                 nvme_rdma_wr_error(cq, wc, "SIG");
1392 }
1393
1394 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1395                 struct nvme_rdma_request *req, struct nvme_command *c,
1396                 int count, int pi_count)
1397 {
1398         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1399         struct ib_reg_wr *wr = &req->reg_wr;
1400         struct request *rq = blk_mq_rq_from_pdu(req);
1401         struct nvme_ns *ns = rq->q->queuedata;
1402         struct bio *bio = rq->bio;
1403         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1404         int nr;
1405
1406         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1407         if (WARN_ON_ONCE(!req->mr))
1408                 return -EAGAIN;
1409
1410         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1411                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1412                              SZ_4K);
1413         if (unlikely(nr))
1414                 goto mr_put;
1415
1416         nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1417                                 req->mr->sig_attrs, ns->pi_type);
1418         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1419
1420         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1421
1422         req->reg_cqe.done = nvme_rdma_sig_done;
1423         memset(wr, 0, sizeof(*wr));
1424         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1425         wr->wr.wr_cqe = &req->reg_cqe;
1426         wr->wr.num_sge = 0;
1427         wr->wr.send_flags = 0;
1428         wr->mr = req->mr;
1429         wr->key = req->mr->rkey;
1430         wr->access = IB_ACCESS_LOCAL_WRITE |
1431                      IB_ACCESS_REMOTE_READ |
1432                      IB_ACCESS_REMOTE_WRITE;
1433
1434         sg->addr = cpu_to_le64(req->mr->iova);
1435         put_unaligned_le24(req->mr->length, sg->length);
1436         put_unaligned_le32(req->mr->rkey, sg->key);
1437         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1438
1439         return 0;
1440
1441 mr_put:
1442         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1443         req->mr = NULL;
1444         if (nr < 0)
1445                 return nr;
1446         return -EINVAL;
1447 }
1448
1449 static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq,
1450                 int *count, int *pi_count)
1451 {
1452         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1453         int ret;
1454
1455         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1456         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1457                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1458                         NVME_INLINE_SG_CNT);
1459         if (ret)
1460                 return -ENOMEM;
1461
1462         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1463                                             req->data_sgl.sg_table.sgl);
1464
1465         *count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1466                                req->data_sgl.nents, rq_dma_dir(rq));
1467         if (unlikely(*count <= 0)) {
1468                 ret = -EIO;
1469                 goto out_free_table;
1470         }
1471
1472         if (blk_integrity_rq(rq)) {
1473                 req->metadata_sgl->sg_table.sgl =
1474                         (struct scatterlist *)(req->metadata_sgl + 1);
1475                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1476                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1477                                 req->metadata_sgl->sg_table.sgl,
1478                                 NVME_INLINE_METADATA_SG_CNT);
1479                 if (unlikely(ret)) {
1480                         ret = -ENOMEM;
1481                         goto out_unmap_sg;
1482                 }
1483
1484                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1485                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1486                 *pi_count = ib_dma_map_sg(ibdev,
1487                                           req->metadata_sgl->sg_table.sgl,
1488                                           req->metadata_sgl->nents,
1489                                           rq_dma_dir(rq));
1490                 if (unlikely(*pi_count <= 0)) {
1491                         ret = -EIO;
1492                         goto out_free_pi_table;
1493                 }
1494         }
1495
1496         return 0;
1497
1498 out_free_pi_table:
1499         sg_free_table_chained(&req->metadata_sgl->sg_table,
1500                               NVME_INLINE_METADATA_SG_CNT);
1501 out_unmap_sg:
1502         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1503                         rq_dma_dir(rq));
1504 out_free_table:
1505         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1506         return ret;
1507 }
1508
1509 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1510                 struct request *rq, struct nvme_command *c)
1511 {
1512         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1513         struct nvme_rdma_device *dev = queue->device;
1514         struct ib_device *ibdev = dev->dev;
1515         int pi_count = 0;
1516         int count, ret;
1517
1518         req->num_sge = 1;
1519         refcount_set(&req->ref, 2); /* send and recv completions */
1520
1521         c->common.flags |= NVME_CMD_SGL_METABUF;
1522
1523         if (!blk_rq_nr_phys_segments(rq))
1524                 return nvme_rdma_set_sg_null(c);
1525
1526         ret = nvme_rdma_dma_map_req(ibdev, rq, &count, &pi_count);
1527         if (unlikely(ret))
1528                 return ret;
1529
1530         if (req->use_sig_mr) {
1531                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1532                 goto out;
1533         }
1534
1535         if (count <= dev->num_inline_segments) {
1536                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1537                     queue->ctrl->use_inline_data &&
1538                     blk_rq_payload_bytes(rq) <=
1539                                 nvme_rdma_inline_data_size(queue)) {
1540                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1541                         goto out;
1542                 }
1543
1544                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1545                         ret = nvme_rdma_map_sg_single(queue, req, c);
1546                         goto out;
1547                 }
1548         }
1549
1550         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1551 out:
1552         if (unlikely(ret))
1553                 goto out_dma_unmap_req;
1554
1555         return 0;
1556
1557 out_dma_unmap_req:
1558         nvme_rdma_dma_unmap_req(ibdev, rq);
1559         return ret;
1560 }
1561
1562 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1563 {
1564         struct nvme_rdma_qe *qe =
1565                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1566         struct nvme_rdma_request *req =
1567                 container_of(qe, struct nvme_rdma_request, sqe);
1568
1569         if (unlikely(wc->status != IB_WC_SUCCESS))
1570                 nvme_rdma_wr_error(cq, wc, "SEND");
1571         else
1572                 nvme_rdma_end_request(req);
1573 }
1574
1575 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1576                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1577                 struct ib_send_wr *first)
1578 {
1579         struct ib_send_wr wr;
1580         int ret;
1581
1582         sge->addr   = qe->dma;
1583         sge->length = sizeof(struct nvme_command);
1584         sge->lkey   = queue->device->pd->local_dma_lkey;
1585
1586         wr.next       = NULL;
1587         wr.wr_cqe     = &qe->cqe;
1588         wr.sg_list    = sge;
1589         wr.num_sge    = num_sge;
1590         wr.opcode     = IB_WR_SEND;
1591         wr.send_flags = IB_SEND_SIGNALED;
1592
1593         if (first)
1594                 first->next = &wr;
1595         else
1596                 first = &wr;
1597
1598         ret = ib_post_send(queue->qp, first, NULL);
1599         if (unlikely(ret)) {
1600                 dev_err(queue->ctrl->ctrl.device,
1601                              "%s failed with error code %d\n", __func__, ret);
1602         }
1603         return ret;
1604 }
1605
1606 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1607                 struct nvme_rdma_qe *qe)
1608 {
1609         struct ib_recv_wr wr;
1610         struct ib_sge list;
1611         int ret;
1612
1613         list.addr   = qe->dma;
1614         list.length = sizeof(struct nvme_completion);
1615         list.lkey   = queue->device->pd->local_dma_lkey;
1616
1617         qe->cqe.done = nvme_rdma_recv_done;
1618
1619         wr.next     = NULL;
1620         wr.wr_cqe   = &qe->cqe;
1621         wr.sg_list  = &list;
1622         wr.num_sge  = 1;
1623
1624         ret = ib_post_recv(queue->qp, &wr, NULL);
1625         if (unlikely(ret)) {
1626                 dev_err(queue->ctrl->ctrl.device,
1627                         "%s failed with error code %d\n", __func__, ret);
1628         }
1629         return ret;
1630 }
1631
1632 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1633 {
1634         u32 queue_idx = nvme_rdma_queue_idx(queue);
1635
1636         if (queue_idx == 0)
1637                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1638         return queue->ctrl->tag_set.tags[queue_idx - 1];
1639 }
1640
1641 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1642 {
1643         if (unlikely(wc->status != IB_WC_SUCCESS))
1644                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1645 }
1646
1647 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1648 {
1649         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1650         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1651         struct ib_device *dev = queue->device->dev;
1652         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1653         struct nvme_command *cmd = sqe->data;
1654         struct ib_sge sge;
1655         int ret;
1656
1657         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1658
1659         memset(cmd, 0, sizeof(*cmd));
1660         cmd->common.opcode = nvme_admin_async_event;
1661         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1662         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1663         nvme_rdma_set_sg_null(cmd);
1664
1665         sqe->cqe.done = nvme_rdma_async_done;
1666
1667         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1668                         DMA_TO_DEVICE);
1669
1670         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1671         WARN_ON_ONCE(ret);
1672 }
1673
1674 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1675                 struct nvme_completion *cqe, struct ib_wc *wc)
1676 {
1677         struct request *rq;
1678         struct nvme_rdma_request *req;
1679
1680         rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1681         if (!rq) {
1682                 dev_err(queue->ctrl->ctrl.device,
1683                         "got bad command_id %#x on QP %#x\n",
1684                         cqe->command_id, queue->qp->qp_num);
1685                 nvme_rdma_error_recovery(queue->ctrl);
1686                 return;
1687         }
1688         req = blk_mq_rq_to_pdu(rq);
1689
1690         req->status = cqe->status;
1691         req->result = cqe->result;
1692
1693         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1694                 if (unlikely(!req->mr ||
1695                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1696                         dev_err(queue->ctrl->ctrl.device,
1697                                 "Bogus remote invalidation for rkey %#x\n",
1698                                 req->mr ? req->mr->rkey : 0);
1699                         nvme_rdma_error_recovery(queue->ctrl);
1700                 }
1701         } else if (req->mr) {
1702                 int ret;
1703
1704                 ret = nvme_rdma_inv_rkey(queue, req);
1705                 if (unlikely(ret < 0)) {
1706                         dev_err(queue->ctrl->ctrl.device,
1707                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1708                                 req->mr->rkey, ret);
1709                         nvme_rdma_error_recovery(queue->ctrl);
1710                 }
1711                 /* the local invalidation completion will end the request */
1712                 return;
1713         }
1714
1715         nvme_rdma_end_request(req);
1716 }
1717
1718 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1719 {
1720         struct nvme_rdma_qe *qe =
1721                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1722         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1723         struct ib_device *ibdev = queue->device->dev;
1724         struct nvme_completion *cqe = qe->data;
1725         const size_t len = sizeof(struct nvme_completion);
1726
1727         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1728                 nvme_rdma_wr_error(cq, wc, "RECV");
1729                 return;
1730         }
1731
1732         /* sanity checking for received data length */
1733         if (unlikely(wc->byte_len < len)) {
1734                 dev_err(queue->ctrl->ctrl.device,
1735                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1736                 nvme_rdma_error_recovery(queue->ctrl);
1737                 return;
1738         }
1739
1740         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1741         /*
1742          * AEN requests are special as they don't time out and can
1743          * survive any kind of queue freeze and often don't respond to
1744          * aborts.  We don't even bother to allocate a struct request
1745          * for them but rather special case them here.
1746          */
1747         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1748                                      cqe->command_id)))
1749                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1750                                 &cqe->result);
1751         else
1752                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1753         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1754
1755         nvme_rdma_post_recv(queue, qe);
1756 }
1757
1758 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1759 {
1760         int ret, i;
1761
1762         for (i = 0; i < queue->queue_size; i++) {
1763                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1764                 if (ret)
1765                         return ret;
1766         }
1767
1768         return 0;
1769 }
1770
1771 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1772                 struct rdma_cm_event *ev)
1773 {
1774         struct rdma_cm_id *cm_id = queue->cm_id;
1775         int status = ev->status;
1776         const char *rej_msg;
1777         const struct nvme_rdma_cm_rej *rej_data;
1778         u8 rej_data_len;
1779
1780         rej_msg = rdma_reject_msg(cm_id, status);
1781         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1782
1783         if (rej_data && rej_data_len >= sizeof(u16)) {
1784                 u16 sts = le16_to_cpu(rej_data->sts);
1785
1786                 dev_err(queue->ctrl->ctrl.device,
1787                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1788                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1789         } else {
1790                 dev_err(queue->ctrl->ctrl.device,
1791                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1792         }
1793
1794         return -ECONNRESET;
1795 }
1796
1797 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1798 {
1799         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1800         int ret;
1801
1802         ret = nvme_rdma_create_queue_ib(queue);
1803         if (ret)
1804                 return ret;
1805
1806         if (ctrl->opts->tos >= 0)
1807                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1808         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CM_TIMEOUT_MS);
1809         if (ret) {
1810                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1811                         queue->cm_error);
1812                 goto out_destroy_queue;
1813         }
1814
1815         return 0;
1816
1817 out_destroy_queue:
1818         nvme_rdma_destroy_queue_ib(queue);
1819         return ret;
1820 }
1821
1822 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1823 {
1824         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1825         struct rdma_conn_param param = { };
1826         struct nvme_rdma_cm_req priv = { };
1827         int ret;
1828
1829         param.qp_num = queue->qp->qp_num;
1830         param.flow_control = 1;
1831
1832         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1833         /* maximum retry count */
1834         param.retry_count = 7;
1835         param.rnr_retry_count = 7;
1836         param.private_data = &priv;
1837         param.private_data_len = sizeof(priv);
1838
1839         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1840         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1841         /*
1842          * set the admin queue depth to the minimum size
1843          * specified by the Fabrics standard.
1844          */
1845         if (priv.qid == 0) {
1846                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1847                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1848         } else {
1849                 /*
1850                  * current interpretation of the fabrics spec
1851                  * is at minimum you make hrqsize sqsize+1, or a
1852                  * 1's based representation of sqsize.
1853                  */
1854                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1855                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1856         }
1857
1858         ret = rdma_connect_locked(queue->cm_id, &param);
1859         if (ret) {
1860                 dev_err(ctrl->ctrl.device,
1861                         "rdma_connect_locked failed (%d).\n", ret);
1862                 return ret;
1863         }
1864
1865         return 0;
1866 }
1867
1868 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1869                 struct rdma_cm_event *ev)
1870 {
1871         struct nvme_rdma_queue *queue = cm_id->context;
1872         int cm_error = 0;
1873
1874         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1875                 rdma_event_msg(ev->event), ev->event,
1876                 ev->status, cm_id);
1877
1878         switch (ev->event) {
1879         case RDMA_CM_EVENT_ADDR_RESOLVED:
1880                 cm_error = nvme_rdma_addr_resolved(queue);
1881                 break;
1882         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1883                 cm_error = nvme_rdma_route_resolved(queue);
1884                 break;
1885         case RDMA_CM_EVENT_ESTABLISHED:
1886                 queue->cm_error = nvme_rdma_conn_established(queue);
1887                 /* complete cm_done regardless of success/failure */
1888                 complete(&queue->cm_done);
1889                 return 0;
1890         case RDMA_CM_EVENT_REJECTED:
1891                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1892                 break;
1893         case RDMA_CM_EVENT_ROUTE_ERROR:
1894         case RDMA_CM_EVENT_CONNECT_ERROR:
1895         case RDMA_CM_EVENT_UNREACHABLE:
1896         case RDMA_CM_EVENT_ADDR_ERROR:
1897                 dev_dbg(queue->ctrl->ctrl.device,
1898                         "CM error event %d\n", ev->event);
1899                 cm_error = -ECONNRESET;
1900                 break;
1901         case RDMA_CM_EVENT_DISCONNECTED:
1902         case RDMA_CM_EVENT_ADDR_CHANGE:
1903         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1904                 dev_dbg(queue->ctrl->ctrl.device,
1905                         "disconnect received - connection closed\n");
1906                 nvme_rdma_error_recovery(queue->ctrl);
1907                 break;
1908         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1909                 /* device removal is handled via the ib_client API */
1910                 break;
1911         default:
1912                 dev_err(queue->ctrl->ctrl.device,
1913                         "Unexpected RDMA CM event (%d)\n", ev->event);
1914                 nvme_rdma_error_recovery(queue->ctrl);
1915                 break;
1916         }
1917
1918         if (cm_error) {
1919                 queue->cm_error = cm_error;
1920                 complete(&queue->cm_done);
1921         }
1922
1923         return 0;
1924 }
1925
1926 static void nvme_rdma_complete_timed_out(struct request *rq)
1927 {
1928         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1929         struct nvme_rdma_queue *queue = req->queue;
1930
1931         nvme_rdma_stop_queue(queue);
1932         nvmf_complete_timed_out_request(rq);
1933 }
1934
1935 static enum blk_eh_timer_return nvme_rdma_timeout(struct request *rq)
1936 {
1937         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1938         struct nvme_rdma_queue *queue = req->queue;
1939         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1940
1941         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1942                  rq->tag, nvme_rdma_queue_idx(queue));
1943
1944         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1945                 /*
1946                  * If we are resetting, connecting or deleting we should
1947                  * complete immediately because we may block controller
1948                  * teardown or setup sequence
1949                  * - ctrl disable/shutdown fabrics requests
1950                  * - connect requests
1951                  * - initialization admin requests
1952                  * - I/O requests that entered after unquiescing and
1953                  *   the controller stopped responding
1954                  *
1955                  * All other requests should be cancelled by the error
1956                  * recovery work, so it's fine that we fail it here.
1957                  */
1958                 nvme_rdma_complete_timed_out(rq);
1959                 return BLK_EH_DONE;
1960         }
1961
1962         /*
1963          * LIVE state should trigger the normal error recovery which will
1964          * handle completing this request.
1965          */
1966         nvme_rdma_error_recovery(ctrl);
1967         return BLK_EH_RESET_TIMER;
1968 }
1969
1970 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1971                 const struct blk_mq_queue_data *bd)
1972 {
1973         struct nvme_ns *ns = hctx->queue->queuedata;
1974         struct nvme_rdma_queue *queue = hctx->driver_data;
1975         struct request *rq = bd->rq;
1976         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1977         struct nvme_rdma_qe *sqe = &req->sqe;
1978         struct nvme_command *c = nvme_req(rq)->cmd;
1979         struct ib_device *dev;
1980         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1981         blk_status_t ret;
1982         int err;
1983
1984         WARN_ON_ONCE(rq->tag < 0);
1985
1986         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1987                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
1988
1989         dev = queue->device->dev;
1990
1991         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
1992                                          sizeof(struct nvme_command),
1993                                          DMA_TO_DEVICE);
1994         err = ib_dma_mapping_error(dev, req->sqe.dma);
1995         if (unlikely(err))
1996                 return BLK_STS_RESOURCE;
1997
1998         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1999                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2000
2001         ret = nvme_setup_cmd(ns, rq);
2002         if (ret)
2003                 goto unmap_qe;
2004
2005         nvme_start_request(rq);
2006
2007         if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2008             queue->pi_support &&
2009             (c->common.opcode == nvme_cmd_write ||
2010              c->common.opcode == nvme_cmd_read) &&
2011             nvme_ns_has_pi(ns))
2012                 req->use_sig_mr = true;
2013         else
2014                 req->use_sig_mr = false;
2015
2016         err = nvme_rdma_map_data(queue, rq, c);
2017         if (unlikely(err < 0)) {
2018                 dev_err(queue->ctrl->ctrl.device,
2019                              "Failed to map data (%d)\n", err);
2020                 goto err;
2021         }
2022
2023         sqe->cqe.done = nvme_rdma_send_done;
2024
2025         ib_dma_sync_single_for_device(dev, sqe->dma,
2026                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2027
2028         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2029                         req->mr ? &req->reg_wr.wr : NULL);
2030         if (unlikely(err))
2031                 goto err_unmap;
2032
2033         return BLK_STS_OK;
2034
2035 err_unmap:
2036         nvme_rdma_unmap_data(queue, rq);
2037 err:
2038         if (err == -EIO)
2039                 ret = nvme_host_path_error(rq);
2040         else if (err == -ENOMEM || err == -EAGAIN)
2041                 ret = BLK_STS_RESOURCE;
2042         else
2043                 ret = BLK_STS_IOERR;
2044         nvme_cleanup_cmd(rq);
2045 unmap_qe:
2046         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2047                             DMA_TO_DEVICE);
2048         return ret;
2049 }
2050
2051 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2052 {
2053         struct nvme_rdma_queue *queue = hctx->driver_data;
2054
2055         return ib_process_cq_direct(queue->ib_cq, -1);
2056 }
2057
2058 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2059 {
2060         struct request *rq = blk_mq_rq_from_pdu(req);
2061         struct ib_mr_status mr_status;
2062         int ret;
2063
2064         ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2065         if (ret) {
2066                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2067                 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2068                 return;
2069         }
2070
2071         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2072                 switch (mr_status.sig_err.err_type) {
2073                 case IB_SIG_BAD_GUARD:
2074                         nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2075                         break;
2076                 case IB_SIG_BAD_REFTAG:
2077                         nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2078                         break;
2079                 case IB_SIG_BAD_APPTAG:
2080                         nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2081                         break;
2082                 }
2083                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2084                        mr_status.sig_err.err_type, mr_status.sig_err.expected,
2085                        mr_status.sig_err.actual);
2086         }
2087 }
2088
2089 static void nvme_rdma_complete_rq(struct request *rq)
2090 {
2091         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2092         struct nvme_rdma_queue *queue = req->queue;
2093         struct ib_device *ibdev = queue->device->dev;
2094
2095         if (req->use_sig_mr)
2096                 nvme_rdma_check_pi_status(req);
2097
2098         nvme_rdma_unmap_data(queue, rq);
2099         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2100                             DMA_TO_DEVICE);
2101         nvme_complete_rq(rq);
2102 }
2103
2104 static void nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2105 {
2106         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(set->driver_data);
2107
2108         nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2109 }
2110
2111 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2112         .queue_rq       = nvme_rdma_queue_rq,
2113         .complete       = nvme_rdma_complete_rq,
2114         .init_request   = nvme_rdma_init_request,
2115         .exit_request   = nvme_rdma_exit_request,
2116         .init_hctx      = nvme_rdma_init_hctx,
2117         .timeout        = nvme_rdma_timeout,
2118         .map_queues     = nvme_rdma_map_queues,
2119         .poll           = nvme_rdma_poll,
2120 };
2121
2122 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2123         .queue_rq       = nvme_rdma_queue_rq,
2124         .complete       = nvme_rdma_complete_rq,
2125         .init_request   = nvme_rdma_init_request,
2126         .exit_request   = nvme_rdma_exit_request,
2127         .init_hctx      = nvme_rdma_init_admin_hctx,
2128         .timeout        = nvme_rdma_timeout,
2129 };
2130
2131 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2132 {
2133         nvme_rdma_teardown_io_queues(ctrl, shutdown);
2134         nvme_quiesce_admin_queue(&ctrl->ctrl);
2135         nvme_disable_ctrl(&ctrl->ctrl, shutdown);
2136         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2137 }
2138
2139 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2140 {
2141         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2142 }
2143
2144 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2145 {
2146         struct nvme_rdma_ctrl *ctrl =
2147                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2148
2149         nvme_stop_ctrl(&ctrl->ctrl);
2150         nvme_rdma_shutdown_ctrl(ctrl, false);
2151
2152         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2153                 /* state change failure should never happen */
2154                 WARN_ON_ONCE(1);
2155                 return;
2156         }
2157
2158         if (nvme_rdma_setup_ctrl(ctrl, false))
2159                 goto out_fail;
2160
2161         return;
2162
2163 out_fail:
2164         ++ctrl->ctrl.nr_reconnects;
2165         nvme_rdma_reconnect_or_remove(ctrl);
2166 }
2167
2168 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2169         .name                   = "rdma",
2170         .module                 = THIS_MODULE,
2171         .flags                  = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2172         .reg_read32             = nvmf_reg_read32,
2173         .reg_read64             = nvmf_reg_read64,
2174         .reg_write32            = nvmf_reg_write32,
2175         .free_ctrl              = nvme_rdma_free_ctrl,
2176         .submit_async_event     = nvme_rdma_submit_async_event,
2177         .delete_ctrl            = nvme_rdma_delete_ctrl,
2178         .get_address            = nvmf_get_address,
2179         .stop_ctrl              = nvme_rdma_stop_ctrl,
2180 };
2181
2182 /*
2183  * Fails a connection request if it matches an existing controller
2184  * (association) with the same tuple:
2185  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2186  *
2187  * if local address is not specified in the request, it will match an
2188  * existing controller with all the other parameters the same and no
2189  * local port address specified as well.
2190  *
2191  * The ports don't need to be compared as they are intrinsically
2192  * already matched by the port pointers supplied.
2193  */
2194 static bool
2195 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2196 {
2197         struct nvme_rdma_ctrl *ctrl;
2198         bool found = false;
2199
2200         mutex_lock(&nvme_rdma_ctrl_mutex);
2201         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2202                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2203                 if (found)
2204                         break;
2205         }
2206         mutex_unlock(&nvme_rdma_ctrl_mutex);
2207
2208         return found;
2209 }
2210
2211 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2212                 struct nvmf_ctrl_options *opts)
2213 {
2214         struct nvme_rdma_ctrl *ctrl;
2215         int ret;
2216         bool changed;
2217
2218         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2219         if (!ctrl)
2220                 return ERR_PTR(-ENOMEM);
2221         ctrl->ctrl.opts = opts;
2222         INIT_LIST_HEAD(&ctrl->list);
2223
2224         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2225                 opts->trsvcid =
2226                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2227                 if (!opts->trsvcid) {
2228                         ret = -ENOMEM;
2229                         goto out_free_ctrl;
2230                 }
2231                 opts->mask |= NVMF_OPT_TRSVCID;
2232         }
2233
2234         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2235                         opts->traddr, opts->trsvcid, &ctrl->addr);
2236         if (ret) {
2237                 pr_err("malformed address passed: %s:%s\n",
2238                         opts->traddr, opts->trsvcid);
2239                 goto out_free_ctrl;
2240         }
2241
2242         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2243                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2244                         opts->host_traddr, NULL, &ctrl->src_addr);
2245                 if (ret) {
2246                         pr_err("malformed src address passed: %s\n",
2247                                opts->host_traddr);
2248                         goto out_free_ctrl;
2249                 }
2250         }
2251
2252         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2253                 ret = -EALREADY;
2254                 goto out_free_ctrl;
2255         }
2256
2257         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2258                         nvme_rdma_reconnect_ctrl_work);
2259         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2260         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2261
2262         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2263                                 opts->nr_poll_queues + 1;
2264         ctrl->ctrl.sqsize = opts->queue_size - 1;
2265         ctrl->ctrl.kato = opts->kato;
2266
2267         ret = -ENOMEM;
2268         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2269                                 GFP_KERNEL);
2270         if (!ctrl->queues)
2271                 goto out_free_ctrl;
2272
2273         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2274                                 0 /* no quirks, we're perfect! */);
2275         if (ret)
2276                 goto out_kfree_queues;
2277
2278         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2279         WARN_ON_ONCE(!changed);
2280
2281         ret = nvme_rdma_setup_ctrl(ctrl, true);
2282         if (ret)
2283                 goto out_uninit_ctrl;
2284
2285         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2286                 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2287
2288         mutex_lock(&nvme_rdma_ctrl_mutex);
2289         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2290         mutex_unlock(&nvme_rdma_ctrl_mutex);
2291
2292         return &ctrl->ctrl;
2293
2294 out_uninit_ctrl:
2295         nvme_uninit_ctrl(&ctrl->ctrl);
2296         nvme_put_ctrl(&ctrl->ctrl);
2297         if (ret > 0)
2298                 ret = -EIO;
2299         return ERR_PTR(ret);
2300 out_kfree_queues:
2301         kfree(ctrl->queues);
2302 out_free_ctrl:
2303         kfree(ctrl);
2304         return ERR_PTR(ret);
2305 }
2306
2307 static struct nvmf_transport_ops nvme_rdma_transport = {
2308         .name           = "rdma",
2309         .module         = THIS_MODULE,
2310         .required_opts  = NVMF_OPT_TRADDR,
2311         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2312                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2313                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2314                           NVMF_OPT_TOS,
2315         .create_ctrl    = nvme_rdma_create_ctrl,
2316 };
2317
2318 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2319 {
2320         struct nvme_rdma_ctrl *ctrl;
2321         struct nvme_rdma_device *ndev;
2322         bool found = false;
2323
2324         mutex_lock(&device_list_mutex);
2325         list_for_each_entry(ndev, &device_list, entry) {
2326                 if (ndev->dev == ib_device) {
2327                         found = true;
2328                         break;
2329                 }
2330         }
2331         mutex_unlock(&device_list_mutex);
2332
2333         if (!found)
2334                 return;
2335
2336         /* Delete all controllers using this device */
2337         mutex_lock(&nvme_rdma_ctrl_mutex);
2338         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2339                 if (ctrl->device->dev != ib_device)
2340                         continue;
2341                 nvme_delete_ctrl(&ctrl->ctrl);
2342         }
2343         mutex_unlock(&nvme_rdma_ctrl_mutex);
2344
2345         flush_workqueue(nvme_delete_wq);
2346 }
2347
2348 static struct ib_client nvme_rdma_ib_client = {
2349         .name   = "nvme_rdma",
2350         .remove = nvme_rdma_remove_one
2351 };
2352
2353 static int __init nvme_rdma_init_module(void)
2354 {
2355         int ret;
2356
2357         ret = ib_register_client(&nvme_rdma_ib_client);
2358         if (ret)
2359                 return ret;
2360
2361         ret = nvmf_register_transport(&nvme_rdma_transport);
2362         if (ret)
2363                 goto err_unreg_client;
2364
2365         return 0;
2366
2367 err_unreg_client:
2368         ib_unregister_client(&nvme_rdma_ib_client);
2369         return ret;
2370 }
2371
2372 static void __exit nvme_rdma_cleanup_module(void)
2373 {
2374         struct nvme_rdma_ctrl *ctrl;
2375
2376         nvmf_unregister_transport(&nvme_rdma_transport);
2377         ib_unregister_client(&nvme_rdma_ib_client);
2378
2379         mutex_lock(&nvme_rdma_ctrl_mutex);
2380         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2381                 nvme_delete_ctrl(&ctrl->ctrl);
2382         mutex_unlock(&nvme_rdma_ctrl_mutex);
2383         flush_workqueue(nvme_delete_wq);
2384 }
2385
2386 module_init(nvme_rdma_init_module);
2387 module_exit(nvme_rdma_cleanup_module);
2388
2389 MODULE_LICENSE("GPL v2");