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         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
 650 static 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
 660 static 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
 668 static 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
 676 static 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
 697 static 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
 710 out_stop_queues:
 711         for (i--; i >= first; i--)
 712                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 713         return ret;
 714 }
 715
 716 static 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
 747 out_free_queues:
 748         for (i--; i >= 1; i--)
 749                 nvme_rdma_free_queue(&ctrl->queues[i]);
 750
 751         return ret;
 752 }
 753
 754 static 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
 769 static 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
 780 static 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
 844 out_quiesce_queue:
 845         nvme_quiesce_admin_queue(&ctrl->ctrl);
 846         blk_sync_queue(ctrl->ctrl.admin_q);
 847 out_stop_queue:
 848         nvme_rdma_stop_queue(&ctrl->queues[0]);
 849         nvme_cancel_admin_tagset(&ctrl->ctrl);
 850 out_remove_admin_tag_set:
 851         if (new)
 852                 nvme_remove_admin_tag_set(&ctrl->ctrl);
 853 out_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         }
 859 out_free_queue:
 860         nvme_rdma_free_queue(&ctrl->queues[0]);
 861         return error;
 862 }
 863
 864 static 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
 917 out_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);
 921 out_cleanup_tagset:
 922         nvme_cancel_tagset(&ctrl->ctrl);
 923         if (new)
 924                 nvme_remove_io_tag_set(&ctrl->ctrl);
 925 out_free_io_queues:
 926         nvme_rdma_free_io_queues(ctrl);
 927         return ret;
 928 }
 929
 930 static 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
 944 static 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
 960 static 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
 968 static 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);
 980 free_ctrl:
 981         kfree(ctrl->queues);
 982         kfree(ctrl);
 983 }
 984
 985 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl,
 986                                           int status)
 987 {
 988         enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
 989
 990         /* If we are resetting/deleting then do nothing */
 991         if (state != NVME_CTRL_CONNECTING) {
 992                 WARN_ON_ONCE(state == NVME_CTRL_NEW || state == NVME_CTRL_LIVE);
 993                 return;
 994         }
 995
 996         if (nvmf_should_reconnect(&ctrl->ctrl, status)) {
 997                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 998                         ctrl->ctrl.opts->reconnect_delay);
 999                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1000                                 ctrl->ctrl.opts->reconnect_delay * HZ);
1001         } else {
1002                 nvme_delete_ctrl(&ctrl->ctrl);
1003         }
1004 }
1005
1006 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1007 {
1008         int ret;
1009         bool changed;
1010         u16 max_queue_size;
1011
1012         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1013         if (ret)
1014                 return ret;
1015
1016         if (ctrl->ctrl.icdoff) {
1017                 ret = -EOPNOTSUPP;
1018                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1019                 goto destroy_admin;
1020         }
1021
1022         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1023                 ret = -EOPNOTSUPP;
1024                 dev_err(ctrl->ctrl.device,
1025                         "Mandatory keyed sgls are not supported!\n");
1026                 goto destroy_admin;
1027         }
1028
1029         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1030                 dev_warn(ctrl->ctrl.device,
1031                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1032                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1033         }
1034
1035         if (ctrl->ctrl.max_integrity_segments)
1036                 max_queue_size = NVME_RDMA_MAX_METADATA_QUEUE_SIZE;
1037         else
1038                 max_queue_size = NVME_RDMA_MAX_QUEUE_SIZE;
1039
1040         if (ctrl->ctrl.sqsize + 1 > max_queue_size) {
1041                 dev_warn(ctrl->ctrl.device,
1042                          "ctrl sqsize %u > max queue size %u, clamping down\n",
1043                          ctrl->ctrl.sqsize + 1, max_queue_size);
1044                 ctrl->ctrl.sqsize = max_queue_size - 1;
1045         }
1046
1047         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1048                 dev_warn(ctrl->ctrl.device,
1049                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1050                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1051                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1052         }
1053
1054         if (ctrl->ctrl.sgls & (1 << 20))
1055                 ctrl->use_inline_data = true;
1056
1057         if (ctrl->ctrl.queue_count > 1) {
1058                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1059                 if (ret)
1060                         goto destroy_admin;
1061         }
1062
1063         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1064         if (!changed) {
1065                 /*
1066                  * state change failure is ok if we started ctrl delete,
1067                  * unless we're during creation of a new controller to
1068                  * avoid races with teardown flow.
1069                  */
1070                 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
1071
1072                 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
1073                              state != NVME_CTRL_DELETING_NOIO);
1074                 WARN_ON_ONCE(new);
1075                 ret = -EINVAL;
1076                 goto destroy_io;
1077         }
1078
1079         nvme_start_ctrl(&ctrl->ctrl);
1080         return 0;
1081
1082 destroy_io:
1083         if (ctrl->ctrl.queue_count > 1) {
1084                 nvme_quiesce_io_queues(&ctrl->ctrl);
1085                 nvme_sync_io_queues(&ctrl->ctrl);
1086                 nvme_rdma_stop_io_queues(ctrl);
1087                 nvme_cancel_tagset(&ctrl->ctrl);
1088                 if (new)
1089                         nvme_remove_io_tag_set(&ctrl->ctrl);
1090                 nvme_rdma_free_io_queues(ctrl);
1091         }
1092 destroy_admin:
1093         nvme_stop_keep_alive(&ctrl->ctrl);
1094         nvme_quiesce_admin_queue(&ctrl->ctrl);
1095         blk_sync_queue(ctrl->ctrl.admin_q);
1096         nvme_rdma_stop_queue(&ctrl->queues[0]);
1097         nvme_cancel_admin_tagset(&ctrl->ctrl);
1098         if (new)
1099                 nvme_remove_admin_tag_set(&ctrl->ctrl);
1100         nvme_rdma_destroy_admin_queue(ctrl);
1101         return ret;
1102 }
1103
1104 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1105 {
1106         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1107                         struct nvme_rdma_ctrl, reconnect_work);
1108         int ret;
1109
1110         ++ctrl->ctrl.nr_reconnects;
1111
1112         ret = nvme_rdma_setup_ctrl(ctrl, false);
1113         if (ret)
1114                 goto requeue;
1115
1116         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1117                         ctrl->ctrl.nr_reconnects);
1118
1119         ctrl->ctrl.nr_reconnects = 0;
1120
1121         return;
1122
1123 requeue:
1124         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d/%d\n",
1125                  ctrl->ctrl.nr_reconnects, ctrl->ctrl.opts->max_reconnects);
1126         nvme_rdma_reconnect_or_remove(ctrl, ret);
1127 }
1128
1129 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1130 {
1131         struct nvme_rdma_ctrl *ctrl = container_of(work,
1132                         struct nvme_rdma_ctrl, err_work);
1133
1134         nvme_stop_keep_alive(&ctrl->ctrl);
1135         flush_work(&ctrl->ctrl.async_event_work);
1136         nvme_rdma_teardown_io_queues(ctrl, false);
1137         nvme_unquiesce_io_queues(&ctrl->ctrl);
1138         nvme_rdma_teardown_admin_queue(ctrl, false);
1139         nvme_unquiesce_admin_queue(&ctrl->ctrl);
1140         nvme_auth_stop(&ctrl->ctrl);
1141
1142         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1143                 /* state change failure is ok if we started ctrl delete */
1144                 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
1145
1146                 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
1147                              state != NVME_CTRL_DELETING_NOIO);
1148                 return;
1149         }
1150
1151         nvme_rdma_reconnect_or_remove(ctrl, 0);
1152 }
1153
1154 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1155 {
1156         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1157                 return;
1158
1159         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1160         queue_work(nvme_reset_wq, &ctrl->err_work);
1161 }
1162
1163 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1164 {
1165         struct request *rq = blk_mq_rq_from_pdu(req);
1166
1167         if (!refcount_dec_and_test(&req->ref))
1168                 return;
1169         if (!nvme_try_complete_req(rq, req->status, req->result))
1170                 nvme_rdma_complete_rq(rq);
1171 }
1172
1173 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1174                 const char *op)
1175 {
1176         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1177         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1178
1179         if (nvme_ctrl_state(&ctrl->ctrl) == NVME_CTRL_LIVE)
1180                 dev_info(ctrl->ctrl.device,
1181                              "%s for CQE 0x%p failed with status %s (%d)\n",
1182                              op, wc->wr_cqe,
1183                              ib_wc_status_msg(wc->status), wc->status);
1184         nvme_rdma_error_recovery(ctrl);
1185 }
1186
1187 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1188 {
1189         if (unlikely(wc->status != IB_WC_SUCCESS))
1190                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1191 }
1192
1193 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1194 {
1195         struct nvme_rdma_request *req =
1196                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1197
1198         if (unlikely(wc->status != IB_WC_SUCCESS))
1199                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1200         else
1201                 nvme_rdma_end_request(req);
1202 }
1203
1204 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1205                 struct nvme_rdma_request *req)
1206 {
1207         struct ib_send_wr wr = {
1208                 .opcode             = IB_WR_LOCAL_INV,
1209                 .next               = NULL,
1210                 .num_sge            = 0,
1211                 .send_flags         = IB_SEND_SIGNALED,
1212                 .ex.invalidate_rkey = req->mr->rkey,
1213         };
1214
1215         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1216         wr.wr_cqe = &req->reg_cqe;
1217
1218         return ib_post_send(queue->qp, &wr, NULL);
1219 }
1220
1221 static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq)
1222 {
1223         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1224
1225         if (blk_integrity_rq(rq)) {
1226                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1227                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1228                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1229                                       NVME_INLINE_METADATA_SG_CNT);
1230         }
1231
1232         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1233                         rq_dma_dir(rq));
1234         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1235 }
1236
1237 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1238                 struct request *rq)
1239 {
1240         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1241         struct nvme_rdma_device *dev = queue->device;
1242         struct ib_device *ibdev = dev->dev;
1243         struct list_head *pool = &queue->qp->rdma_mrs;
1244
1245         if (!blk_rq_nr_phys_segments(rq))
1246                 return;
1247
1248         if (req->use_sig_mr)
1249                 pool = &queue->qp->sig_mrs;
1250
1251         if (req->mr) {
1252                 ib_mr_pool_put(queue->qp, pool, req->mr);
1253                 req->mr = NULL;
1254         }
1255
1256         nvme_rdma_dma_unmap_req(ibdev, rq);
1257 }
1258
1259 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1260 {
1261         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1262
1263         sg->addr = 0;
1264         put_unaligned_le24(0, sg->length);
1265         put_unaligned_le32(0, sg->key);
1266         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1267         return 0;
1268 }
1269
1270 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1271                 struct nvme_rdma_request *req, struct nvme_command *c,
1272                 int count)
1273 {
1274         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1275         struct ib_sge *sge = &req->sge[1];
1276         struct scatterlist *sgl;
1277         u32 len = 0;
1278         int i;
1279
1280         for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1281                 sge->addr = sg_dma_address(sgl);
1282                 sge->length = sg_dma_len(sgl);
1283                 sge->lkey = queue->device->pd->local_dma_lkey;
1284                 len += sge->length;
1285                 sge++;
1286         }
1287
1288         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1289         sg->length = cpu_to_le32(len);
1290         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1291
1292         req->num_sge += count;
1293         return 0;
1294 }
1295
1296 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1297                 struct nvme_rdma_request *req, struct nvme_command *c)
1298 {
1299         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1300
1301         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1302         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1303         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1304         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1305         return 0;
1306 }
1307
1308 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1309                 struct nvme_rdma_request *req, struct nvme_command *c,
1310                 int count)
1311 {
1312         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1313         int nr;
1314
1315         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1316         if (WARN_ON_ONCE(!req->mr))
1317                 return -EAGAIN;
1318
1319         /*
1320          * Align the MR to a 4K page size to match the ctrl page size and
1321          * the block virtual boundary.
1322          */
1323         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1324                           SZ_4K);
1325         if (unlikely(nr < count)) {
1326                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1327                 req->mr = NULL;
1328                 if (nr < 0)
1329                         return nr;
1330                 return -EINVAL;
1331         }
1332
1333         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1334
1335         req->reg_cqe.done = nvme_rdma_memreg_done;
1336         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1337         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1338         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1339         req->reg_wr.wr.num_sge = 0;
1340         req->reg_wr.mr = req->mr;
1341         req->reg_wr.key = req->mr->rkey;
1342         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1343                              IB_ACCESS_REMOTE_READ |
1344                              IB_ACCESS_REMOTE_WRITE;
1345
1346         sg->addr = cpu_to_le64(req->mr->iova);
1347         put_unaligned_le24(req->mr->length, sg->length);
1348         put_unaligned_le32(req->mr->rkey, sg->key);
1349         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1350                         NVME_SGL_FMT_INVALIDATE;
1351
1352         return 0;
1353 }
1354
1355 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1356                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1357                 u16 control, u8 pi_type)
1358 {
1359         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1360         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1361         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1362         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1363         if (control & NVME_RW_PRINFO_PRCHK_REF)
1364                 domain->sig.dif.ref_remap = true;
1365
1366         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1367         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1368         domain->sig.dif.app_escape = true;
1369         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1370                 domain->sig.dif.ref_escape = true;
1371 }
1372
1373 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1374                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1375                 u8 pi_type)
1376 {
1377         u16 control = le16_to_cpu(cmd->rw.control);
1378
1379         memset(sig_attrs, 0, sizeof(*sig_attrs));
1380         if (control & NVME_RW_PRINFO_PRACT) {
1381                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1382                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1383                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1384                                          pi_type);
1385                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1386                 control &= ~NVME_RW_PRINFO_PRACT;
1387                 cmd->rw.control = cpu_to_le16(control);
1388         } else {
1389                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1390                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1391                                          pi_type);
1392                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1393                                          pi_type);
1394         }
1395 }
1396
1397 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1398 {
1399         *mask = 0;
1400         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1401                 *mask |= IB_SIG_CHECK_REFTAG;
1402         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1403                 *mask |= IB_SIG_CHECK_GUARD;
1404 }
1405
1406 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1407 {
1408         if (unlikely(wc->status != IB_WC_SUCCESS))
1409                 nvme_rdma_wr_error(cq, wc, "SIG");
1410 }
1411
1412 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1413                 struct nvme_rdma_request *req, struct nvme_command *c,
1414                 int count, int pi_count)
1415 {
1416         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1417         struct ib_reg_wr *wr = &req->reg_wr;
1418         struct request *rq = blk_mq_rq_from_pdu(req);
1419         struct nvme_ns *ns = rq->q->queuedata;
1420         struct bio *bio = rq->bio;
1421         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1422         struct blk_integrity *bi = blk_get_integrity(bio->bi_bdev->bd_disk);
1423         u32 xfer_len;
1424         int nr;
1425
1426         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1427         if (WARN_ON_ONCE(!req->mr))
1428                 return -EAGAIN;
1429
1430         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1431                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1432                              SZ_4K);
1433         if (unlikely(nr))
1434                 goto mr_put;
1435
1436         nvme_rdma_set_sig_attrs(bi, c, req->mr->sig_attrs, ns->head->pi_type);
1437         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1438
1439         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1440
1441         req->reg_cqe.done = nvme_rdma_sig_done;
1442         memset(wr, 0, sizeof(*wr));
1443         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1444         wr->wr.wr_cqe = &req->reg_cqe;
1445         wr->wr.num_sge = 0;
1446         wr->wr.send_flags = 0;
1447         wr->mr = req->mr;
1448         wr->key = req->mr->rkey;
1449         wr->access = IB_ACCESS_LOCAL_WRITE |
1450                      IB_ACCESS_REMOTE_READ |
1451                      IB_ACCESS_REMOTE_WRITE;
1452
1453         sg->addr = cpu_to_le64(req->mr->iova);
1454         xfer_len = req->mr->length;
1455         /* Check if PI is added by the HW */
1456         if (!pi_count)
1457                 xfer_len += (xfer_len >> bi->interval_exp) * ns->head->pi_size;
1458         put_unaligned_le24(xfer_len, sg->length);
1459         put_unaligned_le32(req->mr->rkey, sg->key);
1460         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1461
1462         return 0;
1463
1464 mr_put:
1465         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1466         req->mr = NULL;
1467         if (nr < 0)
1468                 return nr;
1469         return -EINVAL;
1470 }
1471
1472 static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq,
1473                 int *count, int *pi_count)
1474 {
1475         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1476         int ret;
1477
1478         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1479         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1480                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1481                         NVME_INLINE_SG_CNT);
1482         if (ret)
1483                 return -ENOMEM;
1484
1485         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1486                                             req->data_sgl.sg_table.sgl);
1487
1488         *count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1489                                req->data_sgl.nents, rq_dma_dir(rq));
1490         if (unlikely(*count <= 0)) {
1491                 ret = -EIO;
1492                 goto out_free_table;
1493         }
1494
1495         if (blk_integrity_rq(rq)) {
1496                 req->metadata_sgl->sg_table.sgl =
1497                         (struct scatterlist *)(req->metadata_sgl + 1);
1498                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1499                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1500                                 req->metadata_sgl->sg_table.sgl,
1501                                 NVME_INLINE_METADATA_SG_CNT);
1502                 if (unlikely(ret)) {
1503                         ret = -ENOMEM;
1504                         goto out_unmap_sg;
1505                 }
1506
1507                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1508                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1509                 *pi_count = ib_dma_map_sg(ibdev,
1510                                           req->metadata_sgl->sg_table.sgl,
1511                                           req->metadata_sgl->nents,
1512                                           rq_dma_dir(rq));
1513                 if (unlikely(*pi_count <= 0)) {
1514                         ret = -EIO;
1515                         goto out_free_pi_table;
1516                 }
1517         }
1518
1519         return 0;
1520
1521 out_free_pi_table:
1522         sg_free_table_chained(&req->metadata_sgl->sg_table,
1523                               NVME_INLINE_METADATA_SG_CNT);
1524 out_unmap_sg:
1525         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1526                         rq_dma_dir(rq));
1527 out_free_table:
1528         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1529         return ret;
1530 }
1531
1532 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1533                 struct request *rq, struct nvme_command *c)
1534 {
1535         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1536         struct nvme_rdma_device *dev = queue->device;
1537         struct ib_device *ibdev = dev->dev;
1538         int pi_count = 0;
1539         int count, ret;
1540
1541         req->num_sge = 1;
1542         refcount_set(&req->ref, 2); /* send and recv completions */
1543
1544         c->common.flags |= NVME_CMD_SGL_METABUF;
1545
1546         if (!blk_rq_nr_phys_segments(rq))
1547                 return nvme_rdma_set_sg_null(c);
1548
1549         ret = nvme_rdma_dma_map_req(ibdev, rq, &count, &pi_count);
1550         if (unlikely(ret))
1551                 return ret;
1552
1553         if (req->use_sig_mr) {
1554                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1555                 goto out;
1556         }
1557
1558         if (count <= dev->num_inline_segments) {
1559                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1560                     queue->ctrl->use_inline_data &&
1561                     blk_rq_payload_bytes(rq) <=
1562                                 nvme_rdma_inline_data_size(queue)) {
1563                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1564                         goto out;
1565                 }
1566
1567                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1568                         ret = nvme_rdma_map_sg_single(queue, req, c);
1569                         goto out;
1570                 }
1571         }
1572
1573         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1574 out:
1575         if (unlikely(ret))
1576                 goto out_dma_unmap_req;
1577
1578         return 0;
1579
1580 out_dma_unmap_req:
1581         nvme_rdma_dma_unmap_req(ibdev, rq);
1582         return ret;
1583 }
1584
1585 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1586 {
1587         struct nvme_rdma_qe *qe =
1588                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1589         struct nvme_rdma_request *req =
1590                 container_of(qe, struct nvme_rdma_request, sqe);
1591
1592         if (unlikely(wc->status != IB_WC_SUCCESS))
1593                 nvme_rdma_wr_error(cq, wc, "SEND");
1594         else
1595                 nvme_rdma_end_request(req);
1596 }
1597
1598 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1599                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1600                 struct ib_send_wr *first)
1601 {
1602         struct ib_send_wr wr;
1603         int ret;
1604
1605         sge->addr   = qe->dma;
1606         sge->length = sizeof(struct nvme_command);
1607         sge->lkey   = queue->device->pd->local_dma_lkey;
1608
1609         wr.next       = NULL;
1610         wr.wr_cqe     = &qe->cqe;
1611         wr.sg_list    = sge;
1612         wr.num_sge    = num_sge;
1613         wr.opcode     = IB_WR_SEND;
1614         wr.send_flags = IB_SEND_SIGNALED;
1615
1616         if (first)
1617                 first->next = &wr;
1618         else
1619                 first = &wr;
1620
1621         ret = ib_post_send(queue->qp, first, NULL);
1622         if (unlikely(ret)) {
1623                 dev_err(queue->ctrl->ctrl.device,
1624                              "%s failed with error code %d\n", __func__, ret);
1625         }
1626         return ret;
1627 }
1628
1629 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1630                 struct nvme_rdma_qe *qe)
1631 {
1632         struct ib_recv_wr wr;
1633         struct ib_sge list;
1634         int ret;
1635
1636         list.addr   = qe->dma;
1637         list.length = sizeof(struct nvme_completion);
1638         list.lkey   = queue->device->pd->local_dma_lkey;
1639
1640         qe->cqe.done = nvme_rdma_recv_done;
1641
1642         wr.next     = NULL;
1643         wr.wr_cqe   = &qe->cqe;
1644         wr.sg_list  = &list;
1645         wr.num_sge  = 1;
1646
1647         ret = ib_post_recv(queue->qp, &wr, NULL);
1648         if (unlikely(ret)) {
1649                 dev_err(queue->ctrl->ctrl.device,
1650                         "%s failed with error code %d\n", __func__, ret);
1651         }
1652         return ret;
1653 }
1654
1655 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1656 {
1657         u32 queue_idx = nvme_rdma_queue_idx(queue);
1658
1659         if (queue_idx == 0)
1660                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1661         return queue->ctrl->tag_set.tags[queue_idx - 1];
1662 }
1663
1664 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1665 {
1666         if (unlikely(wc->status != IB_WC_SUCCESS))
1667                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1668 }
1669
1670 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1671 {
1672         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1673         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1674         struct ib_device *dev = queue->device->dev;
1675         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1676         struct nvme_command *cmd = sqe->data;
1677         struct ib_sge sge;
1678         int ret;
1679
1680         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1681
1682         memset(cmd, 0, sizeof(*cmd));
1683         cmd->common.opcode = nvme_admin_async_event;
1684         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1685         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1686         nvme_rdma_set_sg_null(cmd);
1687
1688         sqe->cqe.done = nvme_rdma_async_done;
1689
1690         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1691                         DMA_TO_DEVICE);
1692
1693         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1694         WARN_ON_ONCE(ret);
1695 }
1696
1697 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1698                 struct nvme_completion *cqe, struct ib_wc *wc)
1699 {
1700         struct request *rq;
1701         struct nvme_rdma_request *req;
1702
1703         rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1704         if (!rq) {
1705                 dev_err(queue->ctrl->ctrl.device,
1706                         "got bad command_id %#x on QP %#x\n",
1707                         cqe->command_id, queue->qp->qp_num);
1708                 nvme_rdma_error_recovery(queue->ctrl);
1709                 return;
1710         }
1711         req = blk_mq_rq_to_pdu(rq);
1712
1713         req->status = cqe->status;
1714         req->result = cqe->result;
1715
1716         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1717                 if (unlikely(!req->mr ||
1718                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1719                         dev_err(queue->ctrl->ctrl.device,
1720                                 "Bogus remote invalidation for rkey %#x\n",
1721                                 req->mr ? req->mr->rkey : 0);
1722                         nvme_rdma_error_recovery(queue->ctrl);
1723                 }
1724         } else if (req->mr) {
1725                 int ret;
1726
1727                 ret = nvme_rdma_inv_rkey(queue, req);
1728                 if (unlikely(ret < 0)) {
1729                         dev_err(queue->ctrl->ctrl.device,
1730                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1731                                 req->mr->rkey, ret);
1732                         nvme_rdma_error_recovery(queue->ctrl);
1733                 }
1734                 /* the local invalidation completion will end the request */
1735                 return;
1736         }
1737
1738         nvme_rdma_end_request(req);
1739 }
1740
1741 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1742 {
1743         struct nvme_rdma_qe *qe =
1744                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1745         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1746         struct ib_device *ibdev = queue->device->dev;
1747         struct nvme_completion *cqe = qe->data;
1748         const size_t len = sizeof(struct nvme_completion);
1749
1750         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1751                 nvme_rdma_wr_error(cq, wc, "RECV");
1752                 return;
1753         }
1754
1755         /* sanity checking for received data length */
1756         if (unlikely(wc->byte_len < len)) {
1757                 dev_err(queue->ctrl->ctrl.device,
1758                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1759                 nvme_rdma_error_recovery(queue->ctrl);
1760                 return;
1761         }
1762
1763         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1764         /*
1765          * AEN requests are special as they don't time out and can
1766          * survive any kind of queue freeze and often don't respond to
1767          * aborts.  We don't even bother to allocate a struct request
1768          * for them but rather special case them here.
1769          */
1770         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1771                                      cqe->command_id)))
1772                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1773                                 &cqe->result);
1774         else
1775                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1776         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1777
1778         nvme_rdma_post_recv(queue, qe);
1779 }
1780
1781 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1782 {
1783         int ret, i;
1784
1785         for (i = 0; i < queue->queue_size; i++) {
1786                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1787                 if (ret)
1788                         return ret;
1789         }
1790
1791         return 0;
1792 }
1793
1794 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1795                 struct rdma_cm_event *ev)
1796 {
1797         struct rdma_cm_id *cm_id = queue->cm_id;
1798         int status = ev->status;
1799         const char *rej_msg;
1800         const struct nvme_rdma_cm_rej *rej_data;
1801         u8 rej_data_len;
1802
1803         rej_msg = rdma_reject_msg(cm_id, status);
1804         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1805
1806         if (rej_data && rej_data_len >= sizeof(u16)) {
1807                 u16 sts = le16_to_cpu(rej_data->sts);
1808
1809                 dev_err(queue->ctrl->ctrl.device,
1810                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1811                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1812         } else {
1813                 dev_err(queue->ctrl->ctrl.device,
1814                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1815         }
1816
1817         return -ECONNRESET;
1818 }
1819
1820 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1821 {
1822         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1823         int ret;
1824
1825         ret = nvme_rdma_create_queue_ib(queue);
1826         if (ret)
1827                 return ret;
1828
1829         if (ctrl->opts->tos >= 0)
1830                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1831         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CM_TIMEOUT_MS);
1832         if (ret) {
1833                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1834                         queue->cm_error);
1835                 goto out_destroy_queue;
1836         }
1837
1838         return 0;
1839
1840 out_destroy_queue:
1841         nvme_rdma_destroy_queue_ib(queue);
1842         return ret;
1843 }
1844
1845 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1846 {
1847         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1848         struct rdma_conn_param param = { };
1849         struct nvme_rdma_cm_req priv = { };
1850         int ret;
1851
1852         param.qp_num = queue->qp->qp_num;
1853         param.flow_control = 1;
1854
1855         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1856         /* maximum retry count */
1857         param.retry_count = 7;
1858         param.rnr_retry_count = 7;
1859         param.private_data = &priv;
1860         param.private_data_len = sizeof(priv);
1861
1862         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1863         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1864         /*
1865          * set the admin queue depth to the minimum size
1866          * specified by the Fabrics standard.
1867          */
1868         if (priv.qid == 0) {
1869                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1870                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1871         } else {
1872                 /*
1873                  * current interpretation of the fabrics spec
1874                  * is at minimum you make hrqsize sqsize+1, or a
1875                  * 1's based representation of sqsize.
1876                  */
1877                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1878                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1879         }
1880
1881         ret = rdma_connect_locked(queue->cm_id, &param);
1882         if (ret) {
1883                 dev_err(ctrl->ctrl.device,
1884                         "rdma_connect_locked failed (%d).\n", ret);
1885                 return ret;
1886         }
1887
1888         return 0;
1889 }
1890
1891 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1892                 struct rdma_cm_event *ev)
1893 {
1894         struct nvme_rdma_queue *queue = cm_id->context;
1895         int cm_error = 0;
1896
1897         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1898                 rdma_event_msg(ev->event), ev->event,
1899                 ev->status, cm_id);
1900
1901         switch (ev->event) {
1902         case RDMA_CM_EVENT_ADDR_RESOLVED:
1903                 cm_error = nvme_rdma_addr_resolved(queue);
1904                 break;
1905         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1906                 cm_error = nvme_rdma_route_resolved(queue);
1907                 break;
1908         case RDMA_CM_EVENT_ESTABLISHED:
1909                 queue->cm_error = nvme_rdma_conn_established(queue);
1910                 /* complete cm_done regardless of success/failure */
1911                 complete(&queue->cm_done);
1912                 return 0;
1913         case RDMA_CM_EVENT_REJECTED:
1914                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1915                 break;
1916         case RDMA_CM_EVENT_ROUTE_ERROR:
1917         case RDMA_CM_EVENT_CONNECT_ERROR:
1918         case RDMA_CM_EVENT_UNREACHABLE:
1919         case RDMA_CM_EVENT_ADDR_ERROR:
1920                 dev_dbg(queue->ctrl->ctrl.device,
1921                         "CM error event %d\n", ev->event);
1922                 cm_error = -ECONNRESET;
1923                 break;
1924         case RDMA_CM_EVENT_DISCONNECTED:
1925         case RDMA_CM_EVENT_ADDR_CHANGE:
1926         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1927                 dev_dbg(queue->ctrl->ctrl.device,
1928                         "disconnect received - connection closed\n");
1929                 nvme_rdma_error_recovery(queue->ctrl);
1930                 break;
1931         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1932                 /* device removal is handled via the ib_client API */
1933                 break;
1934         default:
1935                 dev_err(queue->ctrl->ctrl.device,
1936                         "Unexpected RDMA CM event (%d)\n", ev->event);
1937                 nvme_rdma_error_recovery(queue->ctrl);
1938                 break;
1939         }
1940
1941         if (cm_error) {
1942                 queue->cm_error = cm_error;
1943                 complete(&queue->cm_done);
1944         }
1945
1946         return 0;
1947 }
1948
1949 static void nvme_rdma_complete_timed_out(struct request *rq)
1950 {
1951         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1952         struct nvme_rdma_queue *queue = req->queue;
1953
1954         nvme_rdma_stop_queue(queue);
1955         nvmf_complete_timed_out_request(rq);
1956 }
1957
1958 static enum blk_eh_timer_return nvme_rdma_timeout(struct request *rq)
1959 {
1960         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1961         struct nvme_rdma_queue *queue = req->queue;
1962         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1963         struct nvme_command *cmd = req->req.cmd;
1964         int qid = nvme_rdma_queue_idx(queue);
1965
1966         dev_warn(ctrl->ctrl.device,
1967                  "I/O tag %d (%04x) opcode %#x (%s) QID %d timeout\n",
1968                  rq->tag, nvme_cid(rq), cmd->common.opcode,
1969                  nvme_fabrics_opcode_str(qid, cmd), qid);
1970
1971         if (nvme_ctrl_state(&ctrl->ctrl) != NVME_CTRL_LIVE) {
1972                 /*
1973                  * If we are resetting, connecting or deleting we should
1974                  * complete immediately because we may block controller
1975                  * teardown or setup sequence
1976                  * - ctrl disable/shutdown fabrics requests
1977                  * - connect requests
1978                  * - initialization admin requests
1979                  * - I/O requests that entered after unquiescing and
1980                  *   the controller stopped responding
1981                  *
1982                  * All other requests should be cancelled by the error
1983                  * recovery work, so it's fine that we fail it here.
1984                  */
1985                 nvme_rdma_complete_timed_out(rq);
1986                 return BLK_EH_DONE;
1987         }
1988
1989         /*
1990          * LIVE state should trigger the normal error recovery which will
1991          * handle completing this request.
1992          */
1993         nvme_rdma_error_recovery(ctrl);
1994         return BLK_EH_RESET_TIMER;
1995 }
1996
1997 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1998                 const struct blk_mq_queue_data *bd)
1999 {
2000         struct nvme_ns *ns = hctx->queue->queuedata;
2001         struct nvme_rdma_queue *queue = hctx->driver_data;
2002         struct request *rq = bd->rq;
2003         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2004         struct nvme_rdma_qe *sqe = &req->sqe;
2005         struct nvme_command *c = nvme_req(rq)->cmd;
2006         struct ib_device *dev;
2007         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2008         blk_status_t ret;
2009         int err;
2010
2011         WARN_ON_ONCE(rq->tag < 0);
2012
2013         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2014                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2015
2016         dev = queue->device->dev;
2017
2018         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2019                                          sizeof(struct nvme_command),
2020                                          DMA_TO_DEVICE);
2021         err = ib_dma_mapping_error(dev, req->sqe.dma);
2022         if (unlikely(err))
2023                 return BLK_STS_RESOURCE;
2024
2025         ib_dma_sync_single_for_cpu(dev, sqe->dma,
2026                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2027
2028         ret = nvme_setup_cmd(ns, rq);
2029         if (ret)
2030                 goto unmap_qe;
2031
2032         nvme_start_request(rq);
2033
2034         if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2035             queue->pi_support &&
2036             (c->common.opcode == nvme_cmd_write ||
2037              c->common.opcode == nvme_cmd_read) &&
2038             nvme_ns_has_pi(ns->head))
2039                 req->use_sig_mr = true;
2040         else
2041                 req->use_sig_mr = false;
2042
2043         err = nvme_rdma_map_data(queue, rq, c);
2044         if (unlikely(err < 0)) {
2045                 dev_err(queue->ctrl->ctrl.device,
2046                              "Failed to map data (%d)\n", err);
2047                 goto err;
2048         }
2049
2050         sqe->cqe.done = nvme_rdma_send_done;
2051
2052         ib_dma_sync_single_for_device(dev, sqe->dma,
2053                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2054
2055         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2056                         req->mr ? &req->reg_wr.wr : NULL);
2057         if (unlikely(err))
2058                 goto err_unmap;
2059
2060         return BLK_STS_OK;
2061
2062 err_unmap:
2063         nvme_rdma_unmap_data(queue, rq);
2064 err:
2065         if (err == -EIO)
2066                 ret = nvme_host_path_error(rq);
2067         else if (err == -ENOMEM || err == -EAGAIN)
2068                 ret = BLK_STS_RESOURCE;
2069         else
2070                 ret = BLK_STS_IOERR;
2071         nvme_cleanup_cmd(rq);
2072 unmap_qe:
2073         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2074                             DMA_TO_DEVICE);
2075         return ret;
2076 }
2077
2078 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2079 {
2080         struct nvme_rdma_queue *queue = hctx->driver_data;
2081
2082         return ib_process_cq_direct(queue->ib_cq, -1);
2083 }
2084
2085 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2086 {
2087         struct request *rq = blk_mq_rq_from_pdu(req);
2088         struct ib_mr_status mr_status;
2089         int ret;
2090
2091         ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2092         if (ret) {
2093                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2094                 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2095                 return;
2096         }
2097
2098         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2099                 switch (mr_status.sig_err.err_type) {
2100                 case IB_SIG_BAD_GUARD:
2101                         nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2102                         break;
2103                 case IB_SIG_BAD_REFTAG:
2104                         nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2105                         break;
2106                 case IB_SIG_BAD_APPTAG:
2107                         nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2108                         break;
2109                 }
2110                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2111                        mr_status.sig_err.err_type, mr_status.sig_err.expected,
2112                        mr_status.sig_err.actual);
2113         }
2114 }
2115
2116 static void nvme_rdma_complete_rq(struct request *rq)
2117 {
2118         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2119         struct nvme_rdma_queue *queue = req->queue;
2120         struct ib_device *ibdev = queue->device->dev;
2121
2122         if (req->use_sig_mr)
2123                 nvme_rdma_check_pi_status(req);
2124
2125         nvme_rdma_unmap_data(queue, rq);
2126         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2127                             DMA_TO_DEVICE);
2128         nvme_complete_rq(rq);
2129 }
2130
2131 static void nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2132 {
2133         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(set->driver_data);
2134
2135         nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2136 }
2137
2138 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2139         .queue_rq       = nvme_rdma_queue_rq,
2140         .complete       = nvme_rdma_complete_rq,
2141         .init_request   = nvme_rdma_init_request,
2142         .exit_request   = nvme_rdma_exit_request,
2143         .init_hctx      = nvme_rdma_init_hctx,
2144         .timeout        = nvme_rdma_timeout,
2145         .map_queues     = nvme_rdma_map_queues,
2146         .poll           = nvme_rdma_poll,
2147 };
2148
2149 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2150         .queue_rq       = nvme_rdma_queue_rq,
2151         .complete       = nvme_rdma_complete_rq,
2152         .init_request   = nvme_rdma_init_request,
2153         .exit_request   = nvme_rdma_exit_request,
2154         .init_hctx      = nvme_rdma_init_admin_hctx,
2155         .timeout        = nvme_rdma_timeout,
2156 };
2157
2158 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2159 {
2160         nvme_rdma_teardown_io_queues(ctrl, shutdown);
2161         nvme_quiesce_admin_queue(&ctrl->ctrl);
2162         nvme_disable_ctrl(&ctrl->ctrl, shutdown);
2163         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2164 }
2165
2166 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2167 {
2168         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2169 }
2170
2171 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2172 {
2173         struct nvme_rdma_ctrl *ctrl =
2174                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2175         int ret;
2176
2177         nvme_stop_ctrl(&ctrl->ctrl);
2178         nvme_rdma_shutdown_ctrl(ctrl, false);
2179
2180         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2181                 /* state change failure should never happen */
2182                 WARN_ON_ONCE(1);
2183                 return;
2184         }
2185
2186         ret = nvme_rdma_setup_ctrl(ctrl, false);
2187         if (ret)
2188                 goto out_fail;
2189
2190         return;
2191
2192 out_fail:
2193         ++ctrl->ctrl.nr_reconnects;
2194         nvme_rdma_reconnect_or_remove(ctrl, ret);
2195 }
2196
2197 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2198         .name                   = "rdma",
2199         .module                 = THIS_MODULE,
2200         .flags                  = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2201         .reg_read32             = nvmf_reg_read32,
2202         .reg_read64             = nvmf_reg_read64,
2203         .reg_write32            = nvmf_reg_write32,
2204         .free_ctrl              = nvme_rdma_free_ctrl,
2205         .submit_async_event     = nvme_rdma_submit_async_event,
2206         .delete_ctrl            = nvme_rdma_delete_ctrl,
2207         .get_address            = nvmf_get_address,
2208         .stop_ctrl              = nvme_rdma_stop_ctrl,
2209 };
2210
2211 /*
2212  * Fails a connection request if it matches an existing controller
2213  * (association) with the same tuple:
2214  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2215  *
2216  * if local address is not specified in the request, it will match an
2217  * existing controller with all the other parameters the same and no
2218  * local port address specified as well.
2219  *
2220  * The ports don't need to be compared as they are intrinsically
2221  * already matched by the port pointers supplied.
2222  */
2223 static bool
2224 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2225 {
2226         struct nvme_rdma_ctrl *ctrl;
2227         bool found = false;
2228
2229         mutex_lock(&nvme_rdma_ctrl_mutex);
2230         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2231                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2232                 if (found)
2233                         break;
2234         }
2235         mutex_unlock(&nvme_rdma_ctrl_mutex);
2236
2237         return found;
2238 }
2239
2240 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2241                 struct nvmf_ctrl_options *opts)
2242 {
2243         struct nvme_rdma_ctrl *ctrl;
2244         int ret;
2245         bool changed;
2246
2247         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2248         if (!ctrl)
2249                 return ERR_PTR(-ENOMEM);
2250         ctrl->ctrl.opts = opts;
2251         INIT_LIST_HEAD(&ctrl->list);
2252
2253         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2254                 opts->trsvcid =
2255                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2256                 if (!opts->trsvcid) {
2257                         ret = -ENOMEM;
2258                         goto out_free_ctrl;
2259                 }
2260                 opts->mask |= NVMF_OPT_TRSVCID;
2261         }
2262
2263         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2264                         opts->traddr, opts->trsvcid, &ctrl->addr);
2265         if (ret) {
2266                 pr_err("malformed address passed: %s:%s\n",
2267                         opts->traddr, opts->trsvcid);
2268                 goto out_free_ctrl;
2269         }
2270
2271         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2272                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2273                         opts->host_traddr, NULL, &ctrl->src_addr);
2274                 if (ret) {
2275                         pr_err("malformed src address passed: %s\n",
2276                                opts->host_traddr);
2277                         goto out_free_ctrl;
2278                 }
2279         }
2280
2281         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2282                 ret = -EALREADY;
2283                 goto out_free_ctrl;
2284         }
2285
2286         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2287                         nvme_rdma_reconnect_ctrl_work);
2288         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2289         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2290
2291         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2292                                 opts->nr_poll_queues + 1;
2293         ctrl->ctrl.sqsize = opts->queue_size - 1;
2294         ctrl->ctrl.kato = opts->kato;
2295
2296         ret = -ENOMEM;
2297         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2298                                 GFP_KERNEL);
2299         if (!ctrl->queues)
2300                 goto out_free_ctrl;
2301
2302         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2303                                 0 /* no quirks, we're perfect! */);
2304         if (ret)
2305                 goto out_kfree_queues;
2306
2307         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2308         WARN_ON_ONCE(!changed);
2309
2310         ret = nvme_rdma_setup_ctrl(ctrl, true);
2311         if (ret)
2312                 goto out_uninit_ctrl;
2313
2314         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs, hostnqn: %s\n",
2315                 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr, opts->host->nqn);
2316
2317         mutex_lock(&nvme_rdma_ctrl_mutex);
2318         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2319         mutex_unlock(&nvme_rdma_ctrl_mutex);
2320
2321         return &ctrl->ctrl;
2322
2323 out_uninit_ctrl:
2324         nvme_uninit_ctrl(&ctrl->ctrl);
2325         nvme_put_ctrl(&ctrl->ctrl);
2326         if (ret > 0)
2327                 ret = -EIO;
2328         return ERR_PTR(ret);
2329 out_kfree_queues:
2330         kfree(ctrl->queues);
2331 out_free_ctrl:
2332         kfree(ctrl);
2333         return ERR_PTR(ret);
2334 }
2335
2336 static struct nvmf_transport_ops nvme_rdma_transport = {
2337         .name           = "rdma",
2338         .module         = THIS_MODULE,
2339         .required_opts  = NVMF_OPT_TRADDR,
2340         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2341                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2342                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2343                           NVMF_OPT_TOS,
2344         .create_ctrl    = nvme_rdma_create_ctrl,
2345 };
2346
2347 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2348 {
2349         struct nvme_rdma_ctrl *ctrl;
2350         struct nvme_rdma_device *ndev;
2351         bool found = false;
2352
2353         mutex_lock(&device_list_mutex);
2354         list_for_each_entry(ndev, &device_list, entry) {
2355                 if (ndev->dev == ib_device) {
2356                         found = true;
2357                         break;
2358                 }
2359         }
2360         mutex_unlock(&device_list_mutex);
2361
2362         if (!found)
2363                 return;
2364
2365         /* Delete all controllers using this device */
2366         mutex_lock(&nvme_rdma_ctrl_mutex);
2367         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2368                 if (ctrl->device->dev != ib_device)
2369                         continue;
2370                 nvme_delete_ctrl(&ctrl->ctrl);
2371         }
2372         mutex_unlock(&nvme_rdma_ctrl_mutex);
2373
2374         flush_workqueue(nvme_delete_wq);
2375 }
2376
2377 static struct ib_client nvme_rdma_ib_client = {
2378         .name   = "nvme_rdma",
2379         .remove = nvme_rdma_remove_one
2380 };
2381
2382 static int __init nvme_rdma_init_module(void)
2383 {
2384         int ret;
2385
2386         ret = ib_register_client(&nvme_rdma_ib_client);
2387         if (ret)
2388                 return ret;
2389
2390         ret = nvmf_register_transport(&nvme_rdma_transport);
2391         if (ret)
2392                 goto err_unreg_client;
2393
2394         return 0;
2395
2396 err_unreg_client:
2397         ib_unregister_client(&nvme_rdma_ib_client);
2398         return ret;
2399 }
2400
2401 static void __exit nvme_rdma_cleanup_module(void)
2402 {
2403         struct nvme_rdma_ctrl *ctrl;
2404
2405         nvmf_unregister_transport(&nvme_rdma_transport);
2406         ib_unregister_client(&nvme_rdma_ib_client);
2407
2408         mutex_lock(&nvme_rdma_ctrl_mutex);
2409         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2410                 nvme_delete_ctrl(&ctrl->ctrl);
2411         mutex_unlock(&nvme_rdma_ctrl_mutex);
2412         flush_workqueue(nvme_delete_wq);
2413 }
2414
2415 module_init(nvme_rdma_init_module);
2416 module_exit(nvme_rdma_cleanup_module);
2417
2418 MODULE_DESCRIPTION("NVMe host RDMA transport driver");
2419 MODULE_LICENSE("GPL v2");