drivers/nvme/host/rdma.c

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