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