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-mq-rdma.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_CONNECT_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 };
 101
 102 struct nvme_rdma_ctrl {
 103         /* read only in the hot path */
 104         struct nvme_rdma_queue  *queues;
 105
 106         /* other member variables */
 107         struct blk_mq_tag_set   tag_set;
 108         struct work_struct      err_work;
 109
 110         struct nvme_rdma_qe     async_event_sqe;
 111
 112         struct delayed_work     reconnect_work;
 113
 114         struct list_head        list;
 115
 116         struct blk_mq_tag_set   admin_tag_set;
 117         struct nvme_rdma_device *device;
 118
 119         u32                     max_fr_pages;
 120
 121         struct sockaddr_storage addr;
 122         struct sockaddr_storage src_addr;
 123
 124         struct nvme_ctrl        ctrl;
 125         struct mutex            teardown_lock;
 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_timeout(&queue->cm_done,
 251                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 252         if (ret < 0)
 253                 return ret;
 254         if (ret == 0)
 255                 return -ETIMEDOUT;
 256         WARN_ON_ONCE(queue->cm_error > 0);
 257         return queue->cm_error;
 258 }
 259
 260 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 261 {
 262         struct nvme_rdma_device *dev = queue->device;
 263         struct ib_qp_init_attr init_attr;
 264         int ret;
 265
 266         memset(&init_attr, 0, sizeof(init_attr));
 267         init_attr.event_handler = nvme_rdma_qp_event;
 268         /* +1 for drain */
 269         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 270         /* +1 for drain */
 271         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 272         init_attr.cap.max_recv_sge = 1;
 273         init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
 274         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 275         init_attr.qp_type = IB_QPT_RC;
 276         init_attr.send_cq = queue->ib_cq;
 277         init_attr.recv_cq = queue->ib_cq;
 278         if (queue->pi_support)
 279                 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
 280         init_attr.qp_context = queue;
 281
 282         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 283
 284         queue->qp = queue->cm_id->qp;
 285         return ret;
 286 }
 287
 288 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 289                 struct request *rq, unsigned int hctx_idx)
 290 {
 291         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 292
 293         kfree(req->sqe.data);
 294 }
 295
 296 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 297                 struct request *rq, unsigned int hctx_idx,
 298                 unsigned int numa_node)
 299 {
 300         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 301         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 302         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 303         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 304
 305         nvme_req(rq)->ctrl = &ctrl->ctrl;
 306         req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
 307         if (!req->sqe.data)
 308                 return -ENOMEM;
 309
 310         /* metadata nvme_rdma_sgl struct is located after command's data SGL */
 311         if (queue->pi_support)
 312                 req->metadata_sgl = (void *)nvme_req(rq) +
 313                         sizeof(struct nvme_rdma_request) +
 314                         NVME_RDMA_DATA_SGL_SIZE;
 315
 316         req->queue = queue;
 317
 318         return 0;
 319 }
 320
 321 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 322                 unsigned int hctx_idx)
 323 {
 324         struct nvme_rdma_ctrl *ctrl = data;
 325         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 326
 327         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 328
 329         hctx->driver_data = queue;
 330         return 0;
 331 }
 332
 333 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 334                 unsigned int hctx_idx)
 335 {
 336         struct nvme_rdma_ctrl *ctrl = data;
 337         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 338
 339         BUG_ON(hctx_idx != 0);
 340
 341         hctx->driver_data = queue;
 342         return 0;
 343 }
 344
 345 static void nvme_rdma_free_dev(struct kref *ref)
 346 {
 347         struct nvme_rdma_device *ndev =
 348                 container_of(ref, struct nvme_rdma_device, ref);
 349
 350         mutex_lock(&device_list_mutex);
 351         list_del(&ndev->entry);
 352         mutex_unlock(&device_list_mutex);
 353
 354         ib_dealloc_pd(ndev->pd);
 355         kfree(ndev);
 356 }
 357
 358 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 359 {
 360         kref_put(&dev->ref, nvme_rdma_free_dev);
 361 }
 362
 363 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 364 {
 365         return kref_get_unless_zero(&dev->ref);
 366 }
 367
 368 static struct nvme_rdma_device *
 369 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 370 {
 371         struct nvme_rdma_device *ndev;
 372
 373         mutex_lock(&device_list_mutex);
 374         list_for_each_entry(ndev, &device_list, entry) {
 375                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 376                     nvme_rdma_dev_get(ndev))
 377                         goto out_unlock;
 378         }
 379
 380         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 381         if (!ndev)
 382                 goto out_err;
 383
 384         ndev->dev = cm_id->device;
 385         kref_init(&ndev->ref);
 386
 387         ndev->pd = ib_alloc_pd(ndev->dev,
 388                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 389         if (IS_ERR(ndev->pd))
 390                 goto out_free_dev;
 391
 392         if (!(ndev->dev->attrs.device_cap_flags &
 393               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 394                 dev_err(&ndev->dev->dev,
 395                         "Memory registrations not supported.\n");
 396                 goto out_free_pd;
 397         }
 398
 399         ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
 400                                         ndev->dev->attrs.max_send_sge - 1);
 401         list_add(&ndev->entry, &device_list);
 402 out_unlock:
 403         mutex_unlock(&device_list_mutex);
 404         return ndev;
 405
 406 out_free_pd:
 407         ib_dealloc_pd(ndev->pd);
 408 out_free_dev:
 409         kfree(ndev);
 410 out_err:
 411         mutex_unlock(&device_list_mutex);
 412         return NULL;
 413 }
 414
 415 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
 416 {
 417         if (nvme_rdma_poll_queue(queue))
 418                 ib_free_cq(queue->ib_cq);
 419         else
 420                 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
 421 }
 422
 423 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 424 {
 425         struct nvme_rdma_device *dev;
 426         struct ib_device *ibdev;
 427
 428         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 429                 return;
 430
 431         dev = queue->device;
 432         ibdev = dev->dev;
 433
 434         if (queue->pi_support)
 435                 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
 436         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 437
 438         /*
 439          * The cm_id object might have been destroyed during RDMA connection
 440          * establishment error flow to avoid getting other cma events, thus
 441          * the destruction of the QP shouldn't use rdma_cm API.
 442          */
 443         ib_destroy_qp(queue->qp);
 444         nvme_rdma_free_cq(queue);
 445
 446         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 447                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 448
 449         nvme_rdma_dev_put(dev);
 450 }
 451
 452 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
 453 {
 454         u32 max_page_list_len;
 455
 456         if (pi_support)
 457                 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
 458         else
 459                 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
 460
 461         return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
 462 }
 463
 464 static int nvme_rdma_create_cq(struct ib_device *ibdev,
 465                 struct nvme_rdma_queue *queue)
 466 {
 467         int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
 468         enum ib_poll_context poll_ctx;
 469
 470         /*
 471          * Spread I/O queues completion vectors according their queue index.
 472          * Admin queues can always go on completion vector 0.
 473          */
 474         comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
 475
 476         /* Polling queues need direct cq polling context */
 477         if (nvme_rdma_poll_queue(queue)) {
 478                 poll_ctx = IB_POLL_DIRECT;
 479                 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
 480                                            comp_vector, poll_ctx);
 481         } else {
 482                 poll_ctx = IB_POLL_SOFTIRQ;
 483                 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
 484                                               comp_vector, poll_ctx);
 485         }
 486
 487         if (IS_ERR(queue->ib_cq)) {
 488                 ret = PTR_ERR(queue->ib_cq);
 489                 return ret;
 490         }
 491
 492         return 0;
 493 }
 494
 495 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 496 {
 497         struct ib_device *ibdev;
 498         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 499         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 500         int ret, pages_per_mr;
 501
 502         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 503         if (!queue->device) {
 504                 dev_err(queue->cm_id->device->dev.parent,
 505                         "no client data found!\n");
 506                 return -ECONNREFUSED;
 507         }
 508         ibdev = queue->device->dev;
 509
 510         /* +1 for ib_stop_cq */
 511         queue->cq_size = cq_factor * queue->queue_size + 1;
 512
 513         ret = nvme_rdma_create_cq(ibdev, queue);
 514         if (ret)
 515                 goto out_put_dev;
 516
 517         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 518         if (ret)
 519                 goto out_destroy_ib_cq;
 520
 521         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 522                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 523         if (!queue->rsp_ring) {
 524                 ret = -ENOMEM;
 525                 goto out_destroy_qp;
 526         }
 527
 528         /*
 529          * Currently we don't use SG_GAPS MR's so if the first entry is
 530          * misaligned we'll end up using two entries for a single data page,
 531          * so one additional entry is required.
 532          */
 533         pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
 534         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 535                               queue->queue_size,
 536                               IB_MR_TYPE_MEM_REG,
 537                               pages_per_mr, 0);
 538         if (ret) {
 539                 dev_err(queue->ctrl->ctrl.device,
 540                         "failed to initialize MR pool sized %d for QID %d\n",
 541                         queue->queue_size, nvme_rdma_queue_idx(queue));
 542                 goto out_destroy_ring;
 543         }
 544
 545         if (queue->pi_support) {
 546                 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
 547                                       queue->queue_size, IB_MR_TYPE_INTEGRITY,
 548                                       pages_per_mr, pages_per_mr);
 549                 if (ret) {
 550                         dev_err(queue->ctrl->ctrl.device,
 551                                 "failed to initialize PI MR pool sized %d for QID %d\n",
 552                                 queue->queue_size, nvme_rdma_queue_idx(queue));
 553                         goto out_destroy_mr_pool;
 554                 }
 555         }
 556
 557         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 558
 559         return 0;
 560
 561 out_destroy_mr_pool:
 562         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 563 out_destroy_ring:
 564         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 565                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 566 out_destroy_qp:
 567         rdma_destroy_qp(queue->cm_id);
 568 out_destroy_ib_cq:
 569         nvme_rdma_free_cq(queue);
 570 out_put_dev:
 571         nvme_rdma_dev_put(queue->device);
 572         return ret;
 573 }
 574
 575 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 576                 int idx, size_t queue_size)
 577 {
 578         struct nvme_rdma_queue *queue;
 579         struct sockaddr *src_addr = NULL;
 580         int ret;
 581
 582         queue = &ctrl->queues[idx];
 583         queue->ctrl = ctrl;
 584         if (idx && ctrl->ctrl.max_integrity_segments)
 585                 queue->pi_support = true;
 586         else
 587                 queue->pi_support = false;
 588         init_completion(&queue->cm_done);
 589
 590         if (idx > 0)
 591                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 592         else
 593                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 594
 595         queue->queue_size = queue_size;
 596
 597         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 598                         RDMA_PS_TCP, IB_QPT_RC);
 599         if (IS_ERR(queue->cm_id)) {
 600                 dev_info(ctrl->ctrl.device,
 601                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 602                 return PTR_ERR(queue->cm_id);
 603         }
 604
 605         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 606                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 607
 608         queue->cm_error = -ETIMEDOUT;
 609         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 610                         (struct sockaddr *)&ctrl->addr,
 611                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 612         if (ret) {
 613                 dev_info(ctrl->ctrl.device,
 614                         "rdma_resolve_addr failed (%d).\n", ret);
 615                 goto out_destroy_cm_id;
 616         }
 617
 618         ret = nvme_rdma_wait_for_cm(queue);
 619         if (ret) {
 620                 dev_info(ctrl->ctrl.device,
 621                         "rdma connection establishment failed (%d)\n", ret);
 622                 goto out_destroy_cm_id;
 623         }
 624
 625         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 626
 627         return 0;
 628
 629 out_destroy_cm_id:
 630         rdma_destroy_id(queue->cm_id);
 631         nvme_rdma_destroy_queue_ib(queue);
 632         return ret;
 633 }
 634
 635 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 636 {
 637         rdma_disconnect(queue->cm_id);
 638         ib_drain_qp(queue->qp);
 639 }
 640
 641 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 642 {
 643         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 644                 return;
 645         __nvme_rdma_stop_queue(queue);
 646 }
 647
 648 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 649 {
 650         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 651                 return;
 652
 653         nvme_rdma_destroy_queue_ib(queue);
 654         rdma_destroy_id(queue->cm_id);
 655 }
 656
 657 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 658 {
 659         int i;
 660
 661         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 662                 nvme_rdma_free_queue(&ctrl->queues[i]);
 663 }
 664
 665 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 666 {
 667         int i;
 668
 669         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 670                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 671 }
 672
 673 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 674 {
 675         struct nvme_rdma_queue *queue = &ctrl->queues[idx];
 676         bool poll = nvme_rdma_poll_queue(queue);
 677         int ret;
 678
 679         if (idx)
 680                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
 681         else
 682                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 683
 684         if (!ret) {
 685                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 686         } else {
 687                 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 688                         __nvme_rdma_stop_queue(queue);
 689                 dev_info(ctrl->ctrl.device,
 690                         "failed to connect queue: %d ret=%d\n", idx, ret);
 691         }
 692         return ret;
 693 }
 694
 695 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 696 {
 697         int i, ret = 0;
 698
 699         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 700                 ret = nvme_rdma_start_queue(ctrl, i);
 701                 if (ret)
 702                         goto out_stop_queues;
 703         }
 704
 705         return 0;
 706
 707 out_stop_queues:
 708         for (i--; i >= 1; i--)
 709                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 710         return ret;
 711 }
 712
 713 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 714 {
 715         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 716         struct ib_device *ibdev = ctrl->device->dev;
 717         unsigned int nr_io_queues, nr_default_queues;
 718         unsigned int nr_read_queues, nr_poll_queues;
 719         int i, ret;
 720
 721         nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
 722                                 min(opts->nr_io_queues, num_online_cpus()));
 723         nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
 724                                 min(opts->nr_write_queues, num_online_cpus()));
 725         nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
 726         nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
 727
 728         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 729         if (ret)
 730                 return ret;
 731
 732         ctrl->ctrl.queue_count = nr_io_queues + 1;
 733         if (ctrl->ctrl.queue_count < 2)
 734                 return 0;
 735
 736         dev_info(ctrl->ctrl.device,
 737                 "creating %d I/O queues.\n", nr_io_queues);
 738
 739         if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
 740                 /*
 741                  * separate read/write queues
 742                  * hand out dedicated default queues only after we have
 743                  * sufficient read queues.
 744                  */
 745                 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
 746                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
 747                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 748                         min(nr_default_queues, nr_io_queues);
 749                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 750         } else {
 751                 /*
 752                  * shared read/write queues
 753                  * either no write queues were requested, or we don't have
 754                  * sufficient queue count to have dedicated default queues.
 755                  */
 756                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 757                         min(nr_read_queues, nr_io_queues);
 758                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 759         }
 760
 761         if (opts->nr_poll_queues && nr_io_queues) {
 762                 /* map dedicated poll queues only if we have queues left */
 763                 ctrl->io_queues[HCTX_TYPE_POLL] =
 764                         min(nr_poll_queues, nr_io_queues);
 765         }
 766
 767         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 768                 ret = nvme_rdma_alloc_queue(ctrl, i,
 769                                 ctrl->ctrl.sqsize + 1);
 770                 if (ret)
 771                         goto out_free_queues;
 772         }
 773
 774         return 0;
 775
 776 out_free_queues:
 777         for (i--; i >= 1; i--)
 778                 nvme_rdma_free_queue(&ctrl->queues[i]);
 779
 780         return ret;
 781 }
 782
 783 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 784                 bool admin)
 785 {
 786         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 787         struct blk_mq_tag_set *set;
 788         int ret;
 789
 790         if (admin) {
 791                 set = &ctrl->admin_tag_set;
 792                 memset(set, 0, sizeof(*set));
 793                 set->ops = &nvme_rdma_admin_mq_ops;
 794                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 795                 set->reserved_tags = 2; /* connect + keep-alive */
 796                 set->numa_node = nctrl->numa_node;
 797                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 798                                 NVME_RDMA_DATA_SGL_SIZE;
 799                 set->driver_data = ctrl;
 800                 set->nr_hw_queues = 1;
 801                 set->timeout = ADMIN_TIMEOUT;
 802                 set->flags = BLK_MQ_F_NO_SCHED;
 803         } else {
 804                 set = &ctrl->tag_set;
 805                 memset(set, 0, sizeof(*set));
 806                 set->ops = &nvme_rdma_mq_ops;
 807                 set->queue_depth = nctrl->sqsize + 1;
 808                 set->reserved_tags = 1; /* fabric connect */
 809                 set->numa_node = nctrl->numa_node;
 810                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 811                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 812                                 NVME_RDMA_DATA_SGL_SIZE;
 813                 if (nctrl->max_integrity_segments)
 814                         set->cmd_size += sizeof(struct nvme_rdma_sgl) +
 815                                          NVME_RDMA_METADATA_SGL_SIZE;
 816                 set->driver_data = ctrl;
 817                 set->nr_hw_queues = nctrl->queue_count - 1;
 818                 set->timeout = NVME_IO_TIMEOUT;
 819                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
 820         }
 821
 822         ret = blk_mq_alloc_tag_set(set);
 823         if (ret)
 824                 return ERR_PTR(ret);
 825
 826         return set;
 827 }
 828
 829 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 830                 bool remove)
 831 {
 832         if (remove) {
 833                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 834                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 835                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 836         }
 837         if (ctrl->async_event_sqe.data) {
 838                 cancel_work_sync(&ctrl->ctrl.async_event_work);
 839                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 840                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 841                 ctrl->async_event_sqe.data = NULL;
 842         }
 843         nvme_rdma_free_queue(&ctrl->queues[0]);
 844 }
 845
 846 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 847                 bool new)
 848 {
 849         bool pi_capable = false;
 850         int error;
 851
 852         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 853         if (error)
 854                 return error;
 855
 856         ctrl->device = ctrl->queues[0].device;
 857         ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
 858
 859         /* T10-PI support */
 860         if (ctrl->device->dev->attrs.device_cap_flags &
 861             IB_DEVICE_INTEGRITY_HANDOVER)
 862                 pi_capable = true;
 863
 864         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
 865                                                         pi_capable);
 866
 867         /*
 868          * Bind the async event SQE DMA mapping to the admin queue lifetime.
 869          * It's safe, since any chage in the underlying RDMA device will issue
 870          * error recovery and queue re-creation.
 871          */
 872         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 873                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 874         if (error)
 875                 goto out_free_queue;
 876
 877         if (new) {
 878                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 879                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 880                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 881                         goto out_free_async_qe;
 882                 }
 883
 884                 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 885                 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 886                         error = PTR_ERR(ctrl->ctrl.fabrics_q);
 887                         goto out_free_tagset;
 888                 }
 889
 890                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 891                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 892                         error = PTR_ERR(ctrl->ctrl.admin_q);
 893                         goto out_cleanup_fabrics_q;
 894                 }
 895         }
 896
 897         error = nvme_rdma_start_queue(ctrl, 0);
 898         if (error)
 899                 goto out_cleanup_queue;
 900
 901         error = nvme_enable_ctrl(&ctrl->ctrl);
 902         if (error)
 903                 goto out_stop_queue;
 904
 905         ctrl->ctrl.max_segments = ctrl->max_fr_pages;
 906         ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
 907         if (pi_capable)
 908                 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
 909         else
 910                 ctrl->ctrl.max_integrity_segments = 0;
 911
 912         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 913
 914         error = nvme_init_identify(&ctrl->ctrl);
 915         if (error)
 916                 goto out_stop_queue;
 917
 918         return 0;
 919
 920 out_stop_queue:
 921         nvme_rdma_stop_queue(&ctrl->queues[0]);
 922 out_cleanup_queue:
 923         if (new)
 924                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 925 out_cleanup_fabrics_q:
 926         if (new)
 927                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 928 out_free_tagset:
 929         if (new)
 930                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 931 out_free_async_qe:
 932         if (ctrl->async_event_sqe.data) {
 933                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 934                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 935                 ctrl->async_event_sqe.data = NULL;
 936         }
 937 out_free_queue:
 938         nvme_rdma_free_queue(&ctrl->queues[0]);
 939         return error;
 940 }
 941
 942 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 943                 bool remove)
 944 {
 945         if (remove) {
 946                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 947                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 948         }
 949         nvme_rdma_free_io_queues(ctrl);
 950 }
 951
 952 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 953 {
 954         int ret;
 955
 956         ret = nvme_rdma_alloc_io_queues(ctrl);
 957         if (ret)
 958                 return ret;
 959
 960         if (new) {
 961                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 962                 if (IS_ERR(ctrl->ctrl.tagset)) {
 963                         ret = PTR_ERR(ctrl->ctrl.tagset);
 964                         goto out_free_io_queues;
 965                 }
 966
 967                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 968                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 969                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 970                         goto out_free_tag_set;
 971                 }
 972         }
 973
 974         ret = nvme_rdma_start_io_queues(ctrl);
 975         if (ret)
 976                 goto out_cleanup_connect_q;
 977
 978         if (!new) {
 979                 nvme_start_queues(&ctrl->ctrl);
 980                 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
 981                         /*
 982                          * If we timed out waiting for freeze we are likely to
 983                          * be stuck.  Fail the controller initialization just
 984                          * to be safe.
 985                          */
 986                         ret = -ENODEV;
 987                         goto out_wait_freeze_timed_out;
 988                 }
 989                 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
 990                         ctrl->ctrl.queue_count - 1);
 991                 nvme_unfreeze(&ctrl->ctrl);
 992         }
 993
 994         return 0;
 995
 996 out_wait_freeze_timed_out:
 997         nvme_stop_queues(&ctrl->ctrl);
 998         nvme_rdma_stop_io_queues(ctrl);
 999 out_cleanup_connect_q:
1000         if (new)
1001                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1002 out_free_tag_set:
1003         if (new)
1004                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1005 out_free_io_queues:
1006         nvme_rdma_free_io_queues(ctrl);
1007         return ret;
1008 }
1009
1010 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1011                 bool remove)
1012 {
1013         mutex_lock(&ctrl->teardown_lock);
1014         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1015         nvme_rdma_stop_queue(&ctrl->queues[0]);
1016         if (ctrl->ctrl.admin_tagset) {
1017                 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
1018                         nvme_cancel_request, &ctrl->ctrl);
1019                 blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
1020         }
1021         if (remove)
1022                 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1023         nvme_rdma_destroy_admin_queue(ctrl, remove);
1024         mutex_unlock(&ctrl->teardown_lock);
1025 }
1026
1027 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1028                 bool remove)
1029 {
1030         mutex_lock(&ctrl->teardown_lock);
1031         if (ctrl->ctrl.queue_count > 1) {
1032                 nvme_start_freeze(&ctrl->ctrl);
1033                 nvme_stop_queues(&ctrl->ctrl);
1034                 nvme_rdma_stop_io_queues(ctrl);
1035                 if (ctrl->ctrl.tagset) {
1036                         blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
1037                                 nvme_cancel_request, &ctrl->ctrl);
1038                         blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
1039                 }
1040                 if (remove)
1041                         nvme_start_queues(&ctrl->ctrl);
1042                 nvme_rdma_destroy_io_queues(ctrl, remove);
1043         }
1044         mutex_unlock(&ctrl->teardown_lock);
1045 }
1046
1047 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1048 {
1049         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1050
1051         if (list_empty(&ctrl->list))
1052                 goto free_ctrl;
1053
1054         mutex_lock(&nvme_rdma_ctrl_mutex);
1055         list_del(&ctrl->list);
1056         mutex_unlock(&nvme_rdma_ctrl_mutex);
1057
1058         nvmf_free_options(nctrl->opts);
1059 free_ctrl:
1060         kfree(ctrl->queues);
1061         kfree(ctrl);
1062 }
1063
1064 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1065 {
1066         /* If we are resetting/deleting then do nothing */
1067         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1068                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1069                         ctrl->ctrl.state == NVME_CTRL_LIVE);
1070                 return;
1071         }
1072
1073         if (nvmf_should_reconnect(&ctrl->ctrl)) {
1074                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1075                         ctrl->ctrl.opts->reconnect_delay);
1076                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1077                                 ctrl->ctrl.opts->reconnect_delay * HZ);
1078         } else {
1079                 nvme_delete_ctrl(&ctrl->ctrl);
1080         }
1081 }
1082
1083 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1084 {
1085         int ret = -EINVAL;
1086         bool changed;
1087
1088         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1089         if (ret)
1090                 return ret;
1091
1092         if (ctrl->ctrl.icdoff) {
1093                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1094                 goto destroy_admin;
1095         }
1096
1097         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1098                 dev_err(ctrl->ctrl.device,
1099                         "Mandatory keyed sgls are not supported!\n");
1100                 goto destroy_admin;
1101         }
1102
1103         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1104                 dev_warn(ctrl->ctrl.device,
1105                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1106                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1107         }
1108
1109         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1110                 dev_warn(ctrl->ctrl.device,
1111                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1112                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1113                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1114         }
1115
1116         if (ctrl->ctrl.sgls & (1 << 20))
1117                 ctrl->use_inline_data = true;
1118
1119         if (ctrl->ctrl.queue_count > 1) {
1120                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1121                 if (ret)
1122                         goto destroy_admin;
1123         }
1124
1125         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1126         if (!changed) {
1127                 /*
1128                  * state change failure is ok if we started ctrl delete,
1129                  * unless we're during creation of a new controller to
1130                  * avoid races with teardown flow.
1131                  */
1132                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1133                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1134                 WARN_ON_ONCE(new);
1135                 ret = -EINVAL;
1136                 goto destroy_io;
1137         }
1138
1139         nvme_start_ctrl(&ctrl->ctrl);
1140         return 0;
1141
1142 destroy_io:
1143         if (ctrl->ctrl.queue_count > 1)
1144                 nvme_rdma_destroy_io_queues(ctrl, new);
1145 destroy_admin:
1146         nvme_rdma_stop_queue(&ctrl->queues[0]);
1147         nvme_rdma_destroy_admin_queue(ctrl, new);
1148         return ret;
1149 }
1150
1151 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1152 {
1153         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1154                         struct nvme_rdma_ctrl, reconnect_work);
1155
1156         ++ctrl->ctrl.nr_reconnects;
1157
1158         if (nvme_rdma_setup_ctrl(ctrl, false))
1159                 goto requeue;
1160
1161         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1162                         ctrl->ctrl.nr_reconnects);
1163
1164         ctrl->ctrl.nr_reconnects = 0;
1165
1166         return;
1167
1168 requeue:
1169         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1170                         ctrl->ctrl.nr_reconnects);
1171         nvme_rdma_reconnect_or_remove(ctrl);
1172 }
1173
1174 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1175 {
1176         struct nvme_rdma_ctrl *ctrl = container_of(work,
1177                         struct nvme_rdma_ctrl, err_work);
1178
1179         nvme_stop_keep_alive(&ctrl->ctrl);
1180         nvme_rdma_teardown_io_queues(ctrl, false);
1181         nvme_start_queues(&ctrl->ctrl);
1182         nvme_rdma_teardown_admin_queue(ctrl, false);
1183         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1184
1185         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1186                 /* state change failure is ok if we started ctrl delete */
1187                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1188                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1189                 return;
1190         }
1191
1192         nvme_rdma_reconnect_or_remove(ctrl);
1193 }
1194
1195 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1196 {
1197         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1198                 return;
1199
1200         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1201         queue_work(nvme_reset_wq, &ctrl->err_work);
1202 }
1203
1204 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1205 {
1206         struct request *rq = blk_mq_rq_from_pdu(req);
1207
1208         if (!refcount_dec_and_test(&req->ref))
1209                 return;
1210         if (!nvme_try_complete_req(rq, req->status, req->result))
1211                 nvme_rdma_complete_rq(rq);
1212 }
1213
1214 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1215                 const char *op)
1216 {
1217         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1218         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1219
1220         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1221                 dev_info(ctrl->ctrl.device,
1222                              "%s for CQE 0x%p failed with status %s (%d)\n",
1223                              op, wc->wr_cqe,
1224                              ib_wc_status_msg(wc->status), wc->status);
1225         nvme_rdma_error_recovery(ctrl);
1226 }
1227
1228 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1229 {
1230         if (unlikely(wc->status != IB_WC_SUCCESS))
1231                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1232 }
1233
1234 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1235 {
1236         struct nvme_rdma_request *req =
1237                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1238
1239         if (unlikely(wc->status != IB_WC_SUCCESS))
1240                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1241         else
1242                 nvme_rdma_end_request(req);
1243 }
1244
1245 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1246                 struct nvme_rdma_request *req)
1247 {
1248         struct ib_send_wr wr = {
1249                 .opcode             = IB_WR_LOCAL_INV,
1250                 .next               = NULL,
1251                 .num_sge            = 0,
1252                 .send_flags         = IB_SEND_SIGNALED,
1253                 .ex.invalidate_rkey = req->mr->rkey,
1254         };
1255
1256         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1257         wr.wr_cqe = &req->reg_cqe;
1258
1259         return ib_post_send(queue->qp, &wr, NULL);
1260 }
1261
1262 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1263                 struct request *rq)
1264 {
1265         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1266         struct nvme_rdma_device *dev = queue->device;
1267         struct ib_device *ibdev = dev->dev;
1268         struct list_head *pool = &queue->qp->rdma_mrs;
1269
1270         if (!blk_rq_nr_phys_segments(rq))
1271                 return;
1272
1273         if (blk_integrity_rq(rq)) {
1274                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1275                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1276                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1277                                       NVME_INLINE_METADATA_SG_CNT);
1278         }
1279
1280         if (req->use_sig_mr)
1281                 pool = &queue->qp->sig_mrs;
1282
1283         if (req->mr) {
1284                 ib_mr_pool_put(queue->qp, pool, req->mr);
1285                 req->mr = NULL;
1286         }
1287
1288         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1289                         rq_dma_dir(rq));
1290         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1291 }
1292
1293 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1294 {
1295         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1296
1297         sg->addr = 0;
1298         put_unaligned_le24(0, sg->length);
1299         put_unaligned_le32(0, sg->key);
1300         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1301         return 0;
1302 }
1303
1304 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1305                 struct nvme_rdma_request *req, struct nvme_command *c,
1306                 int count)
1307 {
1308         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1309         struct scatterlist *sgl = req->data_sgl.sg_table.sgl;
1310         struct ib_sge *sge = &req->sge[1];
1311         u32 len = 0;
1312         int i;
1313
1314         for (i = 0; i < count; i++, sgl++, sge++) {
1315                 sge->addr = sg_dma_address(sgl);
1316                 sge->length = sg_dma_len(sgl);
1317                 sge->lkey = queue->device->pd->local_dma_lkey;
1318                 len += sge->length;
1319         }
1320
1321         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1322         sg->length = cpu_to_le32(len);
1323         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1324
1325         req->num_sge += count;
1326         return 0;
1327 }
1328
1329 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1330                 struct nvme_rdma_request *req, struct nvme_command *c)
1331 {
1332         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1333
1334         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1335         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1336         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1337         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1338         return 0;
1339 }
1340
1341 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1342                 struct nvme_rdma_request *req, struct nvme_command *c,
1343                 int count)
1344 {
1345         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1346         int nr;
1347
1348         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1349         if (WARN_ON_ONCE(!req->mr))
1350                 return -EAGAIN;
1351
1352         /*
1353          * Align the MR to a 4K page size to match the ctrl page size and
1354          * the block virtual boundary.
1355          */
1356         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1357                           SZ_4K);
1358         if (unlikely(nr < count)) {
1359                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1360                 req->mr = NULL;
1361                 if (nr < 0)
1362                         return nr;
1363                 return -EINVAL;
1364         }
1365
1366         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1367
1368         req->reg_cqe.done = nvme_rdma_memreg_done;
1369         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1370         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1371         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1372         req->reg_wr.wr.num_sge = 0;
1373         req->reg_wr.mr = req->mr;
1374         req->reg_wr.key = req->mr->rkey;
1375         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1376                              IB_ACCESS_REMOTE_READ |
1377                              IB_ACCESS_REMOTE_WRITE;
1378
1379         sg->addr = cpu_to_le64(req->mr->iova);
1380         put_unaligned_le24(req->mr->length, sg->length);
1381         put_unaligned_le32(req->mr->rkey, sg->key);
1382         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1383                         NVME_SGL_FMT_INVALIDATE;
1384
1385         return 0;
1386 }
1387
1388 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1389                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1390                 u16 control, u8 pi_type)
1391 {
1392         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1393         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1394         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1395         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1396         if (control & NVME_RW_PRINFO_PRCHK_REF)
1397                 domain->sig.dif.ref_remap = true;
1398
1399         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1400         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1401         domain->sig.dif.app_escape = true;
1402         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1403                 domain->sig.dif.ref_escape = true;
1404 }
1405
1406 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1407                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1408                 u8 pi_type)
1409 {
1410         u16 control = le16_to_cpu(cmd->rw.control);
1411
1412         memset(sig_attrs, 0, sizeof(*sig_attrs));
1413         if (control & NVME_RW_PRINFO_PRACT) {
1414                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1415                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1416                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1417                                          pi_type);
1418                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1419                 control &= ~NVME_RW_PRINFO_PRACT;
1420                 cmd->rw.control = cpu_to_le16(control);
1421         } else {
1422                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1423                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1424                                          pi_type);
1425                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1426                                          pi_type);
1427         }
1428 }
1429
1430 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1431 {
1432         *mask = 0;
1433         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1434                 *mask |= IB_SIG_CHECK_REFTAG;
1435         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1436                 *mask |= IB_SIG_CHECK_GUARD;
1437 }
1438
1439 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1440 {
1441         if (unlikely(wc->status != IB_WC_SUCCESS))
1442                 nvme_rdma_wr_error(cq, wc, "SIG");
1443 }
1444
1445 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1446                 struct nvme_rdma_request *req, struct nvme_command *c,
1447                 int count, int pi_count)
1448 {
1449         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1450         struct ib_reg_wr *wr = &req->reg_wr;
1451         struct request *rq = blk_mq_rq_from_pdu(req);
1452         struct nvme_ns *ns = rq->q->queuedata;
1453         struct bio *bio = rq->bio;
1454         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1455         int nr;
1456
1457         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1458         if (WARN_ON_ONCE(!req->mr))
1459                 return -EAGAIN;
1460
1461         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1462                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1463                              SZ_4K);
1464         if (unlikely(nr))
1465                 goto mr_put;
1466
1467         nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_disk), c,
1468                                 req->mr->sig_attrs, ns->pi_type);
1469         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1470
1471         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1472
1473         req->reg_cqe.done = nvme_rdma_sig_done;
1474         memset(wr, 0, sizeof(*wr));
1475         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1476         wr->wr.wr_cqe = &req->reg_cqe;
1477         wr->wr.num_sge = 0;
1478         wr->wr.send_flags = 0;
1479         wr->mr = req->mr;
1480         wr->key = req->mr->rkey;
1481         wr->access = IB_ACCESS_LOCAL_WRITE |
1482                      IB_ACCESS_REMOTE_READ |
1483                      IB_ACCESS_REMOTE_WRITE;
1484
1485         sg->addr = cpu_to_le64(req->mr->iova);
1486         put_unaligned_le24(req->mr->length, sg->length);
1487         put_unaligned_le32(req->mr->rkey, sg->key);
1488         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1489
1490         return 0;
1491
1492 mr_put:
1493         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1494         req->mr = NULL;
1495         if (nr < 0)
1496                 return nr;
1497         return -EINVAL;
1498 }
1499
1500 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1501                 struct request *rq, struct nvme_command *c)
1502 {
1503         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1504         struct nvme_rdma_device *dev = queue->device;
1505         struct ib_device *ibdev = dev->dev;
1506         int pi_count = 0;
1507         int count, ret;
1508
1509         req->num_sge = 1;
1510         refcount_set(&req->ref, 2); /* send and recv completions */
1511
1512         c->common.flags |= NVME_CMD_SGL_METABUF;
1513
1514         if (!blk_rq_nr_phys_segments(rq))
1515                 return nvme_rdma_set_sg_null(c);
1516
1517         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1518         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1519                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1520                         NVME_INLINE_SG_CNT);
1521         if (ret)
1522                 return -ENOMEM;
1523
1524         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1525                                             req->data_sgl.sg_table.sgl);
1526
1527         count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1528                               req->data_sgl.nents, rq_dma_dir(rq));
1529         if (unlikely(count <= 0)) {
1530                 ret = -EIO;
1531                 goto out_free_table;
1532         }
1533
1534         if (blk_integrity_rq(rq)) {
1535                 req->metadata_sgl->sg_table.sgl =
1536                         (struct scatterlist *)(req->metadata_sgl + 1);
1537                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1538                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1539                                 req->metadata_sgl->sg_table.sgl,
1540                                 NVME_INLINE_METADATA_SG_CNT);
1541                 if (unlikely(ret)) {
1542                         ret = -ENOMEM;
1543                         goto out_unmap_sg;
1544                 }
1545
1546                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1547                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1548                 pi_count = ib_dma_map_sg(ibdev,
1549                                          req->metadata_sgl->sg_table.sgl,
1550                                          req->metadata_sgl->nents,
1551                                          rq_dma_dir(rq));
1552                 if (unlikely(pi_count <= 0)) {
1553                         ret = -EIO;
1554                         goto out_free_pi_table;
1555                 }
1556         }
1557
1558         if (req->use_sig_mr) {
1559                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1560                 goto out;
1561         }
1562
1563         if (count <= dev->num_inline_segments) {
1564                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1565                     queue->ctrl->use_inline_data &&
1566                     blk_rq_payload_bytes(rq) <=
1567                                 nvme_rdma_inline_data_size(queue)) {
1568                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1569                         goto out;
1570                 }
1571
1572                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1573                         ret = nvme_rdma_map_sg_single(queue, req, c);
1574                         goto out;
1575                 }
1576         }
1577
1578         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1579 out:
1580         if (unlikely(ret))
1581                 goto out_unmap_pi_sg;
1582
1583         return 0;
1584
1585 out_unmap_pi_sg:
1586         if (blk_integrity_rq(rq))
1587                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1588                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1589 out_free_pi_table:
1590         if (blk_integrity_rq(rq))
1591                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1592                                       NVME_INLINE_METADATA_SG_CNT);
1593 out_unmap_sg:
1594         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1595                         rq_dma_dir(rq));
1596 out_free_table:
1597         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1598         return ret;
1599 }
1600
1601 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1602 {
1603         struct nvme_rdma_qe *qe =
1604                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1605         struct nvme_rdma_request *req =
1606                 container_of(qe, struct nvme_rdma_request, sqe);
1607
1608         if (unlikely(wc->status != IB_WC_SUCCESS))
1609                 nvme_rdma_wr_error(cq, wc, "SEND");
1610         else
1611                 nvme_rdma_end_request(req);
1612 }
1613
1614 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1615                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1616                 struct ib_send_wr *first)
1617 {
1618         struct ib_send_wr wr;
1619         int ret;
1620
1621         sge->addr   = qe->dma;
1622         sge->length = sizeof(struct nvme_command);
1623         sge->lkey   = queue->device->pd->local_dma_lkey;
1624
1625         wr.next       = NULL;
1626         wr.wr_cqe     = &qe->cqe;
1627         wr.sg_list    = sge;
1628         wr.num_sge    = num_sge;
1629         wr.opcode     = IB_WR_SEND;
1630         wr.send_flags = IB_SEND_SIGNALED;
1631
1632         if (first)
1633                 first->next = &wr;
1634         else
1635                 first = &wr;
1636
1637         ret = ib_post_send(queue->qp, first, NULL);
1638         if (unlikely(ret)) {
1639                 dev_err(queue->ctrl->ctrl.device,
1640                              "%s failed with error code %d\n", __func__, ret);
1641         }
1642         return ret;
1643 }
1644
1645 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1646                 struct nvme_rdma_qe *qe)
1647 {
1648         struct ib_recv_wr wr;
1649         struct ib_sge list;
1650         int ret;
1651
1652         list.addr   = qe->dma;
1653         list.length = sizeof(struct nvme_completion);
1654         list.lkey   = queue->device->pd->local_dma_lkey;
1655
1656         qe->cqe.done = nvme_rdma_recv_done;
1657
1658         wr.next     = NULL;
1659         wr.wr_cqe   = &qe->cqe;
1660         wr.sg_list  = &list;
1661         wr.num_sge  = 1;
1662
1663         ret = ib_post_recv(queue->qp, &wr, NULL);
1664         if (unlikely(ret)) {
1665                 dev_err(queue->ctrl->ctrl.device,
1666                         "%s failed with error code %d\n", __func__, ret);
1667         }
1668         return ret;
1669 }
1670
1671 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1672 {
1673         u32 queue_idx = nvme_rdma_queue_idx(queue);
1674
1675         if (queue_idx == 0)
1676                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1677         return queue->ctrl->tag_set.tags[queue_idx - 1];
1678 }
1679
1680 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1681 {
1682         if (unlikely(wc->status != IB_WC_SUCCESS))
1683                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1684 }
1685
1686 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1687 {
1688         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1689         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1690         struct ib_device *dev = queue->device->dev;
1691         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1692         struct nvme_command *cmd = sqe->data;
1693         struct ib_sge sge;
1694         int ret;
1695
1696         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1697
1698         memset(cmd, 0, sizeof(*cmd));
1699         cmd->common.opcode = nvme_admin_async_event;
1700         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1701         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1702         nvme_rdma_set_sg_null(cmd);
1703
1704         sqe->cqe.done = nvme_rdma_async_done;
1705
1706         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1707                         DMA_TO_DEVICE);
1708
1709         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1710         WARN_ON_ONCE(ret);
1711 }
1712
1713 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1714                 struct nvme_completion *cqe, struct ib_wc *wc)
1715 {
1716         struct request *rq;
1717         struct nvme_rdma_request *req;
1718
1719         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1720         if (!rq) {
1721                 dev_err(queue->ctrl->ctrl.device,
1722                         "tag 0x%x on QP %#x not found\n",
1723                         cqe->command_id, queue->qp->qp_num);
1724                 nvme_rdma_error_recovery(queue->ctrl);
1725                 return;
1726         }
1727         req = blk_mq_rq_to_pdu(rq);
1728
1729         req->status = cqe->status;
1730         req->result = cqe->result;
1731
1732         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1733                 if (unlikely(!req->mr ||
1734                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1735                         dev_err(queue->ctrl->ctrl.device,
1736                                 "Bogus remote invalidation for rkey %#x\n",
1737                                 req->mr ? req->mr->rkey : 0);
1738                         nvme_rdma_error_recovery(queue->ctrl);
1739                 }
1740         } else if (req->mr) {
1741                 int ret;
1742
1743                 ret = nvme_rdma_inv_rkey(queue, req);
1744                 if (unlikely(ret < 0)) {
1745                         dev_err(queue->ctrl->ctrl.device,
1746                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1747                                 req->mr->rkey, ret);
1748                         nvme_rdma_error_recovery(queue->ctrl);
1749                 }
1750                 /* the local invalidation completion will end the request */
1751                 return;
1752         }
1753
1754         nvme_rdma_end_request(req);
1755 }
1756
1757 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1758 {
1759         struct nvme_rdma_qe *qe =
1760                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1761         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1762         struct ib_device *ibdev = queue->device->dev;
1763         struct nvme_completion *cqe = qe->data;
1764         const size_t len = sizeof(struct nvme_completion);
1765
1766         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1767                 nvme_rdma_wr_error(cq, wc, "RECV");
1768                 return;
1769         }
1770
1771         /* sanity checking for received data length */
1772         if (unlikely(wc->byte_len < len)) {
1773                 dev_err(queue->ctrl->ctrl.device,
1774                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1775                 nvme_rdma_error_recovery(queue->ctrl);
1776                 return;
1777         }
1778
1779         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1780         /*
1781          * AEN requests are special as they don't time out and can
1782          * survive any kind of queue freeze and often don't respond to
1783          * aborts.  We don't even bother to allocate a struct request
1784          * for them but rather special case them here.
1785          */
1786         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1787                                      cqe->command_id)))
1788                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1789                                 &cqe->result);
1790         else
1791                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1792         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1793
1794         nvme_rdma_post_recv(queue, qe);
1795 }
1796
1797 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1798 {
1799         int ret, i;
1800
1801         for (i = 0; i < queue->queue_size; i++) {
1802                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1803                 if (ret)
1804                         goto out_destroy_queue_ib;
1805         }
1806
1807         return 0;
1808
1809 out_destroy_queue_ib:
1810         nvme_rdma_destroy_queue_ib(queue);
1811         return ret;
1812 }
1813
1814 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1815                 struct rdma_cm_event *ev)
1816 {
1817         struct rdma_cm_id *cm_id = queue->cm_id;
1818         int status = ev->status;
1819         const char *rej_msg;
1820         const struct nvme_rdma_cm_rej *rej_data;
1821         u8 rej_data_len;
1822
1823         rej_msg = rdma_reject_msg(cm_id, status);
1824         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1825
1826         if (rej_data && rej_data_len >= sizeof(u16)) {
1827                 u16 sts = le16_to_cpu(rej_data->sts);
1828
1829                 dev_err(queue->ctrl->ctrl.device,
1830                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1831                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1832         } else {
1833                 dev_err(queue->ctrl->ctrl.device,
1834                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1835         }
1836
1837         return -ECONNRESET;
1838 }
1839
1840 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1841 {
1842         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1843         int ret;
1844
1845         ret = nvme_rdma_create_queue_ib(queue);
1846         if (ret)
1847                 return ret;
1848
1849         if (ctrl->opts->tos >= 0)
1850                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1851         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1852         if (ret) {
1853                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1854                         queue->cm_error);
1855                 goto out_destroy_queue;
1856         }
1857
1858         return 0;
1859
1860 out_destroy_queue:
1861         nvme_rdma_destroy_queue_ib(queue);
1862         return ret;
1863 }
1864
1865 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1866 {
1867         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1868         struct rdma_conn_param param = { };
1869         struct nvme_rdma_cm_req priv = { };
1870         int ret;
1871
1872         param.qp_num = queue->qp->qp_num;
1873         param.flow_control = 1;
1874
1875         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1876         /* maximum retry count */
1877         param.retry_count = 7;
1878         param.rnr_retry_count = 7;
1879         param.private_data = &priv;
1880         param.private_data_len = sizeof(priv);
1881
1882         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1883         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1884         /*
1885          * set the admin queue depth to the minimum size
1886          * specified by the Fabrics standard.
1887          */
1888         if (priv.qid == 0) {
1889                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1890                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1891         } else {
1892                 /*
1893                  * current interpretation of the fabrics spec
1894                  * is at minimum you make hrqsize sqsize+1, or a
1895                  * 1's based representation of sqsize.
1896                  */
1897                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1898                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1899         }
1900
1901         ret = rdma_connect_locked(queue->cm_id, &param);
1902         if (ret) {
1903                 dev_err(ctrl->ctrl.device,
1904                         "rdma_connect_locked failed (%d).\n", ret);
1905                 goto out_destroy_queue_ib;
1906         }
1907
1908         return 0;
1909
1910 out_destroy_queue_ib:
1911         nvme_rdma_destroy_queue_ib(queue);
1912         return ret;
1913 }
1914
1915 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1916                 struct rdma_cm_event *ev)
1917 {
1918         struct nvme_rdma_queue *queue = cm_id->context;
1919         int cm_error = 0;
1920
1921         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1922                 rdma_event_msg(ev->event), ev->event,
1923                 ev->status, cm_id);
1924
1925         switch (ev->event) {
1926         case RDMA_CM_EVENT_ADDR_RESOLVED:
1927                 cm_error = nvme_rdma_addr_resolved(queue);
1928                 break;
1929         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1930                 cm_error = nvme_rdma_route_resolved(queue);
1931                 break;
1932         case RDMA_CM_EVENT_ESTABLISHED:
1933                 queue->cm_error = nvme_rdma_conn_established(queue);
1934                 /* complete cm_done regardless of success/failure */
1935                 complete(&queue->cm_done);
1936                 return 0;
1937         case RDMA_CM_EVENT_REJECTED:
1938                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1939                 break;
1940         case RDMA_CM_EVENT_ROUTE_ERROR:
1941         case RDMA_CM_EVENT_CONNECT_ERROR:
1942         case RDMA_CM_EVENT_UNREACHABLE:
1943                 nvme_rdma_destroy_queue_ib(queue);
1944                 fallthrough;
1945         case RDMA_CM_EVENT_ADDR_ERROR:
1946                 dev_dbg(queue->ctrl->ctrl.device,
1947                         "CM error event %d\n", ev->event);
1948                 cm_error = -ECONNRESET;
1949                 break;
1950         case RDMA_CM_EVENT_DISCONNECTED:
1951         case RDMA_CM_EVENT_ADDR_CHANGE:
1952         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1953                 dev_dbg(queue->ctrl->ctrl.device,
1954                         "disconnect received - connection closed\n");
1955                 nvme_rdma_error_recovery(queue->ctrl);
1956                 break;
1957         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1958                 /* device removal is handled via the ib_client API */
1959                 break;
1960         default:
1961                 dev_err(queue->ctrl->ctrl.device,
1962                         "Unexpected RDMA CM event (%d)\n", ev->event);
1963                 nvme_rdma_error_recovery(queue->ctrl);
1964                 break;
1965         }
1966
1967         if (cm_error) {
1968                 queue->cm_error = cm_error;
1969                 complete(&queue->cm_done);
1970         }
1971
1972         return 0;
1973 }
1974
1975 static void nvme_rdma_complete_timed_out(struct request *rq)
1976 {
1977         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1978         struct nvme_rdma_queue *queue = req->queue;
1979         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1980
1981         /* fence other contexts that may complete the command */
1982         mutex_lock(&ctrl->teardown_lock);
1983         nvme_rdma_stop_queue(queue);
1984         if (!blk_mq_request_completed(rq)) {
1985                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
1986                 blk_mq_complete_request(rq);
1987         }
1988         mutex_unlock(&ctrl->teardown_lock);
1989 }
1990
1991 static enum blk_eh_timer_return
1992 nvme_rdma_timeout(struct request *rq, bool reserved)
1993 {
1994         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1995         struct nvme_rdma_queue *queue = req->queue;
1996         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1997
1998         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1999                  rq->tag, nvme_rdma_queue_idx(queue));
2000
2001         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2002                 /*
2003                  * If we are resetting, connecting or deleting we should
2004                  * complete immediately because we may block controller
2005                  * teardown or setup sequence
2006                  * - ctrl disable/shutdown fabrics requests
2007                  * - connect requests
2008                  * - initialization admin requests
2009                  * - I/O requests that entered after unquiescing and
2010                  *   the controller stopped responding
2011                  *
2012                  * All other requests should be cancelled by the error
2013                  * recovery work, so it's fine that we fail it here.
2014                  */
2015                 nvme_rdma_complete_timed_out(rq);
2016                 return BLK_EH_DONE;
2017         }
2018
2019         /*
2020          * LIVE state should trigger the normal error recovery which will
2021          * handle completing this request.
2022          */
2023         nvme_rdma_error_recovery(ctrl);
2024         return BLK_EH_RESET_TIMER;
2025 }
2026
2027 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2028                 const struct blk_mq_queue_data *bd)
2029 {
2030         struct nvme_ns *ns = hctx->queue->queuedata;
2031         struct nvme_rdma_queue *queue = hctx->driver_data;
2032         struct request *rq = bd->rq;
2033         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2034         struct nvme_rdma_qe *sqe = &req->sqe;
2035         struct nvme_command *c = sqe->data;
2036         struct ib_device *dev;
2037         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2038         blk_status_t ret;
2039         int err;
2040
2041         WARN_ON_ONCE(rq->tag < 0);
2042
2043         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2044                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2045
2046         dev = queue->device->dev;
2047
2048         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2049                                          sizeof(struct nvme_command),
2050                                          DMA_TO_DEVICE);
2051         err = ib_dma_mapping_error(dev, req->sqe.dma);
2052         if (unlikely(err))
2053                 return BLK_STS_RESOURCE;
2054
2055         ib_dma_sync_single_for_cpu(dev, sqe->dma,
2056                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2057
2058         ret = nvme_setup_cmd(ns, rq, c);
2059         if (ret)
2060                 goto unmap_qe;
2061
2062         blk_mq_start_request(rq);
2063
2064         if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2065             queue->pi_support &&
2066             (c->common.opcode == nvme_cmd_write ||
2067              c->common.opcode == nvme_cmd_read) &&
2068             nvme_ns_has_pi(ns))
2069                 req->use_sig_mr = true;
2070         else
2071                 req->use_sig_mr = false;
2072
2073         err = nvme_rdma_map_data(queue, rq, c);
2074         if (unlikely(err < 0)) {
2075                 dev_err(queue->ctrl->ctrl.device,
2076                              "Failed to map data (%d)\n", err);
2077                 goto err;
2078         }
2079
2080         sqe->cqe.done = nvme_rdma_send_done;
2081
2082         ib_dma_sync_single_for_device(dev, sqe->dma,
2083                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2084
2085         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2086                         req->mr ? &req->reg_wr.wr : NULL);
2087         if (unlikely(err))
2088                 goto err_unmap;
2089
2090         return BLK_STS_OK;
2091
2092 err_unmap:
2093         nvme_rdma_unmap_data(queue, rq);
2094 err:
2095         if (err == -ENOMEM || err == -EAGAIN)
2096                 ret = BLK_STS_RESOURCE;
2097         else
2098                 ret = BLK_STS_IOERR;
2099         nvme_cleanup_cmd(rq);
2100 unmap_qe:
2101         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2102                             DMA_TO_DEVICE);
2103         return ret;
2104 }
2105
2106 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2107 {
2108         struct nvme_rdma_queue *queue = hctx->driver_data;
2109
2110         return ib_process_cq_direct(queue->ib_cq, -1);
2111 }
2112
2113 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2114 {
2115         struct request *rq = blk_mq_rq_from_pdu(req);
2116         struct ib_mr_status mr_status;
2117         int ret;
2118
2119         ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2120         if (ret) {
2121                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2122                 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2123                 return;
2124         }
2125
2126         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2127                 switch (mr_status.sig_err.err_type) {
2128                 case IB_SIG_BAD_GUARD:
2129                         nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2130                         break;
2131                 case IB_SIG_BAD_REFTAG:
2132                         nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2133                         break;
2134                 case IB_SIG_BAD_APPTAG:
2135                         nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2136                         break;
2137                 }
2138                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2139                        mr_status.sig_err.err_type, mr_status.sig_err.expected,
2140                        mr_status.sig_err.actual);
2141         }
2142 }
2143
2144 static void nvme_rdma_complete_rq(struct request *rq)
2145 {
2146         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2147         struct nvme_rdma_queue *queue = req->queue;
2148         struct ib_device *ibdev = queue->device->dev;
2149
2150         if (req->use_sig_mr)
2151                 nvme_rdma_check_pi_status(req);
2152
2153         nvme_rdma_unmap_data(queue, rq);
2154         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2155                             DMA_TO_DEVICE);
2156         nvme_complete_rq(rq);
2157 }
2158
2159 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2160 {
2161         struct nvme_rdma_ctrl *ctrl = set->driver_data;
2162         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2163
2164         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2165                 /* separate read/write queues */
2166                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2167                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2168                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2169                 set->map[HCTX_TYPE_READ].nr_queues =
2170                         ctrl->io_queues[HCTX_TYPE_READ];
2171                 set->map[HCTX_TYPE_READ].queue_offset =
2172                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2173         } else {
2174                 /* shared read/write queues */
2175                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2176                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2177                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2178                 set->map[HCTX_TYPE_READ].nr_queues =
2179                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2180                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2181         }
2182         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2183                         ctrl->device->dev, 0);
2184         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2185                         ctrl->device->dev, 0);
2186
2187         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2188                 /* map dedicated poll queues only if we have queues left */
2189                 set->map[HCTX_TYPE_POLL].nr_queues =
2190                                 ctrl->io_queues[HCTX_TYPE_POLL];
2191                 set->map[HCTX_TYPE_POLL].queue_offset =
2192                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2193                         ctrl->io_queues[HCTX_TYPE_READ];
2194                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2195         }
2196
2197         dev_info(ctrl->ctrl.device,
2198                 "mapped %d/%d/%d default/read/poll queues.\n",
2199                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2200                 ctrl->io_queues[HCTX_TYPE_READ],
2201                 ctrl->io_queues[HCTX_TYPE_POLL]);
2202
2203         return 0;
2204 }
2205
2206 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2207         .queue_rq       = nvme_rdma_queue_rq,
2208         .complete       = nvme_rdma_complete_rq,
2209         .init_request   = nvme_rdma_init_request,
2210         .exit_request   = nvme_rdma_exit_request,
2211         .init_hctx      = nvme_rdma_init_hctx,
2212         .timeout        = nvme_rdma_timeout,
2213         .map_queues     = nvme_rdma_map_queues,
2214         .poll           = nvme_rdma_poll,
2215 };
2216
2217 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2218         .queue_rq       = nvme_rdma_queue_rq,
2219         .complete       = nvme_rdma_complete_rq,
2220         .init_request   = nvme_rdma_init_request,
2221         .exit_request   = nvme_rdma_exit_request,
2222         .init_hctx      = nvme_rdma_init_admin_hctx,
2223         .timeout        = nvme_rdma_timeout,
2224 };
2225
2226 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2227 {
2228         cancel_work_sync(&ctrl->err_work);
2229         cancel_delayed_work_sync(&ctrl->reconnect_work);
2230
2231         nvme_rdma_teardown_io_queues(ctrl, shutdown);
2232         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2233         if (shutdown)
2234                 nvme_shutdown_ctrl(&ctrl->ctrl);
2235         else
2236                 nvme_disable_ctrl(&ctrl->ctrl);
2237         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2238 }
2239
2240 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2241 {
2242         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2243 }
2244
2245 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2246 {
2247         struct nvme_rdma_ctrl *ctrl =
2248                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2249
2250         nvme_stop_ctrl(&ctrl->ctrl);
2251         nvme_rdma_shutdown_ctrl(ctrl, false);
2252
2253         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2254                 /* state change failure should never happen */
2255                 WARN_ON_ONCE(1);
2256                 return;
2257         }
2258
2259         if (nvme_rdma_setup_ctrl(ctrl, false))
2260                 goto out_fail;
2261
2262         return;
2263
2264 out_fail:
2265         ++ctrl->ctrl.nr_reconnects;
2266         nvme_rdma_reconnect_or_remove(ctrl);
2267 }
2268
2269 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2270         .name                   = "rdma",
2271         .module                 = THIS_MODULE,
2272         .flags                  = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2273         .reg_read32             = nvmf_reg_read32,
2274         .reg_read64             = nvmf_reg_read64,
2275         .reg_write32            = nvmf_reg_write32,
2276         .free_ctrl              = nvme_rdma_free_ctrl,
2277         .submit_async_event     = nvme_rdma_submit_async_event,
2278         .delete_ctrl            = nvme_rdma_delete_ctrl,
2279         .get_address            = nvmf_get_address,
2280 };
2281
2282 /*
2283  * Fails a connection request if it matches an existing controller
2284  * (association) with the same tuple:
2285  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2286  *
2287  * if local address is not specified in the request, it will match an
2288  * existing controller with all the other parameters the same and no
2289  * local port address specified as well.
2290  *
2291  * The ports don't need to be compared as they are intrinsically
2292  * already matched by the port pointers supplied.
2293  */
2294 static bool
2295 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2296 {
2297         struct nvme_rdma_ctrl *ctrl;
2298         bool found = false;
2299
2300         mutex_lock(&nvme_rdma_ctrl_mutex);
2301         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2302                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2303                 if (found)
2304                         break;
2305         }
2306         mutex_unlock(&nvme_rdma_ctrl_mutex);
2307
2308         return found;
2309 }
2310
2311 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2312                 struct nvmf_ctrl_options *opts)
2313 {
2314         struct nvme_rdma_ctrl *ctrl;
2315         int ret;
2316         bool changed;
2317
2318         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2319         if (!ctrl)
2320                 return ERR_PTR(-ENOMEM);
2321         ctrl->ctrl.opts = opts;
2322         INIT_LIST_HEAD(&ctrl->list);
2323         mutex_init(&ctrl->teardown_lock);
2324
2325         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2326                 opts->trsvcid =
2327                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2328                 if (!opts->trsvcid) {
2329                         ret = -ENOMEM;
2330                         goto out_free_ctrl;
2331                 }
2332                 opts->mask |= NVMF_OPT_TRSVCID;
2333         }
2334
2335         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2336                         opts->traddr, opts->trsvcid, &ctrl->addr);
2337         if (ret) {
2338                 pr_err("malformed address passed: %s:%s\n",
2339                         opts->traddr, opts->trsvcid);
2340                 goto out_free_ctrl;
2341         }
2342
2343         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2344                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2345                         opts->host_traddr, NULL, &ctrl->src_addr);
2346                 if (ret) {
2347                         pr_err("malformed src address passed: %s\n",
2348                                opts->host_traddr);
2349                         goto out_free_ctrl;
2350                 }
2351         }
2352
2353         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2354                 ret = -EALREADY;
2355                 goto out_free_ctrl;
2356         }
2357
2358         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2359                         nvme_rdma_reconnect_ctrl_work);
2360         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2361         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2362
2363         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2364                                 opts->nr_poll_queues + 1;
2365         ctrl->ctrl.sqsize = opts->queue_size - 1;
2366         ctrl->ctrl.kato = opts->kato;
2367
2368         ret = -ENOMEM;
2369         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2370                                 GFP_KERNEL);
2371         if (!ctrl->queues)
2372                 goto out_free_ctrl;
2373
2374         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2375                                 0 /* no quirks, we're perfect! */);
2376         if (ret)
2377                 goto out_kfree_queues;
2378
2379         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2380         WARN_ON_ONCE(!changed);
2381
2382         ret = nvme_rdma_setup_ctrl(ctrl, true);
2383         if (ret)
2384                 goto out_uninit_ctrl;
2385
2386         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2387                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2388
2389         mutex_lock(&nvme_rdma_ctrl_mutex);
2390         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2391         mutex_unlock(&nvme_rdma_ctrl_mutex);
2392
2393         return &ctrl->ctrl;
2394
2395 out_uninit_ctrl:
2396         nvme_uninit_ctrl(&ctrl->ctrl);
2397         nvme_put_ctrl(&ctrl->ctrl);
2398         if (ret > 0)
2399                 ret = -EIO;
2400         return ERR_PTR(ret);
2401 out_kfree_queues:
2402         kfree(ctrl->queues);
2403 out_free_ctrl:
2404         kfree(ctrl);
2405         return ERR_PTR(ret);
2406 }
2407
2408 static struct nvmf_transport_ops nvme_rdma_transport = {
2409         .name           = "rdma",
2410         .module         = THIS_MODULE,
2411         .required_opts  = NVMF_OPT_TRADDR,
2412         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2413                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2414                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2415                           NVMF_OPT_TOS,
2416         .create_ctrl    = nvme_rdma_create_ctrl,
2417 };
2418
2419 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2420 {
2421         struct nvme_rdma_ctrl *ctrl;
2422         struct nvme_rdma_device *ndev;
2423         bool found = false;
2424
2425         mutex_lock(&device_list_mutex);
2426         list_for_each_entry(ndev, &device_list, entry) {
2427                 if (ndev->dev == ib_device) {
2428                         found = true;
2429                         break;
2430                 }
2431         }
2432         mutex_unlock(&device_list_mutex);
2433
2434         if (!found)
2435                 return;
2436
2437         /* Delete all controllers using this device */
2438         mutex_lock(&nvme_rdma_ctrl_mutex);
2439         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2440                 if (ctrl->device->dev != ib_device)
2441                         continue;
2442                 nvme_delete_ctrl(&ctrl->ctrl);
2443         }
2444         mutex_unlock(&nvme_rdma_ctrl_mutex);
2445
2446         flush_workqueue(nvme_delete_wq);
2447 }
2448
2449 static struct ib_client nvme_rdma_ib_client = {
2450         .name   = "nvme_rdma",
2451         .remove = nvme_rdma_remove_one
2452 };
2453
2454 static int __init nvme_rdma_init_module(void)
2455 {
2456         int ret;
2457
2458         ret = ib_register_client(&nvme_rdma_ib_client);
2459         if (ret)
2460                 return ret;
2461
2462         ret = nvmf_register_transport(&nvme_rdma_transport);
2463         if (ret)
2464                 goto err_unreg_client;
2465
2466         return 0;
2467
2468 err_unreg_client:
2469         ib_unregister_client(&nvme_rdma_ib_client);
2470         return ret;
2471 }
2472
2473 static void __exit nvme_rdma_cleanup_module(void)
2474 {
2475         struct nvme_rdma_ctrl *ctrl;
2476
2477         nvmf_unregister_transport(&nvme_rdma_transport);
2478         ib_unregister_client(&nvme_rdma_ib_client);
2479
2480         mutex_lock(&nvme_rdma_ctrl_mutex);
2481         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2482                 nvme_delete_ctrl(&ctrl->ctrl);
2483         mutex_unlock(&nvme_rdma_ctrl_mutex);
2484         flush_workqueue(nvme_delete_wq);
2485 }
2486
2487 module_init(nvme_rdma_init_module);
2488 module_exit(nvme_rdma_cleanup_module);
2489
2490 MODULE_LICENSE("GPL v2");