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         struct mutex            queue_lock;
 101 };
 102
 103 struct nvme_rdma_ctrl {
 104         /* read only in the hot path */
 105         struct nvme_rdma_queue  *queues;
 106
 107         /* other member variables */
 108         struct blk_mq_tag_set   tag_set;
 109         struct work_struct      err_work;
 110
 111         struct nvme_rdma_qe     async_event_sqe;
 112
 113         struct delayed_work     reconnect_work;
 114
 115         struct list_head        list;
 116
 117         struct blk_mq_tag_set   admin_tag_set;
 118         struct nvme_rdma_device *device;
 119
 120         u32                     max_fr_pages;
 121
 122         struct sockaddr_storage addr;
 123         struct sockaddr_storage src_addr;
 124
 125         struct nvme_ctrl        ctrl;
 126         bool                    use_inline_data;
 127         u32                     io_queues[HCTX_MAX_TYPES];
 128 };
 129
 130 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 131 {
 132         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 133 }
 134
 135 static LIST_HEAD(device_list);
 136 static DEFINE_MUTEX(device_list_mutex);
 137
 138 static LIST_HEAD(nvme_rdma_ctrl_list);
 139 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 140
 141 /*
 142  * Disabling this option makes small I/O goes faster, but is fundamentally
 143  * unsafe.  With it turned off we will have to register a global rkey that
 144  * allows read and write access to all physical memory.
 145  */
 146 static bool register_always = true;
 147 module_param(register_always, bool, 0444);
 148 MODULE_PARM_DESC(register_always,
 149          "Use memory registration even for contiguous memory regions");
 150
 151 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 152                 struct rdma_cm_event *event);
 153 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 154 static void nvme_rdma_complete_rq(struct request *rq);
 155
 156 static const struct blk_mq_ops nvme_rdma_mq_ops;
 157 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 158
 159 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 160 {
 161         return queue - queue->ctrl->queues;
 162 }
 163
 164 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
 165 {
 166         return nvme_rdma_queue_idx(queue) >
 167                 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
 168                 queue->ctrl->io_queues[HCTX_TYPE_READ];
 169 }
 170
 171 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 172 {
 173         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 174 }
 175
 176 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 177                 size_t capsule_size, enum dma_data_direction dir)
 178 {
 179         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 180         kfree(qe->data);
 181 }
 182
 183 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 184                 size_t capsule_size, enum dma_data_direction dir)
 185 {
 186         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 187         if (!qe->data)
 188                 return -ENOMEM;
 189
 190         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 191         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 192                 kfree(qe->data);
 193                 qe->data = NULL;
 194                 return -ENOMEM;
 195         }
 196
 197         return 0;
 198 }
 199
 200 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 201                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 202                 size_t capsule_size, enum dma_data_direction dir)
 203 {
 204         int i;
 205
 206         for (i = 0; i < ib_queue_size; i++)
 207                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 208         kfree(ring);
 209 }
 210
 211 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 212                 size_t ib_queue_size, size_t capsule_size,
 213                 enum dma_data_direction dir)
 214 {
 215         struct nvme_rdma_qe *ring;
 216         int i;
 217
 218         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 219         if (!ring)
 220                 return NULL;
 221
 222         /*
 223          * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
 224          * lifetime. It's safe, since any chage in the underlying RDMA device
 225          * will issue error recovery and queue re-creation.
 226          */
 227         for (i = 0; i < ib_queue_size; i++) {
 228                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 229                         goto out_free_ring;
 230         }
 231
 232         return ring;
 233
 234 out_free_ring:
 235         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 236         return NULL;
 237 }
 238
 239 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 240 {
 241         pr_debug("QP event %s (%d)\n",
 242                  ib_event_msg(event->event), event->event);
 243
 244 }
 245
 246 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 247 {
 248         int ret;
 249
 250         ret = wait_for_completion_interruptible_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         mutex_init(&queue->queue_lock);
 584         queue->ctrl = ctrl;
 585         if (idx && ctrl->ctrl.max_integrity_segments)
 586                 queue->pi_support = true;
 587         else
 588                 queue->pi_support = false;
 589         init_completion(&queue->cm_done);
 590
 591         if (idx > 0)
 592                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 593         else
 594                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 595
 596         queue->queue_size = queue_size;
 597
 598         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 599                         RDMA_PS_TCP, IB_QPT_RC);
 600         if (IS_ERR(queue->cm_id)) {
 601                 dev_info(ctrl->ctrl.device,
 602                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 603                 ret = PTR_ERR(queue->cm_id);
 604                 goto out_destroy_mutex;
 605         }
 606
 607         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 608                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 609
 610         queue->cm_error = -ETIMEDOUT;
 611         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 612                         (struct sockaddr *)&ctrl->addr,
 613                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 614         if (ret) {
 615                 dev_info(ctrl->ctrl.device,
 616                         "rdma_resolve_addr failed (%d).\n", ret);
 617                 goto out_destroy_cm_id;
 618         }
 619
 620         ret = nvme_rdma_wait_for_cm(queue);
 621         if (ret) {
 622                 dev_info(ctrl->ctrl.device,
 623                         "rdma connection establishment failed (%d)\n", ret);
 624                 goto out_destroy_cm_id;
 625         }
 626
 627         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 628
 629         return 0;
 630
 631 out_destroy_cm_id:
 632         rdma_destroy_id(queue->cm_id);
 633         nvme_rdma_destroy_queue_ib(queue);
 634 out_destroy_mutex:
 635         mutex_destroy(&queue->queue_lock);
 636         return ret;
 637 }
 638
 639 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 640 {
 641         rdma_disconnect(queue->cm_id);
 642         ib_drain_qp(queue->qp);
 643 }
 644
 645 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 646 {
 647         mutex_lock(&queue->queue_lock);
 648         if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 649                 __nvme_rdma_stop_queue(queue);
 650         mutex_unlock(&queue->queue_lock);
 651 }
 652
 653 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 654 {
 655         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 656                 return;
 657
 658         nvme_rdma_destroy_queue_ib(queue);
 659         rdma_destroy_id(queue->cm_id);
 660         mutex_destroy(&queue->queue_lock);
 661 }
 662
 663 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 664 {
 665         int i;
 666
 667         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 668                 nvme_rdma_free_queue(&ctrl->queues[i]);
 669 }
 670
 671 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 672 {
 673         int i;
 674
 675         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 676                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 677 }
 678
 679 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 680 {
 681         struct nvme_rdma_queue *queue = &ctrl->queues[idx];
 682         bool poll = nvme_rdma_poll_queue(queue);
 683         int ret;
 684
 685         if (idx)
 686                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
 687         else
 688                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 689
 690         if (!ret) {
 691                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 692         } else {
 693                 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 694                         __nvme_rdma_stop_queue(queue);
 695                 dev_info(ctrl->ctrl.device,
 696                         "failed to connect queue: %d ret=%d\n", idx, ret);
 697         }
 698         return ret;
 699 }
 700
 701 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 702 {
 703         int i, ret = 0;
 704
 705         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 706                 ret = nvme_rdma_start_queue(ctrl, i);
 707                 if (ret)
 708                         goto out_stop_queues;
 709         }
 710
 711         return 0;
 712
 713 out_stop_queues:
 714         for (i--; i >= 1; i--)
 715                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 716         return ret;
 717 }
 718
 719 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 720 {
 721         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 722         struct ib_device *ibdev = ctrl->device->dev;
 723         unsigned int nr_io_queues, nr_default_queues;
 724         unsigned int nr_read_queues, nr_poll_queues;
 725         int i, ret;
 726
 727         nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
 728                                 min(opts->nr_io_queues, num_online_cpus()));
 729         nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
 730                                 min(opts->nr_write_queues, num_online_cpus()));
 731         nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
 732         nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
 733
 734         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 735         if (ret)
 736                 return ret;
 737
 738         ctrl->ctrl.queue_count = nr_io_queues + 1;
 739         if (ctrl->ctrl.queue_count < 2)
 740                 return 0;
 741
 742         dev_info(ctrl->ctrl.device,
 743                 "creating %d I/O queues.\n", nr_io_queues);
 744
 745         if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
 746                 /*
 747                  * separate read/write queues
 748                  * hand out dedicated default queues only after we have
 749                  * sufficient read queues.
 750                  */
 751                 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
 752                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
 753                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 754                         min(nr_default_queues, nr_io_queues);
 755                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 756         } else {
 757                 /*
 758                  * shared read/write queues
 759                  * either no write queues were requested, or we don't have
 760                  * sufficient queue count to have dedicated default queues.
 761                  */
 762                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 763                         min(nr_read_queues, nr_io_queues);
 764                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 765         }
 766
 767         if (opts->nr_poll_queues && nr_io_queues) {
 768                 /* map dedicated poll queues only if we have queues left */
 769                 ctrl->io_queues[HCTX_TYPE_POLL] =
 770                         min(nr_poll_queues, nr_io_queues);
 771         }
 772
 773         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 774                 ret = nvme_rdma_alloc_queue(ctrl, i,
 775                                 ctrl->ctrl.sqsize + 1);
 776                 if (ret)
 777                         goto out_free_queues;
 778         }
 779
 780         return 0;
 781
 782 out_free_queues:
 783         for (i--; i >= 1; i--)
 784                 nvme_rdma_free_queue(&ctrl->queues[i]);
 785
 786         return ret;
 787 }
 788
 789 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 790                 bool admin)
 791 {
 792         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 793         struct blk_mq_tag_set *set;
 794         int ret;
 795
 796         if (admin) {
 797                 set = &ctrl->admin_tag_set;
 798                 memset(set, 0, sizeof(*set));
 799                 set->ops = &nvme_rdma_admin_mq_ops;
 800                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 801                 set->reserved_tags = 2; /* connect + keep-alive */
 802                 set->numa_node = nctrl->numa_node;
 803                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 804                                 NVME_RDMA_DATA_SGL_SIZE;
 805                 set->driver_data = ctrl;
 806                 set->nr_hw_queues = 1;
 807                 set->timeout = NVME_ADMIN_TIMEOUT;
 808                 set->flags = BLK_MQ_F_NO_SCHED;
 809         } else {
 810                 set = &ctrl->tag_set;
 811                 memset(set, 0, sizeof(*set));
 812                 set->ops = &nvme_rdma_mq_ops;
 813                 set->queue_depth = nctrl->sqsize + 1;
 814                 set->reserved_tags = 1; /* fabric connect */
 815                 set->numa_node = nctrl->numa_node;
 816                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 817                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 818                                 NVME_RDMA_DATA_SGL_SIZE;
 819                 if (nctrl->max_integrity_segments)
 820                         set->cmd_size += sizeof(struct nvme_rdma_sgl) +
 821                                          NVME_RDMA_METADATA_SGL_SIZE;
 822                 set->driver_data = ctrl;
 823                 set->nr_hw_queues = nctrl->queue_count - 1;
 824                 set->timeout = NVME_IO_TIMEOUT;
 825                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
 826         }
 827
 828         ret = blk_mq_alloc_tag_set(set);
 829         if (ret)
 830                 return ERR_PTR(ret);
 831
 832         return set;
 833 }
 834
 835 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 836                 bool remove)
 837 {
 838         if (remove) {
 839                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 840                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 841                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 842         }
 843         if (ctrl->async_event_sqe.data) {
 844                 cancel_work_sync(&ctrl->ctrl.async_event_work);
 845                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 846                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 847                 ctrl->async_event_sqe.data = NULL;
 848         }
 849         nvme_rdma_free_queue(&ctrl->queues[0]);
 850 }
 851
 852 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 853                 bool new)
 854 {
 855         bool pi_capable = false;
 856         int error;
 857
 858         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 859         if (error)
 860                 return error;
 861
 862         ctrl->device = ctrl->queues[0].device;
 863         ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
 864
 865         /* T10-PI support */
 866         if (ctrl->device->dev->attrs.device_cap_flags &
 867             IB_DEVICE_INTEGRITY_HANDOVER)
 868                 pi_capable = true;
 869
 870         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
 871                                                         pi_capable);
 872
 873         /*
 874          * Bind the async event SQE DMA mapping to the admin queue lifetime.
 875          * It's safe, since any chage in the underlying RDMA device will issue
 876          * error recovery and queue re-creation.
 877          */
 878         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 879                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 880         if (error)
 881                 goto out_free_queue;
 882
 883         if (new) {
 884                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 885                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 886                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 887                         goto out_free_async_qe;
 888                 }
 889
 890                 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 891                 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 892                         error = PTR_ERR(ctrl->ctrl.fabrics_q);
 893                         goto out_free_tagset;
 894                 }
 895
 896                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 897                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 898                         error = PTR_ERR(ctrl->ctrl.admin_q);
 899                         goto out_cleanup_fabrics_q;
 900                 }
 901         }
 902
 903         error = nvme_rdma_start_queue(ctrl, 0);
 904         if (error)
 905                 goto out_cleanup_queue;
 906
 907         error = nvme_enable_ctrl(&ctrl->ctrl);
 908         if (error)
 909                 goto out_stop_queue;
 910
 911         ctrl->ctrl.max_segments = ctrl->max_fr_pages;
 912         ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
 913         if (pi_capable)
 914                 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
 915         else
 916                 ctrl->ctrl.max_integrity_segments = 0;
 917
 918         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 919
 920         error = nvme_init_identify(&ctrl->ctrl);
 921         if (error)
 922                 goto out_stop_queue;
 923
 924         return 0;
 925
 926 out_stop_queue:
 927         nvme_rdma_stop_queue(&ctrl->queues[0]);
 928 out_cleanup_queue:
 929         if (new)
 930                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 931 out_cleanup_fabrics_q:
 932         if (new)
 933                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 934 out_free_tagset:
 935         if (new)
 936                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 937 out_free_async_qe:
 938         if (ctrl->async_event_sqe.data) {
 939                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 940                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 941                 ctrl->async_event_sqe.data = NULL;
 942         }
 943 out_free_queue:
 944         nvme_rdma_free_queue(&ctrl->queues[0]);
 945         return error;
 946 }
 947
 948 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 949                 bool remove)
 950 {
 951         if (remove) {
 952                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 953                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 954         }
 955         nvme_rdma_free_io_queues(ctrl);
 956 }
 957
 958 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 959 {
 960         int ret;
 961
 962         ret = nvme_rdma_alloc_io_queues(ctrl);
 963         if (ret)
 964                 return ret;
 965
 966         if (new) {
 967                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 968                 if (IS_ERR(ctrl->ctrl.tagset)) {
 969                         ret = PTR_ERR(ctrl->ctrl.tagset);
 970                         goto out_free_io_queues;
 971                 }
 972
 973                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 974                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 975                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 976                         goto out_free_tag_set;
 977                 }
 978         }
 979
 980         ret = nvme_rdma_start_io_queues(ctrl);
 981         if (ret)
 982                 goto out_cleanup_connect_q;
 983
 984         if (!new) {
 985                 nvme_start_queues(&ctrl->ctrl);
 986                 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
 987                         /*
 988                          * If we timed out waiting for freeze we are likely to
 989                          * be stuck.  Fail the controller initialization just
 990                          * to be safe.
 991                          */
 992                         ret = -ENODEV;
 993                         goto out_wait_freeze_timed_out;
 994                 }
 995                 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
 996                         ctrl->ctrl.queue_count - 1);
 997                 nvme_unfreeze(&ctrl->ctrl);
 998         }
 999
1000         return 0;
1001
1002 out_wait_freeze_timed_out:
1003         nvme_stop_queues(&ctrl->ctrl);
1004         nvme_rdma_stop_io_queues(ctrl);
1005 out_cleanup_connect_q:
1006         if (new)
1007                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1008 out_free_tag_set:
1009         if (new)
1010                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1011 out_free_io_queues:
1012         nvme_rdma_free_io_queues(ctrl);
1013         return ret;
1014 }
1015
1016 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1017                 bool remove)
1018 {
1019         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1020         blk_sync_queue(ctrl->ctrl.admin_q);
1021         nvme_rdma_stop_queue(&ctrl->queues[0]);
1022         if (ctrl->ctrl.admin_tagset) {
1023                 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
1024                         nvme_cancel_request, &ctrl->ctrl);
1025                 blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
1026         }
1027         if (remove)
1028                 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1029         nvme_rdma_destroy_admin_queue(ctrl, remove);
1030 }
1031
1032 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1033                 bool remove)
1034 {
1035         if (ctrl->ctrl.queue_count > 1) {
1036                 nvme_start_freeze(&ctrl->ctrl);
1037                 nvme_stop_queues(&ctrl->ctrl);
1038                 nvme_sync_io_queues(&ctrl->ctrl);
1039                 nvme_rdma_stop_io_queues(ctrl);
1040                 if (ctrl->ctrl.tagset) {
1041                         blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
1042                                 nvme_cancel_request, &ctrl->ctrl);
1043                         blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
1044                 }
1045                 if (remove)
1046                         nvme_start_queues(&ctrl->ctrl);
1047                 nvme_rdma_destroy_io_queues(ctrl, remove);
1048         }
1049 }
1050
1051 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1052 {
1053         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1054
1055         if (list_empty(&ctrl->list))
1056                 goto free_ctrl;
1057
1058         mutex_lock(&nvme_rdma_ctrl_mutex);
1059         list_del(&ctrl->list);
1060         mutex_unlock(&nvme_rdma_ctrl_mutex);
1061
1062         nvmf_free_options(nctrl->opts);
1063 free_ctrl:
1064         kfree(ctrl->queues);
1065         kfree(ctrl);
1066 }
1067
1068 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1069 {
1070         /* If we are resetting/deleting then do nothing */
1071         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1072                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1073                         ctrl->ctrl.state == NVME_CTRL_LIVE);
1074                 return;
1075         }
1076
1077         if (nvmf_should_reconnect(&ctrl->ctrl)) {
1078                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1079                         ctrl->ctrl.opts->reconnect_delay);
1080                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1081                                 ctrl->ctrl.opts->reconnect_delay * HZ);
1082         } else {
1083                 nvme_delete_ctrl(&ctrl->ctrl);
1084         }
1085 }
1086
1087 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1088 {
1089         int ret = -EINVAL;
1090         bool changed;
1091
1092         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1093         if (ret)
1094                 return ret;
1095
1096         if (ctrl->ctrl.icdoff) {
1097                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1098                 goto destroy_admin;
1099         }
1100
1101         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1102                 dev_err(ctrl->ctrl.device,
1103                         "Mandatory keyed sgls are not supported!\n");
1104                 goto destroy_admin;
1105         }
1106
1107         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1108                 dev_warn(ctrl->ctrl.device,
1109                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1110                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1111         }
1112
1113         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1114                 dev_warn(ctrl->ctrl.device,
1115                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1116                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1117                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1118         }
1119
1120         if (ctrl->ctrl.sgls & (1 << 20))
1121                 ctrl->use_inline_data = true;
1122
1123         if (ctrl->ctrl.queue_count > 1) {
1124                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1125                 if (ret)
1126                         goto destroy_admin;
1127         }
1128
1129         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1130         if (!changed) {
1131                 /*
1132                  * state change failure is ok if we started ctrl delete,
1133                  * unless we're during creation of a new controller to
1134                  * avoid races with teardown flow.
1135                  */
1136                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1137                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1138                 WARN_ON_ONCE(new);
1139                 ret = -EINVAL;
1140                 goto destroy_io;
1141         }
1142
1143         nvme_start_ctrl(&ctrl->ctrl);
1144         return 0;
1145
1146 destroy_io:
1147         if (ctrl->ctrl.queue_count > 1)
1148                 nvme_rdma_destroy_io_queues(ctrl, new);
1149 destroy_admin:
1150         nvme_rdma_stop_queue(&ctrl->queues[0]);
1151         nvme_rdma_destroy_admin_queue(ctrl, new);
1152         return ret;
1153 }
1154
1155 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1156 {
1157         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1158                         struct nvme_rdma_ctrl, reconnect_work);
1159
1160         ++ctrl->ctrl.nr_reconnects;
1161
1162         if (nvme_rdma_setup_ctrl(ctrl, false))
1163                 goto requeue;
1164
1165         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1166                         ctrl->ctrl.nr_reconnects);
1167
1168         ctrl->ctrl.nr_reconnects = 0;
1169
1170         return;
1171
1172 requeue:
1173         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1174                         ctrl->ctrl.nr_reconnects);
1175         nvme_rdma_reconnect_or_remove(ctrl);
1176 }
1177
1178 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1179 {
1180         struct nvme_rdma_ctrl *ctrl = container_of(work,
1181                         struct nvme_rdma_ctrl, err_work);
1182
1183         nvme_stop_keep_alive(&ctrl->ctrl);
1184         nvme_rdma_teardown_io_queues(ctrl, false);
1185         nvme_start_queues(&ctrl->ctrl);
1186         nvme_rdma_teardown_admin_queue(ctrl, false);
1187         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1188
1189         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1190                 /* state change failure is ok if we started ctrl delete */
1191                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1192                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1193                 return;
1194         }
1195
1196         nvme_rdma_reconnect_or_remove(ctrl);
1197 }
1198
1199 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1200 {
1201         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1202                 return;
1203
1204         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1205         queue_work(nvme_reset_wq, &ctrl->err_work);
1206 }
1207
1208 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1209 {
1210         struct request *rq = blk_mq_rq_from_pdu(req);
1211
1212         if (!refcount_dec_and_test(&req->ref))
1213                 return;
1214         if (!nvme_try_complete_req(rq, req->status, req->result))
1215                 nvme_rdma_complete_rq(rq);
1216 }
1217
1218 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1219                 const char *op)
1220 {
1221         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1222         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1223
1224         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1225                 dev_info(ctrl->ctrl.device,
1226                              "%s for CQE 0x%p failed with status %s (%d)\n",
1227                              op, wc->wr_cqe,
1228                              ib_wc_status_msg(wc->status), wc->status);
1229         nvme_rdma_error_recovery(ctrl);
1230 }
1231
1232 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1233 {
1234         if (unlikely(wc->status != IB_WC_SUCCESS))
1235                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1236 }
1237
1238 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1239 {
1240         struct nvme_rdma_request *req =
1241                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1242
1243         if (unlikely(wc->status != IB_WC_SUCCESS))
1244                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1245         else
1246                 nvme_rdma_end_request(req);
1247 }
1248
1249 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1250                 struct nvme_rdma_request *req)
1251 {
1252         struct ib_send_wr wr = {
1253                 .opcode             = IB_WR_LOCAL_INV,
1254                 .next               = NULL,
1255                 .num_sge            = 0,
1256                 .send_flags         = IB_SEND_SIGNALED,
1257                 .ex.invalidate_rkey = req->mr->rkey,
1258         };
1259
1260         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1261         wr.wr_cqe = &req->reg_cqe;
1262
1263         return ib_post_send(queue->qp, &wr, NULL);
1264 }
1265
1266 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1267                 struct request *rq)
1268 {
1269         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1270         struct nvme_rdma_device *dev = queue->device;
1271         struct ib_device *ibdev = dev->dev;
1272         struct list_head *pool = &queue->qp->rdma_mrs;
1273
1274         if (!blk_rq_nr_phys_segments(rq))
1275                 return;
1276
1277         if (blk_integrity_rq(rq)) {
1278                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1279                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1280                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1281                                       NVME_INLINE_METADATA_SG_CNT);
1282         }
1283
1284         if (req->use_sig_mr)
1285                 pool = &queue->qp->sig_mrs;
1286
1287         if (req->mr) {
1288                 ib_mr_pool_put(queue->qp, pool, req->mr);
1289                 req->mr = NULL;
1290         }
1291
1292         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1293                         rq_dma_dir(rq));
1294         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1295 }
1296
1297 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1298 {
1299         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1300
1301         sg->addr = 0;
1302         put_unaligned_le24(0, sg->length);
1303         put_unaligned_le32(0, sg->key);
1304         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1305         return 0;
1306 }
1307
1308 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1309                 struct nvme_rdma_request *req, struct nvme_command *c,
1310                 int count)
1311 {
1312         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1313         struct scatterlist *sgl = req->data_sgl.sg_table.sgl;
1314         struct ib_sge *sge = &req->sge[1];
1315         u32 len = 0;
1316         int i;
1317
1318         for (i = 0; i < count; i++, sgl++, sge++) {
1319                 sge->addr = sg_dma_address(sgl);
1320                 sge->length = sg_dma_len(sgl);
1321                 sge->lkey = queue->device->pd->local_dma_lkey;
1322                 len += sge->length;
1323         }
1324
1325         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1326         sg->length = cpu_to_le32(len);
1327         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1328
1329         req->num_sge += count;
1330         return 0;
1331 }
1332
1333 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1334                 struct nvme_rdma_request *req, struct nvme_command *c)
1335 {
1336         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1337
1338         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1339         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1340         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1341         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1342         return 0;
1343 }
1344
1345 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1346                 struct nvme_rdma_request *req, struct nvme_command *c,
1347                 int count)
1348 {
1349         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1350         int nr;
1351
1352         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1353         if (WARN_ON_ONCE(!req->mr))
1354                 return -EAGAIN;
1355
1356         /*
1357          * Align the MR to a 4K page size to match the ctrl page size and
1358          * the block virtual boundary.
1359          */
1360         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1361                           SZ_4K);
1362         if (unlikely(nr < count)) {
1363                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1364                 req->mr = NULL;
1365                 if (nr < 0)
1366                         return nr;
1367                 return -EINVAL;
1368         }
1369
1370         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1371
1372         req->reg_cqe.done = nvme_rdma_memreg_done;
1373         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1374         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1375         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1376         req->reg_wr.wr.num_sge = 0;
1377         req->reg_wr.mr = req->mr;
1378         req->reg_wr.key = req->mr->rkey;
1379         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1380                              IB_ACCESS_REMOTE_READ |
1381                              IB_ACCESS_REMOTE_WRITE;
1382
1383         sg->addr = cpu_to_le64(req->mr->iova);
1384         put_unaligned_le24(req->mr->length, sg->length);
1385         put_unaligned_le32(req->mr->rkey, sg->key);
1386         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1387                         NVME_SGL_FMT_INVALIDATE;
1388
1389         return 0;
1390 }
1391
1392 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1393                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1394                 u16 control, u8 pi_type)
1395 {
1396         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1397         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1398         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1399         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1400         if (control & NVME_RW_PRINFO_PRCHK_REF)
1401                 domain->sig.dif.ref_remap = true;
1402
1403         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1404         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1405         domain->sig.dif.app_escape = true;
1406         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1407                 domain->sig.dif.ref_escape = true;
1408 }
1409
1410 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1411                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1412                 u8 pi_type)
1413 {
1414         u16 control = le16_to_cpu(cmd->rw.control);
1415
1416         memset(sig_attrs, 0, sizeof(*sig_attrs));
1417         if (control & NVME_RW_PRINFO_PRACT) {
1418                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1419                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1420                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1421                                          pi_type);
1422                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1423                 control &= ~NVME_RW_PRINFO_PRACT;
1424                 cmd->rw.control = cpu_to_le16(control);
1425         } else {
1426                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1427                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1428                                          pi_type);
1429                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1430                                          pi_type);
1431         }
1432 }
1433
1434 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1435 {
1436         *mask = 0;
1437         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1438                 *mask |= IB_SIG_CHECK_REFTAG;
1439         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1440                 *mask |= IB_SIG_CHECK_GUARD;
1441 }
1442
1443 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1444 {
1445         if (unlikely(wc->status != IB_WC_SUCCESS))
1446                 nvme_rdma_wr_error(cq, wc, "SIG");
1447 }
1448
1449 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1450                 struct nvme_rdma_request *req, struct nvme_command *c,
1451                 int count, int pi_count)
1452 {
1453         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1454         struct ib_reg_wr *wr = &req->reg_wr;
1455         struct request *rq = blk_mq_rq_from_pdu(req);
1456         struct nvme_ns *ns = rq->q->queuedata;
1457         struct bio *bio = rq->bio;
1458         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1459         int nr;
1460
1461         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1462         if (WARN_ON_ONCE(!req->mr))
1463                 return -EAGAIN;
1464
1465         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1466                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1467                              SZ_4K);
1468         if (unlikely(nr))
1469                 goto mr_put;
1470
1471         nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_disk), c,
1472                                 req->mr->sig_attrs, ns->pi_type);
1473         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1474
1475         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1476
1477         req->reg_cqe.done = nvme_rdma_sig_done;
1478         memset(wr, 0, sizeof(*wr));
1479         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1480         wr->wr.wr_cqe = &req->reg_cqe;
1481         wr->wr.num_sge = 0;
1482         wr->wr.send_flags = 0;
1483         wr->mr = req->mr;
1484         wr->key = req->mr->rkey;
1485         wr->access = IB_ACCESS_LOCAL_WRITE |
1486                      IB_ACCESS_REMOTE_READ |
1487                      IB_ACCESS_REMOTE_WRITE;
1488
1489         sg->addr = cpu_to_le64(req->mr->iova);
1490         put_unaligned_le24(req->mr->length, sg->length);
1491         put_unaligned_le32(req->mr->rkey, sg->key);
1492         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1493
1494         return 0;
1495
1496 mr_put:
1497         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1498         req->mr = NULL;
1499         if (nr < 0)
1500                 return nr;
1501         return -EINVAL;
1502 }
1503
1504 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1505                 struct request *rq, struct nvme_command *c)
1506 {
1507         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1508         struct nvme_rdma_device *dev = queue->device;
1509         struct ib_device *ibdev = dev->dev;
1510         int pi_count = 0;
1511         int count, ret;
1512
1513         req->num_sge = 1;
1514         refcount_set(&req->ref, 2); /* send and recv completions */
1515
1516         c->common.flags |= NVME_CMD_SGL_METABUF;
1517
1518         if (!blk_rq_nr_phys_segments(rq))
1519                 return nvme_rdma_set_sg_null(c);
1520
1521         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1522         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1523                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1524                         NVME_INLINE_SG_CNT);
1525         if (ret)
1526                 return -ENOMEM;
1527
1528         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1529                                             req->data_sgl.sg_table.sgl);
1530
1531         count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1532                               req->data_sgl.nents, rq_dma_dir(rq));
1533         if (unlikely(count <= 0)) {
1534                 ret = -EIO;
1535                 goto out_free_table;
1536         }
1537
1538         if (blk_integrity_rq(rq)) {
1539                 req->metadata_sgl->sg_table.sgl =
1540                         (struct scatterlist *)(req->metadata_sgl + 1);
1541                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1542                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1543                                 req->metadata_sgl->sg_table.sgl,
1544                                 NVME_INLINE_METADATA_SG_CNT);
1545                 if (unlikely(ret)) {
1546                         ret = -ENOMEM;
1547                         goto out_unmap_sg;
1548                 }
1549
1550                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1551                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1552                 pi_count = ib_dma_map_sg(ibdev,
1553                                          req->metadata_sgl->sg_table.sgl,
1554                                          req->metadata_sgl->nents,
1555                                          rq_dma_dir(rq));
1556                 if (unlikely(pi_count <= 0)) {
1557                         ret = -EIO;
1558                         goto out_free_pi_table;
1559                 }
1560         }
1561
1562         if (req->use_sig_mr) {
1563                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1564                 goto out;
1565         }
1566
1567         if (count <= dev->num_inline_segments) {
1568                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1569                     queue->ctrl->use_inline_data &&
1570                     blk_rq_payload_bytes(rq) <=
1571                                 nvme_rdma_inline_data_size(queue)) {
1572                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1573                         goto out;
1574                 }
1575
1576                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1577                         ret = nvme_rdma_map_sg_single(queue, req, c);
1578                         goto out;
1579                 }
1580         }
1581
1582         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1583 out:
1584         if (unlikely(ret))
1585                 goto out_unmap_pi_sg;
1586
1587         return 0;
1588
1589 out_unmap_pi_sg:
1590         if (blk_integrity_rq(rq))
1591                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1592                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1593 out_free_pi_table:
1594         if (blk_integrity_rq(rq))
1595                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1596                                       NVME_INLINE_METADATA_SG_CNT);
1597 out_unmap_sg:
1598         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1599                         rq_dma_dir(rq));
1600 out_free_table:
1601         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1602         return ret;
1603 }
1604
1605 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1606 {
1607         struct nvme_rdma_qe *qe =
1608                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1609         struct nvme_rdma_request *req =
1610                 container_of(qe, struct nvme_rdma_request, sqe);
1611
1612         if (unlikely(wc->status != IB_WC_SUCCESS))
1613                 nvme_rdma_wr_error(cq, wc, "SEND");
1614         else
1615                 nvme_rdma_end_request(req);
1616 }
1617
1618 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1619                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1620                 struct ib_send_wr *first)
1621 {
1622         struct ib_send_wr wr;
1623         int ret;
1624
1625         sge->addr   = qe->dma;
1626         sge->length = sizeof(struct nvme_command);
1627         sge->lkey   = queue->device->pd->local_dma_lkey;
1628
1629         wr.next       = NULL;
1630         wr.wr_cqe     = &qe->cqe;
1631         wr.sg_list    = sge;
1632         wr.num_sge    = num_sge;
1633         wr.opcode     = IB_WR_SEND;
1634         wr.send_flags = IB_SEND_SIGNALED;
1635
1636         if (first)
1637                 first->next = &wr;
1638         else
1639                 first = &wr;
1640
1641         ret = ib_post_send(queue->qp, first, NULL);
1642         if (unlikely(ret)) {
1643                 dev_err(queue->ctrl->ctrl.device,
1644                              "%s failed with error code %d\n", __func__, ret);
1645         }
1646         return ret;
1647 }
1648
1649 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1650                 struct nvme_rdma_qe *qe)
1651 {
1652         struct ib_recv_wr wr;
1653         struct ib_sge list;
1654         int ret;
1655
1656         list.addr   = qe->dma;
1657         list.length = sizeof(struct nvme_completion);
1658         list.lkey   = queue->device->pd->local_dma_lkey;
1659
1660         qe->cqe.done = nvme_rdma_recv_done;
1661
1662         wr.next     = NULL;
1663         wr.wr_cqe   = &qe->cqe;
1664         wr.sg_list  = &list;
1665         wr.num_sge  = 1;
1666
1667         ret = ib_post_recv(queue->qp, &wr, NULL);
1668         if (unlikely(ret)) {
1669                 dev_err(queue->ctrl->ctrl.device,
1670                         "%s failed with error code %d\n", __func__, ret);
1671         }
1672         return ret;
1673 }
1674
1675 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1676 {
1677         u32 queue_idx = nvme_rdma_queue_idx(queue);
1678
1679         if (queue_idx == 0)
1680                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1681         return queue->ctrl->tag_set.tags[queue_idx - 1];
1682 }
1683
1684 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1685 {
1686         if (unlikely(wc->status != IB_WC_SUCCESS))
1687                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1688 }
1689
1690 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1691 {
1692         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1693         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1694         struct ib_device *dev = queue->device->dev;
1695         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1696         struct nvme_command *cmd = sqe->data;
1697         struct ib_sge sge;
1698         int ret;
1699
1700         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1701
1702         memset(cmd, 0, sizeof(*cmd));
1703         cmd->common.opcode = nvme_admin_async_event;
1704         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1705         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1706         nvme_rdma_set_sg_null(cmd);
1707
1708         sqe->cqe.done = nvme_rdma_async_done;
1709
1710         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1711                         DMA_TO_DEVICE);
1712
1713         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1714         WARN_ON_ONCE(ret);
1715 }
1716
1717 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1718                 struct nvme_completion *cqe, struct ib_wc *wc)
1719 {
1720         struct request *rq;
1721         struct nvme_rdma_request *req;
1722
1723         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1724         if (!rq) {
1725                 dev_err(queue->ctrl->ctrl.device,
1726                         "tag 0x%x on QP %#x not found\n",
1727                         cqe->command_id, queue->qp->qp_num);
1728                 nvme_rdma_error_recovery(queue->ctrl);
1729                 return;
1730         }
1731         req = blk_mq_rq_to_pdu(rq);
1732
1733         req->status = cqe->status;
1734         req->result = cqe->result;
1735
1736         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1737                 if (unlikely(!req->mr ||
1738                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1739                         dev_err(queue->ctrl->ctrl.device,
1740                                 "Bogus remote invalidation for rkey %#x\n",
1741                                 req->mr ? req->mr->rkey : 0);
1742                         nvme_rdma_error_recovery(queue->ctrl);
1743                 }
1744         } else if (req->mr) {
1745                 int ret;
1746
1747                 ret = nvme_rdma_inv_rkey(queue, req);
1748                 if (unlikely(ret < 0)) {
1749                         dev_err(queue->ctrl->ctrl.device,
1750                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1751                                 req->mr->rkey, ret);
1752                         nvme_rdma_error_recovery(queue->ctrl);
1753                 }
1754                 /* the local invalidation completion will end the request */
1755                 return;
1756         }
1757
1758         nvme_rdma_end_request(req);
1759 }
1760
1761 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1762 {
1763         struct nvme_rdma_qe *qe =
1764                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1765         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1766         struct ib_device *ibdev = queue->device->dev;
1767         struct nvme_completion *cqe = qe->data;
1768         const size_t len = sizeof(struct nvme_completion);
1769
1770         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1771                 nvme_rdma_wr_error(cq, wc, "RECV");
1772                 return;
1773         }
1774
1775         /* sanity checking for received data length */
1776         if (unlikely(wc->byte_len < len)) {
1777                 dev_err(queue->ctrl->ctrl.device,
1778                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1779                 nvme_rdma_error_recovery(queue->ctrl);
1780                 return;
1781         }
1782
1783         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1784         /*
1785          * AEN requests are special as they don't time out and can
1786          * survive any kind of queue freeze and often don't respond to
1787          * aborts.  We don't even bother to allocate a struct request
1788          * for them but rather special case them here.
1789          */
1790         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1791                                      cqe->command_id)))
1792                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1793                                 &cqe->result);
1794         else
1795                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1796         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1797
1798         nvme_rdma_post_recv(queue, qe);
1799 }
1800
1801 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1802 {
1803         int ret, i;
1804
1805         for (i = 0; i < queue->queue_size; i++) {
1806                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1807                 if (ret)
1808                         goto out_destroy_queue_ib;
1809         }
1810
1811         return 0;
1812
1813 out_destroy_queue_ib:
1814         nvme_rdma_destroy_queue_ib(queue);
1815         return ret;
1816 }
1817
1818 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1819                 struct rdma_cm_event *ev)
1820 {
1821         struct rdma_cm_id *cm_id = queue->cm_id;
1822         int status = ev->status;
1823         const char *rej_msg;
1824         const struct nvme_rdma_cm_rej *rej_data;
1825         u8 rej_data_len;
1826
1827         rej_msg = rdma_reject_msg(cm_id, status);
1828         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1829
1830         if (rej_data && rej_data_len >= sizeof(u16)) {
1831                 u16 sts = le16_to_cpu(rej_data->sts);
1832
1833                 dev_err(queue->ctrl->ctrl.device,
1834                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1835                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1836         } else {
1837                 dev_err(queue->ctrl->ctrl.device,
1838                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1839         }
1840
1841         return -ECONNRESET;
1842 }
1843
1844 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1845 {
1846         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1847         int ret;
1848
1849         ret = nvme_rdma_create_queue_ib(queue);
1850         if (ret)
1851                 return ret;
1852
1853         if (ctrl->opts->tos >= 0)
1854                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1855         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1856         if (ret) {
1857                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1858                         queue->cm_error);
1859                 goto out_destroy_queue;
1860         }
1861
1862         return 0;
1863
1864 out_destroy_queue:
1865         nvme_rdma_destroy_queue_ib(queue);
1866         return ret;
1867 }
1868
1869 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1870 {
1871         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1872         struct rdma_conn_param param = { };
1873         struct nvme_rdma_cm_req priv = { };
1874         int ret;
1875
1876         param.qp_num = queue->qp->qp_num;
1877         param.flow_control = 1;
1878
1879         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1880         /* maximum retry count */
1881         param.retry_count = 7;
1882         param.rnr_retry_count = 7;
1883         param.private_data = &priv;
1884         param.private_data_len = sizeof(priv);
1885
1886         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1887         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1888         /*
1889          * set the admin queue depth to the minimum size
1890          * specified by the Fabrics standard.
1891          */
1892         if (priv.qid == 0) {
1893                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1894                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1895         } else {
1896                 /*
1897                  * current interpretation of the fabrics spec
1898                  * is at minimum you make hrqsize sqsize+1, or a
1899                  * 1's based representation of sqsize.
1900                  */
1901                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1902                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1903         }
1904
1905         ret = rdma_connect_locked(queue->cm_id, &param);
1906         if (ret) {
1907                 dev_err(ctrl->ctrl.device,
1908                         "rdma_connect_locked failed (%d).\n", ret);
1909                 goto out_destroy_queue_ib;
1910         }
1911
1912         return 0;
1913
1914 out_destroy_queue_ib:
1915         nvme_rdma_destroy_queue_ib(queue);
1916         return ret;
1917 }
1918
1919 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1920                 struct rdma_cm_event *ev)
1921 {
1922         struct nvme_rdma_queue *queue = cm_id->context;
1923         int cm_error = 0;
1924
1925         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1926                 rdma_event_msg(ev->event), ev->event,
1927                 ev->status, cm_id);
1928
1929         switch (ev->event) {
1930         case RDMA_CM_EVENT_ADDR_RESOLVED:
1931                 cm_error = nvme_rdma_addr_resolved(queue);
1932                 break;
1933         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1934                 cm_error = nvme_rdma_route_resolved(queue);
1935                 break;
1936         case RDMA_CM_EVENT_ESTABLISHED:
1937                 queue->cm_error = nvme_rdma_conn_established(queue);
1938                 /* complete cm_done regardless of success/failure */
1939                 complete(&queue->cm_done);
1940                 return 0;
1941         case RDMA_CM_EVENT_REJECTED:
1942                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1943                 break;
1944         case RDMA_CM_EVENT_ROUTE_ERROR:
1945         case RDMA_CM_EVENT_CONNECT_ERROR:
1946         case RDMA_CM_EVENT_UNREACHABLE:
1947                 nvme_rdma_destroy_queue_ib(queue);
1948                 fallthrough;
1949         case RDMA_CM_EVENT_ADDR_ERROR:
1950                 dev_dbg(queue->ctrl->ctrl.device,
1951                         "CM error event %d\n", ev->event);
1952                 cm_error = -ECONNRESET;
1953                 break;
1954         case RDMA_CM_EVENT_DISCONNECTED:
1955         case RDMA_CM_EVENT_ADDR_CHANGE:
1956         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1957                 dev_dbg(queue->ctrl->ctrl.device,
1958                         "disconnect received - connection closed\n");
1959                 nvme_rdma_error_recovery(queue->ctrl);
1960                 break;
1961         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1962                 /* device removal is handled via the ib_client API */
1963                 break;
1964         default:
1965                 dev_err(queue->ctrl->ctrl.device,
1966                         "Unexpected RDMA CM event (%d)\n", ev->event);
1967                 nvme_rdma_error_recovery(queue->ctrl);
1968                 break;
1969         }
1970
1971         if (cm_error) {
1972                 queue->cm_error = cm_error;
1973                 complete(&queue->cm_done);
1974         }
1975
1976         return 0;
1977 }
1978
1979 static void nvme_rdma_complete_timed_out(struct request *rq)
1980 {
1981         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1982         struct nvme_rdma_queue *queue = req->queue;
1983
1984         nvme_rdma_stop_queue(queue);
1985         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
1986                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
1987                 blk_mq_complete_request(rq);
1988         }
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
2324         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2325                 opts->trsvcid =
2326                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2327                 if (!opts->trsvcid) {
2328                         ret = -ENOMEM;
2329                         goto out_free_ctrl;
2330                 }
2331                 opts->mask |= NVMF_OPT_TRSVCID;
2332         }
2333
2334         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2335                         opts->traddr, opts->trsvcid, &ctrl->addr);
2336         if (ret) {
2337                 pr_err("malformed address passed: %s:%s\n",
2338                         opts->traddr, opts->trsvcid);
2339                 goto out_free_ctrl;
2340         }
2341
2342         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2343                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2344                         opts->host_traddr, NULL, &ctrl->src_addr);
2345                 if (ret) {
2346                         pr_err("malformed src address passed: %s\n",
2347                                opts->host_traddr);
2348                         goto out_free_ctrl;
2349                 }
2350         }
2351
2352         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2353                 ret = -EALREADY;
2354                 goto out_free_ctrl;
2355         }
2356
2357         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2358                         nvme_rdma_reconnect_ctrl_work);
2359         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2360         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2361
2362         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2363                                 opts->nr_poll_queues + 1;
2364         ctrl->ctrl.sqsize = opts->queue_size - 1;
2365         ctrl->ctrl.kato = opts->kato;
2366
2367         ret = -ENOMEM;
2368         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2369                                 GFP_KERNEL);
2370         if (!ctrl->queues)
2371                 goto out_free_ctrl;
2372
2373         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2374                                 0 /* no quirks, we're perfect! */);
2375         if (ret)
2376                 goto out_kfree_queues;
2377
2378         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2379         WARN_ON_ONCE(!changed);
2380
2381         ret = nvme_rdma_setup_ctrl(ctrl, true);
2382         if (ret)
2383                 goto out_uninit_ctrl;
2384
2385         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2386                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2387
2388         mutex_lock(&nvme_rdma_ctrl_mutex);
2389         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2390         mutex_unlock(&nvme_rdma_ctrl_mutex);
2391
2392         return &ctrl->ctrl;
2393
2394 out_uninit_ctrl:
2395         nvme_uninit_ctrl(&ctrl->ctrl);
2396         nvme_put_ctrl(&ctrl->ctrl);
2397         if (ret > 0)
2398                 ret = -EIO;
2399         return ERR_PTR(ret);
2400 out_kfree_queues:
2401         kfree(ctrl->queues);
2402 out_free_ctrl:
2403         kfree(ctrl);
2404         return ERR_PTR(ret);
2405 }
2406
2407 static struct nvmf_transport_ops nvme_rdma_transport = {
2408         .name           = "rdma",
2409         .module         = THIS_MODULE,
2410         .required_opts  = NVMF_OPT_TRADDR,
2411         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2412                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2413                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2414                           NVMF_OPT_TOS,
2415         .create_ctrl    = nvme_rdma_create_ctrl,
2416 };
2417
2418 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2419 {
2420         struct nvme_rdma_ctrl *ctrl;
2421         struct nvme_rdma_device *ndev;
2422         bool found = false;
2423
2424         mutex_lock(&device_list_mutex);
2425         list_for_each_entry(ndev, &device_list, entry) {
2426                 if (ndev->dev == ib_device) {
2427                         found = true;
2428                         break;
2429                 }
2430         }
2431         mutex_unlock(&device_list_mutex);
2432
2433         if (!found)
2434                 return;
2435
2436         /* Delete all controllers using this device */
2437         mutex_lock(&nvme_rdma_ctrl_mutex);
2438         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2439                 if (ctrl->device->dev != ib_device)
2440                         continue;
2441                 nvme_delete_ctrl(&ctrl->ctrl);
2442         }
2443         mutex_unlock(&nvme_rdma_ctrl_mutex);
2444
2445         flush_workqueue(nvme_delete_wq);
2446 }
2447
2448 static struct ib_client nvme_rdma_ib_client = {
2449         .name   = "nvme_rdma",
2450         .remove = nvme_rdma_remove_one
2451 };
2452
2453 static int __init nvme_rdma_init_module(void)
2454 {
2455         int ret;
2456
2457         ret = ib_register_client(&nvme_rdma_ib_client);
2458         if (ret)
2459                 return ret;
2460
2461         ret = nvmf_register_transport(&nvme_rdma_transport);
2462         if (ret)
2463                 goto err_unreg_client;
2464
2465         return 0;
2466
2467 err_unreg_client:
2468         ib_unregister_client(&nvme_rdma_ib_client);
2469         return ret;
2470 }
2471
2472 static void __exit nvme_rdma_cleanup_module(void)
2473 {
2474         struct nvme_rdma_ctrl *ctrl;
2475
2476         nvmf_unregister_transport(&nvme_rdma_transport);
2477         ib_unregister_client(&nvme_rdma_ib_client);
2478
2479         mutex_lock(&nvme_rdma_ctrl_mutex);
2480         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2481                 nvme_delete_ctrl(&ctrl->ctrl);
2482         mutex_unlock(&nvme_rdma_ctrl_mutex);
2483         flush_workqueue(nvme_delete_wq);
2484 }
2485
2486 module_init(nvme_rdma_init_module);
2487 module_exit(nvme_rdma_cleanup_module);
2488
2489 MODULE_LICENSE("GPL v2");