drivers/nvme/host/rdma.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics RDMA host code.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/slab.h>
  10 #include <rdma/mr_pool.h>
  11 #include <linux/err.h>
  12 #include <linux/string.h>
  13 #include <linux/atomic.h>
  14 #include <linux/blk-mq.h>
  15 #include <linux/blk-mq-rdma.h>
  16 #include <linux/types.h>
  17 #include <linux/list.h>
  18 #include <linux/mutex.h>
  19 #include <linux/scatterlist.h>
  20 #include <linux/nvme.h>
  21 #include <asm/unaligned.h>
  22
  23 #include <rdma/ib_verbs.h>
  24 #include <rdma/rdma_cm.h>
  25 #include <linux/nvme-rdma.h>
  26
  27 #include "nvme.h"
  28 #include "fabrics.h"
  29
  30
  31 #define NVME_RDMA_CONNECT_TIMEOUT_MS    3000            /* 3 second */
  32
  33 #define NVME_RDMA_MAX_SEGMENTS          256
  34
  35 #define NVME_RDMA_MAX_INLINE_SEGMENTS   4
  36
  37 #define NVME_RDMA_DATA_SGL_SIZE \
  38         (sizeof(struct scatterlist) * NVME_INLINE_SG_CNT)
  39 #define NVME_RDMA_METADATA_SGL_SIZE \
  40         (sizeof(struct scatterlist) * NVME_INLINE_METADATA_SG_CNT)
  41
  42 struct nvme_rdma_device {
  43         struct ib_device        *dev;
  44         struct ib_pd            *pd;
  45         struct kref             ref;
  46         struct list_head        entry;
  47         unsigned int            num_inline_segments;
  48 };
  49
  50 struct nvme_rdma_qe {
  51         struct ib_cqe           cqe;
  52         void                    *data;
  53         u64                     dma;
  54 };
  55
  56 struct nvme_rdma_sgl {
  57         int                     nents;
  58         struct sg_table         sg_table;
  59 };
  60
  61 struct nvme_rdma_queue;
  62 struct nvme_rdma_request {
  63         struct nvme_request     req;
  64         struct ib_mr            *mr;
  65         struct nvme_rdma_qe     sqe;
  66         union nvme_result       result;
  67         __le16                  status;
  68         refcount_t              ref;
  69         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  70         u32                     num_sge;
  71         struct ib_reg_wr        reg_wr;
  72         struct ib_cqe           reg_cqe;
  73         struct nvme_rdma_queue  *queue;
  74         struct nvme_rdma_sgl    data_sgl;
  75         struct nvme_rdma_sgl    *metadata_sgl;
  76         bool                    use_sig_mr;
  77 };
  78
  79 enum nvme_rdma_queue_flags {
  80         NVME_RDMA_Q_ALLOCATED           = 0,
  81         NVME_RDMA_Q_LIVE                = 1,
  82         NVME_RDMA_Q_TR_READY            = 2,
  83 };
  84
  85 struct nvme_rdma_queue {
  86         struct nvme_rdma_qe     *rsp_ring;
  87         int                     queue_size;
  88         size_t                  cmnd_capsule_len;
  89         struct nvme_rdma_ctrl   *ctrl;
  90         struct nvme_rdma_device *device;
  91         struct ib_cq            *ib_cq;
  92         struct ib_qp            *qp;
  93
  94         unsigned long           flags;
  95         struct rdma_cm_id       *cm_id;
  96         int                     cm_error;
  97         struct completion       cm_done;
  98         bool                    pi_support;
  99         int                     cq_size;
 100 };
 101
 102 struct nvme_rdma_ctrl {
 103         /* read only in the hot path */
 104         struct nvme_rdma_queue  *queues;
 105
 106         /* other member variables */
 107         struct blk_mq_tag_set   tag_set;
 108         struct work_struct      err_work;
 109
 110         struct nvme_rdma_qe     async_event_sqe;
 111
 112         struct delayed_work     reconnect_work;
 113
 114         struct list_head        list;
 115
 116         struct blk_mq_tag_set   admin_tag_set;
 117         struct nvme_rdma_device *device;
 118
 119         u32                     max_fr_pages;
 120
 121         struct sockaddr_storage addr;
 122         struct sockaddr_storage src_addr;
 123
 124         struct nvme_ctrl        ctrl;
 125         bool                    use_inline_data;
 126         u32                     io_queues[HCTX_MAX_TYPES];
 127 };
 128
 129 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 130 {
 131         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 132 }
 133
 134 static LIST_HEAD(device_list);
 135 static DEFINE_MUTEX(device_list_mutex);
 136
 137 static LIST_HEAD(nvme_rdma_ctrl_list);
 138 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 139
 140 /*
 141  * Disabling this option makes small I/O goes faster, but is fundamentally
 142  * unsafe.  With it turned off we will have to register a global rkey that
 143  * allows read and write access to all physical memory.
 144  */
 145 static bool register_always = true;
 146 module_param(register_always, bool, 0444);
 147 MODULE_PARM_DESC(register_always,
 148          "Use memory registration even for contiguous memory regions");
 149
 150 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 151                 struct rdma_cm_event *event);
 152 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 153 static void nvme_rdma_complete_rq(struct request *rq);
 154
 155 static const struct blk_mq_ops nvme_rdma_mq_ops;
 156 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 157
 158 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 159 {
 160         return queue - queue->ctrl->queues;
 161 }
 162
 163 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
 164 {
 165         return nvme_rdma_queue_idx(queue) >
 166                 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
 167                 queue->ctrl->io_queues[HCTX_TYPE_READ];
 168 }
 169
 170 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 171 {
 172         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 173 }
 174
 175 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 176                 size_t capsule_size, enum dma_data_direction dir)
 177 {
 178         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 179         kfree(qe->data);
 180 }
 181
 182 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 183                 size_t capsule_size, enum dma_data_direction dir)
 184 {
 185         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 186         if (!qe->data)
 187                 return -ENOMEM;
 188
 189         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 190         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 191                 kfree(qe->data);
 192                 qe->data = NULL;
 193                 return -ENOMEM;
 194         }
 195
 196         return 0;
 197 }
 198
 199 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 200                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 201                 size_t capsule_size, enum dma_data_direction dir)
 202 {
 203         int i;
 204
 205         for (i = 0; i < ib_queue_size; i++)
 206                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 207         kfree(ring);
 208 }
 209
 210 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 211                 size_t ib_queue_size, size_t capsule_size,
 212                 enum dma_data_direction dir)
 213 {
 214         struct nvme_rdma_qe *ring;
 215         int i;
 216
 217         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 218         if (!ring)
 219                 return NULL;
 220
 221         /*
 222          * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
 223          * lifetime. It's safe, since any chage in the underlying RDMA device
 224          * will issue error recovery and queue re-creation.
 225          */
 226         for (i = 0; i < ib_queue_size; i++) {
 227                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 228                         goto out_free_ring;
 229         }
 230
 231         return ring;
 232
 233 out_free_ring:
 234         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 235         return NULL;
 236 }
 237
 238 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 239 {
 240         pr_debug("QP event %s (%d)\n",
 241                  ib_event_msg(event->event), event->event);
 242
 243 }
 244
 245 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 246 {
 247         int ret;
 248
 249         ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
 250                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 251         if (ret < 0)
 252                 return ret;
 253         if (ret == 0)
 254                 return -ETIMEDOUT;
 255         WARN_ON_ONCE(queue->cm_error > 0);
 256         return queue->cm_error;
 257 }
 258
 259 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 260 {
 261         struct nvme_rdma_device *dev = queue->device;
 262         struct ib_qp_init_attr init_attr;
 263         int ret;
 264
 265         memset(&init_attr, 0, sizeof(init_attr));
 266         init_attr.event_handler = nvme_rdma_qp_event;
 267         /* +1 for drain */
 268         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 269         /* +1 for drain */
 270         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 271         init_attr.cap.max_recv_sge = 1;
 272         init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
 273         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 274         init_attr.qp_type = IB_QPT_RC;
 275         init_attr.send_cq = queue->ib_cq;
 276         init_attr.recv_cq = queue->ib_cq;
 277         if (queue->pi_support)
 278                 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
 279         init_attr.qp_context = queue;
 280
 281         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 282
 283         queue->qp = queue->cm_id->qp;
 284         return ret;
 285 }
 286
 287 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 288                 struct request *rq, unsigned int hctx_idx)
 289 {
 290         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 291
 292         kfree(req->sqe.data);
 293 }
 294
 295 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 296                 struct request *rq, unsigned int hctx_idx,
 297                 unsigned int numa_node)
 298 {
 299         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 300         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 301         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 302         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 303
 304         nvme_req(rq)->ctrl = &ctrl->ctrl;
 305         req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
 306         if (!req->sqe.data)
 307                 return -ENOMEM;
 308
 309         /* metadata nvme_rdma_sgl struct is located after command's data SGL */
 310         if (queue->pi_support)
 311                 req->metadata_sgl = (void *)nvme_req(rq) +
 312                         sizeof(struct nvme_rdma_request) +
 313                         NVME_RDMA_DATA_SGL_SIZE;
 314
 315         req->queue = queue;
 316
 317         return 0;
 318 }
 319
 320 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 321                 unsigned int hctx_idx)
 322 {
 323         struct nvme_rdma_ctrl *ctrl = data;
 324         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 325
 326         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 327
 328         hctx->driver_data = queue;
 329         return 0;
 330 }
 331
 332 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 333                 unsigned int hctx_idx)
 334 {
 335         struct nvme_rdma_ctrl *ctrl = data;
 336         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 337
 338         BUG_ON(hctx_idx != 0);
 339
 340         hctx->driver_data = queue;
 341         return 0;
 342 }
 343
 344 static void nvme_rdma_free_dev(struct kref *ref)
 345 {
 346         struct nvme_rdma_device *ndev =
 347                 container_of(ref, struct nvme_rdma_device, ref);
 348
 349         mutex_lock(&device_list_mutex);
 350         list_del(&ndev->entry);
 351         mutex_unlock(&device_list_mutex);
 352
 353         ib_dealloc_pd(ndev->pd);
 354         kfree(ndev);
 355 }
 356
 357 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 358 {
 359         kref_put(&dev->ref, nvme_rdma_free_dev);
 360 }
 361
 362 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 363 {
 364         return kref_get_unless_zero(&dev->ref);
 365 }
 366
 367 static struct nvme_rdma_device *
 368 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 369 {
 370         struct nvme_rdma_device *ndev;
 371
 372         mutex_lock(&device_list_mutex);
 373         list_for_each_entry(ndev, &device_list, entry) {
 374                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 375                     nvme_rdma_dev_get(ndev))
 376                         goto out_unlock;
 377         }
 378
 379         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 380         if (!ndev)
 381                 goto out_err;
 382
 383         ndev->dev = cm_id->device;
 384         kref_init(&ndev->ref);
 385
 386         ndev->pd = ib_alloc_pd(ndev->dev,
 387                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 388         if (IS_ERR(ndev->pd))
 389                 goto out_free_dev;
 390
 391         if (!(ndev->dev->attrs.device_cap_flags &
 392               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 393                 dev_err(&ndev->dev->dev,
 394                         "Memory registrations not supported.\n");
 395                 goto out_free_pd;
 396         }
 397
 398         ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
 399                                         ndev->dev->attrs.max_send_sge - 1);
 400         list_add(&ndev->entry, &device_list);
 401 out_unlock:
 402         mutex_unlock(&device_list_mutex);
 403         return ndev;
 404
 405 out_free_pd:
 406         ib_dealloc_pd(ndev->pd);
 407 out_free_dev:
 408         kfree(ndev);
 409 out_err:
 410         mutex_unlock(&device_list_mutex);
 411         return NULL;
 412 }
 413
 414 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
 415 {
 416         if (nvme_rdma_poll_queue(queue))
 417                 ib_free_cq(queue->ib_cq);
 418         else
 419                 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
 420 }
 421
 422 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 423 {
 424         struct nvme_rdma_device *dev;
 425         struct ib_device *ibdev;
 426
 427         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 428                 return;
 429
 430         dev = queue->device;
 431         ibdev = dev->dev;
 432
 433         if (queue->pi_support)
 434                 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
 435         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 436
 437         /*
 438          * The cm_id object might have been destroyed during RDMA connection
 439          * establishment error flow to avoid getting other cma events, thus
 440          * the destruction of the QP shouldn't use rdma_cm API.
 441          */
 442         ib_destroy_qp(queue->qp);
 443         nvme_rdma_free_cq(queue);
 444
 445         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 446                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 447
 448         nvme_rdma_dev_put(dev);
 449 }
 450
 451 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
 452 {
 453         u32 max_page_list_len;
 454
 455         if (pi_support)
 456                 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
 457         else
 458                 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
 459
 460         return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
 461 }
 462
 463 static int nvme_rdma_create_cq(struct ib_device *ibdev,
 464                 struct nvme_rdma_queue *queue)
 465 {
 466         int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
 467         enum ib_poll_context poll_ctx;
 468
 469         /*
 470          * Spread I/O queues completion vectors according their queue index.
 471          * Admin queues can always go on completion vector 0.
 472          */
 473         comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
 474
 475         /* Polling queues need direct cq polling context */
 476         if (nvme_rdma_poll_queue(queue)) {
 477                 poll_ctx = IB_POLL_DIRECT;
 478                 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
 479                                            comp_vector, poll_ctx);
 480         } else {
 481                 poll_ctx = IB_POLL_SOFTIRQ;
 482                 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
 483                                               comp_vector, poll_ctx);
 484         }
 485
 486         if (IS_ERR(queue->ib_cq)) {
 487                 ret = PTR_ERR(queue->ib_cq);
 488                 return ret;
 489         }
 490
 491         return 0;
 492 }
 493
 494 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 495 {
 496         struct ib_device *ibdev;
 497         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 498         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 499         int ret, pages_per_mr;
 500
 501         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 502         if (!queue->device) {
 503                 dev_err(queue->cm_id->device->dev.parent,
 504                         "no client data found!\n");
 505                 return -ECONNREFUSED;
 506         }
 507         ibdev = queue->device->dev;
 508
 509         /* +1 for ib_stop_cq */
 510         queue->cq_size = cq_factor * queue->queue_size + 1;
 511
 512         ret = nvme_rdma_create_cq(ibdev, queue);
 513         if (ret)
 514                 goto out_put_dev;
 515
 516         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 517         if (ret)
 518                 goto out_destroy_ib_cq;
 519
 520         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 521                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 522         if (!queue->rsp_ring) {
 523                 ret = -ENOMEM;
 524                 goto out_destroy_qp;
 525         }
 526
 527         /*
 528          * Currently we don't use SG_GAPS MR's so if the first entry is
 529          * misaligned we'll end up using two entries for a single data page,
 530          * so one additional entry is required.
 531          */
 532         pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
 533         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 534                               queue->queue_size,
 535                               IB_MR_TYPE_MEM_REG,
 536                               pages_per_mr, 0);
 537         if (ret) {
 538                 dev_err(queue->ctrl->ctrl.device,
 539                         "failed to initialize MR pool sized %d for QID %d\n",
 540                         queue->queue_size, nvme_rdma_queue_idx(queue));
 541                 goto out_destroy_ring;
 542         }
 543
 544         if (queue->pi_support) {
 545                 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
 546                                       queue->queue_size, IB_MR_TYPE_INTEGRITY,
 547                                       pages_per_mr, pages_per_mr);
 548                 if (ret) {
 549                         dev_err(queue->ctrl->ctrl.device,
 550                                 "failed to initialize PI MR pool sized %d for QID %d\n",
 551                                 queue->queue_size, nvme_rdma_queue_idx(queue));
 552                         goto out_destroy_mr_pool;
 553                 }
 554         }
 555
 556         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 557
 558         return 0;
 559
 560 out_destroy_mr_pool:
 561         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 562 out_destroy_ring:
 563         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 564                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 565 out_destroy_qp:
 566         rdma_destroy_qp(queue->cm_id);
 567 out_destroy_ib_cq:
 568         nvme_rdma_free_cq(queue);
 569 out_put_dev:
 570         nvme_rdma_dev_put(queue->device);
 571         return ret;
 572 }
 573
 574 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 575                 int idx, size_t queue_size)
 576 {
 577         struct nvme_rdma_queue *queue;
 578         struct sockaddr *src_addr = NULL;
 579         int ret;
 580
 581         queue = &ctrl->queues[idx];
 582         queue->ctrl = ctrl;
 583         if (idx && ctrl->ctrl.max_integrity_segments)
 584                 queue->pi_support = true;
 585         else
 586                 queue->pi_support = false;
 587         init_completion(&queue->cm_done);
 588
 589         if (idx > 0)
 590                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 591         else
 592                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 593
 594         queue->queue_size = queue_size;
 595
 596         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 597                         RDMA_PS_TCP, IB_QPT_RC);
 598         if (IS_ERR(queue->cm_id)) {
 599                 dev_info(ctrl->ctrl.device,
 600                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 601                 return PTR_ERR(queue->cm_id);
 602         }
 603
 604         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 605                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 606
 607         queue->cm_error = -ETIMEDOUT;
 608         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 609                         (struct sockaddr *)&ctrl->addr,
 610                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 611         if (ret) {
 612                 dev_info(ctrl->ctrl.device,
 613                         "rdma_resolve_addr failed (%d).\n", ret);
 614                 goto out_destroy_cm_id;
 615         }
 616
 617         ret = nvme_rdma_wait_for_cm(queue);
 618         if (ret) {
 619                 dev_info(ctrl->ctrl.device,
 620                         "rdma connection establishment failed (%d)\n", ret);
 621                 goto out_destroy_cm_id;
 622         }
 623
 624         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 625
 626         return 0;
 627
 628 out_destroy_cm_id:
 629         rdma_destroy_id(queue->cm_id);
 630         nvme_rdma_destroy_queue_ib(queue);
 631         return ret;
 632 }
 633
 634 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 635 {
 636         rdma_disconnect(queue->cm_id);
 637         ib_drain_qp(queue->qp);
 638 }
 639
 640 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 641 {
 642         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 643                 return;
 644         __nvme_rdma_stop_queue(queue);
 645 }
 646
 647 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 648 {
 649         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 650                 return;
 651
 652         nvme_rdma_destroy_queue_ib(queue);
 653         rdma_destroy_id(queue->cm_id);
 654 }
 655
 656 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 657 {
 658         int i;
 659
 660         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 661                 nvme_rdma_free_queue(&ctrl->queues[i]);
 662 }
 663
 664 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 665 {
 666         int i;
 667
 668         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 669                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 670 }
 671
 672 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 673 {
 674         struct nvme_rdma_queue *queue = &ctrl->queues[idx];
 675         bool poll = nvme_rdma_poll_queue(queue);
 676         int ret;
 677
 678         if (idx)
 679                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
 680         else
 681                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 682
 683         if (!ret) {
 684                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 685         } else {
 686                 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 687                         __nvme_rdma_stop_queue(queue);
 688                 dev_info(ctrl->ctrl.device,
 689                         "failed to connect queue: %d ret=%d\n", idx, ret);
 690         }
 691         return ret;
 692 }
 693
 694 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 695 {
 696         int i, ret = 0;
 697
 698         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 699                 ret = nvme_rdma_start_queue(ctrl, i);
 700                 if (ret)
 701                         goto out_stop_queues;
 702         }
 703
 704         return 0;
 705
 706 out_stop_queues:
 707         for (i--; i >= 1; i--)
 708                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 709         return ret;
 710 }
 711
 712 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 713 {
 714         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 715         struct ib_device *ibdev = ctrl->device->dev;
 716         unsigned int nr_io_queues, nr_default_queues;
 717         unsigned int nr_read_queues, nr_poll_queues;
 718         int i, ret;
 719
 720         nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
 721                                 min(opts->nr_io_queues, num_online_cpus()));
 722         nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
 723                                 min(opts->nr_write_queues, num_online_cpus()));
 724         nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
 725         nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
 726
 727         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 728         if (ret)
 729                 return ret;
 730
 731         ctrl->ctrl.queue_count = nr_io_queues + 1;
 732         if (ctrl->ctrl.queue_count < 2)
 733                 return 0;
 734
 735         dev_info(ctrl->ctrl.device,
 736                 "creating %d I/O queues.\n", nr_io_queues);
 737
 738         if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
 739                 /*
 740                  * separate read/write queues
 741                  * hand out dedicated default queues only after we have
 742                  * sufficient read queues.
 743                  */
 744                 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
 745                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
 746                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 747                         min(nr_default_queues, nr_io_queues);
 748                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 749         } else {
 750                 /*
 751                  * shared read/write queues
 752                  * either no write queues were requested, or we don't have
 753                  * sufficient queue count to have dedicated default queues.
 754                  */
 755                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 756                         min(nr_read_queues, nr_io_queues);
 757                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 758         }
 759
 760         if (opts->nr_poll_queues && nr_io_queues) {
 761                 /* map dedicated poll queues only if we have queues left */
 762                 ctrl->io_queues[HCTX_TYPE_POLL] =
 763                         min(nr_poll_queues, nr_io_queues);
 764         }
 765
 766         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 767                 ret = nvme_rdma_alloc_queue(ctrl, i,
 768                                 ctrl->ctrl.sqsize + 1);
 769                 if (ret)
 770                         goto out_free_queues;
 771         }
 772
 773         return 0;
 774
 775 out_free_queues:
 776         for (i--; i >= 1; i--)
 777                 nvme_rdma_free_queue(&ctrl->queues[i]);
 778
 779         return ret;
 780 }
 781
 782 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 783                 bool admin)
 784 {
 785         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 786         struct blk_mq_tag_set *set;
 787         int ret;
 788
 789         if (admin) {
 790                 set = &ctrl->admin_tag_set;
 791                 memset(set, 0, sizeof(*set));
 792                 set->ops = &nvme_rdma_admin_mq_ops;
 793                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 794                 set->reserved_tags = 2; /* connect + keep-alive */
 795                 set->numa_node = nctrl->numa_node;
 796                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 797                                 NVME_RDMA_DATA_SGL_SIZE;
 798                 set->driver_data = ctrl;
 799                 set->nr_hw_queues = 1;
 800                 set->timeout = NVME_ADMIN_TIMEOUT;
 801                 set->flags = BLK_MQ_F_NO_SCHED;
 802         } else {
 803                 set = &ctrl->tag_set;
 804                 memset(set, 0, sizeof(*set));
 805                 set->ops = &nvme_rdma_mq_ops;
 806                 set->queue_depth = nctrl->sqsize + 1;
 807                 set->reserved_tags = 1; /* fabric connect */
 808                 set->numa_node = nctrl->numa_node;
 809                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 810                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 811                                 NVME_RDMA_DATA_SGL_SIZE;
 812                 if (nctrl->max_integrity_segments)
 813                         set->cmd_size += sizeof(struct nvme_rdma_sgl) +
 814                                          NVME_RDMA_METADATA_SGL_SIZE;
 815                 set->driver_data = ctrl;
 816                 set->nr_hw_queues = nctrl->queue_count - 1;
 817                 set->timeout = NVME_IO_TIMEOUT;
 818                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
 819         }
 820
 821         ret = blk_mq_alloc_tag_set(set);
 822         if (ret)
 823                 return ERR_PTR(ret);
 824
 825         return set;
 826 }
 827
 828 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 829                 bool remove)
 830 {
 831         if (remove) {
 832                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 833                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 834                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 835         }
 836         if (ctrl->async_event_sqe.data) {
 837                 cancel_work_sync(&ctrl->ctrl.async_event_work);
 838                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 839                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 840                 ctrl->async_event_sqe.data = NULL;
 841         }
 842         nvme_rdma_free_queue(&ctrl->queues[0]);
 843 }
 844
 845 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 846                 bool new)
 847 {
 848         bool pi_capable = false;
 849         int error;
 850
 851         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 852         if (error)
 853                 return error;
 854
 855         ctrl->device = ctrl->queues[0].device;
 856         ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
 857
 858         /* T10-PI support */
 859         if (ctrl->device->dev->attrs.device_cap_flags &
 860             IB_DEVICE_INTEGRITY_HANDOVER)
 861                 pi_capable = true;
 862
 863         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
 864                                                         pi_capable);
 865
 866         /*
 867          * Bind the async event SQE DMA mapping to the admin queue lifetime.
 868          * It's safe, since any chage in the underlying RDMA device will issue
 869          * error recovery and queue re-creation.
 870          */
 871         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 872                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 873         if (error)
 874                 goto out_free_queue;
 875
 876         if (new) {
 877                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 878                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 879                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 880                         goto out_free_async_qe;
 881                 }
 882
 883                 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 884                 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 885                         error = PTR_ERR(ctrl->ctrl.fabrics_q);
 886                         goto out_free_tagset;
 887                 }
 888
 889                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 890                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 891                         error = PTR_ERR(ctrl->ctrl.admin_q);
 892                         goto out_cleanup_fabrics_q;
 893                 }
 894         }
 895
 896         error = nvme_rdma_start_queue(ctrl, 0);
 897         if (error)
 898                 goto out_cleanup_queue;
 899
 900         error = nvme_enable_ctrl(&ctrl->ctrl);
 901         if (error)
 902                 goto out_stop_queue;
 903
 904         ctrl->ctrl.max_segments = ctrl->max_fr_pages;
 905         ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
 906         if (pi_capable)
 907                 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
 908         else
 909                 ctrl->ctrl.max_integrity_segments = 0;
 910
 911         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 912
 913         error = nvme_init_identify(&ctrl->ctrl);
 914         if (error)
 915                 goto out_stop_queue;
 916
 917         return 0;
 918
 919 out_stop_queue:
 920         nvme_rdma_stop_queue(&ctrl->queues[0]);
 921 out_cleanup_queue:
 922         if (new)
 923                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 924 out_cleanup_fabrics_q:
 925         if (new)
 926                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 927 out_free_tagset:
 928         if (new)
 929                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 930 out_free_async_qe:
 931         if (ctrl->async_event_sqe.data) {
 932                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 933                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 934                 ctrl->async_event_sqe.data = NULL;
 935         }
 936 out_free_queue:
 937         nvme_rdma_free_queue(&ctrl->queues[0]);
 938         return error;
 939 }
 940
 941 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 942                 bool remove)
 943 {
 944         if (remove) {
 945                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 946                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 947         }
 948         nvme_rdma_free_io_queues(ctrl);
 949 }
 950
 951 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 952 {
 953         int ret;
 954
 955         ret = nvme_rdma_alloc_io_queues(ctrl);
 956         if (ret)
 957                 return ret;
 958
 959         if (new) {
 960                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 961                 if (IS_ERR(ctrl->ctrl.tagset)) {
 962                         ret = PTR_ERR(ctrl->ctrl.tagset);
 963                         goto out_free_io_queues;
 964                 }
 965
 966                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 967                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 968                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 969                         goto out_free_tag_set;
 970                 }
 971         }
 972
 973         ret = nvme_rdma_start_io_queues(ctrl);
 974         if (ret)
 975                 goto out_cleanup_connect_q;
 976
 977         if (!new) {
 978                 nvme_start_queues(&ctrl->ctrl);
 979                 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
 980                         /*
 981                          * If we timed out waiting for freeze we are likely to
 982                          * be stuck.  Fail the controller initialization just
 983                          * to be safe.
 984                          */
 985                         ret = -ENODEV;
 986                         goto out_wait_freeze_timed_out;
 987                 }
 988                 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
 989                         ctrl->ctrl.queue_count - 1);
 990                 nvme_unfreeze(&ctrl->ctrl);
 991         }
 992
 993         return 0;
 994
 995 out_wait_freeze_timed_out:
 996         nvme_stop_queues(&ctrl->ctrl);
 997         nvme_rdma_stop_io_queues(ctrl);
 998 out_cleanup_connect_q:
 999         if (new)
1000                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1001 out_free_tag_set:
1002         if (new)
1003                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1004 out_free_io_queues:
1005         nvme_rdma_free_io_queues(ctrl);
1006         return ret;
1007 }
1008
1009 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1010                 bool remove)
1011 {
1012         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1013         blk_sync_queue(ctrl->ctrl.admin_q);
1014         nvme_rdma_stop_queue(&ctrl->queues[0]);
1015         if (ctrl->ctrl.admin_tagset) {
1016                 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
1017                         nvme_cancel_request, &ctrl->ctrl);
1018                 blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
1019         }
1020         if (remove)
1021                 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1022         nvme_rdma_destroy_admin_queue(ctrl, remove);
1023 }
1024
1025 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1026                 bool remove)
1027 {
1028         if (ctrl->ctrl.queue_count > 1) {
1029                 nvme_start_freeze(&ctrl->ctrl);
1030                 nvme_stop_queues(&ctrl->ctrl);
1031                 nvme_sync_io_queues(&ctrl->ctrl);
1032                 nvme_rdma_stop_io_queues(ctrl);
1033                 if (ctrl->ctrl.tagset) {
1034                         blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
1035                                 nvme_cancel_request, &ctrl->ctrl);
1036                         blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
1037                 }
1038                 if (remove)
1039                         nvme_start_queues(&ctrl->ctrl);
1040                 nvme_rdma_destroy_io_queues(ctrl, remove);
1041         }
1042 }
1043
1044 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1045 {
1046         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1047
1048         if (list_empty(&ctrl->list))
1049                 goto free_ctrl;
1050
1051         mutex_lock(&nvme_rdma_ctrl_mutex);
1052         list_del(&ctrl->list);
1053         mutex_unlock(&nvme_rdma_ctrl_mutex);
1054
1055         nvmf_free_options(nctrl->opts);
1056 free_ctrl:
1057         kfree(ctrl->queues);
1058         kfree(ctrl);
1059 }
1060
1061 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1062 {
1063         /* If we are resetting/deleting then do nothing */
1064         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1065                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1066                         ctrl->ctrl.state == NVME_CTRL_LIVE);
1067                 return;
1068         }
1069
1070         if (nvmf_should_reconnect(&ctrl->ctrl)) {
1071                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1072                         ctrl->ctrl.opts->reconnect_delay);
1073                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1074                                 ctrl->ctrl.opts->reconnect_delay * HZ);
1075         } else {
1076                 nvme_delete_ctrl(&ctrl->ctrl);
1077         }
1078 }
1079
1080 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1081 {
1082         int ret = -EINVAL;
1083         bool changed;
1084
1085         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1086         if (ret)
1087                 return ret;
1088
1089         if (ctrl->ctrl.icdoff) {
1090                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1091                 goto destroy_admin;
1092         }
1093
1094         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1095                 dev_err(ctrl->ctrl.device,
1096                         "Mandatory keyed sgls are not supported!\n");
1097                 goto destroy_admin;
1098         }
1099
1100         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1101                 dev_warn(ctrl->ctrl.device,
1102                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1103                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1104         }
1105
1106         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1107                 dev_warn(ctrl->ctrl.device,
1108                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1109                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1110                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1111         }
1112
1113         if (ctrl->ctrl.sgls & (1 << 20))
1114                 ctrl->use_inline_data = true;
1115
1116         if (ctrl->ctrl.queue_count > 1) {
1117                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1118                 if (ret)
1119                         goto destroy_admin;
1120         }
1121
1122         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1123         if (!changed) {
1124                 /*
1125                  * state change failure is ok if we started ctrl delete,
1126                  * unless we're during creation of a new controller to
1127                  * avoid races with teardown flow.
1128                  */
1129                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1130                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1131                 WARN_ON_ONCE(new);
1132                 ret = -EINVAL;
1133                 goto destroy_io;
1134         }
1135
1136         nvme_start_ctrl(&ctrl->ctrl);
1137         return 0;
1138
1139 destroy_io:
1140         if (ctrl->ctrl.queue_count > 1)
1141                 nvme_rdma_destroy_io_queues(ctrl, new);
1142 destroy_admin:
1143         nvme_rdma_stop_queue(&ctrl->queues[0]);
1144         nvme_rdma_destroy_admin_queue(ctrl, new);
1145         return ret;
1146 }
1147
1148 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1149 {
1150         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1151                         struct nvme_rdma_ctrl, reconnect_work);
1152
1153         ++ctrl->ctrl.nr_reconnects;
1154
1155         if (nvme_rdma_setup_ctrl(ctrl, false))
1156                 goto requeue;
1157
1158         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1159                         ctrl->ctrl.nr_reconnects);
1160
1161         ctrl->ctrl.nr_reconnects = 0;
1162
1163         return;
1164
1165 requeue:
1166         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1167                         ctrl->ctrl.nr_reconnects);
1168         nvme_rdma_reconnect_or_remove(ctrl);
1169 }
1170
1171 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1172 {
1173         struct nvme_rdma_ctrl *ctrl = container_of(work,
1174                         struct nvme_rdma_ctrl, err_work);
1175
1176         nvme_stop_keep_alive(&ctrl->ctrl);
1177         nvme_rdma_teardown_io_queues(ctrl, false);
1178         nvme_start_queues(&ctrl->ctrl);
1179         nvme_rdma_teardown_admin_queue(ctrl, false);
1180         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1181
1182         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1183                 /* state change failure is ok if we started ctrl delete */
1184                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1185                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1186                 return;
1187         }
1188
1189         nvme_rdma_reconnect_or_remove(ctrl);
1190 }
1191
1192 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1193 {
1194         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1195                 return;
1196
1197         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1198         queue_work(nvme_reset_wq, &ctrl->err_work);
1199 }
1200
1201 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1202 {
1203         struct request *rq = blk_mq_rq_from_pdu(req);
1204
1205         if (!refcount_dec_and_test(&req->ref))
1206                 return;
1207         if (!nvme_try_complete_req(rq, req->status, req->result))
1208                 nvme_rdma_complete_rq(rq);
1209 }
1210
1211 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1212                 const char *op)
1213 {
1214         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1215         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1216
1217         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1218                 dev_info(ctrl->ctrl.device,
1219                              "%s for CQE 0x%p failed with status %s (%d)\n",
1220                              op, wc->wr_cqe,
1221                              ib_wc_status_msg(wc->status), wc->status);
1222         nvme_rdma_error_recovery(ctrl);
1223 }
1224
1225 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1226 {
1227         if (unlikely(wc->status != IB_WC_SUCCESS))
1228                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1229 }
1230
1231 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1232 {
1233         struct nvme_rdma_request *req =
1234                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1235
1236         if (unlikely(wc->status != IB_WC_SUCCESS))
1237                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1238         else
1239                 nvme_rdma_end_request(req);
1240 }
1241
1242 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1243                 struct nvme_rdma_request *req)
1244 {
1245         struct ib_send_wr wr = {
1246                 .opcode             = IB_WR_LOCAL_INV,
1247                 .next               = NULL,
1248                 .num_sge            = 0,
1249                 .send_flags         = IB_SEND_SIGNALED,
1250                 .ex.invalidate_rkey = req->mr->rkey,
1251         };
1252
1253         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1254         wr.wr_cqe = &req->reg_cqe;
1255
1256         return ib_post_send(queue->qp, &wr, NULL);
1257 }
1258
1259 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1260                 struct request *rq)
1261 {
1262         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1263         struct nvme_rdma_device *dev = queue->device;
1264         struct ib_device *ibdev = dev->dev;
1265         struct list_head *pool = &queue->qp->rdma_mrs;
1266
1267         if (!blk_rq_nr_phys_segments(rq))
1268                 return;
1269
1270         if (blk_integrity_rq(rq)) {
1271                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1272                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1273                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1274                                       NVME_INLINE_METADATA_SG_CNT);
1275         }
1276
1277         if (req->use_sig_mr)
1278                 pool = &queue->qp->sig_mrs;
1279
1280         if (req->mr) {
1281                 ib_mr_pool_put(queue->qp, pool, req->mr);
1282                 req->mr = NULL;
1283         }
1284
1285         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1286                         rq_dma_dir(rq));
1287         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1288 }
1289
1290 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1291 {
1292         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1293
1294         sg->addr = 0;
1295         put_unaligned_le24(0, sg->length);
1296         put_unaligned_le32(0, sg->key);
1297         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1298         return 0;
1299 }
1300
1301 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1302                 struct nvme_rdma_request *req, struct nvme_command *c,
1303                 int count)
1304 {
1305         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1306         struct scatterlist *sgl = req->data_sgl.sg_table.sgl;
1307         struct ib_sge *sge = &req->sge[1];
1308         u32 len = 0;
1309         int i;
1310
1311         for (i = 0; i < count; i++, sgl++, sge++) {
1312                 sge->addr = sg_dma_address(sgl);
1313                 sge->length = sg_dma_len(sgl);
1314                 sge->lkey = queue->device->pd->local_dma_lkey;
1315                 len += sge->length;
1316         }
1317
1318         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1319         sg->length = cpu_to_le32(len);
1320         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1321
1322         req->num_sge += count;
1323         return 0;
1324 }
1325
1326 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1327                 struct nvme_rdma_request *req, struct nvme_command *c)
1328 {
1329         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1330
1331         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1332         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1333         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1334         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1335         return 0;
1336 }
1337
1338 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1339                 struct nvme_rdma_request *req, struct nvme_command *c,
1340                 int count)
1341 {
1342         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1343         int nr;
1344
1345         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1346         if (WARN_ON_ONCE(!req->mr))
1347                 return -EAGAIN;
1348
1349         /*
1350          * Align the MR to a 4K page size to match the ctrl page size and
1351          * the block virtual boundary.
1352          */
1353         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1354                           SZ_4K);
1355         if (unlikely(nr < count)) {
1356                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1357                 req->mr = NULL;
1358                 if (nr < 0)
1359                         return nr;
1360                 return -EINVAL;
1361         }
1362
1363         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1364
1365         req->reg_cqe.done = nvme_rdma_memreg_done;
1366         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1367         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1368         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1369         req->reg_wr.wr.num_sge = 0;
1370         req->reg_wr.mr = req->mr;
1371         req->reg_wr.key = req->mr->rkey;
1372         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1373                              IB_ACCESS_REMOTE_READ |
1374                              IB_ACCESS_REMOTE_WRITE;
1375
1376         sg->addr = cpu_to_le64(req->mr->iova);
1377         put_unaligned_le24(req->mr->length, sg->length);
1378         put_unaligned_le32(req->mr->rkey, sg->key);
1379         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1380                         NVME_SGL_FMT_INVALIDATE;
1381
1382         return 0;
1383 }
1384
1385 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1386                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1387                 u16 control, u8 pi_type)
1388 {
1389         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1390         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1391         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1392         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1393         if (control & NVME_RW_PRINFO_PRCHK_REF)
1394                 domain->sig.dif.ref_remap = true;
1395
1396         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1397         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1398         domain->sig.dif.app_escape = true;
1399         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1400                 domain->sig.dif.ref_escape = true;
1401 }
1402
1403 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1404                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1405                 u8 pi_type)
1406 {
1407         u16 control = le16_to_cpu(cmd->rw.control);
1408
1409         memset(sig_attrs, 0, sizeof(*sig_attrs));
1410         if (control & NVME_RW_PRINFO_PRACT) {
1411                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1412                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1413                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1414                                          pi_type);
1415                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1416                 control &= ~NVME_RW_PRINFO_PRACT;
1417                 cmd->rw.control = cpu_to_le16(control);
1418         } else {
1419                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1420                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1421                                          pi_type);
1422                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1423                                          pi_type);
1424         }
1425 }
1426
1427 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1428 {
1429         *mask = 0;
1430         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1431                 *mask |= IB_SIG_CHECK_REFTAG;
1432         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1433                 *mask |= IB_SIG_CHECK_GUARD;
1434 }
1435
1436 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1437 {
1438         if (unlikely(wc->status != IB_WC_SUCCESS))
1439                 nvme_rdma_wr_error(cq, wc, "SIG");
1440 }
1441
1442 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1443                 struct nvme_rdma_request *req, struct nvme_command *c,
1444                 int count, int pi_count)
1445 {
1446         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1447         struct ib_reg_wr *wr = &req->reg_wr;
1448         struct request *rq = blk_mq_rq_from_pdu(req);
1449         struct nvme_ns *ns = rq->q->queuedata;
1450         struct bio *bio = rq->bio;
1451         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1452         int nr;
1453
1454         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1455         if (WARN_ON_ONCE(!req->mr))
1456                 return -EAGAIN;
1457
1458         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1459                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1460                              SZ_4K);
1461         if (unlikely(nr))
1462                 goto mr_put;
1463
1464         nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_disk), c,
1465                                 req->mr->sig_attrs, ns->pi_type);
1466         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1467
1468         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1469
1470         req->reg_cqe.done = nvme_rdma_sig_done;
1471         memset(wr, 0, sizeof(*wr));
1472         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1473         wr->wr.wr_cqe = &req->reg_cqe;
1474         wr->wr.num_sge = 0;
1475         wr->wr.send_flags = 0;
1476         wr->mr = req->mr;
1477         wr->key = req->mr->rkey;
1478         wr->access = IB_ACCESS_LOCAL_WRITE |
1479                      IB_ACCESS_REMOTE_READ |
1480                      IB_ACCESS_REMOTE_WRITE;
1481
1482         sg->addr = cpu_to_le64(req->mr->iova);
1483         put_unaligned_le24(req->mr->length, sg->length);
1484         put_unaligned_le32(req->mr->rkey, sg->key);
1485         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1486
1487         return 0;
1488
1489 mr_put:
1490         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1491         req->mr = NULL;
1492         if (nr < 0)
1493                 return nr;
1494         return -EINVAL;
1495 }
1496
1497 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1498                 struct request *rq, struct nvme_command *c)
1499 {
1500         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1501         struct nvme_rdma_device *dev = queue->device;
1502         struct ib_device *ibdev = dev->dev;
1503         int pi_count = 0;
1504         int count, ret;
1505
1506         req->num_sge = 1;
1507         refcount_set(&req->ref, 2); /* send and recv completions */
1508
1509         c->common.flags |= NVME_CMD_SGL_METABUF;
1510
1511         if (!blk_rq_nr_phys_segments(rq))
1512                 return nvme_rdma_set_sg_null(c);
1513
1514         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1515         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1516                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1517                         NVME_INLINE_SG_CNT);
1518         if (ret)
1519                 return -ENOMEM;
1520
1521         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1522                                             req->data_sgl.sg_table.sgl);
1523
1524         count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1525                               req->data_sgl.nents, rq_dma_dir(rq));
1526         if (unlikely(count <= 0)) {
1527                 ret = -EIO;
1528                 goto out_free_table;
1529         }
1530
1531         if (blk_integrity_rq(rq)) {
1532                 req->metadata_sgl->sg_table.sgl =
1533                         (struct scatterlist *)(req->metadata_sgl + 1);
1534                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1535                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1536                                 req->metadata_sgl->sg_table.sgl,
1537                                 NVME_INLINE_METADATA_SG_CNT);
1538                 if (unlikely(ret)) {
1539                         ret = -ENOMEM;
1540                         goto out_unmap_sg;
1541                 }
1542
1543                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1544                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1545                 pi_count = ib_dma_map_sg(ibdev,
1546                                          req->metadata_sgl->sg_table.sgl,
1547                                          req->metadata_sgl->nents,
1548                                          rq_dma_dir(rq));
1549                 if (unlikely(pi_count <= 0)) {
1550                         ret = -EIO;
1551                         goto out_free_pi_table;
1552                 }
1553         }
1554
1555         if (req->use_sig_mr) {
1556                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1557                 goto out;
1558         }
1559
1560         if (count <= dev->num_inline_segments) {
1561                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1562                     queue->ctrl->use_inline_data &&
1563                     blk_rq_payload_bytes(rq) <=
1564                                 nvme_rdma_inline_data_size(queue)) {
1565                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1566                         goto out;
1567                 }
1568
1569                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1570                         ret = nvme_rdma_map_sg_single(queue, req, c);
1571                         goto out;
1572                 }
1573         }
1574
1575         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1576 out:
1577         if (unlikely(ret))
1578                 goto out_unmap_pi_sg;
1579
1580         return 0;
1581
1582 out_unmap_pi_sg:
1583         if (blk_integrity_rq(rq))
1584                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1585                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1586 out_free_pi_table:
1587         if (blk_integrity_rq(rq))
1588                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1589                                       NVME_INLINE_METADATA_SG_CNT);
1590 out_unmap_sg:
1591         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1592                         rq_dma_dir(rq));
1593 out_free_table:
1594         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1595         return ret;
1596 }
1597
1598 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1599 {
1600         struct nvme_rdma_qe *qe =
1601                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1602         struct nvme_rdma_request *req =
1603                 container_of(qe, struct nvme_rdma_request, sqe);
1604
1605         if (unlikely(wc->status != IB_WC_SUCCESS))
1606                 nvme_rdma_wr_error(cq, wc, "SEND");
1607         else
1608                 nvme_rdma_end_request(req);
1609 }
1610
1611 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1612                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1613                 struct ib_send_wr *first)
1614 {
1615         struct ib_send_wr wr;
1616         int ret;
1617
1618         sge->addr   = qe->dma;
1619         sge->length = sizeof(struct nvme_command);
1620         sge->lkey   = queue->device->pd->local_dma_lkey;
1621
1622         wr.next       = NULL;
1623         wr.wr_cqe     = &qe->cqe;
1624         wr.sg_list    = sge;
1625         wr.num_sge    = num_sge;
1626         wr.opcode     = IB_WR_SEND;
1627         wr.send_flags = IB_SEND_SIGNALED;
1628
1629         if (first)
1630                 first->next = &wr;
1631         else
1632                 first = &wr;
1633
1634         ret = ib_post_send(queue->qp, first, NULL);
1635         if (unlikely(ret)) {
1636                 dev_err(queue->ctrl->ctrl.device,
1637                              "%s failed with error code %d\n", __func__, ret);
1638         }
1639         return ret;
1640 }
1641
1642 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1643                 struct nvme_rdma_qe *qe)
1644 {
1645         struct ib_recv_wr wr;
1646         struct ib_sge list;
1647         int ret;
1648
1649         list.addr   = qe->dma;
1650         list.length = sizeof(struct nvme_completion);
1651         list.lkey   = queue->device->pd->local_dma_lkey;
1652
1653         qe->cqe.done = nvme_rdma_recv_done;
1654
1655         wr.next     = NULL;
1656         wr.wr_cqe   = &qe->cqe;
1657         wr.sg_list  = &list;
1658         wr.num_sge  = 1;
1659
1660         ret = ib_post_recv(queue->qp, &wr, NULL);
1661         if (unlikely(ret)) {
1662                 dev_err(queue->ctrl->ctrl.device,
1663                         "%s failed with error code %d\n", __func__, ret);
1664         }
1665         return ret;
1666 }
1667
1668 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1669 {
1670         u32 queue_idx = nvme_rdma_queue_idx(queue);
1671
1672         if (queue_idx == 0)
1673                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1674         return queue->ctrl->tag_set.tags[queue_idx - 1];
1675 }
1676
1677 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1678 {
1679         if (unlikely(wc->status != IB_WC_SUCCESS))
1680                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1681 }
1682
1683 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1684 {
1685         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1686         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1687         struct ib_device *dev = queue->device->dev;
1688         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1689         struct nvme_command *cmd = sqe->data;
1690         struct ib_sge sge;
1691         int ret;
1692
1693         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1694
1695         memset(cmd, 0, sizeof(*cmd));
1696         cmd->common.opcode = nvme_admin_async_event;
1697         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1698         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1699         nvme_rdma_set_sg_null(cmd);
1700
1701         sqe->cqe.done = nvme_rdma_async_done;
1702
1703         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1704                         DMA_TO_DEVICE);
1705
1706         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1707         WARN_ON_ONCE(ret);
1708 }
1709
1710 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1711                 struct nvme_completion *cqe, struct ib_wc *wc)
1712 {
1713         struct request *rq;
1714         struct nvme_rdma_request *req;
1715
1716         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1717         if (!rq) {
1718                 dev_err(queue->ctrl->ctrl.device,
1719                         "tag 0x%x on QP %#x not found\n",
1720                         cqe->command_id, queue->qp->qp_num);
1721                 nvme_rdma_error_recovery(queue->ctrl);
1722                 return;
1723         }
1724         req = blk_mq_rq_to_pdu(rq);
1725
1726         req->status = cqe->status;
1727         req->result = cqe->result;
1728
1729         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1730                 if (unlikely(!req->mr ||
1731                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1732                         dev_err(queue->ctrl->ctrl.device,
1733                                 "Bogus remote invalidation for rkey %#x\n",
1734                                 req->mr ? req->mr->rkey : 0);
1735                         nvme_rdma_error_recovery(queue->ctrl);
1736                 }
1737         } else if (req->mr) {
1738                 int ret;
1739
1740                 ret = nvme_rdma_inv_rkey(queue, req);
1741                 if (unlikely(ret < 0)) {
1742                         dev_err(queue->ctrl->ctrl.device,
1743                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1744                                 req->mr->rkey, ret);
1745                         nvme_rdma_error_recovery(queue->ctrl);
1746                 }
1747                 /* the local invalidation completion will end the request */
1748                 return;
1749         }
1750
1751         nvme_rdma_end_request(req);
1752 }
1753
1754 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1755 {
1756         struct nvme_rdma_qe *qe =
1757                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1758         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1759         struct ib_device *ibdev = queue->device->dev;
1760         struct nvme_completion *cqe = qe->data;
1761         const size_t len = sizeof(struct nvme_completion);
1762
1763         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1764                 nvme_rdma_wr_error(cq, wc, "RECV");
1765                 return;
1766         }
1767
1768         /* sanity checking for received data length */
1769         if (unlikely(wc->byte_len < len)) {
1770                 dev_err(queue->ctrl->ctrl.device,
1771                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1772                 nvme_rdma_error_recovery(queue->ctrl);
1773                 return;
1774         }
1775
1776         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1777         /*
1778          * AEN requests are special as they don't time out and can
1779          * survive any kind of queue freeze and often don't respond to
1780          * aborts.  We don't even bother to allocate a struct request
1781          * for them but rather special case them here.
1782          */
1783         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1784                                      cqe->command_id)))
1785                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1786                                 &cqe->result);
1787         else
1788                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1789         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1790
1791         nvme_rdma_post_recv(queue, qe);
1792 }
1793
1794 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1795 {
1796         int ret, i;
1797
1798         for (i = 0; i < queue->queue_size; i++) {
1799                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1800                 if (ret)
1801                         goto out_destroy_queue_ib;
1802         }
1803
1804         return 0;
1805
1806 out_destroy_queue_ib:
1807         nvme_rdma_destroy_queue_ib(queue);
1808         return ret;
1809 }
1810
1811 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1812                 struct rdma_cm_event *ev)
1813 {
1814         struct rdma_cm_id *cm_id = queue->cm_id;
1815         int status = ev->status;
1816         const char *rej_msg;
1817         const struct nvme_rdma_cm_rej *rej_data;
1818         u8 rej_data_len;
1819
1820         rej_msg = rdma_reject_msg(cm_id, status);
1821         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1822
1823         if (rej_data && rej_data_len >= sizeof(u16)) {
1824                 u16 sts = le16_to_cpu(rej_data->sts);
1825
1826                 dev_err(queue->ctrl->ctrl.device,
1827                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1828                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1829         } else {
1830                 dev_err(queue->ctrl->ctrl.device,
1831                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1832         }
1833
1834         return -ECONNRESET;
1835 }
1836
1837 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1838 {
1839         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1840         int ret;
1841
1842         ret = nvme_rdma_create_queue_ib(queue);
1843         if (ret)
1844                 return ret;
1845
1846         if (ctrl->opts->tos >= 0)
1847                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1848         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1849         if (ret) {
1850                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1851                         queue->cm_error);
1852                 goto out_destroy_queue;
1853         }
1854
1855         return 0;
1856
1857 out_destroy_queue:
1858         nvme_rdma_destroy_queue_ib(queue);
1859         return ret;
1860 }
1861
1862 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1863 {
1864         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1865         struct rdma_conn_param param = { };
1866         struct nvme_rdma_cm_req priv = { };
1867         int ret;
1868
1869         param.qp_num = queue->qp->qp_num;
1870         param.flow_control = 1;
1871
1872         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1873         /* maximum retry count */
1874         param.retry_count = 7;
1875         param.rnr_retry_count = 7;
1876         param.private_data = &priv;
1877         param.private_data_len = sizeof(priv);
1878
1879         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1880         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1881         /*
1882          * set the admin queue depth to the minimum size
1883          * specified by the Fabrics standard.
1884          */
1885         if (priv.qid == 0) {
1886                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1887                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1888         } else {
1889                 /*
1890                  * current interpretation of the fabrics spec
1891                  * is at minimum you make hrqsize sqsize+1, or a
1892                  * 1's based representation of sqsize.
1893                  */
1894                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1895                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1896         }
1897
1898         ret = rdma_connect_locked(queue->cm_id, &param);
1899         if (ret) {
1900                 dev_err(ctrl->ctrl.device,
1901                         "rdma_connect_locked failed (%d).\n", ret);
1902                 goto out_destroy_queue_ib;
1903         }
1904
1905         return 0;
1906
1907 out_destroy_queue_ib:
1908         nvme_rdma_destroy_queue_ib(queue);
1909         return ret;
1910 }
1911
1912 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1913                 struct rdma_cm_event *ev)
1914 {
1915         struct nvme_rdma_queue *queue = cm_id->context;
1916         int cm_error = 0;
1917
1918         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1919                 rdma_event_msg(ev->event), ev->event,
1920                 ev->status, cm_id);
1921
1922         switch (ev->event) {
1923         case RDMA_CM_EVENT_ADDR_RESOLVED:
1924                 cm_error = nvme_rdma_addr_resolved(queue);
1925                 break;
1926         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1927                 cm_error = nvme_rdma_route_resolved(queue);
1928                 break;
1929         case RDMA_CM_EVENT_ESTABLISHED:
1930                 queue->cm_error = nvme_rdma_conn_established(queue);
1931                 /* complete cm_done regardless of success/failure */
1932                 complete(&queue->cm_done);
1933                 return 0;
1934         case RDMA_CM_EVENT_REJECTED:
1935                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1936                 break;
1937         case RDMA_CM_EVENT_ROUTE_ERROR:
1938         case RDMA_CM_EVENT_CONNECT_ERROR:
1939         case RDMA_CM_EVENT_UNREACHABLE:
1940                 nvme_rdma_destroy_queue_ib(queue);
1941                 fallthrough;
1942         case RDMA_CM_EVENT_ADDR_ERROR:
1943                 dev_dbg(queue->ctrl->ctrl.device,
1944                         "CM error event %d\n", ev->event);
1945                 cm_error = -ECONNRESET;
1946                 break;
1947         case RDMA_CM_EVENT_DISCONNECTED:
1948         case RDMA_CM_EVENT_ADDR_CHANGE:
1949         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1950                 dev_dbg(queue->ctrl->ctrl.device,
1951                         "disconnect received - connection closed\n");
1952                 nvme_rdma_error_recovery(queue->ctrl);
1953                 break;
1954         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1955                 /* device removal is handled via the ib_client API */
1956                 break;
1957         default:
1958                 dev_err(queue->ctrl->ctrl.device,
1959                         "Unexpected RDMA CM event (%d)\n", ev->event);
1960                 nvme_rdma_error_recovery(queue->ctrl);
1961                 break;
1962         }
1963
1964         if (cm_error) {
1965                 queue->cm_error = cm_error;
1966                 complete(&queue->cm_done);
1967         }
1968
1969         return 0;
1970 }
1971
1972 static void nvme_rdma_complete_timed_out(struct request *rq)
1973 {
1974         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1975         struct nvme_rdma_queue *queue = req->queue;
1976
1977         nvme_rdma_stop_queue(queue);
1978         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
1979                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
1980                 blk_mq_complete_request(rq);
1981         }
1982 }
1983
1984 static enum blk_eh_timer_return
1985 nvme_rdma_timeout(struct request *rq, bool reserved)
1986 {
1987         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1988         struct nvme_rdma_queue *queue = req->queue;
1989         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1990
1991         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1992                  rq->tag, nvme_rdma_queue_idx(queue));
1993
1994         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1995                 /*
1996                  * If we are resetting, connecting or deleting we should
1997                  * complete immediately because we may block controller
1998                  * teardown or setup sequence
1999                  * - ctrl disable/shutdown fabrics requests
2000                  * - connect requests
2001                  * - initialization admin requests
2002                  * - I/O requests that entered after unquiescing and
2003                  *   the controller stopped responding
2004                  *
2005                  * All other requests should be cancelled by the error
2006                  * recovery work, so it's fine that we fail it here.
2007                  */
2008                 nvme_rdma_complete_timed_out(rq);
2009                 return BLK_EH_DONE;
2010         }
2011
2012         /*
2013          * LIVE state should trigger the normal error recovery which will
2014          * handle completing this request.
2015          */
2016         nvme_rdma_error_recovery(ctrl);
2017         return BLK_EH_RESET_TIMER;
2018 }
2019
2020 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2021                 const struct blk_mq_queue_data *bd)
2022 {
2023         struct nvme_ns *ns = hctx->queue->queuedata;
2024         struct nvme_rdma_queue *queue = hctx->driver_data;
2025         struct request *rq = bd->rq;
2026         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2027         struct nvme_rdma_qe *sqe = &req->sqe;
2028         struct nvme_command *c = sqe->data;
2029         struct ib_device *dev;
2030         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2031         blk_status_t ret;
2032         int err;
2033
2034         WARN_ON_ONCE(rq->tag < 0);
2035
2036         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2037                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2038
2039         dev = queue->device->dev;
2040
2041         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2042                                          sizeof(struct nvme_command),
2043                                          DMA_TO_DEVICE);
2044         err = ib_dma_mapping_error(dev, req->sqe.dma);
2045         if (unlikely(err))
2046                 return BLK_STS_RESOURCE;
2047
2048         ib_dma_sync_single_for_cpu(dev, sqe->dma,
2049                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2050
2051         ret = nvme_setup_cmd(ns, rq, c);
2052         if (ret)
2053                 goto unmap_qe;
2054
2055         blk_mq_start_request(rq);
2056
2057         if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2058             queue->pi_support &&
2059             (c->common.opcode == nvme_cmd_write ||
2060              c->common.opcode == nvme_cmd_read) &&
2061             nvme_ns_has_pi(ns))
2062                 req->use_sig_mr = true;
2063         else
2064                 req->use_sig_mr = false;
2065
2066         err = nvme_rdma_map_data(queue, rq, c);
2067         if (unlikely(err < 0)) {
2068                 dev_err(queue->ctrl->ctrl.device,
2069                              "Failed to map data (%d)\n", err);
2070                 goto err;
2071         }
2072
2073         sqe->cqe.done = nvme_rdma_send_done;
2074
2075         ib_dma_sync_single_for_device(dev, sqe->dma,
2076                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2077
2078         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2079                         req->mr ? &req->reg_wr.wr : NULL);
2080         if (unlikely(err))
2081                 goto err_unmap;
2082
2083         return BLK_STS_OK;
2084
2085 err_unmap:
2086         nvme_rdma_unmap_data(queue, rq);
2087 err:
2088         if (err == -ENOMEM || err == -EAGAIN)
2089                 ret = BLK_STS_RESOURCE;
2090         else
2091                 ret = BLK_STS_IOERR;
2092         nvme_cleanup_cmd(rq);
2093 unmap_qe:
2094         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2095                             DMA_TO_DEVICE);
2096         return ret;
2097 }
2098
2099 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2100 {
2101         struct nvme_rdma_queue *queue = hctx->driver_data;
2102
2103         return ib_process_cq_direct(queue->ib_cq, -1);
2104 }
2105
2106 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2107 {
2108         struct request *rq = blk_mq_rq_from_pdu(req);
2109         struct ib_mr_status mr_status;
2110         int ret;
2111
2112         ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2113         if (ret) {
2114                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2115                 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2116                 return;
2117         }
2118
2119         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2120                 switch (mr_status.sig_err.err_type) {
2121                 case IB_SIG_BAD_GUARD:
2122                         nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2123                         break;
2124                 case IB_SIG_BAD_REFTAG:
2125                         nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2126                         break;
2127                 case IB_SIG_BAD_APPTAG:
2128                         nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2129                         break;
2130                 }
2131                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2132                        mr_status.sig_err.err_type, mr_status.sig_err.expected,
2133                        mr_status.sig_err.actual);
2134         }
2135 }
2136
2137 static void nvme_rdma_complete_rq(struct request *rq)
2138 {
2139         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2140         struct nvme_rdma_queue *queue = req->queue;
2141         struct ib_device *ibdev = queue->device->dev;
2142
2143         if (req->use_sig_mr)
2144                 nvme_rdma_check_pi_status(req);
2145
2146         nvme_rdma_unmap_data(queue, rq);
2147         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2148                             DMA_TO_DEVICE);
2149         nvme_complete_rq(rq);
2150 }
2151
2152 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2153 {
2154         struct nvme_rdma_ctrl *ctrl = set->driver_data;
2155         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2156
2157         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2158                 /* separate read/write queues */
2159                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2160                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2161                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2162                 set->map[HCTX_TYPE_READ].nr_queues =
2163                         ctrl->io_queues[HCTX_TYPE_READ];
2164                 set->map[HCTX_TYPE_READ].queue_offset =
2165                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2166         } else {
2167                 /* shared read/write queues */
2168                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2169                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2170                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2171                 set->map[HCTX_TYPE_READ].nr_queues =
2172                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2173                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2174         }
2175         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2176                         ctrl->device->dev, 0);
2177         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2178                         ctrl->device->dev, 0);
2179
2180         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2181                 /* map dedicated poll queues only if we have queues left */
2182                 set->map[HCTX_TYPE_POLL].nr_queues =
2183                                 ctrl->io_queues[HCTX_TYPE_POLL];
2184                 set->map[HCTX_TYPE_POLL].queue_offset =
2185                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2186                         ctrl->io_queues[HCTX_TYPE_READ];
2187                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2188         }
2189
2190         dev_info(ctrl->ctrl.device,
2191                 "mapped %d/%d/%d default/read/poll queues.\n",
2192                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2193                 ctrl->io_queues[HCTX_TYPE_READ],
2194                 ctrl->io_queues[HCTX_TYPE_POLL]);
2195
2196         return 0;
2197 }
2198
2199 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2200         .queue_rq       = nvme_rdma_queue_rq,
2201         .complete       = nvme_rdma_complete_rq,
2202         .init_request   = nvme_rdma_init_request,
2203         .exit_request   = nvme_rdma_exit_request,
2204         .init_hctx      = nvme_rdma_init_hctx,
2205         .timeout        = nvme_rdma_timeout,
2206         .map_queues     = nvme_rdma_map_queues,
2207         .poll           = nvme_rdma_poll,
2208 };
2209
2210 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2211         .queue_rq       = nvme_rdma_queue_rq,
2212         .complete       = nvme_rdma_complete_rq,
2213         .init_request   = nvme_rdma_init_request,
2214         .exit_request   = nvme_rdma_exit_request,
2215         .init_hctx      = nvme_rdma_init_admin_hctx,
2216         .timeout        = nvme_rdma_timeout,
2217 };
2218
2219 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2220 {
2221         cancel_work_sync(&ctrl->err_work);
2222         cancel_delayed_work_sync(&ctrl->reconnect_work);
2223
2224         nvme_rdma_teardown_io_queues(ctrl, shutdown);
2225         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2226         if (shutdown)
2227                 nvme_shutdown_ctrl(&ctrl->ctrl);
2228         else
2229                 nvme_disable_ctrl(&ctrl->ctrl);
2230         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2231 }
2232
2233 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2234 {
2235         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2236 }
2237
2238 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2239 {
2240         struct nvme_rdma_ctrl *ctrl =
2241                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2242
2243         nvme_stop_ctrl(&ctrl->ctrl);
2244         nvme_rdma_shutdown_ctrl(ctrl, false);
2245
2246         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2247                 /* state change failure should never happen */
2248                 WARN_ON_ONCE(1);
2249                 return;
2250         }
2251
2252         if (nvme_rdma_setup_ctrl(ctrl, false))
2253                 goto out_fail;
2254
2255         return;
2256
2257 out_fail:
2258         ++ctrl->ctrl.nr_reconnects;
2259         nvme_rdma_reconnect_or_remove(ctrl);
2260 }
2261
2262 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2263         .name                   = "rdma",
2264         .module                 = THIS_MODULE,
2265         .flags                  = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2266         .reg_read32             = nvmf_reg_read32,
2267         .reg_read64             = nvmf_reg_read64,
2268         .reg_write32            = nvmf_reg_write32,
2269         .free_ctrl              = nvme_rdma_free_ctrl,
2270         .submit_async_event     = nvme_rdma_submit_async_event,
2271         .delete_ctrl            = nvme_rdma_delete_ctrl,
2272         .get_address            = nvmf_get_address,
2273 };
2274
2275 /*
2276  * Fails a connection request if it matches an existing controller
2277  * (association) with the same tuple:
2278  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2279  *
2280  * if local address is not specified in the request, it will match an
2281  * existing controller with all the other parameters the same and no
2282  * local port address specified as well.
2283  *
2284  * The ports don't need to be compared as they are intrinsically
2285  * already matched by the port pointers supplied.
2286  */
2287 static bool
2288 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2289 {
2290         struct nvme_rdma_ctrl *ctrl;
2291         bool found = false;
2292
2293         mutex_lock(&nvme_rdma_ctrl_mutex);
2294         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2295                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2296                 if (found)
2297                         break;
2298         }
2299         mutex_unlock(&nvme_rdma_ctrl_mutex);
2300
2301         return found;
2302 }
2303
2304 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2305                 struct nvmf_ctrl_options *opts)
2306 {
2307         struct nvme_rdma_ctrl *ctrl;
2308         int ret;
2309         bool changed;
2310
2311         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2312         if (!ctrl)
2313                 return ERR_PTR(-ENOMEM);
2314         ctrl->ctrl.opts = opts;
2315         INIT_LIST_HEAD(&ctrl->list);
2316
2317         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2318                 opts->trsvcid =
2319                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2320                 if (!opts->trsvcid) {
2321                         ret = -ENOMEM;
2322                         goto out_free_ctrl;
2323                 }
2324                 opts->mask |= NVMF_OPT_TRSVCID;
2325         }
2326
2327         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2328                         opts->traddr, opts->trsvcid, &ctrl->addr);
2329         if (ret) {
2330                 pr_err("malformed address passed: %s:%s\n",
2331                         opts->traddr, opts->trsvcid);
2332                 goto out_free_ctrl;
2333         }
2334
2335         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2336                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2337                         opts->host_traddr, NULL, &ctrl->src_addr);
2338                 if (ret) {
2339                         pr_err("malformed src address passed: %s\n",
2340                                opts->host_traddr);
2341                         goto out_free_ctrl;
2342                 }
2343         }
2344
2345         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2346                 ret = -EALREADY;
2347                 goto out_free_ctrl;
2348         }
2349
2350         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2351                         nvme_rdma_reconnect_ctrl_work);
2352         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2353         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2354
2355         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2356                                 opts->nr_poll_queues + 1;
2357         ctrl->ctrl.sqsize = opts->queue_size - 1;
2358         ctrl->ctrl.kato = opts->kato;
2359
2360         ret = -ENOMEM;
2361         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2362                                 GFP_KERNEL);
2363         if (!ctrl->queues)
2364                 goto out_free_ctrl;
2365
2366         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2367                                 0 /* no quirks, we're perfect! */);
2368         if (ret)
2369                 goto out_kfree_queues;
2370
2371         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2372         WARN_ON_ONCE(!changed);
2373
2374         ret = nvme_rdma_setup_ctrl(ctrl, true);
2375         if (ret)
2376                 goto out_uninit_ctrl;
2377
2378         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2379                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2380
2381         mutex_lock(&nvme_rdma_ctrl_mutex);
2382         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2383         mutex_unlock(&nvme_rdma_ctrl_mutex);
2384
2385         return &ctrl->ctrl;
2386
2387 out_uninit_ctrl:
2388         nvme_uninit_ctrl(&ctrl->ctrl);
2389         nvme_put_ctrl(&ctrl->ctrl);
2390         if (ret > 0)
2391                 ret = -EIO;
2392         return ERR_PTR(ret);
2393 out_kfree_queues:
2394         kfree(ctrl->queues);
2395 out_free_ctrl:
2396         kfree(ctrl);
2397         return ERR_PTR(ret);
2398 }
2399
2400 static struct nvmf_transport_ops nvme_rdma_transport = {
2401         .name           = "rdma",
2402         .module         = THIS_MODULE,
2403         .required_opts  = NVMF_OPT_TRADDR,
2404         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2405                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2406                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2407                           NVMF_OPT_TOS,
2408         .create_ctrl    = nvme_rdma_create_ctrl,
2409 };
2410
2411 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2412 {
2413         struct nvme_rdma_ctrl *ctrl;
2414         struct nvme_rdma_device *ndev;
2415         bool found = false;
2416
2417         mutex_lock(&device_list_mutex);
2418         list_for_each_entry(ndev, &device_list, entry) {
2419                 if (ndev->dev == ib_device) {
2420                         found = true;
2421                         break;
2422                 }
2423         }
2424         mutex_unlock(&device_list_mutex);
2425
2426         if (!found)
2427                 return;
2428
2429         /* Delete all controllers using this device */
2430         mutex_lock(&nvme_rdma_ctrl_mutex);
2431         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2432                 if (ctrl->device->dev != ib_device)
2433                         continue;
2434                 nvme_delete_ctrl(&ctrl->ctrl);
2435         }
2436         mutex_unlock(&nvme_rdma_ctrl_mutex);
2437
2438         flush_workqueue(nvme_delete_wq);
2439 }
2440
2441 static struct ib_client nvme_rdma_ib_client = {
2442         .name   = "nvme_rdma",
2443         .remove = nvme_rdma_remove_one
2444 };
2445
2446 static int __init nvme_rdma_init_module(void)
2447 {
2448         int ret;
2449
2450         ret = ib_register_client(&nvme_rdma_ib_client);
2451         if (ret)
2452                 return ret;
2453
2454         ret = nvmf_register_transport(&nvme_rdma_transport);
2455         if (ret)
2456                 goto err_unreg_client;
2457
2458         return 0;
2459
2460 err_unreg_client:
2461         ib_unregister_client(&nvme_rdma_ib_client);
2462         return ret;
2463 }
2464
2465 static void __exit nvme_rdma_cleanup_module(void)
2466 {
2467         struct nvme_rdma_ctrl *ctrl;
2468
2469         nvmf_unregister_transport(&nvme_rdma_transport);
2470         ib_unregister_client(&nvme_rdma_ib_client);
2471
2472         mutex_lock(&nvme_rdma_ctrl_mutex);
2473         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2474                 nvme_delete_ctrl(&ctrl->ctrl);
2475         mutex_unlock(&nvme_rdma_ctrl_mutex);
2476         flush_workqueue(nvme_delete_wq);
2477 }
2478
2479 module_init(nvme_rdma_init_module);
2480 module_exit(nvme_rdma_cleanup_module);
2481
2482 MODULE_LICENSE("GPL v2");