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