drivers/nvme/host/rdma.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics RDMA host code.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/slab.h>
  10 #include <rdma/mr_pool.h>
  11 #include <linux/err.h>
  12 #include <linux/string.h>
  13 #include <linux/atomic.h>
  14 #include <linux/blk-mq.h>
  15 #include <linux/blk-mq-rdma.h>
  16 #include <linux/types.h>
  17 #include <linux/list.h>
  18 #include <linux/mutex.h>
  19 #include <linux/scatterlist.h>
  20 #include <linux/nvme.h>
  21 #include <asm/unaligned.h>
  22
  23 #include <rdma/ib_verbs.h>
  24 #include <rdma/rdma_cm.h>
  25 #include <linux/nvme-rdma.h>
  26
  27 #include "nvme.h"
  28 #include "fabrics.h"
  29
  30
  31 #define NVME_RDMA_CONNECT_TIMEOUT_MS    3000            /* 3 second */
  32
  33 #define NVME_RDMA_MAX_SEGMENTS          256
  34
  35 #define NVME_RDMA_MAX_INLINE_SEGMENTS   4
  36
  37 #define NVME_RDMA_DATA_SGL_SIZE \
  38         (sizeof(struct scatterlist) * NVME_INLINE_SG_CNT)
  39 #define NVME_RDMA_METADATA_SGL_SIZE \
  40         (sizeof(struct scatterlist) * NVME_INLINE_METADATA_SG_CNT)
  41
  42 struct nvme_rdma_device {
  43         struct ib_device        *dev;
  44         struct ib_pd            *pd;
  45         struct kref             ref;
  46         struct list_head        entry;
  47         unsigned int            num_inline_segments;
  48 };
  49
  50 struct nvme_rdma_qe {
  51         struct ib_cqe           cqe;
  52         void                    *data;
  53         u64                     dma;
  54 };
  55
  56 struct nvme_rdma_sgl {
  57         int                     nents;
  58         struct sg_table         sg_table;
  59 };
  60
  61 struct nvme_rdma_queue;
  62 struct nvme_rdma_request {
  63         struct nvme_request     req;
  64         struct ib_mr            *mr;
  65         struct nvme_rdma_qe     sqe;
  66         union nvme_result       result;
  67         __le16                  status;
  68         refcount_t              ref;
  69         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  70         u32                     num_sge;
  71         struct ib_reg_wr        reg_wr;
  72         struct ib_cqe           reg_cqe;
  73         struct nvme_rdma_queue  *queue;
  74         struct nvme_rdma_sgl    data_sgl;
  75         struct nvme_rdma_sgl    *metadata_sgl;
  76         bool                    use_sig_mr;
  77 };
  78
  79 enum nvme_rdma_queue_flags {
  80         NVME_RDMA_Q_ALLOCATED           = 0,
  81         NVME_RDMA_Q_LIVE                = 1,
  82         NVME_RDMA_Q_TR_READY            = 2,
  83 };
  84
  85 struct nvme_rdma_queue {
  86         struct nvme_rdma_qe     *rsp_ring;
  87         int                     queue_size;
  88         size_t                  cmnd_capsule_len;
  89         struct nvme_rdma_ctrl   *ctrl;
  90         struct nvme_rdma_device *device;
  91         struct ib_cq            *ib_cq;
  92         struct ib_qp            *qp;
  93
  94         unsigned long           flags;
  95         struct rdma_cm_id       *cm_id;
  96         int                     cm_error;
  97         struct completion       cm_done;
  98         bool                    pi_support;
  99         int                     cq_size;
 100         struct mutex            queue_lock;
 101 };
 102
 103 struct nvme_rdma_ctrl {
 104         /* read only in the hot path */
 105         struct nvme_rdma_queue  *queues;
 106
 107         /* other member variables */
 108         struct blk_mq_tag_set   tag_set;
 109         struct work_struct      err_work;
 110
 111         struct nvme_rdma_qe     async_event_sqe;
 112
 113         struct delayed_work     reconnect_work;
 114
 115         struct list_head        list;
 116
 117         struct blk_mq_tag_set   admin_tag_set;
 118         struct nvme_rdma_device *device;
 119
 120         u32                     max_fr_pages;
 121
 122         struct sockaddr_storage addr;
 123         struct sockaddr_storage src_addr;
 124
 125         struct nvme_ctrl        ctrl;
 126         bool                    use_inline_data;
 127         u32                     io_queues[HCTX_MAX_TYPES];
 128 };
 129
 130 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 131 {
 132         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 133 }
 134
 135 static LIST_HEAD(device_list);
 136 static DEFINE_MUTEX(device_list_mutex);
 137
 138 static LIST_HEAD(nvme_rdma_ctrl_list);
 139 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 140
 141 /*
 142  * Disabling this option makes small I/O goes faster, but is fundamentally
 143  * unsafe.  With it turned off we will have to register a global rkey that
 144  * allows read and write access to all physical memory.
 145  */
 146 static bool register_always = true;
 147 module_param(register_always, bool, 0444);
 148 MODULE_PARM_DESC(register_always,
 149          "Use memory registration even for contiguous memory regions");
 150
 151 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 152                 struct rdma_cm_event *event);
 153 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 154 static void nvme_rdma_complete_rq(struct request *rq);
 155
 156 static const struct blk_mq_ops nvme_rdma_mq_ops;
 157 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 158
 159 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 160 {
 161         return queue - queue->ctrl->queues;
 162 }
 163
 164 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
 165 {
 166         return nvme_rdma_queue_idx(queue) >
 167                 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
 168                 queue->ctrl->io_queues[HCTX_TYPE_READ];
 169 }
 170
 171 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 172 {
 173         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 174 }
 175
 176 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 177                 size_t capsule_size, enum dma_data_direction dir)
 178 {
 179         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 180         kfree(qe->data);
 181 }
 182
 183 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 184                 size_t capsule_size, enum dma_data_direction dir)
 185 {
 186         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 187         if (!qe->data)
 188                 return -ENOMEM;
 189
 190         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 191         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 192                 kfree(qe->data);
 193                 qe->data = NULL;
 194                 return -ENOMEM;
 195         }
 196
 197         return 0;
 198 }
 199
 200 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 201                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 202                 size_t capsule_size, enum dma_data_direction dir)
 203 {
 204         int i;
 205
 206         for (i = 0; i < ib_queue_size; i++)
 207                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 208         kfree(ring);
 209 }
 210
 211 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 212                 size_t ib_queue_size, size_t capsule_size,
 213                 enum dma_data_direction dir)
 214 {
 215         struct nvme_rdma_qe *ring;
 216         int i;
 217
 218         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 219         if (!ring)
 220                 return NULL;
 221
 222         /*
 223          * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
 224          * lifetime. It's safe, since any chage in the underlying RDMA device
 225          * will issue error recovery and queue re-creation.
 226          */
 227         for (i = 0; i < ib_queue_size; i++) {
 228                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 229                         goto out_free_ring;
 230         }
 231
 232         return ring;
 233
 234 out_free_ring:
 235         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 236         return NULL;
 237 }
 238
 239 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 240 {
 241         pr_debug("QP event %s (%d)\n",
 242                  ib_event_msg(event->event), event->event);
 243
 244 }
 245
 246 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 247 {
 248         int ret;
 249
 250         ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
 251                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 252         if (ret < 0)
 253                 return ret;
 254         if (ret == 0)
 255                 return -ETIMEDOUT;
 256         WARN_ON_ONCE(queue->cm_error > 0);
 257         return queue->cm_error;
 258 }
 259
 260 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 261 {
 262         struct nvme_rdma_device *dev = queue->device;
 263         struct ib_qp_init_attr init_attr;
 264         int ret;
 265
 266         memset(&init_attr, 0, sizeof(init_attr));
 267         init_attr.event_handler = nvme_rdma_qp_event;
 268         /* +1 for drain */
 269         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 270         /* +1 for drain */
 271         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 272         init_attr.cap.max_recv_sge = 1;
 273         init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
 274         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 275         init_attr.qp_type = IB_QPT_RC;
 276         init_attr.send_cq = queue->ib_cq;
 277         init_attr.recv_cq = queue->ib_cq;
 278         if (queue->pi_support)
 279                 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
 280         init_attr.qp_context = queue;
 281
 282         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 283
 284         queue->qp = queue->cm_id->qp;
 285         return ret;
 286 }
 287
 288 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 289                 struct request *rq, unsigned int hctx_idx)
 290 {
 291         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 292
 293         kfree(req->sqe.data);
 294 }
 295
 296 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 297                 struct request *rq, unsigned int hctx_idx,
 298                 unsigned int numa_node)
 299 {
 300         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 301         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 302         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 303         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 304
 305         nvme_req(rq)->ctrl = &ctrl->ctrl;
 306         req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
 307         if (!req->sqe.data)
 308                 return -ENOMEM;
 309
 310         /* metadata nvme_rdma_sgl struct is located after command's data SGL */
 311         if (queue->pi_support)
 312                 req->metadata_sgl = (void *)nvme_req(rq) +
 313                         sizeof(struct nvme_rdma_request) +
 314                         NVME_RDMA_DATA_SGL_SIZE;
 315
 316         req->queue = queue;
 317         nvme_req(rq)->cmd = req->sqe.data;
 318
 319         return 0;
 320 }
 321
 322 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 323                 unsigned int hctx_idx)
 324 {
 325         struct nvme_rdma_ctrl *ctrl = data;
 326         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 327
 328         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 329
 330         hctx->driver_data = queue;
 331         return 0;
 332 }
 333
 334 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 335                 unsigned int hctx_idx)
 336 {
 337         struct nvme_rdma_ctrl *ctrl = data;
 338         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 339
 340         BUG_ON(hctx_idx != 0);
 341
 342         hctx->driver_data = queue;
 343         return 0;
 344 }
 345
 346 static void nvme_rdma_free_dev(struct kref *ref)
 347 {
 348         struct nvme_rdma_device *ndev =
 349                 container_of(ref, struct nvme_rdma_device, ref);
 350
 351         mutex_lock(&device_list_mutex);
 352         list_del(&ndev->entry);
 353         mutex_unlock(&device_list_mutex);
 354
 355         ib_dealloc_pd(ndev->pd);
 356         kfree(ndev);
 357 }
 358
 359 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 360 {
 361         kref_put(&dev->ref, nvme_rdma_free_dev);
 362 }
 363
 364 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 365 {
 366         return kref_get_unless_zero(&dev->ref);
 367 }
 368
 369 static struct nvme_rdma_device *
 370 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 371 {
 372         struct nvme_rdma_device *ndev;
 373
 374         mutex_lock(&device_list_mutex);
 375         list_for_each_entry(ndev, &device_list, entry) {
 376                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 377                     nvme_rdma_dev_get(ndev))
 378                         goto out_unlock;
 379         }
 380
 381         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 382         if (!ndev)
 383                 goto out_err;
 384
 385         ndev->dev = cm_id->device;
 386         kref_init(&ndev->ref);
 387
 388         ndev->pd = ib_alloc_pd(ndev->dev,
 389                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 390         if (IS_ERR(ndev->pd))
 391                 goto out_free_dev;
 392
 393         if (!(ndev->dev->attrs.device_cap_flags &
 394               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 395                 dev_err(&ndev->dev->dev,
 396                         "Memory registrations not supported.\n");
 397                 goto out_free_pd;
 398         }
 399
 400         ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
 401                                         ndev->dev->attrs.max_send_sge - 1);
 402         list_add(&ndev->entry, &device_list);
 403 out_unlock:
 404         mutex_unlock(&device_list_mutex);
 405         return ndev;
 406
 407 out_free_pd:
 408         ib_dealloc_pd(ndev->pd);
 409 out_free_dev:
 410         kfree(ndev);
 411 out_err:
 412         mutex_unlock(&device_list_mutex);
 413         return NULL;
 414 }
 415
 416 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
 417 {
 418         if (nvme_rdma_poll_queue(queue))
 419                 ib_free_cq(queue->ib_cq);
 420         else
 421                 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
 422 }
 423
 424 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 425 {
 426         struct nvme_rdma_device *dev;
 427         struct ib_device *ibdev;
 428
 429         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 430                 return;
 431
 432         dev = queue->device;
 433         ibdev = dev->dev;
 434
 435         if (queue->pi_support)
 436                 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
 437         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 438
 439         /*
 440          * The cm_id object might have been destroyed during RDMA connection
 441          * establishment error flow to avoid getting other cma events, thus
 442          * the destruction of the QP shouldn't use rdma_cm API.
 443          */
 444         ib_destroy_qp(queue->qp);
 445         nvme_rdma_free_cq(queue);
 446
 447         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 448                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 449
 450         nvme_rdma_dev_put(dev);
 451 }
 452
 453 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
 454 {
 455         u32 max_page_list_len;
 456
 457         if (pi_support)
 458                 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
 459         else
 460                 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
 461
 462         return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
 463 }
 464
 465 static int nvme_rdma_create_cq(struct ib_device *ibdev,
 466                 struct nvme_rdma_queue *queue)
 467 {
 468         int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
 469         enum ib_poll_context poll_ctx;
 470
 471         /*
 472          * Spread I/O queues completion vectors according their queue index.
 473          * Admin queues can always go on completion vector 0.
 474          */
 475         comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
 476
 477         /* Polling queues need direct cq polling context */
 478         if (nvme_rdma_poll_queue(queue)) {
 479                 poll_ctx = IB_POLL_DIRECT;
 480                 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
 481                                            comp_vector, poll_ctx);
 482         } else {
 483                 poll_ctx = IB_POLL_SOFTIRQ;
 484                 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
 485                                               comp_vector, poll_ctx);
 486         }
 487
 488         if (IS_ERR(queue->ib_cq)) {
 489                 ret = PTR_ERR(queue->ib_cq);
 490                 return ret;
 491         }
 492
 493         return 0;
 494 }
 495
 496 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 497 {
 498         struct ib_device *ibdev;
 499         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 500         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 501         int ret, pages_per_mr;
 502
 503         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 504         if (!queue->device) {
 505                 dev_err(queue->cm_id->device->dev.parent,
 506                         "no client data found!\n");
 507                 return -ECONNREFUSED;
 508         }
 509         ibdev = queue->device->dev;
 510
 511         /* +1 for ib_stop_cq */
 512         queue->cq_size = cq_factor * queue->queue_size + 1;
 513
 514         ret = nvme_rdma_create_cq(ibdev, queue);
 515         if (ret)
 516                 goto out_put_dev;
 517
 518         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 519         if (ret)
 520                 goto out_destroy_ib_cq;
 521
 522         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 523                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 524         if (!queue->rsp_ring) {
 525                 ret = -ENOMEM;
 526                 goto out_destroy_qp;
 527         }
 528
 529         /*
 530          * Currently we don't use SG_GAPS MR's so if the first entry is
 531          * misaligned we'll end up using two entries for a single data page,
 532          * so one additional entry is required.
 533          */
 534         pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
 535         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 536                               queue->queue_size,
 537                               IB_MR_TYPE_MEM_REG,
 538                               pages_per_mr, 0);
 539         if (ret) {
 540                 dev_err(queue->ctrl->ctrl.device,
 541                         "failed to initialize MR pool sized %d for QID %d\n",
 542                         queue->queue_size, nvme_rdma_queue_idx(queue));
 543                 goto out_destroy_ring;
 544         }
 545
 546         if (queue->pi_support) {
 547                 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
 548                                       queue->queue_size, IB_MR_TYPE_INTEGRITY,
 549                                       pages_per_mr, pages_per_mr);
 550                 if (ret) {
 551                         dev_err(queue->ctrl->ctrl.device,
 552                                 "failed to initialize PI MR pool sized %d for QID %d\n",
 553                                 queue->queue_size, nvme_rdma_queue_idx(queue));
 554                         goto out_destroy_mr_pool;
 555                 }
 556         }
 557
 558         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 559
 560         return 0;
 561
 562 out_destroy_mr_pool:
 563         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 564 out_destroy_ring:
 565         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 566                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 567 out_destroy_qp:
 568         rdma_destroy_qp(queue->cm_id);
 569 out_destroy_ib_cq:
 570         nvme_rdma_free_cq(queue);
 571 out_put_dev:
 572         nvme_rdma_dev_put(queue->device);
 573         return ret;
 574 }
 575
 576 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 577                 int idx, size_t queue_size)
 578 {
 579         struct nvme_rdma_queue *queue;
 580         struct sockaddr *src_addr = NULL;
 581         int ret;
 582
 583         queue = &ctrl->queues[idx];
 584         mutex_init(&queue->queue_lock);
 585         queue->ctrl = ctrl;
 586         if (idx && ctrl->ctrl.max_integrity_segments)
 587                 queue->pi_support = true;
 588         else
 589                 queue->pi_support = false;
 590         init_completion(&queue->cm_done);
 591
 592         if (idx > 0)
 593                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 594         else
 595                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 596
 597         queue->queue_size = queue_size;
 598
 599         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 600                         RDMA_PS_TCP, IB_QPT_RC);
 601         if (IS_ERR(queue->cm_id)) {
 602                 dev_info(ctrl->ctrl.device,
 603                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 604                 ret = PTR_ERR(queue->cm_id);
 605                 goto out_destroy_mutex;
 606         }
 607
 608         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 609                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 610
 611         queue->cm_error = -ETIMEDOUT;
 612         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 613                         (struct sockaddr *)&ctrl->addr,
 614                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 615         if (ret) {
 616                 dev_info(ctrl->ctrl.device,
 617                         "rdma_resolve_addr failed (%d).\n", ret);
 618                 goto out_destroy_cm_id;
 619         }
 620
 621         ret = nvme_rdma_wait_for_cm(queue);
 622         if (ret) {
 623                 dev_info(ctrl->ctrl.device,
 624                         "rdma connection establishment failed (%d)\n", ret);
 625                 goto out_destroy_cm_id;
 626         }
 627
 628         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 629
 630         return 0;
 631
 632 out_destroy_cm_id:
 633         rdma_destroy_id(queue->cm_id);
 634         nvme_rdma_destroy_queue_ib(queue);
 635 out_destroy_mutex:
 636         mutex_destroy(&queue->queue_lock);
 637         return ret;
 638 }
 639
 640 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 641 {
 642         rdma_disconnect(queue->cm_id);
 643         ib_drain_qp(queue->qp);
 644 }
 645
 646 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 647 {
 648         mutex_lock(&queue->queue_lock);
 649         if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 650                 __nvme_rdma_stop_queue(queue);
 651         mutex_unlock(&queue->queue_lock);
 652 }
 653
 654 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 655 {
 656         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 657                 return;
 658
 659         nvme_rdma_destroy_queue_ib(queue);
 660         rdma_destroy_id(queue->cm_id);
 661         mutex_destroy(&queue->queue_lock);
 662 }
 663
 664 static void nvme_rdma_free_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_free_queue(&ctrl->queues[i]);
 670 }
 671
 672 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 673 {
 674         int i;
 675
 676         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 677                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 678 }
 679
 680 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 681 {
 682         struct nvme_rdma_queue *queue = &ctrl->queues[idx];
 683         int ret;
 684
 685         if (idx)
 686                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 687         else
 688                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 689
 690         if (!ret) {
 691                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 692         } else {
 693                 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 694                         __nvme_rdma_stop_queue(queue);
 695                 dev_info(ctrl->ctrl.device,
 696                         "failed to connect queue: %d ret=%d\n", idx, ret);
 697         }
 698         return ret;
 699 }
 700
 701 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 702 {
 703         int i, ret = 0;
 704
 705         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 706                 ret = nvme_rdma_start_queue(ctrl, i);
 707                 if (ret)
 708                         goto out_stop_queues;
 709         }
 710
 711         return 0;
 712
 713 out_stop_queues:
 714         for (i--; i >= 1; i--)
 715                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 716         return ret;
 717 }
 718
 719 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 720 {
 721         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 722         struct ib_device *ibdev = ctrl->device->dev;
 723         unsigned int nr_io_queues, nr_default_queues;
 724         unsigned int nr_read_queues, nr_poll_queues;
 725         int i, ret;
 726
 727         nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
 728                                 min(opts->nr_io_queues, num_online_cpus()));
 729         nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
 730                                 min(opts->nr_write_queues, num_online_cpus()));
 731         nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
 732         nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
 733
 734         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 735         if (ret)
 736                 return ret;
 737
 738         if (nr_io_queues == 0) {
 739                 dev_err(ctrl->ctrl.device,
 740                         "unable to set any I/O queues\n");
 741                 return -ENOMEM;
 742         }
 743
 744         ctrl->ctrl.queue_count = nr_io_queues + 1;
 745         dev_info(ctrl->ctrl.device,
 746                 "creating %d I/O queues.\n", nr_io_queues);
 747
 748         if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
 749                 /*
 750                  * separate read/write queues
 751                  * hand out dedicated default queues only after we have
 752                  * sufficient read queues.
 753                  */
 754                 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
 755                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
 756                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 757                         min(nr_default_queues, nr_io_queues);
 758                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 759         } else {
 760                 /*
 761                  * shared read/write queues
 762                  * either no write queues were requested, or we don't have
 763                  * sufficient queue count to have dedicated default queues.
 764                  */
 765                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 766                         min(nr_read_queues, nr_io_queues);
 767                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 768         }
 769
 770         if (opts->nr_poll_queues && nr_io_queues) {
 771                 /* map dedicated poll queues only if we have queues left */
 772                 ctrl->io_queues[HCTX_TYPE_POLL] =
 773                         min(nr_poll_queues, nr_io_queues);
 774         }
 775
 776         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 777                 ret = nvme_rdma_alloc_queue(ctrl, i,
 778                                 ctrl->ctrl.sqsize + 1);
 779                 if (ret)
 780                         goto out_free_queues;
 781         }
 782
 783         return 0;
 784
 785 out_free_queues:
 786         for (i--; i >= 1; i--)
 787                 nvme_rdma_free_queue(&ctrl->queues[i]);
 788
 789         return ret;
 790 }
 791
 792 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 793                 bool admin)
 794 {
 795         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 796         struct blk_mq_tag_set *set;
 797         int ret;
 798
 799         if (admin) {
 800                 set = &ctrl->admin_tag_set;
 801                 memset(set, 0, sizeof(*set));
 802                 set->ops = &nvme_rdma_admin_mq_ops;
 803                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 804                 set->reserved_tags = NVMF_RESERVED_TAGS;
 805                 set->numa_node = nctrl->numa_node;
 806                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 807                                 NVME_RDMA_DATA_SGL_SIZE;
 808                 set->driver_data = ctrl;
 809                 set->nr_hw_queues = 1;
 810                 set->timeout = NVME_ADMIN_TIMEOUT;
 811                 set->flags = BLK_MQ_F_NO_SCHED;
 812         } else {
 813                 set = &ctrl->tag_set;
 814                 memset(set, 0, sizeof(*set));
 815                 set->ops = &nvme_rdma_mq_ops;
 816                 set->queue_depth = nctrl->sqsize + 1;
 817                 set->reserved_tags = NVMF_RESERVED_TAGS;
 818                 set->numa_node = nctrl->numa_node;
 819                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 820                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 821                                 NVME_RDMA_DATA_SGL_SIZE;
 822                 if (nctrl->max_integrity_segments)
 823                         set->cmd_size += sizeof(struct nvme_rdma_sgl) +
 824                                          NVME_RDMA_METADATA_SGL_SIZE;
 825                 set->driver_data = ctrl;
 826                 set->nr_hw_queues = nctrl->queue_count - 1;
 827                 set->timeout = NVME_IO_TIMEOUT;
 828                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
 829         }
 830
 831         ret = blk_mq_alloc_tag_set(set);
 832         if (ret)
 833                 return ERR_PTR(ret);
 834
 835         return set;
 836 }
 837
 838 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 839                 bool remove)
 840 {
 841         if (remove) {
 842                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 843                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 844                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 845         }
 846         if (ctrl->async_event_sqe.data) {
 847                 cancel_work_sync(&ctrl->ctrl.async_event_work);
 848                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 849                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 850                 ctrl->async_event_sqe.data = NULL;
 851         }
 852         nvme_rdma_free_queue(&ctrl->queues[0]);
 853 }
 854
 855 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 856                 bool new)
 857 {
 858         bool pi_capable = false;
 859         int error;
 860
 861         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 862         if (error)
 863                 return error;
 864
 865         ctrl->device = ctrl->queues[0].device;
 866         ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
 867
 868         /* T10-PI support */
 869         if (ctrl->device->dev->attrs.device_cap_flags &
 870             IB_DEVICE_INTEGRITY_HANDOVER)
 871                 pi_capable = true;
 872
 873         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
 874                                                         pi_capable);
 875
 876         /*
 877          * Bind the async event SQE DMA mapping to the admin queue lifetime.
 878          * It's safe, since any chage in the underlying RDMA device will issue
 879          * error recovery and queue re-creation.
 880          */
 881         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 882                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 883         if (error)
 884                 goto out_free_queue;
 885
 886         if (new) {
 887                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 888                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 889                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 890                         goto out_free_async_qe;
 891                 }
 892
 893                 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 894                 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 895                         error = PTR_ERR(ctrl->ctrl.fabrics_q);
 896                         goto out_free_tagset;
 897                 }
 898
 899                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 900                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 901                         error = PTR_ERR(ctrl->ctrl.admin_q);
 902                         goto out_cleanup_fabrics_q;
 903                 }
 904         }
 905
 906         error = nvme_rdma_start_queue(ctrl, 0);
 907         if (error)
 908                 goto out_cleanup_queue;
 909
 910         error = nvme_enable_ctrl(&ctrl->ctrl);
 911         if (error)
 912                 goto out_stop_queue;
 913
 914         ctrl->ctrl.max_segments = ctrl->max_fr_pages;
 915         ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
 916         if (pi_capable)
 917                 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
 918         else
 919                 ctrl->ctrl.max_integrity_segments = 0;
 920
 921         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 922
 923         error = nvme_init_ctrl_finish(&ctrl->ctrl);
 924         if (error)
 925                 goto out_quiesce_queue;
 926
 927         return 0;
 928
 929 out_quiesce_queue:
 930         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 931         blk_sync_queue(ctrl->ctrl.admin_q);
 932 out_stop_queue:
 933         nvme_rdma_stop_queue(&ctrl->queues[0]);
 934         nvme_cancel_admin_tagset(&ctrl->ctrl);
 935 out_cleanup_queue:
 936         if (new)
 937                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 938 out_cleanup_fabrics_q:
 939         if (new)
 940                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 941 out_free_tagset:
 942         if (new)
 943                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 944 out_free_async_qe:
 945         if (ctrl->async_event_sqe.data) {
 946                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 947                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 948                 ctrl->async_event_sqe.data = NULL;
 949         }
 950 out_free_queue:
 951         nvme_rdma_free_queue(&ctrl->queues[0]);
 952         return error;
 953 }
 954
 955 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 956                 bool remove)
 957 {
 958         if (remove) {
 959                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 960                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 961         }
 962         nvme_rdma_free_io_queues(ctrl);
 963 }
 964
 965 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 966 {
 967         int ret;
 968
 969         ret = nvme_rdma_alloc_io_queues(ctrl);
 970         if (ret)
 971                 return ret;
 972
 973         if (new) {
 974                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 975                 if (IS_ERR(ctrl->ctrl.tagset)) {
 976                         ret = PTR_ERR(ctrl->ctrl.tagset);
 977                         goto out_free_io_queues;
 978                 }
 979
 980                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 981                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 982                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 983                         goto out_free_tag_set;
 984                 }
 985         }
 986
 987         ret = nvme_rdma_start_io_queues(ctrl);
 988         if (ret)
 989                 goto out_cleanup_connect_q;
 990
 991         if (!new) {
 992                 nvme_start_queues(&ctrl->ctrl);
 993                 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
 994                         /*
 995                          * If we timed out waiting for freeze we are likely to
 996                          * be stuck.  Fail the controller initialization just
 997                          * to be safe.
 998                          */
 999                         ret = -ENODEV;
1000                         goto out_wait_freeze_timed_out;
1001                 }
1002                 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
1003                         ctrl->ctrl.queue_count - 1);
1004                 nvme_unfreeze(&ctrl->ctrl);
1005         }
1006
1007         return 0;
1008
1009 out_wait_freeze_timed_out:
1010         nvme_stop_queues(&ctrl->ctrl);
1011         nvme_sync_io_queues(&ctrl->ctrl);
1012         nvme_rdma_stop_io_queues(ctrl);
1013 out_cleanup_connect_q:
1014         nvme_cancel_tagset(&ctrl->ctrl);
1015         if (new)
1016                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1017 out_free_tag_set:
1018         if (new)
1019                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1020 out_free_io_queues:
1021         nvme_rdma_free_io_queues(ctrl);
1022         return ret;
1023 }
1024
1025 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1026                 bool remove)
1027 {
1028         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1029         blk_sync_queue(ctrl->ctrl.admin_q);
1030         nvme_rdma_stop_queue(&ctrl->queues[0]);
1031         nvme_cancel_admin_tagset(&ctrl->ctrl);
1032         if (remove)
1033                 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1034         nvme_rdma_destroy_admin_queue(ctrl, remove);
1035 }
1036
1037 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1038                 bool remove)
1039 {
1040         if (ctrl->ctrl.queue_count > 1) {
1041                 nvme_start_freeze(&ctrl->ctrl);
1042                 nvme_stop_queues(&ctrl->ctrl);
1043                 nvme_sync_io_queues(&ctrl->ctrl);
1044                 nvme_rdma_stop_io_queues(ctrl);
1045                 nvme_cancel_tagset(&ctrl->ctrl);
1046                 if (remove)
1047                         nvme_start_queues(&ctrl->ctrl);
1048                 nvme_rdma_destroy_io_queues(ctrl, remove);
1049         }
1050 }
1051
1052 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1053 {
1054         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1055
1056         if (list_empty(&ctrl->list))
1057                 goto free_ctrl;
1058
1059         mutex_lock(&nvme_rdma_ctrl_mutex);
1060         list_del(&ctrl->list);
1061         mutex_unlock(&nvme_rdma_ctrl_mutex);
1062
1063         nvmf_free_options(nctrl->opts);
1064 free_ctrl:
1065         kfree(ctrl->queues);
1066         kfree(ctrl);
1067 }
1068
1069 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1070 {
1071         /* If we are resetting/deleting then do nothing */
1072         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1073                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1074                         ctrl->ctrl.state == NVME_CTRL_LIVE);
1075                 return;
1076         }
1077
1078         if (nvmf_should_reconnect(&ctrl->ctrl)) {
1079                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1080                         ctrl->ctrl.opts->reconnect_delay);
1081                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1082                                 ctrl->ctrl.opts->reconnect_delay * HZ);
1083         } else {
1084                 nvme_delete_ctrl(&ctrl->ctrl);
1085         }
1086 }
1087
1088 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1089 {
1090         int ret;
1091         bool changed;
1092
1093         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1094         if (ret)
1095                 return ret;
1096
1097         if (ctrl->ctrl.icdoff) {
1098                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1099                 goto destroy_admin;
1100         }
1101
1102         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1103                 dev_err(ctrl->ctrl.device,
1104                         "Mandatory keyed sgls are not supported!\n");
1105                 goto destroy_admin;
1106         }
1107
1108         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1109                 dev_warn(ctrl->ctrl.device,
1110                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1111                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1112         }
1113
1114         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1115                 dev_warn(ctrl->ctrl.device,
1116                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1117                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1118                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1119         }
1120
1121         if (ctrl->ctrl.sgls & (1 << 20))
1122                 ctrl->use_inline_data = true;
1123
1124         if (ctrl->ctrl.queue_count > 1) {
1125                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1126                 if (ret)
1127                         goto destroy_admin;
1128         }
1129
1130         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1131         if (!changed) {
1132                 /*
1133                  * state change failure is ok if we started ctrl delete,
1134                  * unless we're during creation of a new controller to
1135                  * avoid races with teardown flow.
1136                  */
1137                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1138                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1139                 WARN_ON_ONCE(new);
1140                 ret = -EINVAL;
1141                 goto destroy_io;
1142         }
1143
1144         nvme_start_ctrl(&ctrl->ctrl);
1145         return 0;
1146
1147 destroy_io:
1148         if (ctrl->ctrl.queue_count > 1) {
1149                 nvme_stop_queues(&ctrl->ctrl);
1150                 nvme_sync_io_queues(&ctrl->ctrl);
1151                 nvme_rdma_stop_io_queues(ctrl);
1152                 nvme_cancel_tagset(&ctrl->ctrl);
1153                 nvme_rdma_destroy_io_queues(ctrl, new);
1154         }
1155 destroy_admin:
1156         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1157         blk_sync_queue(ctrl->ctrl.admin_q);
1158         nvme_rdma_stop_queue(&ctrl->queues[0]);
1159         nvme_cancel_admin_tagset(&ctrl->ctrl);
1160         nvme_rdma_destroy_admin_queue(ctrl, new);
1161         return ret;
1162 }
1163
1164 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1165 {
1166         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1167                         struct nvme_rdma_ctrl, reconnect_work);
1168
1169         ++ctrl->ctrl.nr_reconnects;
1170
1171         if (nvme_rdma_setup_ctrl(ctrl, false))
1172                 goto requeue;
1173
1174         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1175                         ctrl->ctrl.nr_reconnects);
1176
1177         ctrl->ctrl.nr_reconnects = 0;
1178
1179         return;
1180
1181 requeue:
1182         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1183                         ctrl->ctrl.nr_reconnects);
1184         nvme_rdma_reconnect_or_remove(ctrl);
1185 }
1186
1187 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1188 {
1189         struct nvme_rdma_ctrl *ctrl = container_of(work,
1190                         struct nvme_rdma_ctrl, err_work);
1191
1192         nvme_stop_keep_alive(&ctrl->ctrl);
1193         nvme_rdma_teardown_io_queues(ctrl, false);
1194         nvme_start_queues(&ctrl->ctrl);
1195         nvme_rdma_teardown_admin_queue(ctrl, false);
1196         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1197
1198         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1199                 /* state change failure is ok if we started ctrl delete */
1200                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1201                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1202                 return;
1203         }
1204
1205         nvme_rdma_reconnect_or_remove(ctrl);
1206 }
1207
1208 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1209 {
1210         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1211                 return;
1212
1213         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1214         queue_work(nvme_reset_wq, &ctrl->err_work);
1215 }
1216
1217 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1218 {
1219         struct request *rq = blk_mq_rq_from_pdu(req);
1220
1221         if (!refcount_dec_and_test(&req->ref))
1222                 return;
1223         if (!nvme_try_complete_req(rq, req->status, req->result))
1224                 nvme_rdma_complete_rq(rq);
1225 }
1226
1227 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1228                 const char *op)
1229 {
1230         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1231         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1232
1233         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1234                 dev_info(ctrl->ctrl.device,
1235                              "%s for CQE 0x%p failed with status %s (%d)\n",
1236                              op, wc->wr_cqe,
1237                              ib_wc_status_msg(wc->status), wc->status);
1238         nvme_rdma_error_recovery(ctrl);
1239 }
1240
1241 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1242 {
1243         if (unlikely(wc->status != IB_WC_SUCCESS))
1244                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1245 }
1246
1247 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1248 {
1249         struct nvme_rdma_request *req =
1250                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1251
1252         if (unlikely(wc->status != IB_WC_SUCCESS))
1253                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1254         else
1255                 nvme_rdma_end_request(req);
1256 }
1257
1258 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1259                 struct nvme_rdma_request *req)
1260 {
1261         struct ib_send_wr wr = {
1262                 .opcode             = IB_WR_LOCAL_INV,
1263                 .next               = NULL,
1264                 .num_sge            = 0,
1265                 .send_flags         = IB_SEND_SIGNALED,
1266                 .ex.invalidate_rkey = req->mr->rkey,
1267         };
1268
1269         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1270         wr.wr_cqe = &req->reg_cqe;
1271
1272         return ib_post_send(queue->qp, &wr, NULL);
1273 }
1274
1275 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1276                 struct request *rq)
1277 {
1278         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1279         struct nvme_rdma_device *dev = queue->device;
1280         struct ib_device *ibdev = dev->dev;
1281         struct list_head *pool = &queue->qp->rdma_mrs;
1282
1283         if (!blk_rq_nr_phys_segments(rq))
1284                 return;
1285
1286         if (blk_integrity_rq(rq)) {
1287                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1288                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1289                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1290                                       NVME_INLINE_METADATA_SG_CNT);
1291         }
1292
1293         if (req->use_sig_mr)
1294                 pool = &queue->qp->sig_mrs;
1295
1296         if (req->mr) {
1297                 ib_mr_pool_put(queue->qp, pool, req->mr);
1298                 req->mr = NULL;
1299         }
1300
1301         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1302                         rq_dma_dir(rq));
1303         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1304 }
1305
1306 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1307 {
1308         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1309
1310         sg->addr = 0;
1311         put_unaligned_le24(0, sg->length);
1312         put_unaligned_le32(0, sg->key);
1313         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1314         return 0;
1315 }
1316
1317 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1318                 struct nvme_rdma_request *req, struct nvme_command *c,
1319                 int count)
1320 {
1321         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1322         struct ib_sge *sge = &req->sge[1];
1323         struct scatterlist *sgl;
1324         u32 len = 0;
1325         int i;
1326
1327         for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1328                 sge->addr = sg_dma_address(sgl);
1329                 sge->length = sg_dma_len(sgl);
1330                 sge->lkey = queue->device->pd->local_dma_lkey;
1331                 len += sge->length;
1332                 sge++;
1333         }
1334
1335         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1336         sg->length = cpu_to_le32(len);
1337         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1338
1339         req->num_sge += count;
1340         return 0;
1341 }
1342
1343 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1344                 struct nvme_rdma_request *req, struct nvme_command *c)
1345 {
1346         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1347
1348         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1349         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1350         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1351         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1352         return 0;
1353 }
1354
1355 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1356                 struct nvme_rdma_request *req, struct nvme_command *c,
1357                 int count)
1358 {
1359         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1360         int nr;
1361
1362         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1363         if (WARN_ON_ONCE(!req->mr))
1364                 return -EAGAIN;
1365
1366         /*
1367          * Align the MR to a 4K page size to match the ctrl page size and
1368          * the block virtual boundary.
1369          */
1370         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1371                           SZ_4K);
1372         if (unlikely(nr < count)) {
1373                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1374                 req->mr = NULL;
1375                 if (nr < 0)
1376                         return nr;
1377                 return -EINVAL;
1378         }
1379
1380         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1381
1382         req->reg_cqe.done = nvme_rdma_memreg_done;
1383         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1384         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1385         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1386         req->reg_wr.wr.num_sge = 0;
1387         req->reg_wr.mr = req->mr;
1388         req->reg_wr.key = req->mr->rkey;
1389         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1390                              IB_ACCESS_REMOTE_READ |
1391                              IB_ACCESS_REMOTE_WRITE;
1392
1393         sg->addr = cpu_to_le64(req->mr->iova);
1394         put_unaligned_le24(req->mr->length, sg->length);
1395         put_unaligned_le32(req->mr->rkey, sg->key);
1396         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1397                         NVME_SGL_FMT_INVALIDATE;
1398
1399         return 0;
1400 }
1401
1402 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1403                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1404                 u16 control, u8 pi_type)
1405 {
1406         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1407         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1408         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1409         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1410         if (control & NVME_RW_PRINFO_PRCHK_REF)
1411                 domain->sig.dif.ref_remap = true;
1412
1413         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1414         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1415         domain->sig.dif.app_escape = true;
1416         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1417                 domain->sig.dif.ref_escape = true;
1418 }
1419
1420 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1421                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1422                 u8 pi_type)
1423 {
1424         u16 control = le16_to_cpu(cmd->rw.control);
1425
1426         memset(sig_attrs, 0, sizeof(*sig_attrs));
1427         if (control & NVME_RW_PRINFO_PRACT) {
1428                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1429                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1430                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1431                                          pi_type);
1432                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1433                 control &= ~NVME_RW_PRINFO_PRACT;
1434                 cmd->rw.control = cpu_to_le16(control);
1435         } else {
1436                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1437                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1438                                          pi_type);
1439                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1440                                          pi_type);
1441         }
1442 }
1443
1444 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1445 {
1446         *mask = 0;
1447         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1448                 *mask |= IB_SIG_CHECK_REFTAG;
1449         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1450                 *mask |= IB_SIG_CHECK_GUARD;
1451 }
1452
1453 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1454 {
1455         if (unlikely(wc->status != IB_WC_SUCCESS))
1456                 nvme_rdma_wr_error(cq, wc, "SIG");
1457 }
1458
1459 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1460                 struct nvme_rdma_request *req, struct nvme_command *c,
1461                 int count, int pi_count)
1462 {
1463         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1464         struct ib_reg_wr *wr = &req->reg_wr;
1465         struct request *rq = blk_mq_rq_from_pdu(req);
1466         struct nvme_ns *ns = rq->q->queuedata;
1467         struct bio *bio = rq->bio;
1468         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1469         int nr;
1470
1471         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1472         if (WARN_ON_ONCE(!req->mr))
1473                 return -EAGAIN;
1474
1475         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1476                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1477                              SZ_4K);
1478         if (unlikely(nr))
1479                 goto mr_put;
1480
1481         nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1482                                 req->mr->sig_attrs, ns->pi_type);
1483         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1484
1485         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1486
1487         req->reg_cqe.done = nvme_rdma_sig_done;
1488         memset(wr, 0, sizeof(*wr));
1489         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1490         wr->wr.wr_cqe = &req->reg_cqe;
1491         wr->wr.num_sge = 0;
1492         wr->wr.send_flags = 0;
1493         wr->mr = req->mr;
1494         wr->key = req->mr->rkey;
1495         wr->access = IB_ACCESS_LOCAL_WRITE |
1496                      IB_ACCESS_REMOTE_READ |
1497                      IB_ACCESS_REMOTE_WRITE;
1498
1499         sg->addr = cpu_to_le64(req->mr->iova);
1500         put_unaligned_le24(req->mr->length, sg->length);
1501         put_unaligned_le32(req->mr->rkey, sg->key);
1502         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1503
1504         return 0;
1505
1506 mr_put:
1507         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1508         req->mr = NULL;
1509         if (nr < 0)
1510                 return nr;
1511         return -EINVAL;
1512 }
1513
1514 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1515                 struct request *rq, struct nvme_command *c)
1516 {
1517         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1518         struct nvme_rdma_device *dev = queue->device;
1519         struct ib_device *ibdev = dev->dev;
1520         int pi_count = 0;
1521         int count, ret;
1522
1523         req->num_sge = 1;
1524         refcount_set(&req->ref, 2); /* send and recv completions */
1525
1526         c->common.flags |= NVME_CMD_SGL_METABUF;
1527
1528         if (!blk_rq_nr_phys_segments(rq))
1529                 return nvme_rdma_set_sg_null(c);
1530
1531         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1532         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1533                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1534                         NVME_INLINE_SG_CNT);
1535         if (ret)
1536                 return -ENOMEM;
1537
1538         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1539                                             req->data_sgl.sg_table.sgl);
1540
1541         count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1542                               req->data_sgl.nents, rq_dma_dir(rq));
1543         if (unlikely(count <= 0)) {
1544                 ret = -EIO;
1545                 goto out_free_table;
1546         }
1547
1548         if (blk_integrity_rq(rq)) {
1549                 req->metadata_sgl->sg_table.sgl =
1550                         (struct scatterlist *)(req->metadata_sgl + 1);
1551                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1552                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1553                                 req->metadata_sgl->sg_table.sgl,
1554                                 NVME_INLINE_METADATA_SG_CNT);
1555                 if (unlikely(ret)) {
1556                         ret = -ENOMEM;
1557                         goto out_unmap_sg;
1558                 }
1559
1560                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1561                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1562                 pi_count = ib_dma_map_sg(ibdev,
1563                                          req->metadata_sgl->sg_table.sgl,
1564                                          req->metadata_sgl->nents,
1565                                          rq_dma_dir(rq));
1566                 if (unlikely(pi_count <= 0)) {
1567                         ret = -EIO;
1568                         goto out_free_pi_table;
1569                 }
1570         }
1571
1572         if (req->use_sig_mr) {
1573                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1574                 goto out;
1575         }
1576
1577         if (count <= dev->num_inline_segments) {
1578                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1579                     queue->ctrl->use_inline_data &&
1580                     blk_rq_payload_bytes(rq) <=
1581                                 nvme_rdma_inline_data_size(queue)) {
1582                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1583                         goto out;
1584                 }
1585
1586                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1587                         ret = nvme_rdma_map_sg_single(queue, req, c);
1588                         goto out;
1589                 }
1590         }
1591
1592         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1593 out:
1594         if (unlikely(ret))
1595                 goto out_unmap_pi_sg;
1596
1597         return 0;
1598
1599 out_unmap_pi_sg:
1600         if (blk_integrity_rq(rq))
1601                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1602                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1603 out_free_pi_table:
1604         if (blk_integrity_rq(rq))
1605                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1606                                       NVME_INLINE_METADATA_SG_CNT);
1607 out_unmap_sg:
1608         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1609                         rq_dma_dir(rq));
1610 out_free_table:
1611         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1612         return ret;
1613 }
1614
1615 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1616 {
1617         struct nvme_rdma_qe *qe =
1618                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1619         struct nvme_rdma_request *req =
1620                 container_of(qe, struct nvme_rdma_request, sqe);
1621
1622         if (unlikely(wc->status != IB_WC_SUCCESS))
1623                 nvme_rdma_wr_error(cq, wc, "SEND");
1624         else
1625                 nvme_rdma_end_request(req);
1626 }
1627
1628 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1629                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1630                 struct ib_send_wr *first)
1631 {
1632         struct ib_send_wr wr;
1633         int ret;
1634
1635         sge->addr   = qe->dma;
1636         sge->length = sizeof(struct nvme_command);
1637         sge->lkey   = queue->device->pd->local_dma_lkey;
1638
1639         wr.next       = NULL;
1640         wr.wr_cqe     = &qe->cqe;
1641         wr.sg_list    = sge;
1642         wr.num_sge    = num_sge;
1643         wr.opcode     = IB_WR_SEND;
1644         wr.send_flags = IB_SEND_SIGNALED;
1645
1646         if (first)
1647                 first->next = &wr;
1648         else
1649                 first = &wr;
1650
1651         ret = ib_post_send(queue->qp, first, NULL);
1652         if (unlikely(ret)) {
1653                 dev_err(queue->ctrl->ctrl.device,
1654                              "%s failed with error code %d\n", __func__, ret);
1655         }
1656         return ret;
1657 }
1658
1659 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1660                 struct nvme_rdma_qe *qe)
1661 {
1662         struct ib_recv_wr wr;
1663         struct ib_sge list;
1664         int ret;
1665
1666         list.addr   = qe->dma;
1667         list.length = sizeof(struct nvme_completion);
1668         list.lkey   = queue->device->pd->local_dma_lkey;
1669
1670         qe->cqe.done = nvme_rdma_recv_done;
1671
1672         wr.next     = NULL;
1673         wr.wr_cqe   = &qe->cqe;
1674         wr.sg_list  = &list;
1675         wr.num_sge  = 1;
1676
1677         ret = ib_post_recv(queue->qp, &wr, NULL);
1678         if (unlikely(ret)) {
1679                 dev_err(queue->ctrl->ctrl.device,
1680                         "%s failed with error code %d\n", __func__, ret);
1681         }
1682         return ret;
1683 }
1684
1685 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1686 {
1687         u32 queue_idx = nvme_rdma_queue_idx(queue);
1688
1689         if (queue_idx == 0)
1690                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1691         return queue->ctrl->tag_set.tags[queue_idx - 1];
1692 }
1693
1694 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1695 {
1696         if (unlikely(wc->status != IB_WC_SUCCESS))
1697                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1698 }
1699
1700 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1701 {
1702         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1703         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1704         struct ib_device *dev = queue->device->dev;
1705         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1706         struct nvme_command *cmd = sqe->data;
1707         struct ib_sge sge;
1708         int ret;
1709
1710         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1711
1712         memset(cmd, 0, sizeof(*cmd));
1713         cmd->common.opcode = nvme_admin_async_event;
1714         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1715         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1716         nvme_rdma_set_sg_null(cmd);
1717
1718         sqe->cqe.done = nvme_rdma_async_done;
1719
1720         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1721                         DMA_TO_DEVICE);
1722
1723         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1724         WARN_ON_ONCE(ret);
1725 }
1726
1727 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1728                 struct nvme_completion *cqe, struct ib_wc *wc)
1729 {
1730         struct request *rq;
1731         struct nvme_rdma_request *req;
1732
1733         rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1734         if (!rq) {
1735                 dev_err(queue->ctrl->ctrl.device,
1736                         "got bad command_id %#x on QP %#x\n",
1737                         cqe->command_id, queue->qp->qp_num);
1738                 nvme_rdma_error_recovery(queue->ctrl);
1739                 return;
1740         }
1741         req = blk_mq_rq_to_pdu(rq);
1742
1743         req->status = cqe->status;
1744         req->result = cqe->result;
1745
1746         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1747                 if (unlikely(!req->mr ||
1748                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1749                         dev_err(queue->ctrl->ctrl.device,
1750                                 "Bogus remote invalidation for rkey %#x\n",
1751                                 req->mr ? req->mr->rkey : 0);
1752                         nvme_rdma_error_recovery(queue->ctrl);
1753                 }
1754         } else if (req->mr) {
1755                 int ret;
1756
1757                 ret = nvme_rdma_inv_rkey(queue, req);
1758                 if (unlikely(ret < 0)) {
1759                         dev_err(queue->ctrl->ctrl.device,
1760                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1761                                 req->mr->rkey, ret);
1762                         nvme_rdma_error_recovery(queue->ctrl);
1763                 }
1764                 /* the local invalidation completion will end the request */
1765                 return;
1766         }
1767
1768         nvme_rdma_end_request(req);
1769 }
1770
1771 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1772 {
1773         struct nvme_rdma_qe *qe =
1774                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1775         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1776         struct ib_device *ibdev = queue->device->dev;
1777         struct nvme_completion *cqe = qe->data;
1778         const size_t len = sizeof(struct nvme_completion);
1779
1780         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1781                 nvme_rdma_wr_error(cq, wc, "RECV");
1782                 return;
1783         }
1784
1785         /* sanity checking for received data length */
1786         if (unlikely(wc->byte_len < len)) {
1787                 dev_err(queue->ctrl->ctrl.device,
1788                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1789                 nvme_rdma_error_recovery(queue->ctrl);
1790                 return;
1791         }
1792
1793         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1794         /*
1795          * AEN requests are special as they don't time out and can
1796          * survive any kind of queue freeze and often don't respond to
1797          * aborts.  We don't even bother to allocate a struct request
1798          * for them but rather special case them here.
1799          */
1800         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1801                                      cqe->command_id)))
1802                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1803                                 &cqe->result);
1804         else
1805                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1806         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1807
1808         nvme_rdma_post_recv(queue, qe);
1809 }
1810
1811 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1812 {
1813         int ret, i;
1814
1815         for (i = 0; i < queue->queue_size; i++) {
1816                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1817                 if (ret)
1818                         goto out_destroy_queue_ib;
1819         }
1820
1821         return 0;
1822
1823 out_destroy_queue_ib:
1824         nvme_rdma_destroy_queue_ib(queue);
1825         return ret;
1826 }
1827
1828 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1829                 struct rdma_cm_event *ev)
1830 {
1831         struct rdma_cm_id *cm_id = queue->cm_id;
1832         int status = ev->status;
1833         const char *rej_msg;
1834         const struct nvme_rdma_cm_rej *rej_data;
1835         u8 rej_data_len;
1836
1837         rej_msg = rdma_reject_msg(cm_id, status);
1838         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1839
1840         if (rej_data && rej_data_len >= sizeof(u16)) {
1841                 u16 sts = le16_to_cpu(rej_data->sts);
1842
1843                 dev_err(queue->ctrl->ctrl.device,
1844                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1845                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1846         } else {
1847                 dev_err(queue->ctrl->ctrl.device,
1848                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1849         }
1850
1851         return -ECONNRESET;
1852 }
1853
1854 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1855 {
1856         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1857         int ret;
1858
1859         ret = nvme_rdma_create_queue_ib(queue);
1860         if (ret)
1861                 return ret;
1862
1863         if (ctrl->opts->tos >= 0)
1864                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1865         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1866         if (ret) {
1867                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1868                         queue->cm_error);
1869                 goto out_destroy_queue;
1870         }
1871
1872         return 0;
1873
1874 out_destroy_queue:
1875         nvme_rdma_destroy_queue_ib(queue);
1876         return ret;
1877 }
1878
1879 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1880 {
1881         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1882         struct rdma_conn_param param = { };
1883         struct nvme_rdma_cm_req priv = { };
1884         int ret;
1885
1886         param.qp_num = queue->qp->qp_num;
1887         param.flow_control = 1;
1888
1889         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1890         /* maximum retry count */
1891         param.retry_count = 7;
1892         param.rnr_retry_count = 7;
1893         param.private_data = &priv;
1894         param.private_data_len = sizeof(priv);
1895
1896         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1897         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1898         /*
1899          * set the admin queue depth to the minimum size
1900          * specified by the Fabrics standard.
1901          */
1902         if (priv.qid == 0) {
1903                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1904                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1905         } else {
1906                 /*
1907                  * current interpretation of the fabrics spec
1908                  * is at minimum you make hrqsize sqsize+1, or a
1909                  * 1's based representation of sqsize.
1910                  */
1911                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1912                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1913         }
1914
1915         ret = rdma_connect_locked(queue->cm_id, &param);
1916         if (ret) {
1917                 dev_err(ctrl->ctrl.device,
1918                         "rdma_connect_locked failed (%d).\n", ret);
1919                 goto out_destroy_queue_ib;
1920         }
1921
1922         return 0;
1923
1924 out_destroy_queue_ib:
1925         nvme_rdma_destroy_queue_ib(queue);
1926         return ret;
1927 }
1928
1929 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1930                 struct rdma_cm_event *ev)
1931 {
1932         struct nvme_rdma_queue *queue = cm_id->context;
1933         int cm_error = 0;
1934
1935         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1936                 rdma_event_msg(ev->event), ev->event,
1937                 ev->status, cm_id);
1938
1939         switch (ev->event) {
1940         case RDMA_CM_EVENT_ADDR_RESOLVED:
1941                 cm_error = nvme_rdma_addr_resolved(queue);
1942                 break;
1943         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1944                 cm_error = nvme_rdma_route_resolved(queue);
1945                 break;
1946         case RDMA_CM_EVENT_ESTABLISHED:
1947                 queue->cm_error = nvme_rdma_conn_established(queue);
1948                 /* complete cm_done regardless of success/failure */
1949                 complete(&queue->cm_done);
1950                 return 0;
1951         case RDMA_CM_EVENT_REJECTED:
1952                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1953                 break;
1954         case RDMA_CM_EVENT_ROUTE_ERROR:
1955         case RDMA_CM_EVENT_CONNECT_ERROR:
1956         case RDMA_CM_EVENT_UNREACHABLE:
1957                 nvme_rdma_destroy_queue_ib(queue);
1958                 fallthrough;
1959         case RDMA_CM_EVENT_ADDR_ERROR:
1960                 dev_dbg(queue->ctrl->ctrl.device,
1961                         "CM error event %d\n", ev->event);
1962                 cm_error = -ECONNRESET;
1963                 break;
1964         case RDMA_CM_EVENT_DISCONNECTED:
1965         case RDMA_CM_EVENT_ADDR_CHANGE:
1966         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1967                 dev_dbg(queue->ctrl->ctrl.device,
1968                         "disconnect received - connection closed\n");
1969                 nvme_rdma_error_recovery(queue->ctrl);
1970                 break;
1971         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1972                 /* device removal is handled via the ib_client API */
1973                 break;
1974         default:
1975                 dev_err(queue->ctrl->ctrl.device,
1976                         "Unexpected RDMA CM event (%d)\n", ev->event);
1977                 nvme_rdma_error_recovery(queue->ctrl);
1978                 break;
1979         }
1980
1981         if (cm_error) {
1982                 queue->cm_error = cm_error;
1983                 complete(&queue->cm_done);
1984         }
1985
1986         return 0;
1987 }
1988
1989 static void nvme_rdma_complete_timed_out(struct request *rq)
1990 {
1991         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1992         struct nvme_rdma_queue *queue = req->queue;
1993
1994         nvme_rdma_stop_queue(queue);
1995         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
1996                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
1997                 blk_mq_complete_request(rq);
1998         }
1999 }
2000
2001 static enum blk_eh_timer_return
2002 nvme_rdma_timeout(struct request *rq, bool reserved)
2003 {
2004         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2005         struct nvme_rdma_queue *queue = req->queue;
2006         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
2007
2008         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
2009                  rq->tag, nvme_rdma_queue_idx(queue));
2010
2011         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2012                 /*
2013                  * If we are resetting, connecting or deleting we should
2014                  * complete immediately because we may block controller
2015                  * teardown or setup sequence
2016                  * - ctrl disable/shutdown fabrics requests
2017                  * - connect requests
2018                  * - initialization admin requests
2019                  * - I/O requests that entered after unquiescing and
2020                  *   the controller stopped responding
2021                  *
2022                  * All other requests should be cancelled by the error
2023                  * recovery work, so it's fine that we fail it here.
2024                  */
2025                 nvme_rdma_complete_timed_out(rq);
2026                 return BLK_EH_DONE;
2027         }
2028
2029         /*
2030          * LIVE state should trigger the normal error recovery which will
2031          * handle completing this request.
2032          */
2033         nvme_rdma_error_recovery(ctrl);
2034         return BLK_EH_RESET_TIMER;
2035 }
2036
2037 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2038                 const struct blk_mq_queue_data *bd)
2039 {
2040         struct nvme_ns *ns = hctx->queue->queuedata;
2041         struct nvme_rdma_queue *queue = hctx->driver_data;
2042         struct request *rq = bd->rq;
2043         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2044         struct nvme_rdma_qe *sqe = &req->sqe;
2045         struct nvme_command *c = nvme_req(rq)->cmd;
2046         struct ib_device *dev;
2047         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2048         blk_status_t ret;
2049         int err;
2050
2051         WARN_ON_ONCE(rq->tag < 0);
2052
2053         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2054                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2055
2056         dev = queue->device->dev;
2057
2058         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2059                                          sizeof(struct nvme_command),
2060                                          DMA_TO_DEVICE);
2061         err = ib_dma_mapping_error(dev, req->sqe.dma);
2062         if (unlikely(err))
2063                 return BLK_STS_RESOURCE;
2064
2065         ib_dma_sync_single_for_cpu(dev, sqe->dma,
2066                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2067
2068         ret = nvme_setup_cmd(ns, rq);
2069         if (ret)
2070                 goto unmap_qe;
2071
2072         blk_mq_start_request(rq);
2073
2074         if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2075             queue->pi_support &&
2076             (c->common.opcode == nvme_cmd_write ||
2077              c->common.opcode == nvme_cmd_read) &&
2078             nvme_ns_has_pi(ns))
2079                 req->use_sig_mr = true;
2080         else
2081                 req->use_sig_mr = false;
2082
2083         err = nvme_rdma_map_data(queue, rq, c);
2084         if (unlikely(err < 0)) {
2085                 dev_err(queue->ctrl->ctrl.device,
2086                              "Failed to map data (%d)\n", err);
2087                 goto err;
2088         }
2089
2090         sqe->cqe.done = nvme_rdma_send_done;
2091
2092         ib_dma_sync_single_for_device(dev, sqe->dma,
2093                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2094
2095         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2096                         req->mr ? &req->reg_wr.wr : NULL);
2097         if (unlikely(err))
2098                 goto err_unmap;
2099
2100         return BLK_STS_OK;
2101
2102 err_unmap:
2103         nvme_rdma_unmap_data(queue, rq);
2104 err:
2105         if (err == -EIO)
2106                 ret = nvme_host_path_error(rq);
2107         else if (err == -ENOMEM || err == -EAGAIN)
2108                 ret = BLK_STS_RESOURCE;
2109         else
2110                 ret = BLK_STS_IOERR;
2111         nvme_cleanup_cmd(rq);
2112 unmap_qe:
2113         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2114                             DMA_TO_DEVICE);
2115         return ret;
2116 }
2117
2118 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2119 {
2120         struct nvme_rdma_queue *queue = hctx->driver_data;
2121
2122         return ib_process_cq_direct(queue->ib_cq, -1);
2123 }
2124
2125 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2126 {
2127         struct request *rq = blk_mq_rq_from_pdu(req);
2128         struct ib_mr_status mr_status;
2129         int ret;
2130
2131         ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2132         if (ret) {
2133                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2134                 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2135                 return;
2136         }
2137
2138         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2139                 switch (mr_status.sig_err.err_type) {
2140                 case IB_SIG_BAD_GUARD:
2141                         nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2142                         break;
2143                 case IB_SIG_BAD_REFTAG:
2144                         nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2145                         break;
2146                 case IB_SIG_BAD_APPTAG:
2147                         nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2148                         break;
2149                 }
2150                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2151                        mr_status.sig_err.err_type, mr_status.sig_err.expected,
2152                        mr_status.sig_err.actual);
2153         }
2154 }
2155
2156 static void nvme_rdma_complete_rq(struct request *rq)
2157 {
2158         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2159         struct nvme_rdma_queue *queue = req->queue;
2160         struct ib_device *ibdev = queue->device->dev;
2161
2162         if (req->use_sig_mr)
2163                 nvme_rdma_check_pi_status(req);
2164
2165         nvme_rdma_unmap_data(queue, rq);
2166         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2167                             DMA_TO_DEVICE);
2168         nvme_complete_rq(rq);
2169 }
2170
2171 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2172 {
2173         struct nvme_rdma_ctrl *ctrl = set->driver_data;
2174         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2175
2176         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2177                 /* separate read/write queues */
2178                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2179                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2180                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2181                 set->map[HCTX_TYPE_READ].nr_queues =
2182                         ctrl->io_queues[HCTX_TYPE_READ];
2183                 set->map[HCTX_TYPE_READ].queue_offset =
2184                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2185         } else {
2186                 /* shared read/write queues */
2187                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2188                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2189                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2190                 set->map[HCTX_TYPE_READ].nr_queues =
2191                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2192                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2193         }
2194         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2195                         ctrl->device->dev, 0);
2196         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2197                         ctrl->device->dev, 0);
2198
2199         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2200                 /* map dedicated poll queues only if we have queues left */
2201                 set->map[HCTX_TYPE_POLL].nr_queues =
2202                                 ctrl->io_queues[HCTX_TYPE_POLL];
2203                 set->map[HCTX_TYPE_POLL].queue_offset =
2204                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2205                         ctrl->io_queues[HCTX_TYPE_READ];
2206                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2207         }
2208
2209         dev_info(ctrl->ctrl.device,
2210                 "mapped %d/%d/%d default/read/poll queues.\n",
2211                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2212                 ctrl->io_queues[HCTX_TYPE_READ],
2213                 ctrl->io_queues[HCTX_TYPE_POLL]);
2214
2215         return 0;
2216 }
2217
2218 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2219         .queue_rq       = nvme_rdma_queue_rq,
2220         .complete       = nvme_rdma_complete_rq,
2221         .init_request   = nvme_rdma_init_request,
2222         .exit_request   = nvme_rdma_exit_request,
2223         .init_hctx      = nvme_rdma_init_hctx,
2224         .timeout        = nvme_rdma_timeout,
2225         .map_queues     = nvme_rdma_map_queues,
2226         .poll           = nvme_rdma_poll,
2227 };
2228
2229 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2230         .queue_rq       = nvme_rdma_queue_rq,
2231         .complete       = nvme_rdma_complete_rq,
2232         .init_request   = nvme_rdma_init_request,
2233         .exit_request   = nvme_rdma_exit_request,
2234         .init_hctx      = nvme_rdma_init_admin_hctx,
2235         .timeout        = nvme_rdma_timeout,
2236 };
2237
2238 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2239 {
2240         cancel_work_sync(&ctrl->err_work);
2241         cancel_delayed_work_sync(&ctrl->reconnect_work);
2242
2243         nvme_rdma_teardown_io_queues(ctrl, shutdown);
2244         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2245         if (shutdown)
2246                 nvme_shutdown_ctrl(&ctrl->ctrl);
2247         else
2248                 nvme_disable_ctrl(&ctrl->ctrl);
2249         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2250 }
2251
2252 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2253 {
2254         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2255 }
2256
2257 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2258 {
2259         struct nvme_rdma_ctrl *ctrl =
2260                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2261
2262         nvme_stop_ctrl(&ctrl->ctrl);
2263         nvme_rdma_shutdown_ctrl(ctrl, false);
2264
2265         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2266                 /* state change failure should never happen */
2267                 WARN_ON_ONCE(1);
2268                 return;
2269         }
2270
2271         if (nvme_rdma_setup_ctrl(ctrl, false))
2272                 goto out_fail;
2273
2274         return;
2275
2276 out_fail:
2277         ++ctrl->ctrl.nr_reconnects;
2278         nvme_rdma_reconnect_or_remove(ctrl);
2279 }
2280
2281 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2282         .name                   = "rdma",
2283         .module                 = THIS_MODULE,
2284         .flags                  = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2285         .reg_read32             = nvmf_reg_read32,
2286         .reg_read64             = nvmf_reg_read64,
2287         .reg_write32            = nvmf_reg_write32,
2288         .free_ctrl              = nvme_rdma_free_ctrl,
2289         .submit_async_event     = nvme_rdma_submit_async_event,
2290         .delete_ctrl            = nvme_rdma_delete_ctrl,
2291         .get_address            = nvmf_get_address,
2292 };
2293
2294 /*
2295  * Fails a connection request if it matches an existing controller
2296  * (association) with the same tuple:
2297  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2298  *
2299  * if local address is not specified in the request, it will match an
2300  * existing controller with all the other parameters the same and no
2301  * local port address specified as well.
2302  *
2303  * The ports don't need to be compared as they are intrinsically
2304  * already matched by the port pointers supplied.
2305  */
2306 static bool
2307 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2308 {
2309         struct nvme_rdma_ctrl *ctrl;
2310         bool found = false;
2311
2312         mutex_lock(&nvme_rdma_ctrl_mutex);
2313         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2314                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2315                 if (found)
2316                         break;
2317         }
2318         mutex_unlock(&nvme_rdma_ctrl_mutex);
2319
2320         return found;
2321 }
2322
2323 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2324                 struct nvmf_ctrl_options *opts)
2325 {
2326         struct nvme_rdma_ctrl *ctrl;
2327         int ret;
2328         bool changed;
2329
2330         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2331         if (!ctrl)
2332                 return ERR_PTR(-ENOMEM);
2333         ctrl->ctrl.opts = opts;
2334         INIT_LIST_HEAD(&ctrl->list);
2335
2336         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2337                 opts->trsvcid =
2338                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2339                 if (!opts->trsvcid) {
2340                         ret = -ENOMEM;
2341                         goto out_free_ctrl;
2342                 }
2343                 opts->mask |= NVMF_OPT_TRSVCID;
2344         }
2345
2346         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2347                         opts->traddr, opts->trsvcid, &ctrl->addr);
2348         if (ret) {
2349                 pr_err("malformed address passed: %s:%s\n",
2350                         opts->traddr, opts->trsvcid);
2351                 goto out_free_ctrl;
2352         }
2353
2354         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2355                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2356                         opts->host_traddr, NULL, &ctrl->src_addr);
2357                 if (ret) {
2358                         pr_err("malformed src address passed: %s\n",
2359                                opts->host_traddr);
2360                         goto out_free_ctrl;
2361                 }
2362         }
2363
2364         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2365                 ret = -EALREADY;
2366                 goto out_free_ctrl;
2367         }
2368
2369         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2370                         nvme_rdma_reconnect_ctrl_work);
2371         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2372         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2373
2374         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2375                                 opts->nr_poll_queues + 1;
2376         ctrl->ctrl.sqsize = opts->queue_size - 1;
2377         ctrl->ctrl.kato = opts->kato;
2378
2379         ret = -ENOMEM;
2380         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2381                                 GFP_KERNEL);
2382         if (!ctrl->queues)
2383                 goto out_free_ctrl;
2384
2385         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2386                                 0 /* no quirks, we're perfect! */);
2387         if (ret)
2388                 goto out_kfree_queues;
2389
2390         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2391         WARN_ON_ONCE(!changed);
2392
2393         ret = nvme_rdma_setup_ctrl(ctrl, true);
2394         if (ret)
2395                 goto out_uninit_ctrl;
2396
2397         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2398                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2399
2400         mutex_lock(&nvme_rdma_ctrl_mutex);
2401         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2402         mutex_unlock(&nvme_rdma_ctrl_mutex);
2403
2404         return &ctrl->ctrl;
2405
2406 out_uninit_ctrl:
2407         nvme_uninit_ctrl(&ctrl->ctrl);
2408         nvme_put_ctrl(&ctrl->ctrl);
2409         if (ret > 0)
2410                 ret = -EIO;
2411         return ERR_PTR(ret);
2412 out_kfree_queues:
2413         kfree(ctrl->queues);
2414 out_free_ctrl:
2415         kfree(ctrl);
2416         return ERR_PTR(ret);
2417 }
2418
2419 static struct nvmf_transport_ops nvme_rdma_transport = {
2420         .name           = "rdma",
2421         .module         = THIS_MODULE,
2422         .required_opts  = NVMF_OPT_TRADDR,
2423         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2424                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2425                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2426                           NVMF_OPT_TOS,
2427         .create_ctrl    = nvme_rdma_create_ctrl,
2428 };
2429
2430 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2431 {
2432         struct nvme_rdma_ctrl *ctrl;
2433         struct nvme_rdma_device *ndev;
2434         bool found = false;
2435
2436         mutex_lock(&device_list_mutex);
2437         list_for_each_entry(ndev, &device_list, entry) {
2438                 if (ndev->dev == ib_device) {
2439                         found = true;
2440                         break;
2441                 }
2442         }
2443         mutex_unlock(&device_list_mutex);
2444
2445         if (!found)
2446                 return;
2447
2448         /* Delete all controllers using this device */
2449         mutex_lock(&nvme_rdma_ctrl_mutex);
2450         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2451                 if (ctrl->device->dev != ib_device)
2452                         continue;
2453                 nvme_delete_ctrl(&ctrl->ctrl);
2454         }
2455         mutex_unlock(&nvme_rdma_ctrl_mutex);
2456
2457         flush_workqueue(nvme_delete_wq);
2458 }
2459
2460 static struct ib_client nvme_rdma_ib_client = {
2461         .name   = "nvme_rdma",
2462         .remove = nvme_rdma_remove_one
2463 };
2464
2465 static int __init nvme_rdma_init_module(void)
2466 {
2467         int ret;
2468
2469         ret = ib_register_client(&nvme_rdma_ib_client);
2470         if (ret)
2471                 return ret;
2472
2473         ret = nvmf_register_transport(&nvme_rdma_transport);
2474         if (ret)
2475                 goto err_unreg_client;
2476
2477         return 0;
2478
2479 err_unreg_client:
2480         ib_unregister_client(&nvme_rdma_ib_client);
2481         return ret;
2482 }
2483
2484 static void __exit nvme_rdma_cleanup_module(void)
2485 {
2486         struct nvme_rdma_ctrl *ctrl;
2487
2488         nvmf_unregister_transport(&nvme_rdma_transport);
2489         ib_unregister_client(&nvme_rdma_ib_client);
2490
2491         mutex_lock(&nvme_rdma_ctrl_mutex);
2492         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2493                 nvme_delete_ctrl(&ctrl->ctrl);
2494         mutex_unlock(&nvme_rdma_ctrl_mutex);
2495         flush_workqueue(nvme_delete_wq);
2496 }
2497
2498 module_init(nvme_rdma_init_module);
2499 module_exit(nvme_rdma_cleanup_module);
2500
2501 MODULE_LICENSE("GPL v2");