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         bool poll = nvme_rdma_poll_queue(queue);
 684         int ret;
 685
 686         if (idx)
 687                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
 688         else
 689                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 690
 691         if (!ret) {
 692                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 693         } else {
 694                 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 695                         __nvme_rdma_stop_queue(queue);
 696                 dev_info(ctrl->ctrl.device,
 697                         "failed to connect queue: %d ret=%d\n", idx, ret);
 698         }
 699         return ret;
 700 }
 701
 702 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 703 {
 704         int i, ret = 0;
 705
 706         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 707                 ret = nvme_rdma_start_queue(ctrl, i);
 708                 if (ret)
 709                         goto out_stop_queues;
 710         }
 711
 712         return 0;
 713
 714 out_stop_queues:
 715         for (i--; i >= 1; i--)
 716                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 717         return ret;
 718 }
 719
 720 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 721 {
 722         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 723         struct ib_device *ibdev = ctrl->device->dev;
 724         unsigned int nr_io_queues, nr_default_queues;
 725         unsigned int nr_read_queues, nr_poll_queues;
 726         int i, ret;
 727
 728         nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
 729                                 min(opts->nr_io_queues, num_online_cpus()));
 730         nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
 731                                 min(opts->nr_write_queues, num_online_cpus()));
 732         nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
 733         nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
 734
 735         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 736         if (ret)
 737                 return ret;
 738
 739         ctrl->ctrl.queue_count = nr_io_queues + 1;
 740         if (ctrl->ctrl.queue_count < 2) {
 741                 dev_err(ctrl->ctrl.device,
 742                         "unable to set any I/O queues\n");
 743                 return -ENOMEM;
 744         }
 745
 746         dev_info(ctrl->ctrl.device,
 747                 "creating %d I/O queues.\n", nr_io_queues);
 748
 749         if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
 750                 /*
 751                  * separate read/write queues
 752                  * hand out dedicated default queues only after we have
 753                  * sufficient read queues.
 754                  */
 755                 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
 756                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
 757                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 758                         min(nr_default_queues, nr_io_queues);
 759                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 760         } else {
 761                 /*
 762                  * shared read/write queues
 763                  * either no write queues were requested, or we don't have
 764                  * sufficient queue count to have dedicated default queues.
 765                  */
 766                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 767                         min(nr_read_queues, nr_io_queues);
 768                 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
 769         }
 770
 771         if (opts->nr_poll_queues && nr_io_queues) {
 772                 /* map dedicated poll queues only if we have queues left */
 773                 ctrl->io_queues[HCTX_TYPE_POLL] =
 774                         min(nr_poll_queues, nr_io_queues);
 775         }
 776
 777         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 778                 ret = nvme_rdma_alloc_queue(ctrl, i,
 779                                 ctrl->ctrl.sqsize + 1);
 780                 if (ret)
 781                         goto out_free_queues;
 782         }
 783
 784         return 0;
 785
 786 out_free_queues:
 787         for (i--; i >= 1; i--)
 788                 nvme_rdma_free_queue(&ctrl->queues[i]);
 789
 790         return ret;
 791 }
 792
 793 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 794                 bool admin)
 795 {
 796         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 797         struct blk_mq_tag_set *set;
 798         int ret;
 799
 800         if (admin) {
 801                 set = &ctrl->admin_tag_set;
 802                 memset(set, 0, sizeof(*set));
 803                 set->ops = &nvme_rdma_admin_mq_ops;
 804                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 805                 set->reserved_tags = NVMF_RESERVED_TAGS;
 806                 set->numa_node = nctrl->numa_node;
 807                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 808                                 NVME_RDMA_DATA_SGL_SIZE;
 809                 set->driver_data = ctrl;
 810                 set->nr_hw_queues = 1;
 811                 set->timeout = NVME_ADMIN_TIMEOUT;
 812                 set->flags = BLK_MQ_F_NO_SCHED;
 813         } else {
 814                 set = &ctrl->tag_set;
 815                 memset(set, 0, sizeof(*set));
 816                 set->ops = &nvme_rdma_mq_ops;
 817                 set->queue_depth = nctrl->sqsize + 1;
 818                 set->reserved_tags = NVMF_RESERVED_TAGS;
 819                 set->numa_node = nctrl->numa_node;
 820                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 821                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 822                                 NVME_RDMA_DATA_SGL_SIZE;
 823                 if (nctrl->max_integrity_segments)
 824                         set->cmd_size += sizeof(struct nvme_rdma_sgl) +
 825                                          NVME_RDMA_METADATA_SGL_SIZE;
 826                 set->driver_data = ctrl;
 827                 set->nr_hw_queues = nctrl->queue_count - 1;
 828                 set->timeout = NVME_IO_TIMEOUT;
 829                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
 830         }
 831
 832         ret = blk_mq_alloc_tag_set(set);
 833         if (ret)
 834                 return ERR_PTR(ret);
 835
 836         return set;
 837 }
 838
 839 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 840                 bool remove)
 841 {
 842         if (remove) {
 843                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 844                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 845                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 846         }
 847         if (ctrl->async_event_sqe.data) {
 848                 cancel_work_sync(&ctrl->ctrl.async_event_work);
 849                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 850                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 851                 ctrl->async_event_sqe.data = NULL;
 852         }
 853         nvme_rdma_free_queue(&ctrl->queues[0]);
 854 }
 855
 856 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 857                 bool new)
 858 {
 859         bool pi_capable = false;
 860         int error;
 861
 862         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 863         if (error)
 864                 return error;
 865
 866         ctrl->device = ctrl->queues[0].device;
 867         ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
 868
 869         /* T10-PI support */
 870         if (ctrl->device->dev->attrs.device_cap_flags &
 871             IB_DEVICE_INTEGRITY_HANDOVER)
 872                 pi_capable = true;
 873
 874         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
 875                                                         pi_capable);
 876
 877         /*
 878          * Bind the async event SQE DMA mapping to the admin queue lifetime.
 879          * It's safe, since any chage in the underlying RDMA device will issue
 880          * error recovery and queue re-creation.
 881          */
 882         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 883                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 884         if (error)
 885                 goto out_free_queue;
 886
 887         if (new) {
 888                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 889                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 890                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 891                         goto out_free_async_qe;
 892                 }
 893
 894                 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 895                 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
 896                         error = PTR_ERR(ctrl->ctrl.fabrics_q);
 897                         goto out_free_tagset;
 898                 }
 899
 900                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 901                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 902                         error = PTR_ERR(ctrl->ctrl.admin_q);
 903                         goto out_cleanup_fabrics_q;
 904                 }
 905         }
 906
 907         error = nvme_rdma_start_queue(ctrl, 0);
 908         if (error)
 909                 goto out_cleanup_queue;
 910
 911         error = nvme_enable_ctrl(&ctrl->ctrl);
 912         if (error)
 913                 goto out_stop_queue;
 914
 915         ctrl->ctrl.max_segments = ctrl->max_fr_pages;
 916         ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
 917         if (pi_capable)
 918                 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
 919         else
 920                 ctrl->ctrl.max_integrity_segments = 0;
 921
 922         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 923
 924         error = nvme_init_ctrl_finish(&ctrl->ctrl);
 925         if (error)
 926                 goto out_quiesce_queue;
 927
 928         return 0;
 929
 930 out_quiesce_queue:
 931         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 932         blk_sync_queue(ctrl->ctrl.admin_q);
 933 out_stop_queue:
 934         nvme_rdma_stop_queue(&ctrl->queues[0]);
 935         nvme_cancel_admin_tagset(&ctrl->ctrl);
 936 out_cleanup_queue:
 937         if (new)
 938                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 939 out_cleanup_fabrics_q:
 940         if (new)
 941                 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
 942 out_free_tagset:
 943         if (new)
 944                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 945 out_free_async_qe:
 946         if (ctrl->async_event_sqe.data) {
 947                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 948                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 949                 ctrl->async_event_sqe.data = NULL;
 950         }
 951 out_free_queue:
 952         nvme_rdma_free_queue(&ctrl->queues[0]);
 953         return error;
 954 }
 955
 956 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 957                 bool remove)
 958 {
 959         if (remove) {
 960                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 961                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 962         }
 963         nvme_rdma_free_io_queues(ctrl);
 964 }
 965
 966 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 967 {
 968         int ret;
 969
 970         ret = nvme_rdma_alloc_io_queues(ctrl);
 971         if (ret)
 972                 return ret;
 973
 974         if (new) {
 975                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 976                 if (IS_ERR(ctrl->ctrl.tagset)) {
 977                         ret = PTR_ERR(ctrl->ctrl.tagset);
 978                         goto out_free_io_queues;
 979                 }
 980
 981                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 982                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 983                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 984                         goto out_free_tag_set;
 985                 }
 986         }
 987
 988         ret = nvme_rdma_start_io_queues(ctrl);
 989         if (ret)
 990                 goto out_cleanup_connect_q;
 991
 992         if (!new) {
 993                 nvme_start_queues(&ctrl->ctrl);
 994                 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
 995                         /*
 996                          * If we timed out waiting for freeze we are likely to
 997                          * be stuck.  Fail the controller initialization just
 998                          * to be safe.
 999                          */
1000                         ret = -ENODEV;
1001                         goto out_wait_freeze_timed_out;
1002                 }
1003                 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
1004                         ctrl->ctrl.queue_count - 1);
1005                 nvme_unfreeze(&ctrl->ctrl);
1006         }
1007
1008         return 0;
1009
1010 out_wait_freeze_timed_out:
1011         nvme_stop_queues(&ctrl->ctrl);
1012         nvme_sync_io_queues(&ctrl->ctrl);
1013         nvme_rdma_stop_io_queues(ctrl);
1014 out_cleanup_connect_q:
1015         nvme_cancel_tagset(&ctrl->ctrl);
1016         if (new)
1017                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1018 out_free_tag_set:
1019         if (new)
1020                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1021 out_free_io_queues:
1022         nvme_rdma_free_io_queues(ctrl);
1023         return ret;
1024 }
1025
1026 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1027                 bool remove)
1028 {
1029         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1030         blk_sync_queue(ctrl->ctrl.admin_q);
1031         nvme_rdma_stop_queue(&ctrl->queues[0]);
1032         nvme_cancel_admin_tagset(&ctrl->ctrl);
1033         if (remove)
1034                 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1035         nvme_rdma_destroy_admin_queue(ctrl, remove);
1036 }
1037
1038 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1039                 bool remove)
1040 {
1041         if (ctrl->ctrl.queue_count > 1) {
1042                 nvme_start_freeze(&ctrl->ctrl);
1043                 nvme_stop_queues(&ctrl->ctrl);
1044                 nvme_sync_io_queues(&ctrl->ctrl);
1045                 nvme_rdma_stop_io_queues(ctrl);
1046                 nvme_cancel_tagset(&ctrl->ctrl);
1047                 if (remove)
1048                         nvme_start_queues(&ctrl->ctrl);
1049                 nvme_rdma_destroy_io_queues(ctrl, remove);
1050         }
1051 }
1052
1053 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1054 {
1055         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1056
1057         if (list_empty(&ctrl->list))
1058                 goto free_ctrl;
1059
1060         mutex_lock(&nvme_rdma_ctrl_mutex);
1061         list_del(&ctrl->list);
1062         mutex_unlock(&nvme_rdma_ctrl_mutex);
1063
1064         nvmf_free_options(nctrl->opts);
1065 free_ctrl:
1066         kfree(ctrl->queues);
1067         kfree(ctrl);
1068 }
1069
1070 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1071 {
1072         /* If we are resetting/deleting then do nothing */
1073         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1074                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1075                         ctrl->ctrl.state == NVME_CTRL_LIVE);
1076                 return;
1077         }
1078
1079         if (nvmf_should_reconnect(&ctrl->ctrl)) {
1080                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1081                         ctrl->ctrl.opts->reconnect_delay);
1082                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1083                                 ctrl->ctrl.opts->reconnect_delay * HZ);
1084         } else {
1085                 nvme_delete_ctrl(&ctrl->ctrl);
1086         }
1087 }
1088
1089 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1090 {
1091         int ret = -EINVAL;
1092         bool changed;
1093
1094         ret = nvme_rdma_configure_admin_queue(ctrl, new);
1095         if (ret)
1096                 return ret;
1097
1098         if (ctrl->ctrl.icdoff) {
1099                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1100                 goto destroy_admin;
1101         }
1102
1103         if (!(ctrl->ctrl.sgls & (1 << 2))) {
1104                 dev_err(ctrl->ctrl.device,
1105                         "Mandatory keyed sgls are not supported!\n");
1106                 goto destroy_admin;
1107         }
1108
1109         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1110                 dev_warn(ctrl->ctrl.device,
1111                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1112                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1113         }
1114
1115         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1116                 dev_warn(ctrl->ctrl.device,
1117                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
1118                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1119                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1120         }
1121
1122         if (ctrl->ctrl.sgls & (1 << 20))
1123                 ctrl->use_inline_data = true;
1124
1125         if (ctrl->ctrl.queue_count > 1) {
1126                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1127                 if (ret)
1128                         goto destroy_admin;
1129         }
1130
1131         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1132         if (!changed) {
1133                 /*
1134                  * state change failure is ok if we started ctrl delete,
1135                  * unless we're during creation of a new controller to
1136                  * avoid races with teardown flow.
1137                  */
1138                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1139                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1140                 WARN_ON_ONCE(new);
1141                 ret = -EINVAL;
1142                 goto destroy_io;
1143         }
1144
1145         nvme_start_ctrl(&ctrl->ctrl);
1146         return 0;
1147
1148 destroy_io:
1149         if (ctrl->ctrl.queue_count > 1) {
1150                 nvme_stop_queues(&ctrl->ctrl);
1151                 nvme_sync_io_queues(&ctrl->ctrl);
1152                 nvme_rdma_stop_io_queues(ctrl);
1153                 nvme_cancel_tagset(&ctrl->ctrl);
1154                 nvme_rdma_destroy_io_queues(ctrl, new);
1155         }
1156 destroy_admin:
1157         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1158         blk_sync_queue(ctrl->ctrl.admin_q);
1159         nvme_rdma_stop_queue(&ctrl->queues[0]);
1160         nvme_cancel_admin_tagset(&ctrl->ctrl);
1161         nvme_rdma_destroy_admin_queue(ctrl, new);
1162         return ret;
1163 }
1164
1165 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1166 {
1167         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1168                         struct nvme_rdma_ctrl, reconnect_work);
1169
1170         ++ctrl->ctrl.nr_reconnects;
1171
1172         if (nvme_rdma_setup_ctrl(ctrl, false))
1173                 goto requeue;
1174
1175         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1176                         ctrl->ctrl.nr_reconnects);
1177
1178         ctrl->ctrl.nr_reconnects = 0;
1179
1180         return;
1181
1182 requeue:
1183         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1184                         ctrl->ctrl.nr_reconnects);
1185         nvme_rdma_reconnect_or_remove(ctrl);
1186 }
1187
1188 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1189 {
1190         struct nvme_rdma_ctrl *ctrl = container_of(work,
1191                         struct nvme_rdma_ctrl, err_work);
1192
1193         nvme_stop_keep_alive(&ctrl->ctrl);
1194         nvme_rdma_teardown_io_queues(ctrl, false);
1195         nvme_start_queues(&ctrl->ctrl);
1196         nvme_rdma_teardown_admin_queue(ctrl, false);
1197         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1198
1199         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1200                 /* state change failure is ok if we started ctrl delete */
1201                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1202                              ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1203                 return;
1204         }
1205
1206         nvme_rdma_reconnect_or_remove(ctrl);
1207 }
1208
1209 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1210 {
1211         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1212                 return;
1213
1214         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1215         queue_work(nvme_reset_wq, &ctrl->err_work);
1216 }
1217
1218 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1219 {
1220         struct request *rq = blk_mq_rq_from_pdu(req);
1221
1222         if (!refcount_dec_and_test(&req->ref))
1223                 return;
1224         if (!nvme_try_complete_req(rq, req->status, req->result))
1225                 nvme_rdma_complete_rq(rq);
1226 }
1227
1228 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1229                 const char *op)
1230 {
1231         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1232         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1233
1234         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1235                 dev_info(ctrl->ctrl.device,
1236                              "%s for CQE 0x%p failed with status %s (%d)\n",
1237                              op, wc->wr_cqe,
1238                              ib_wc_status_msg(wc->status), wc->status);
1239         nvme_rdma_error_recovery(ctrl);
1240 }
1241
1242 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1243 {
1244         if (unlikely(wc->status != IB_WC_SUCCESS))
1245                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1246 }
1247
1248 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1249 {
1250         struct nvme_rdma_request *req =
1251                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1252
1253         if (unlikely(wc->status != IB_WC_SUCCESS))
1254                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1255         else
1256                 nvme_rdma_end_request(req);
1257 }
1258
1259 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1260                 struct nvme_rdma_request *req)
1261 {
1262         struct ib_send_wr wr = {
1263                 .opcode             = IB_WR_LOCAL_INV,
1264                 .next               = NULL,
1265                 .num_sge            = 0,
1266                 .send_flags         = IB_SEND_SIGNALED,
1267                 .ex.invalidate_rkey = req->mr->rkey,
1268         };
1269
1270         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1271         wr.wr_cqe = &req->reg_cqe;
1272
1273         return ib_post_send(queue->qp, &wr, NULL);
1274 }
1275
1276 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1277                 struct request *rq)
1278 {
1279         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1280         struct nvme_rdma_device *dev = queue->device;
1281         struct ib_device *ibdev = dev->dev;
1282         struct list_head *pool = &queue->qp->rdma_mrs;
1283
1284         if (!blk_rq_nr_phys_segments(rq))
1285                 return;
1286
1287         if (blk_integrity_rq(rq)) {
1288                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1289                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1290                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1291                                       NVME_INLINE_METADATA_SG_CNT);
1292         }
1293
1294         if (req->use_sig_mr)
1295                 pool = &queue->qp->sig_mrs;
1296
1297         if (req->mr) {
1298                 ib_mr_pool_put(queue->qp, pool, req->mr);
1299                 req->mr = NULL;
1300         }
1301
1302         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1303                         rq_dma_dir(rq));
1304         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1305 }
1306
1307 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1308 {
1309         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1310
1311         sg->addr = 0;
1312         put_unaligned_le24(0, sg->length);
1313         put_unaligned_le32(0, sg->key);
1314         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1315         return 0;
1316 }
1317
1318 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1319                 struct nvme_rdma_request *req, struct nvme_command *c,
1320                 int count)
1321 {
1322         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1323         struct ib_sge *sge = &req->sge[1];
1324         struct scatterlist *sgl;
1325         u32 len = 0;
1326         int i;
1327
1328         for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1329                 sge->addr = sg_dma_address(sgl);
1330                 sge->length = sg_dma_len(sgl);
1331                 sge->lkey = queue->device->pd->local_dma_lkey;
1332                 len += sge->length;
1333                 sge++;
1334         }
1335
1336         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1337         sg->length = cpu_to_le32(len);
1338         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1339
1340         req->num_sge += count;
1341         return 0;
1342 }
1343
1344 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1345                 struct nvme_rdma_request *req, struct nvme_command *c)
1346 {
1347         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1348
1349         sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1350         put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1351         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1352         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1353         return 0;
1354 }
1355
1356 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1357                 struct nvme_rdma_request *req, struct nvme_command *c,
1358                 int count)
1359 {
1360         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1361         int nr;
1362
1363         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1364         if (WARN_ON_ONCE(!req->mr))
1365                 return -EAGAIN;
1366
1367         /*
1368          * Align the MR to a 4K page size to match the ctrl page size and
1369          * the block virtual boundary.
1370          */
1371         nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1372                           SZ_4K);
1373         if (unlikely(nr < count)) {
1374                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1375                 req->mr = NULL;
1376                 if (nr < 0)
1377                         return nr;
1378                 return -EINVAL;
1379         }
1380
1381         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1382
1383         req->reg_cqe.done = nvme_rdma_memreg_done;
1384         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1385         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1386         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1387         req->reg_wr.wr.num_sge = 0;
1388         req->reg_wr.mr = req->mr;
1389         req->reg_wr.key = req->mr->rkey;
1390         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1391                              IB_ACCESS_REMOTE_READ |
1392                              IB_ACCESS_REMOTE_WRITE;
1393
1394         sg->addr = cpu_to_le64(req->mr->iova);
1395         put_unaligned_le24(req->mr->length, sg->length);
1396         put_unaligned_le32(req->mr->rkey, sg->key);
1397         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1398                         NVME_SGL_FMT_INVALIDATE;
1399
1400         return 0;
1401 }
1402
1403 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1404                 struct nvme_command *cmd, struct ib_sig_domain *domain,
1405                 u16 control, u8 pi_type)
1406 {
1407         domain->sig_type = IB_SIG_TYPE_T10_DIF;
1408         domain->sig.dif.bg_type = IB_T10DIF_CRC;
1409         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1410         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1411         if (control & NVME_RW_PRINFO_PRCHK_REF)
1412                 domain->sig.dif.ref_remap = true;
1413
1414         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1415         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1416         domain->sig.dif.app_escape = true;
1417         if (pi_type == NVME_NS_DPS_PI_TYPE3)
1418                 domain->sig.dif.ref_escape = true;
1419 }
1420
1421 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1422                 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1423                 u8 pi_type)
1424 {
1425         u16 control = le16_to_cpu(cmd->rw.control);
1426
1427         memset(sig_attrs, 0, sizeof(*sig_attrs));
1428         if (control & NVME_RW_PRINFO_PRACT) {
1429                 /* for WRITE_INSERT/READ_STRIP no memory domain */
1430                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1431                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1432                                          pi_type);
1433                 /* Clear the PRACT bit since HCA will generate/verify the PI */
1434                 control &= ~NVME_RW_PRINFO_PRACT;
1435                 cmd->rw.control = cpu_to_le16(control);
1436         } else {
1437                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1438                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1439                                          pi_type);
1440                 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1441                                          pi_type);
1442         }
1443 }
1444
1445 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1446 {
1447         *mask = 0;
1448         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1449                 *mask |= IB_SIG_CHECK_REFTAG;
1450         if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1451                 *mask |= IB_SIG_CHECK_GUARD;
1452 }
1453
1454 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1455 {
1456         if (unlikely(wc->status != IB_WC_SUCCESS))
1457                 nvme_rdma_wr_error(cq, wc, "SIG");
1458 }
1459
1460 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1461                 struct nvme_rdma_request *req, struct nvme_command *c,
1462                 int count, int pi_count)
1463 {
1464         struct nvme_rdma_sgl *sgl = &req->data_sgl;
1465         struct ib_reg_wr *wr = &req->reg_wr;
1466         struct request *rq = blk_mq_rq_from_pdu(req);
1467         struct nvme_ns *ns = rq->q->queuedata;
1468         struct bio *bio = rq->bio;
1469         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1470         int nr;
1471
1472         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1473         if (WARN_ON_ONCE(!req->mr))
1474                 return -EAGAIN;
1475
1476         nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1477                              req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1478                              SZ_4K);
1479         if (unlikely(nr))
1480                 goto mr_put;
1481
1482         nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1483                                 req->mr->sig_attrs, ns->pi_type);
1484         nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1485
1486         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1487
1488         req->reg_cqe.done = nvme_rdma_sig_done;
1489         memset(wr, 0, sizeof(*wr));
1490         wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1491         wr->wr.wr_cqe = &req->reg_cqe;
1492         wr->wr.num_sge = 0;
1493         wr->wr.send_flags = 0;
1494         wr->mr = req->mr;
1495         wr->key = req->mr->rkey;
1496         wr->access = IB_ACCESS_LOCAL_WRITE |
1497                      IB_ACCESS_REMOTE_READ |
1498                      IB_ACCESS_REMOTE_WRITE;
1499
1500         sg->addr = cpu_to_le64(req->mr->iova);
1501         put_unaligned_le24(req->mr->length, sg->length);
1502         put_unaligned_le32(req->mr->rkey, sg->key);
1503         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1504
1505         return 0;
1506
1507 mr_put:
1508         ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1509         req->mr = NULL;
1510         if (nr < 0)
1511                 return nr;
1512         return -EINVAL;
1513 }
1514
1515 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1516                 struct request *rq, struct nvme_command *c)
1517 {
1518         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1519         struct nvme_rdma_device *dev = queue->device;
1520         struct ib_device *ibdev = dev->dev;
1521         int pi_count = 0;
1522         int count, ret;
1523
1524         req->num_sge = 1;
1525         refcount_set(&req->ref, 2); /* send and recv completions */
1526
1527         c->common.flags |= NVME_CMD_SGL_METABUF;
1528
1529         if (!blk_rq_nr_phys_segments(rq))
1530                 return nvme_rdma_set_sg_null(c);
1531
1532         req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1533         ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1534                         blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1535                         NVME_INLINE_SG_CNT);
1536         if (ret)
1537                 return -ENOMEM;
1538
1539         req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1540                                             req->data_sgl.sg_table.sgl);
1541
1542         count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1543                               req->data_sgl.nents, rq_dma_dir(rq));
1544         if (unlikely(count <= 0)) {
1545                 ret = -EIO;
1546                 goto out_free_table;
1547         }
1548
1549         if (blk_integrity_rq(rq)) {
1550                 req->metadata_sgl->sg_table.sgl =
1551                         (struct scatterlist *)(req->metadata_sgl + 1);
1552                 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1553                                 blk_rq_count_integrity_sg(rq->q, rq->bio),
1554                                 req->metadata_sgl->sg_table.sgl,
1555                                 NVME_INLINE_METADATA_SG_CNT);
1556                 if (unlikely(ret)) {
1557                         ret = -ENOMEM;
1558                         goto out_unmap_sg;
1559                 }
1560
1561                 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1562                                 rq->bio, req->metadata_sgl->sg_table.sgl);
1563                 pi_count = ib_dma_map_sg(ibdev,
1564                                          req->metadata_sgl->sg_table.sgl,
1565                                          req->metadata_sgl->nents,
1566                                          rq_dma_dir(rq));
1567                 if (unlikely(pi_count <= 0)) {
1568                         ret = -EIO;
1569                         goto out_free_pi_table;
1570                 }
1571         }
1572
1573         if (req->use_sig_mr) {
1574                 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1575                 goto out;
1576         }
1577
1578         if (count <= dev->num_inline_segments) {
1579                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1580                     queue->ctrl->use_inline_data &&
1581                     blk_rq_payload_bytes(rq) <=
1582                                 nvme_rdma_inline_data_size(queue)) {
1583                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1584                         goto out;
1585                 }
1586
1587                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1588                         ret = nvme_rdma_map_sg_single(queue, req, c);
1589                         goto out;
1590                 }
1591         }
1592
1593         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1594 out:
1595         if (unlikely(ret))
1596                 goto out_unmap_pi_sg;
1597
1598         return 0;
1599
1600 out_unmap_pi_sg:
1601         if (blk_integrity_rq(rq))
1602                 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1603                                 req->metadata_sgl->nents, rq_dma_dir(rq));
1604 out_free_pi_table:
1605         if (blk_integrity_rq(rq))
1606                 sg_free_table_chained(&req->metadata_sgl->sg_table,
1607                                       NVME_INLINE_METADATA_SG_CNT);
1608 out_unmap_sg:
1609         ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1610                         rq_dma_dir(rq));
1611 out_free_table:
1612         sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1613         return ret;
1614 }
1615
1616 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1617 {
1618         struct nvme_rdma_qe *qe =
1619                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1620         struct nvme_rdma_request *req =
1621                 container_of(qe, struct nvme_rdma_request, sqe);
1622
1623         if (unlikely(wc->status != IB_WC_SUCCESS))
1624                 nvme_rdma_wr_error(cq, wc, "SEND");
1625         else
1626                 nvme_rdma_end_request(req);
1627 }
1628
1629 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1630                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1631                 struct ib_send_wr *first)
1632 {
1633         struct ib_send_wr wr;
1634         int ret;
1635
1636         sge->addr   = qe->dma;
1637         sge->length = sizeof(struct nvme_command);
1638         sge->lkey   = queue->device->pd->local_dma_lkey;
1639
1640         wr.next       = NULL;
1641         wr.wr_cqe     = &qe->cqe;
1642         wr.sg_list    = sge;
1643         wr.num_sge    = num_sge;
1644         wr.opcode     = IB_WR_SEND;
1645         wr.send_flags = IB_SEND_SIGNALED;
1646
1647         if (first)
1648                 first->next = &wr;
1649         else
1650                 first = &wr;
1651
1652         ret = ib_post_send(queue->qp, first, NULL);
1653         if (unlikely(ret)) {
1654                 dev_err(queue->ctrl->ctrl.device,
1655                              "%s failed with error code %d\n", __func__, ret);
1656         }
1657         return ret;
1658 }
1659
1660 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1661                 struct nvme_rdma_qe *qe)
1662 {
1663         struct ib_recv_wr wr;
1664         struct ib_sge list;
1665         int ret;
1666
1667         list.addr   = qe->dma;
1668         list.length = sizeof(struct nvme_completion);
1669         list.lkey   = queue->device->pd->local_dma_lkey;
1670
1671         qe->cqe.done = nvme_rdma_recv_done;
1672
1673         wr.next     = NULL;
1674         wr.wr_cqe   = &qe->cqe;
1675         wr.sg_list  = &list;
1676         wr.num_sge  = 1;
1677
1678         ret = ib_post_recv(queue->qp, &wr, NULL);
1679         if (unlikely(ret)) {
1680                 dev_err(queue->ctrl->ctrl.device,
1681                         "%s failed with error code %d\n", __func__, ret);
1682         }
1683         return ret;
1684 }
1685
1686 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1687 {
1688         u32 queue_idx = nvme_rdma_queue_idx(queue);
1689
1690         if (queue_idx == 0)
1691                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1692         return queue->ctrl->tag_set.tags[queue_idx - 1];
1693 }
1694
1695 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1696 {
1697         if (unlikely(wc->status != IB_WC_SUCCESS))
1698                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1699 }
1700
1701 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1702 {
1703         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1704         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1705         struct ib_device *dev = queue->device->dev;
1706         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1707         struct nvme_command *cmd = sqe->data;
1708         struct ib_sge sge;
1709         int ret;
1710
1711         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1712
1713         memset(cmd, 0, sizeof(*cmd));
1714         cmd->common.opcode = nvme_admin_async_event;
1715         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1716         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1717         nvme_rdma_set_sg_null(cmd);
1718
1719         sqe->cqe.done = nvme_rdma_async_done;
1720
1721         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1722                         DMA_TO_DEVICE);
1723
1724         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1725         WARN_ON_ONCE(ret);
1726 }
1727
1728 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1729                 struct nvme_completion *cqe, struct ib_wc *wc)
1730 {
1731         struct request *rq;
1732         struct nvme_rdma_request *req;
1733
1734         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1735         if (!rq) {
1736                 dev_err(queue->ctrl->ctrl.device,
1737                         "tag 0x%x on QP %#x not found\n",
1738                         cqe->command_id, queue->qp->qp_num);
1739                 nvme_rdma_error_recovery(queue->ctrl);
1740                 return;
1741         }
1742         req = blk_mq_rq_to_pdu(rq);
1743
1744         req->status = cqe->status;
1745         req->result = cqe->result;
1746
1747         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1748                 if (unlikely(!req->mr ||
1749                              wc->ex.invalidate_rkey != req->mr->rkey)) {
1750                         dev_err(queue->ctrl->ctrl.device,
1751                                 "Bogus remote invalidation for rkey %#x\n",
1752                                 req->mr ? req->mr->rkey : 0);
1753                         nvme_rdma_error_recovery(queue->ctrl);
1754                 }
1755         } else if (req->mr) {
1756                 int ret;
1757
1758                 ret = nvme_rdma_inv_rkey(queue, req);
1759                 if (unlikely(ret < 0)) {
1760                         dev_err(queue->ctrl->ctrl.device,
1761                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1762                                 req->mr->rkey, ret);
1763                         nvme_rdma_error_recovery(queue->ctrl);
1764                 }
1765                 /* the local invalidation completion will end the request */
1766                 return;
1767         }
1768
1769         nvme_rdma_end_request(req);
1770 }
1771
1772 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1773 {
1774         struct nvme_rdma_qe *qe =
1775                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1776         struct nvme_rdma_queue *queue = wc->qp->qp_context;
1777         struct ib_device *ibdev = queue->device->dev;
1778         struct nvme_completion *cqe = qe->data;
1779         const size_t len = sizeof(struct nvme_completion);
1780
1781         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1782                 nvme_rdma_wr_error(cq, wc, "RECV");
1783                 return;
1784         }
1785
1786         /* sanity checking for received data length */
1787         if (unlikely(wc->byte_len < len)) {
1788                 dev_err(queue->ctrl->ctrl.device,
1789                         "Unexpected nvme completion length(%d)\n", wc->byte_len);
1790                 nvme_rdma_error_recovery(queue->ctrl);
1791                 return;
1792         }
1793
1794         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1795         /*
1796          * AEN requests are special as they don't time out and can
1797          * survive any kind of queue freeze and often don't respond to
1798          * aborts.  We don't even bother to allocate a struct request
1799          * for them but rather special case them here.
1800          */
1801         if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1802                                      cqe->command_id)))
1803                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1804                                 &cqe->result);
1805         else
1806                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1807         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1808
1809         nvme_rdma_post_recv(queue, qe);
1810 }
1811
1812 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1813 {
1814         int ret, i;
1815
1816         for (i = 0; i < queue->queue_size; i++) {
1817                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1818                 if (ret)
1819                         goto out_destroy_queue_ib;
1820         }
1821
1822         return 0;
1823
1824 out_destroy_queue_ib:
1825         nvme_rdma_destroy_queue_ib(queue);
1826         return ret;
1827 }
1828
1829 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1830                 struct rdma_cm_event *ev)
1831 {
1832         struct rdma_cm_id *cm_id = queue->cm_id;
1833         int status = ev->status;
1834         const char *rej_msg;
1835         const struct nvme_rdma_cm_rej *rej_data;
1836         u8 rej_data_len;
1837
1838         rej_msg = rdma_reject_msg(cm_id, status);
1839         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1840
1841         if (rej_data && rej_data_len >= sizeof(u16)) {
1842                 u16 sts = le16_to_cpu(rej_data->sts);
1843
1844                 dev_err(queue->ctrl->ctrl.device,
1845                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1846                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1847         } else {
1848                 dev_err(queue->ctrl->ctrl.device,
1849                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1850         }
1851
1852         return -ECONNRESET;
1853 }
1854
1855 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1856 {
1857         struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1858         int ret;
1859
1860         ret = nvme_rdma_create_queue_ib(queue);
1861         if (ret)
1862                 return ret;
1863
1864         if (ctrl->opts->tos >= 0)
1865                 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1866         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1867         if (ret) {
1868                 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1869                         queue->cm_error);
1870                 goto out_destroy_queue;
1871         }
1872
1873         return 0;
1874
1875 out_destroy_queue:
1876         nvme_rdma_destroy_queue_ib(queue);
1877         return ret;
1878 }
1879
1880 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1881 {
1882         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1883         struct rdma_conn_param param = { };
1884         struct nvme_rdma_cm_req priv = { };
1885         int ret;
1886
1887         param.qp_num = queue->qp->qp_num;
1888         param.flow_control = 1;
1889
1890         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1891         /* maximum retry count */
1892         param.retry_count = 7;
1893         param.rnr_retry_count = 7;
1894         param.private_data = &priv;
1895         param.private_data_len = sizeof(priv);
1896
1897         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1898         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1899         /*
1900          * set the admin queue depth to the minimum size
1901          * specified by the Fabrics standard.
1902          */
1903         if (priv.qid == 0) {
1904                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1905                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1906         } else {
1907                 /*
1908                  * current interpretation of the fabrics spec
1909                  * is at minimum you make hrqsize sqsize+1, or a
1910                  * 1's based representation of sqsize.
1911                  */
1912                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1913                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1914         }
1915
1916         ret = rdma_connect_locked(queue->cm_id, &param);
1917         if (ret) {
1918                 dev_err(ctrl->ctrl.device,
1919                         "rdma_connect_locked failed (%d).\n", ret);
1920                 goto out_destroy_queue_ib;
1921         }
1922
1923         return 0;
1924
1925 out_destroy_queue_ib:
1926         nvme_rdma_destroy_queue_ib(queue);
1927         return ret;
1928 }
1929
1930 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1931                 struct rdma_cm_event *ev)
1932 {
1933         struct nvme_rdma_queue *queue = cm_id->context;
1934         int cm_error = 0;
1935
1936         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1937                 rdma_event_msg(ev->event), ev->event,
1938                 ev->status, cm_id);
1939
1940         switch (ev->event) {
1941         case RDMA_CM_EVENT_ADDR_RESOLVED:
1942                 cm_error = nvme_rdma_addr_resolved(queue);
1943                 break;
1944         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1945                 cm_error = nvme_rdma_route_resolved(queue);
1946                 break;
1947         case RDMA_CM_EVENT_ESTABLISHED:
1948                 queue->cm_error = nvme_rdma_conn_established(queue);
1949                 /* complete cm_done regardless of success/failure */
1950                 complete(&queue->cm_done);
1951                 return 0;
1952         case RDMA_CM_EVENT_REJECTED:
1953                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1954                 break;
1955         case RDMA_CM_EVENT_ROUTE_ERROR:
1956         case RDMA_CM_EVENT_CONNECT_ERROR:
1957         case RDMA_CM_EVENT_UNREACHABLE:
1958                 nvme_rdma_destroy_queue_ib(queue);
1959                 fallthrough;
1960         case RDMA_CM_EVENT_ADDR_ERROR:
1961                 dev_dbg(queue->ctrl->ctrl.device,
1962                         "CM error event %d\n", ev->event);
1963                 cm_error = -ECONNRESET;
1964                 break;
1965         case RDMA_CM_EVENT_DISCONNECTED:
1966         case RDMA_CM_EVENT_ADDR_CHANGE:
1967         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1968                 dev_dbg(queue->ctrl->ctrl.device,
1969                         "disconnect received - connection closed\n");
1970                 nvme_rdma_error_recovery(queue->ctrl);
1971                 break;
1972         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1973                 /* device removal is handled via the ib_client API */
1974                 break;
1975         default:
1976                 dev_err(queue->ctrl->ctrl.device,
1977                         "Unexpected RDMA CM event (%d)\n", ev->event);
1978                 nvme_rdma_error_recovery(queue->ctrl);
1979                 break;
1980         }
1981
1982         if (cm_error) {
1983                 queue->cm_error = cm_error;
1984                 complete(&queue->cm_done);
1985         }
1986
1987         return 0;
1988 }
1989
1990 static void nvme_rdma_complete_timed_out(struct request *rq)
1991 {
1992         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1993         struct nvme_rdma_queue *queue = req->queue;
1994
1995         nvme_rdma_stop_queue(queue);
1996         if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
1997                 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
1998                 blk_mq_complete_request(rq);
1999         }
2000 }
2001
2002 static enum blk_eh_timer_return
2003 nvme_rdma_timeout(struct request *rq, bool reserved)
2004 {
2005         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2006         struct nvme_rdma_queue *queue = req->queue;
2007         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
2008
2009         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
2010                  rq->tag, nvme_rdma_queue_idx(queue));
2011
2012         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2013                 /*
2014                  * If we are resetting, connecting or deleting we should
2015                  * complete immediately because we may block controller
2016                  * teardown or setup sequence
2017                  * - ctrl disable/shutdown fabrics requests
2018                  * - connect requests
2019                  * - initialization admin requests
2020                  * - I/O requests that entered after unquiescing and
2021                  *   the controller stopped responding
2022                  *
2023                  * All other requests should be cancelled by the error
2024                  * recovery work, so it's fine that we fail it here.
2025                  */
2026                 nvme_rdma_complete_timed_out(rq);
2027                 return BLK_EH_DONE;
2028         }
2029
2030         /*
2031          * LIVE state should trigger the normal error recovery which will
2032          * handle completing this request.
2033          */
2034         nvme_rdma_error_recovery(ctrl);
2035         return BLK_EH_RESET_TIMER;
2036 }
2037
2038 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2039                 const struct blk_mq_queue_data *bd)
2040 {
2041         struct nvme_ns *ns = hctx->queue->queuedata;
2042         struct nvme_rdma_queue *queue = hctx->driver_data;
2043         struct request *rq = bd->rq;
2044         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2045         struct nvme_rdma_qe *sqe = &req->sqe;
2046         struct nvme_command *c = nvme_req(rq)->cmd;
2047         struct ib_device *dev;
2048         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2049         blk_status_t ret;
2050         int err;
2051
2052         WARN_ON_ONCE(rq->tag < 0);
2053
2054         if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2055                 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2056
2057         dev = queue->device->dev;
2058
2059         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2060                                          sizeof(struct nvme_command),
2061                                          DMA_TO_DEVICE);
2062         err = ib_dma_mapping_error(dev, req->sqe.dma);
2063         if (unlikely(err))
2064                 return BLK_STS_RESOURCE;
2065
2066         ib_dma_sync_single_for_cpu(dev, sqe->dma,
2067                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2068
2069         ret = nvme_setup_cmd(ns, rq);
2070         if (ret)
2071                 goto unmap_qe;
2072
2073         blk_mq_start_request(rq);
2074
2075         if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2076             queue->pi_support &&
2077             (c->common.opcode == nvme_cmd_write ||
2078              c->common.opcode == nvme_cmd_read) &&
2079             nvme_ns_has_pi(ns))
2080                 req->use_sig_mr = true;
2081         else
2082                 req->use_sig_mr = false;
2083
2084         err = nvme_rdma_map_data(queue, rq, c);
2085         if (unlikely(err < 0)) {
2086                 dev_err(queue->ctrl->ctrl.device,
2087                              "Failed to map data (%d)\n", err);
2088                 goto err;
2089         }
2090
2091         sqe->cqe.done = nvme_rdma_send_done;
2092
2093         ib_dma_sync_single_for_device(dev, sqe->dma,
2094                         sizeof(struct nvme_command), DMA_TO_DEVICE);
2095
2096         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2097                         req->mr ? &req->reg_wr.wr : NULL);
2098         if (unlikely(err))
2099                 goto err_unmap;
2100
2101         return BLK_STS_OK;
2102
2103 err_unmap:
2104         nvme_rdma_unmap_data(queue, rq);
2105 err:
2106         if (err == -EIO)
2107                 ret = nvme_host_path_error(rq);
2108         else if (err == -ENOMEM || err == -EAGAIN)
2109                 ret = BLK_STS_RESOURCE;
2110         else
2111                 ret = BLK_STS_IOERR;
2112         nvme_cleanup_cmd(rq);
2113 unmap_qe:
2114         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2115                             DMA_TO_DEVICE);
2116         return ret;
2117 }
2118
2119 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2120 {
2121         struct nvme_rdma_queue *queue = hctx->driver_data;
2122
2123         return ib_process_cq_direct(queue->ib_cq, -1);
2124 }
2125
2126 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2127 {
2128         struct request *rq = blk_mq_rq_from_pdu(req);
2129         struct ib_mr_status mr_status;
2130         int ret;
2131
2132         ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2133         if (ret) {
2134                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2135                 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2136                 return;
2137         }
2138
2139         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2140                 switch (mr_status.sig_err.err_type) {
2141                 case IB_SIG_BAD_GUARD:
2142                         nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2143                         break;
2144                 case IB_SIG_BAD_REFTAG:
2145                         nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2146                         break;
2147                 case IB_SIG_BAD_APPTAG:
2148                         nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2149                         break;
2150                 }
2151                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2152                        mr_status.sig_err.err_type, mr_status.sig_err.expected,
2153                        mr_status.sig_err.actual);
2154         }
2155 }
2156
2157 static void nvme_rdma_complete_rq(struct request *rq)
2158 {
2159         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2160         struct nvme_rdma_queue *queue = req->queue;
2161         struct ib_device *ibdev = queue->device->dev;
2162
2163         if (req->use_sig_mr)
2164                 nvme_rdma_check_pi_status(req);
2165
2166         nvme_rdma_unmap_data(queue, rq);
2167         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2168                             DMA_TO_DEVICE);
2169         nvme_complete_rq(rq);
2170 }
2171
2172 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2173 {
2174         struct nvme_rdma_ctrl *ctrl = set->driver_data;
2175         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2176
2177         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2178                 /* separate read/write queues */
2179                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2180                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2181                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2182                 set->map[HCTX_TYPE_READ].nr_queues =
2183                         ctrl->io_queues[HCTX_TYPE_READ];
2184                 set->map[HCTX_TYPE_READ].queue_offset =
2185                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2186         } else {
2187                 /* shared read/write queues */
2188                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2189                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2190                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2191                 set->map[HCTX_TYPE_READ].nr_queues =
2192                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
2193                 set->map[HCTX_TYPE_READ].queue_offset = 0;
2194         }
2195         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2196                         ctrl->device->dev, 0);
2197         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2198                         ctrl->device->dev, 0);
2199
2200         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2201                 /* map dedicated poll queues only if we have queues left */
2202                 set->map[HCTX_TYPE_POLL].nr_queues =
2203                                 ctrl->io_queues[HCTX_TYPE_POLL];
2204                 set->map[HCTX_TYPE_POLL].queue_offset =
2205                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2206                         ctrl->io_queues[HCTX_TYPE_READ];
2207                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2208         }
2209
2210         dev_info(ctrl->ctrl.device,
2211                 "mapped %d/%d/%d default/read/poll queues.\n",
2212                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2213                 ctrl->io_queues[HCTX_TYPE_READ],
2214                 ctrl->io_queues[HCTX_TYPE_POLL]);
2215
2216         return 0;
2217 }
2218
2219 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2220         .queue_rq       = nvme_rdma_queue_rq,
2221         .complete       = nvme_rdma_complete_rq,
2222         .init_request   = nvme_rdma_init_request,
2223         .exit_request   = nvme_rdma_exit_request,
2224         .init_hctx      = nvme_rdma_init_hctx,
2225         .timeout        = nvme_rdma_timeout,
2226         .map_queues     = nvme_rdma_map_queues,
2227         .poll           = nvme_rdma_poll,
2228 };
2229
2230 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2231         .queue_rq       = nvme_rdma_queue_rq,
2232         .complete       = nvme_rdma_complete_rq,
2233         .init_request   = nvme_rdma_init_request,
2234         .exit_request   = nvme_rdma_exit_request,
2235         .init_hctx      = nvme_rdma_init_admin_hctx,
2236         .timeout        = nvme_rdma_timeout,
2237 };
2238
2239 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2240 {
2241         cancel_work_sync(&ctrl->err_work);
2242         cancel_delayed_work_sync(&ctrl->reconnect_work);
2243
2244         nvme_rdma_teardown_io_queues(ctrl, shutdown);
2245         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2246         if (shutdown)
2247                 nvme_shutdown_ctrl(&ctrl->ctrl);
2248         else
2249                 nvme_disable_ctrl(&ctrl->ctrl);
2250         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2251 }
2252
2253 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2254 {
2255         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2256 }
2257
2258 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2259 {
2260         struct nvme_rdma_ctrl *ctrl =
2261                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2262
2263         nvme_stop_ctrl(&ctrl->ctrl);
2264         nvme_rdma_shutdown_ctrl(ctrl, false);
2265
2266         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2267                 /* state change failure should never happen */
2268                 WARN_ON_ONCE(1);
2269                 return;
2270         }
2271
2272         if (nvme_rdma_setup_ctrl(ctrl, false))
2273                 goto out_fail;
2274
2275         return;
2276
2277 out_fail:
2278         ++ctrl->ctrl.nr_reconnects;
2279         nvme_rdma_reconnect_or_remove(ctrl);
2280 }
2281
2282 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2283         .name                   = "rdma",
2284         .module                 = THIS_MODULE,
2285         .flags                  = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2286         .reg_read32             = nvmf_reg_read32,
2287         .reg_read64             = nvmf_reg_read64,
2288         .reg_write32            = nvmf_reg_write32,
2289         .free_ctrl              = nvme_rdma_free_ctrl,
2290         .submit_async_event     = nvme_rdma_submit_async_event,
2291         .delete_ctrl            = nvme_rdma_delete_ctrl,
2292         .get_address            = nvmf_get_address,
2293 };
2294
2295 /*
2296  * Fails a connection request if it matches an existing controller
2297  * (association) with the same tuple:
2298  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2299  *
2300  * if local address is not specified in the request, it will match an
2301  * existing controller with all the other parameters the same and no
2302  * local port address specified as well.
2303  *
2304  * The ports don't need to be compared as they are intrinsically
2305  * already matched by the port pointers supplied.
2306  */
2307 static bool
2308 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2309 {
2310         struct nvme_rdma_ctrl *ctrl;
2311         bool found = false;
2312
2313         mutex_lock(&nvme_rdma_ctrl_mutex);
2314         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2315                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2316                 if (found)
2317                         break;
2318         }
2319         mutex_unlock(&nvme_rdma_ctrl_mutex);
2320
2321         return found;
2322 }
2323
2324 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2325                 struct nvmf_ctrl_options *opts)
2326 {
2327         struct nvme_rdma_ctrl *ctrl;
2328         int ret;
2329         bool changed;
2330
2331         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2332         if (!ctrl)
2333                 return ERR_PTR(-ENOMEM);
2334         ctrl->ctrl.opts = opts;
2335         INIT_LIST_HEAD(&ctrl->list);
2336
2337         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2338                 opts->trsvcid =
2339                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2340                 if (!opts->trsvcid) {
2341                         ret = -ENOMEM;
2342                         goto out_free_ctrl;
2343                 }
2344                 opts->mask |= NVMF_OPT_TRSVCID;
2345         }
2346
2347         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2348                         opts->traddr, opts->trsvcid, &ctrl->addr);
2349         if (ret) {
2350                 pr_err("malformed address passed: %s:%s\n",
2351                         opts->traddr, opts->trsvcid);
2352                 goto out_free_ctrl;
2353         }
2354
2355         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2356                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2357                         opts->host_traddr, NULL, &ctrl->src_addr);
2358                 if (ret) {
2359                         pr_err("malformed src address passed: %s\n",
2360                                opts->host_traddr);
2361                         goto out_free_ctrl;
2362                 }
2363         }
2364
2365         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2366                 ret = -EALREADY;
2367                 goto out_free_ctrl;
2368         }
2369
2370         INIT_DELAYED_WORK(&ctrl->reconnect_work,
2371                         nvme_rdma_reconnect_ctrl_work);
2372         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2373         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2374
2375         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2376                                 opts->nr_poll_queues + 1;
2377         ctrl->ctrl.sqsize = opts->queue_size - 1;
2378         ctrl->ctrl.kato = opts->kato;
2379
2380         ret = -ENOMEM;
2381         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2382                                 GFP_KERNEL);
2383         if (!ctrl->queues)
2384                 goto out_free_ctrl;
2385
2386         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2387                                 0 /* no quirks, we're perfect! */);
2388         if (ret)
2389                 goto out_kfree_queues;
2390
2391         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2392         WARN_ON_ONCE(!changed);
2393
2394         ret = nvme_rdma_setup_ctrl(ctrl, true);
2395         if (ret)
2396                 goto out_uninit_ctrl;
2397
2398         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2399                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2400
2401         mutex_lock(&nvme_rdma_ctrl_mutex);
2402         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2403         mutex_unlock(&nvme_rdma_ctrl_mutex);
2404
2405         return &ctrl->ctrl;
2406
2407 out_uninit_ctrl:
2408         nvme_uninit_ctrl(&ctrl->ctrl);
2409         nvme_put_ctrl(&ctrl->ctrl);
2410         if (ret > 0)
2411                 ret = -EIO;
2412         return ERR_PTR(ret);
2413 out_kfree_queues:
2414         kfree(ctrl->queues);
2415 out_free_ctrl:
2416         kfree(ctrl);
2417         return ERR_PTR(ret);
2418 }
2419
2420 static struct nvmf_transport_ops nvme_rdma_transport = {
2421         .name           = "rdma",
2422         .module         = THIS_MODULE,
2423         .required_opts  = NVMF_OPT_TRADDR,
2424         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2425                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2426                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2427                           NVMF_OPT_TOS,
2428         .create_ctrl    = nvme_rdma_create_ctrl,
2429 };
2430
2431 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2432 {
2433         struct nvme_rdma_ctrl *ctrl;
2434         struct nvme_rdma_device *ndev;
2435         bool found = false;
2436
2437         mutex_lock(&device_list_mutex);
2438         list_for_each_entry(ndev, &device_list, entry) {
2439                 if (ndev->dev == ib_device) {
2440                         found = true;
2441                         break;
2442                 }
2443         }
2444         mutex_unlock(&device_list_mutex);
2445
2446         if (!found)
2447                 return;
2448
2449         /* Delete all controllers using this device */
2450         mutex_lock(&nvme_rdma_ctrl_mutex);
2451         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2452                 if (ctrl->device->dev != ib_device)
2453                         continue;
2454                 nvme_delete_ctrl(&ctrl->ctrl);
2455         }
2456         mutex_unlock(&nvme_rdma_ctrl_mutex);
2457
2458         flush_workqueue(nvme_delete_wq);
2459 }
2460
2461 static struct ib_client nvme_rdma_ib_client = {
2462         .name   = "nvme_rdma",
2463         .remove = nvme_rdma_remove_one
2464 };
2465
2466 static int __init nvme_rdma_init_module(void)
2467 {
2468         int ret;
2469
2470         ret = ib_register_client(&nvme_rdma_ib_client);
2471         if (ret)
2472                 return ret;
2473
2474         ret = nvmf_register_transport(&nvme_rdma_transport);
2475         if (ret)
2476                 goto err_unreg_client;
2477
2478         return 0;
2479
2480 err_unreg_client:
2481         ib_unregister_client(&nvme_rdma_ib_client);
2482         return ret;
2483 }
2484
2485 static void __exit nvme_rdma_cleanup_module(void)
2486 {
2487         struct nvme_rdma_ctrl *ctrl;
2488
2489         nvmf_unregister_transport(&nvme_rdma_transport);
2490         ib_unregister_client(&nvme_rdma_ib_client);
2491
2492         mutex_lock(&nvme_rdma_ctrl_mutex);
2493         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2494                 nvme_delete_ctrl(&ctrl->ctrl);
2495         mutex_unlock(&nvme_rdma_ctrl_mutex);
2496         flush_workqueue(nvme_delete_wq);
2497 }
2498
2499 module_init(nvme_rdma_init_module);
2500 module_exit(nvme_rdma_cleanup_module);
2501
2502 MODULE_LICENSE("GPL v2");