drivers/nvme/host/rdma.c

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