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), 0);
 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_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
 908         }
 909         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 910         nvme_rdma_destroy_admin_queue(ctrl, remove);
 911 }
 912
 913 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
 914                 bool remove)
 915 {
 916         if (ctrl->ctrl.queue_count > 1) {
 917                 nvme_stop_queues(&ctrl->ctrl);
 918                 nvme_rdma_stop_io_queues(ctrl);
 919                 if (ctrl->ctrl.tagset) {
 920                         blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
 921                                 nvme_cancel_request, &ctrl->ctrl);
 922                         blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
 923                 }
 924                 if (remove)
 925                         nvme_start_queues(&ctrl->ctrl);
 926                 nvme_rdma_destroy_io_queues(ctrl, remove);
 927         }
 928 }
 929
 930 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 931 {
 932         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 933
 934         if (list_empty(&ctrl->list))
 935                 goto free_ctrl;
 936
 937         mutex_lock(&nvme_rdma_ctrl_mutex);
 938         list_del(&ctrl->list);
 939         mutex_unlock(&nvme_rdma_ctrl_mutex);
 940
 941         nvmf_free_options(nctrl->opts);
 942 free_ctrl:
 943         kfree(ctrl->queues);
 944         kfree(ctrl);
 945 }
 946
 947 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 948 {
 949         /* If we are resetting/deleting then do nothing */
 950         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 951                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 952                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 953                 return;
 954         }
 955
 956         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 957                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 958                         ctrl->ctrl.opts->reconnect_delay);
 959                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 960                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 961         } else {
 962                 nvme_delete_ctrl(&ctrl->ctrl);
 963         }
 964 }
 965
 966 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
 967 {
 968         int ret = -EINVAL;
 969         bool changed;
 970
 971         ret = nvme_rdma_configure_admin_queue(ctrl, new);
 972         if (ret)
 973                 return ret;
 974
 975         if (ctrl->ctrl.icdoff) {
 976                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
 977                 goto destroy_admin;
 978         }
 979
 980         if (!(ctrl->ctrl.sgls & (1 << 2))) {
 981                 dev_err(ctrl->ctrl.device,
 982                         "Mandatory keyed sgls are not supported!\n");
 983                 goto destroy_admin;
 984         }
 985
 986         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
 987                 dev_warn(ctrl->ctrl.device,
 988                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
 989                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
 990         }
 991
 992         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
 993                 dev_warn(ctrl->ctrl.device,
 994                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
 995                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
 996                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
 997         }
 998
 999         if (ctrl->ctrl.sgls & (1 << 20))
1000                 ctrl->use_inline_data = true;
1001
1002         if (ctrl->ctrl.queue_count > 1) {
1003                 ret = nvme_rdma_configure_io_queues(ctrl, new);
1004                 if (ret)
1005                         goto destroy_admin;
1006         }
1007
1008         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1009         if (!changed) {
1010                 /* state change failure is ok if we're in DELETING state */
1011                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1012                 ret = -EINVAL;
1013                 goto destroy_io;
1014         }
1015
1016         nvme_start_ctrl(&ctrl->ctrl);
1017         return 0;
1018
1019 destroy_io:
1020         if (ctrl->ctrl.queue_count > 1)
1021                 nvme_rdma_destroy_io_queues(ctrl, new);
1022 destroy_admin:
1023         nvme_rdma_stop_queue(&ctrl->queues[0]);
1024         nvme_rdma_destroy_admin_queue(ctrl, new);
1025         return ret;
1026 }
1027
1028 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1029 {
1030         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1031                         struct nvme_rdma_ctrl, reconnect_work);
1032
1033         ++ctrl->ctrl.nr_reconnects;
1034
1035         if (nvme_rdma_setup_ctrl(ctrl, false))
1036                 goto requeue;
1037
1038         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1039                         ctrl->ctrl.nr_reconnects);
1040
1041         ctrl->ctrl.nr_reconnects = 0;
1042
1043         return;
1044
1045 requeue:
1046         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1047                         ctrl->ctrl.nr_reconnects);
1048         nvme_rdma_reconnect_or_remove(ctrl);
1049 }
1050
1051 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1052 {
1053         struct nvme_rdma_ctrl *ctrl = container_of(work,
1054                         struct nvme_rdma_ctrl, err_work);
1055
1056         nvme_stop_keep_alive(&ctrl->ctrl);
1057         nvme_rdma_teardown_io_queues(ctrl, false);
1058         nvme_start_queues(&ctrl->ctrl);
1059         nvme_rdma_teardown_admin_queue(ctrl, false);
1060
1061         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1062                 /* state change failure is ok if we're in DELETING state */
1063                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1064                 return;
1065         }
1066
1067         nvme_rdma_reconnect_or_remove(ctrl);
1068 }
1069
1070 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1071 {
1072         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1073                 return;
1074
1075         queue_work(nvme_wq, &ctrl->err_work);
1076 }
1077
1078 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1079                 const char *op)
1080 {
1081         struct nvme_rdma_queue *queue = cq->cq_context;
1082         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1083
1084         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1085                 dev_info(ctrl->ctrl.device,
1086                              "%s for CQE 0x%p failed with status %s (%d)\n",
1087                              op, wc->wr_cqe,
1088                              ib_wc_status_msg(wc->status), wc->status);
1089         nvme_rdma_error_recovery(ctrl);
1090 }
1091
1092 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1093 {
1094         if (unlikely(wc->status != IB_WC_SUCCESS))
1095                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1096 }
1097
1098 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1099 {
1100         struct nvme_rdma_request *req =
1101                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1102         struct request *rq = blk_mq_rq_from_pdu(req);
1103
1104         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1105                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1106                 return;
1107         }
1108
1109         if (refcount_dec_and_test(&req->ref))
1110                 nvme_end_request(rq, req->status, req->result);
1111
1112 }
1113
1114 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1115                 struct nvme_rdma_request *req)
1116 {
1117         struct ib_send_wr wr = {
1118                 .opcode             = IB_WR_LOCAL_INV,
1119                 .next               = NULL,
1120                 .num_sge            = 0,
1121                 .send_flags         = IB_SEND_SIGNALED,
1122                 .ex.invalidate_rkey = req->mr->rkey,
1123         };
1124
1125         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1126         wr.wr_cqe = &req->reg_cqe;
1127
1128         return ib_post_send(queue->qp, &wr, NULL);
1129 }
1130
1131 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1132                 struct request *rq)
1133 {
1134         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1135         struct nvme_rdma_device *dev = queue->device;
1136         struct ib_device *ibdev = dev->dev;
1137
1138         if (!blk_rq_nr_phys_segments(rq))
1139                 return;
1140
1141         if (req->mr) {
1142                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1143                 req->mr = NULL;
1144         }
1145
1146         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1147                         req->nents, rq_data_dir(rq) ==
1148                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1149
1150         nvme_cleanup_cmd(rq);
1151         sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE);
1152 }
1153
1154 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1155 {
1156         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1157
1158         sg->addr = 0;
1159         put_unaligned_le24(0, sg->length);
1160         put_unaligned_le32(0, sg->key);
1161         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1162         return 0;
1163 }
1164
1165 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1166                 struct nvme_rdma_request *req, struct nvme_command *c,
1167                 int count)
1168 {
1169         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1170         struct scatterlist *sgl = req->sg_table.sgl;
1171         struct ib_sge *sge = &req->sge[1];
1172         u32 len = 0;
1173         int i;
1174
1175         for (i = 0; i < count; i++, sgl++, sge++) {
1176                 sge->addr = sg_dma_address(sgl);
1177                 sge->length = sg_dma_len(sgl);
1178                 sge->lkey = queue->device->pd->local_dma_lkey;
1179                 len += sge->length;
1180         }
1181
1182         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1183         sg->length = cpu_to_le32(len);
1184         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1185
1186         req->num_sge += count;
1187         return 0;
1188 }
1189
1190 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1191                 struct nvme_rdma_request *req, struct nvme_command *c)
1192 {
1193         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1194
1195         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1196         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1197         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1198         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1199         return 0;
1200 }
1201
1202 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1203                 struct nvme_rdma_request *req, struct nvme_command *c,
1204                 int count)
1205 {
1206         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1207         int nr;
1208
1209         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1210         if (WARN_ON_ONCE(!req->mr))
1211                 return -EAGAIN;
1212
1213         /*
1214          * Align the MR to a 4K page size to match the ctrl page size and
1215          * the block virtual boundary.
1216          */
1217         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1218         if (unlikely(nr < count)) {
1219                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1220                 req->mr = NULL;
1221                 if (nr < 0)
1222                         return nr;
1223                 return -EINVAL;
1224         }
1225
1226         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1227
1228         req->reg_cqe.done = nvme_rdma_memreg_done;
1229         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1230         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1231         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1232         req->reg_wr.wr.num_sge = 0;
1233         req->reg_wr.mr = req->mr;
1234         req->reg_wr.key = req->mr->rkey;
1235         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1236                              IB_ACCESS_REMOTE_READ |
1237                              IB_ACCESS_REMOTE_WRITE;
1238
1239         sg->addr = cpu_to_le64(req->mr->iova);
1240         put_unaligned_le24(req->mr->length, sg->length);
1241         put_unaligned_le32(req->mr->rkey, sg->key);
1242         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1243                         NVME_SGL_FMT_INVALIDATE;
1244
1245         return 0;
1246 }
1247
1248 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1249                 struct request *rq, struct nvme_command *c)
1250 {
1251         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1252         struct nvme_rdma_device *dev = queue->device;
1253         struct ib_device *ibdev = dev->dev;
1254         int count, ret;
1255
1256         req->num_sge = 1;
1257         refcount_set(&req->ref, 2); /* send and recv completions */
1258
1259         c->common.flags |= NVME_CMD_SGL_METABUF;
1260
1261         if (!blk_rq_nr_phys_segments(rq))
1262                 return nvme_rdma_set_sg_null(c);
1263
1264         req->sg_table.sgl = req->first_sgl;
1265         ret = sg_alloc_table_chained(&req->sg_table,
1266                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl,
1267                         SG_CHUNK_SIZE);
1268         if (ret)
1269                 return -ENOMEM;
1270
1271         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1272
1273         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1274                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1275         if (unlikely(count <= 0)) {
1276                 ret = -EIO;
1277                 goto out_free_table;
1278         }
1279
1280         if (count <= dev->num_inline_segments) {
1281                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1282                     queue->ctrl->use_inline_data &&
1283                     blk_rq_payload_bytes(rq) <=
1284                                 nvme_rdma_inline_data_size(queue)) {
1285                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1286                         goto out;
1287                 }
1288
1289                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1290                         ret = nvme_rdma_map_sg_single(queue, req, c);
1291                         goto out;
1292                 }
1293         }
1294
1295         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1296 out:
1297         if (unlikely(ret))
1298                 goto out_unmap_sg;
1299
1300         return 0;
1301
1302 out_unmap_sg:
1303         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1304                         req->nents, rq_data_dir(rq) ==
1305                         WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1306 out_free_table:
1307         sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE);
1308         return ret;
1309 }
1310
1311 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1312 {
1313         struct nvme_rdma_qe *qe =
1314                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1315         struct nvme_rdma_request *req =
1316                 container_of(qe, struct nvme_rdma_request, sqe);
1317         struct request *rq = blk_mq_rq_from_pdu(req);
1318
1319         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1320                 nvme_rdma_wr_error(cq, wc, "SEND");
1321                 return;
1322         }
1323
1324         if (refcount_dec_and_test(&req->ref))
1325                 nvme_end_request(rq, req->status, req->result);
1326 }
1327
1328 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1329                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1330                 struct ib_send_wr *first)
1331 {
1332         struct ib_send_wr wr;
1333         int ret;
1334
1335         sge->addr   = qe->dma;
1336         sge->length = sizeof(struct nvme_command),
1337         sge->lkey   = queue->device->pd->local_dma_lkey;
1338
1339         wr.next       = NULL;
1340         wr.wr_cqe     = &qe->cqe;
1341         wr.sg_list    = sge;
1342         wr.num_sge    = num_sge;
1343         wr.opcode     = IB_WR_SEND;
1344         wr.send_flags = IB_SEND_SIGNALED;
1345
1346         if (first)
1347                 first->next = &wr;
1348         else
1349                 first = &wr;
1350
1351         ret = ib_post_send(queue->qp, first, NULL);
1352         if (unlikely(ret)) {
1353                 dev_err(queue->ctrl->ctrl.device,
1354                              "%s failed with error code %d\n", __func__, ret);
1355         }
1356         return ret;
1357 }
1358
1359 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1360                 struct nvme_rdma_qe *qe)
1361 {
1362         struct ib_recv_wr wr;
1363         struct ib_sge list;
1364         int ret;
1365
1366         list.addr   = qe->dma;
1367         list.length = sizeof(struct nvme_completion);
1368         list.lkey   = queue->device->pd->local_dma_lkey;
1369
1370         qe->cqe.done = nvme_rdma_recv_done;
1371
1372         wr.next     = NULL;
1373         wr.wr_cqe   = &qe->cqe;
1374         wr.sg_list  = &list;
1375         wr.num_sge  = 1;
1376
1377         ret = ib_post_recv(queue->qp, &wr, NULL);
1378         if (unlikely(ret)) {
1379                 dev_err(queue->ctrl->ctrl.device,
1380                         "%s failed with error code %d\n", __func__, ret);
1381         }
1382         return ret;
1383 }
1384
1385 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1386 {
1387         u32 queue_idx = nvme_rdma_queue_idx(queue);
1388
1389         if (queue_idx == 0)
1390                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1391         return queue->ctrl->tag_set.tags[queue_idx - 1];
1392 }
1393
1394 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1395 {
1396         if (unlikely(wc->status != IB_WC_SUCCESS))
1397                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1398 }
1399
1400 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1401 {
1402         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1403         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1404         struct ib_device *dev = queue->device->dev;
1405         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1406         struct nvme_command *cmd = sqe->data;
1407         struct ib_sge sge;
1408         int ret;
1409
1410         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1411
1412         memset(cmd, 0, sizeof(*cmd));
1413         cmd->common.opcode = nvme_admin_async_event;
1414         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1415         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1416         nvme_rdma_set_sg_null(cmd);
1417
1418         sqe->cqe.done = nvme_rdma_async_done;
1419
1420         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1421                         DMA_TO_DEVICE);
1422
1423         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1424         WARN_ON_ONCE(ret);
1425 }
1426
1427 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1428                 struct nvme_completion *cqe, struct ib_wc *wc)
1429 {
1430         struct request *rq;
1431         struct nvme_rdma_request *req;
1432
1433         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1434         if (!rq) {
1435                 dev_err(queue->ctrl->ctrl.device,
1436                         "tag 0x%x on QP %#x not found\n",
1437                         cqe->command_id, queue->qp->qp_num);
1438                 nvme_rdma_error_recovery(queue->ctrl);
1439                 return;
1440         }
1441         req = blk_mq_rq_to_pdu(rq);
1442
1443         req->status = cqe->status;
1444         req->result = cqe->result;
1445
1446         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1447                 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1448                         dev_err(queue->ctrl->ctrl.device,
1449                                 "Bogus remote invalidation for rkey %#x\n",
1450                                 req->mr->rkey);
1451                         nvme_rdma_error_recovery(queue->ctrl);
1452                 }
1453         } else if (req->mr) {
1454                 int ret;
1455
1456                 ret = nvme_rdma_inv_rkey(queue, req);
1457                 if (unlikely(ret < 0)) {
1458                         dev_err(queue->ctrl->ctrl.device,
1459                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1460                                 req->mr->rkey, ret);
1461                         nvme_rdma_error_recovery(queue->ctrl);
1462                 }
1463                 /* the local invalidation completion will end the request */
1464                 return;
1465         }
1466
1467         if (refcount_dec_and_test(&req->ref))
1468                 nvme_end_request(rq, req->status, req->result);
1469 }
1470
1471 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1472 {
1473         struct nvme_rdma_qe *qe =
1474                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1475         struct nvme_rdma_queue *queue = cq->cq_context;
1476         struct ib_device *ibdev = queue->device->dev;
1477         struct nvme_completion *cqe = qe->data;
1478         const size_t len = sizeof(struct nvme_completion);
1479
1480         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1481                 nvme_rdma_wr_error(cq, wc, "RECV");
1482                 return;
1483         }
1484
1485         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1486         /*
1487          * AEN requests are special as they don't time out and can
1488          * survive any kind of queue freeze and often don't respond to
1489          * aborts.  We don't even bother to allocate a struct request
1490          * for them but rather special case them here.
1491          */
1492         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1493                         cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1494                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1495                                 &cqe->result);
1496         else
1497                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1498         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1499
1500         nvme_rdma_post_recv(queue, qe);
1501 }
1502
1503 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1504 {
1505         int ret, i;
1506
1507         for (i = 0; i < queue->queue_size; i++) {
1508                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1509                 if (ret)
1510                         goto out_destroy_queue_ib;
1511         }
1512
1513         return 0;
1514
1515 out_destroy_queue_ib:
1516         nvme_rdma_destroy_queue_ib(queue);
1517         return ret;
1518 }
1519
1520 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1521                 struct rdma_cm_event *ev)
1522 {
1523         struct rdma_cm_id *cm_id = queue->cm_id;
1524         int status = ev->status;
1525         const char *rej_msg;
1526         const struct nvme_rdma_cm_rej *rej_data;
1527         u8 rej_data_len;
1528
1529         rej_msg = rdma_reject_msg(cm_id, status);
1530         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1531
1532         if (rej_data && rej_data_len >= sizeof(u16)) {
1533                 u16 sts = le16_to_cpu(rej_data->sts);
1534
1535                 dev_err(queue->ctrl->ctrl.device,
1536                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1537                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1538         } else {
1539                 dev_err(queue->ctrl->ctrl.device,
1540                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1541         }
1542
1543         return -ECONNRESET;
1544 }
1545
1546 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1547 {
1548         int ret;
1549
1550         ret = nvme_rdma_create_queue_ib(queue);
1551         if (ret)
1552                 return ret;
1553
1554         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1555         if (ret) {
1556                 dev_err(queue->ctrl->ctrl.device,
1557                         "rdma_resolve_route failed (%d).\n",
1558                         queue->cm_error);
1559                 goto out_destroy_queue;
1560         }
1561
1562         return 0;
1563
1564 out_destroy_queue:
1565         nvme_rdma_destroy_queue_ib(queue);
1566         return ret;
1567 }
1568
1569 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1570 {
1571         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1572         struct rdma_conn_param param = { };
1573         struct nvme_rdma_cm_req priv = { };
1574         int ret;
1575
1576         param.qp_num = queue->qp->qp_num;
1577         param.flow_control = 1;
1578
1579         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1580         /* maximum retry count */
1581         param.retry_count = 7;
1582         param.rnr_retry_count = 7;
1583         param.private_data = &priv;
1584         param.private_data_len = sizeof(priv);
1585
1586         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1587         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1588         /*
1589          * set the admin queue depth to the minimum size
1590          * specified by the Fabrics standard.
1591          */
1592         if (priv.qid == 0) {
1593                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1594                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1595         } else {
1596                 /*
1597                  * current interpretation of the fabrics spec
1598                  * is at minimum you make hrqsize sqsize+1, or a
1599                  * 1's based representation of sqsize.
1600                  */
1601                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1602                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1603         }
1604
1605         ret = rdma_connect(queue->cm_id, &param);
1606         if (ret) {
1607                 dev_err(ctrl->ctrl.device,
1608                         "rdma_connect failed (%d).\n", ret);
1609                 goto out_destroy_queue_ib;
1610         }
1611
1612         return 0;
1613
1614 out_destroy_queue_ib:
1615         nvme_rdma_destroy_queue_ib(queue);
1616         return ret;
1617 }
1618
1619 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1620                 struct rdma_cm_event *ev)
1621 {
1622         struct nvme_rdma_queue *queue = cm_id->context;
1623         int cm_error = 0;
1624
1625         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1626                 rdma_event_msg(ev->event), ev->event,
1627                 ev->status, cm_id);
1628
1629         switch (ev->event) {
1630         case RDMA_CM_EVENT_ADDR_RESOLVED:
1631                 cm_error = nvme_rdma_addr_resolved(queue);
1632                 break;
1633         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1634                 cm_error = nvme_rdma_route_resolved(queue);
1635                 break;
1636         case RDMA_CM_EVENT_ESTABLISHED:
1637                 queue->cm_error = nvme_rdma_conn_established(queue);
1638                 /* complete cm_done regardless of success/failure */
1639                 complete(&queue->cm_done);
1640                 return 0;
1641         case RDMA_CM_EVENT_REJECTED:
1642                 nvme_rdma_destroy_queue_ib(queue);
1643                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1644                 break;
1645         case RDMA_CM_EVENT_ROUTE_ERROR:
1646         case RDMA_CM_EVENT_CONNECT_ERROR:
1647         case RDMA_CM_EVENT_UNREACHABLE:
1648                 nvme_rdma_destroy_queue_ib(queue);
1649                 /* fall through */
1650         case RDMA_CM_EVENT_ADDR_ERROR:
1651                 dev_dbg(queue->ctrl->ctrl.device,
1652                         "CM error event %d\n", ev->event);
1653                 cm_error = -ECONNRESET;
1654                 break;
1655         case RDMA_CM_EVENT_DISCONNECTED:
1656         case RDMA_CM_EVENT_ADDR_CHANGE:
1657         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1658                 dev_dbg(queue->ctrl->ctrl.device,
1659                         "disconnect received - connection closed\n");
1660                 nvme_rdma_error_recovery(queue->ctrl);
1661                 break;
1662         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1663                 /* device removal is handled via the ib_client API */
1664                 break;
1665         default:
1666                 dev_err(queue->ctrl->ctrl.device,
1667                         "Unexpected RDMA CM event (%d)\n", ev->event);
1668                 nvme_rdma_error_recovery(queue->ctrl);
1669                 break;
1670         }
1671
1672         if (cm_error) {
1673                 queue->cm_error = cm_error;
1674                 complete(&queue->cm_done);
1675         }
1676
1677         return 0;
1678 }
1679
1680 static enum blk_eh_timer_return
1681 nvme_rdma_timeout(struct request *rq, bool reserved)
1682 {
1683         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1684         struct nvme_rdma_queue *queue = req->queue;
1685         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1686
1687         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1688                  rq->tag, nvme_rdma_queue_idx(queue));
1689
1690         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1691                 /*
1692                  * Teardown immediately if controller times out while starting
1693                  * or we are already started error recovery. all outstanding
1694                  * requests are completed on shutdown, so we return BLK_EH_DONE.
1695                  */
1696                 flush_work(&ctrl->err_work);
1697                 nvme_rdma_teardown_io_queues(ctrl, false);
1698                 nvme_rdma_teardown_admin_queue(ctrl, false);
1699                 return BLK_EH_DONE;
1700         }
1701
1702         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1703         nvme_rdma_error_recovery(ctrl);
1704
1705         return BLK_EH_RESET_TIMER;
1706 }
1707
1708 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1709                 const struct blk_mq_queue_data *bd)
1710 {
1711         struct nvme_ns *ns = hctx->queue->queuedata;
1712         struct nvme_rdma_queue *queue = hctx->driver_data;
1713         struct request *rq = bd->rq;
1714         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1715         struct nvme_rdma_qe *sqe = &req->sqe;
1716         struct nvme_command *c = sqe->data;
1717         struct ib_device *dev;
1718         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1719         blk_status_t ret;
1720         int err;
1721
1722         WARN_ON_ONCE(rq->tag < 0);
1723
1724         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1725                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1726
1727         dev = queue->device->dev;
1728
1729         req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
1730                                          sizeof(struct nvme_command),
1731                                          DMA_TO_DEVICE);
1732         err = ib_dma_mapping_error(dev, req->sqe.dma);
1733         if (unlikely(err))
1734                 return BLK_STS_RESOURCE;
1735
1736         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1737                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1738
1739         ret = nvme_setup_cmd(ns, rq, c);
1740         if (ret)
1741                 goto unmap_qe;
1742
1743         blk_mq_start_request(rq);
1744
1745         err = nvme_rdma_map_data(queue, rq, c);
1746         if (unlikely(err < 0)) {
1747                 dev_err(queue->ctrl->ctrl.device,
1748                              "Failed to map data (%d)\n", err);
1749                 nvme_cleanup_cmd(rq);
1750                 goto err;
1751         }
1752
1753         sqe->cqe.done = nvme_rdma_send_done;
1754
1755         ib_dma_sync_single_for_device(dev, sqe->dma,
1756                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1757
1758         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1759                         req->mr ? &req->reg_wr.wr : NULL);
1760         if (unlikely(err)) {
1761                 nvme_rdma_unmap_data(queue, rq);
1762                 goto err;
1763         }
1764
1765         return BLK_STS_OK;
1766
1767 err:
1768         if (err == -ENOMEM || err == -EAGAIN)
1769                 ret = BLK_STS_RESOURCE;
1770         else
1771                 ret = BLK_STS_IOERR;
1772 unmap_qe:
1773         ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
1774                             DMA_TO_DEVICE);
1775         return ret;
1776 }
1777
1778 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
1779 {
1780         struct nvme_rdma_queue *queue = hctx->driver_data;
1781
1782         return ib_process_cq_direct(queue->ib_cq, -1);
1783 }
1784
1785 static void nvme_rdma_complete_rq(struct request *rq)
1786 {
1787         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1788         struct nvme_rdma_queue *queue = req->queue;
1789         struct ib_device *ibdev = queue->device->dev;
1790
1791         nvme_rdma_unmap_data(queue, rq);
1792         ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
1793                             DMA_TO_DEVICE);
1794         nvme_complete_rq(rq);
1795 }
1796
1797 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1798 {
1799         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1800         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1801
1802         if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
1803                 /* separate read/write queues */
1804                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1805                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
1806                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1807                 set->map[HCTX_TYPE_READ].nr_queues =
1808                         ctrl->io_queues[HCTX_TYPE_READ];
1809                 set->map[HCTX_TYPE_READ].queue_offset =
1810                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
1811         } else {
1812                 /* shared read/write queues */
1813                 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1814                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
1815                 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1816                 set->map[HCTX_TYPE_READ].nr_queues =
1817                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
1818                 set->map[HCTX_TYPE_READ].queue_offset = 0;
1819         }
1820         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1821                         ctrl->device->dev, 0);
1822         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1823                         ctrl->device->dev, 0);
1824
1825         if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
1826                 /* map dedicated poll queues only if we have queues left */
1827                 set->map[HCTX_TYPE_POLL].nr_queues =
1828                                 ctrl->io_queues[HCTX_TYPE_POLL];
1829                 set->map[HCTX_TYPE_POLL].queue_offset =
1830                         ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1831                         ctrl->io_queues[HCTX_TYPE_READ];
1832                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1833         }
1834
1835         dev_info(ctrl->ctrl.device,
1836                 "mapped %d/%d/%d default/read/poll queues.\n",
1837                 ctrl->io_queues[HCTX_TYPE_DEFAULT],
1838                 ctrl->io_queues[HCTX_TYPE_READ],
1839                 ctrl->io_queues[HCTX_TYPE_POLL]);
1840
1841         return 0;
1842 }
1843
1844 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1845         .queue_rq       = nvme_rdma_queue_rq,
1846         .complete       = nvme_rdma_complete_rq,
1847         .init_request   = nvme_rdma_init_request,
1848         .exit_request   = nvme_rdma_exit_request,
1849         .init_hctx      = nvme_rdma_init_hctx,
1850         .timeout        = nvme_rdma_timeout,
1851         .map_queues     = nvme_rdma_map_queues,
1852         .poll           = nvme_rdma_poll,
1853 };
1854
1855 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1856         .queue_rq       = nvme_rdma_queue_rq,
1857         .complete       = nvme_rdma_complete_rq,
1858         .init_request   = nvme_rdma_init_request,
1859         .exit_request   = nvme_rdma_exit_request,
1860         .init_hctx      = nvme_rdma_init_admin_hctx,
1861         .timeout        = nvme_rdma_timeout,
1862 };
1863
1864 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1865 {
1866         cancel_work_sync(&ctrl->err_work);
1867         cancel_delayed_work_sync(&ctrl->reconnect_work);
1868
1869         nvme_rdma_teardown_io_queues(ctrl, shutdown);
1870         if (shutdown)
1871                 nvme_shutdown_ctrl(&ctrl->ctrl);
1872         else
1873                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1874         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1875 }
1876
1877 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1878 {
1879         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1880 }
1881
1882 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1883 {
1884         struct nvme_rdma_ctrl *ctrl =
1885                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1886
1887         nvme_stop_ctrl(&ctrl->ctrl);
1888         nvme_rdma_shutdown_ctrl(ctrl, false);
1889
1890         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1891                 /* state change failure should never happen */
1892                 WARN_ON_ONCE(1);
1893                 return;
1894         }
1895
1896         if (nvme_rdma_setup_ctrl(ctrl, false))
1897                 goto out_fail;
1898
1899         return;
1900
1901 out_fail:
1902         ++ctrl->ctrl.nr_reconnects;
1903         nvme_rdma_reconnect_or_remove(ctrl);
1904 }
1905
1906 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1907         .name                   = "rdma",
1908         .module                 = THIS_MODULE,
1909         .flags                  = NVME_F_FABRICS,
1910         .reg_read32             = nvmf_reg_read32,
1911         .reg_read64             = nvmf_reg_read64,
1912         .reg_write32            = nvmf_reg_write32,
1913         .free_ctrl              = nvme_rdma_free_ctrl,
1914         .submit_async_event     = nvme_rdma_submit_async_event,
1915         .delete_ctrl            = nvme_rdma_delete_ctrl,
1916         .get_address            = nvmf_get_address,
1917 };
1918
1919 /*
1920  * Fails a connection request if it matches an existing controller
1921  * (association) with the same tuple:
1922  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1923  *
1924  * if local address is not specified in the request, it will match an
1925  * existing controller with all the other parameters the same and no
1926  * local port address specified as well.
1927  *
1928  * The ports don't need to be compared as they are intrinsically
1929  * already matched by the port pointers supplied.
1930  */
1931 static bool
1932 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1933 {
1934         struct nvme_rdma_ctrl *ctrl;
1935         bool found = false;
1936
1937         mutex_lock(&nvme_rdma_ctrl_mutex);
1938         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1939                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1940                 if (found)
1941                         break;
1942         }
1943         mutex_unlock(&nvme_rdma_ctrl_mutex);
1944
1945         return found;
1946 }
1947
1948 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1949                 struct nvmf_ctrl_options *opts)
1950 {
1951         struct nvme_rdma_ctrl *ctrl;
1952         int ret;
1953         bool changed;
1954
1955         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1956         if (!ctrl)
1957                 return ERR_PTR(-ENOMEM);
1958         ctrl->ctrl.opts = opts;
1959         INIT_LIST_HEAD(&ctrl->list);
1960
1961         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1962                 opts->trsvcid =
1963                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1964                 if (!opts->trsvcid) {
1965                         ret = -ENOMEM;
1966                         goto out_free_ctrl;
1967                 }
1968                 opts->mask |= NVMF_OPT_TRSVCID;
1969         }
1970
1971         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1972                         opts->traddr, opts->trsvcid, &ctrl->addr);
1973         if (ret) {
1974                 pr_err("malformed address passed: %s:%s\n",
1975                         opts->traddr, opts->trsvcid);
1976                 goto out_free_ctrl;
1977         }
1978
1979         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1980                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1981                         opts->host_traddr, NULL, &ctrl->src_addr);
1982                 if (ret) {
1983                         pr_err("malformed src address passed: %s\n",
1984                                opts->host_traddr);
1985                         goto out_free_ctrl;
1986                 }
1987         }
1988
1989         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1990                 ret = -EALREADY;
1991                 goto out_free_ctrl;
1992         }
1993
1994         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1995                         nvme_rdma_reconnect_ctrl_work);
1996         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1997         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1998
1999         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2000                                 opts->nr_poll_queues + 1;
2001         ctrl->ctrl.sqsize = opts->queue_size - 1;
2002         ctrl->ctrl.kato = opts->kato;
2003
2004         ret = -ENOMEM;
2005         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2006                                 GFP_KERNEL);
2007         if (!ctrl->queues)
2008                 goto out_free_ctrl;
2009
2010         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2011                                 0 /* no quirks, we're perfect! */);
2012         if (ret)
2013                 goto out_kfree_queues;
2014
2015         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2016         WARN_ON_ONCE(!changed);
2017
2018         ret = nvme_rdma_setup_ctrl(ctrl, true);
2019         if (ret)
2020                 goto out_uninit_ctrl;
2021
2022         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2023                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2024
2025         nvme_get_ctrl(&ctrl->ctrl);
2026
2027         mutex_lock(&nvme_rdma_ctrl_mutex);
2028         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2029         mutex_unlock(&nvme_rdma_ctrl_mutex);
2030
2031         return &ctrl->ctrl;
2032
2033 out_uninit_ctrl:
2034         nvme_uninit_ctrl(&ctrl->ctrl);
2035         nvme_put_ctrl(&ctrl->ctrl);
2036         if (ret > 0)
2037                 ret = -EIO;
2038         return ERR_PTR(ret);
2039 out_kfree_queues:
2040         kfree(ctrl->queues);
2041 out_free_ctrl:
2042         kfree(ctrl);
2043         return ERR_PTR(ret);
2044 }
2045
2046 static struct nvmf_transport_ops nvme_rdma_transport = {
2047         .name           = "rdma",
2048         .module         = THIS_MODULE,
2049         .required_opts  = NVMF_OPT_TRADDR,
2050         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2051                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2052                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES,
2053         .create_ctrl    = nvme_rdma_create_ctrl,
2054 };
2055
2056 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2057 {
2058         struct nvme_rdma_ctrl *ctrl;
2059         struct nvme_rdma_device *ndev;
2060         bool found = false;
2061
2062         mutex_lock(&device_list_mutex);
2063         list_for_each_entry(ndev, &device_list, entry) {
2064                 if (ndev->dev == ib_device) {
2065                         found = true;
2066                         break;
2067                 }
2068         }
2069         mutex_unlock(&device_list_mutex);
2070
2071         if (!found)
2072                 return;
2073
2074         /* Delete all controllers using this device */
2075         mutex_lock(&nvme_rdma_ctrl_mutex);
2076         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2077                 if (ctrl->device->dev != ib_device)
2078                         continue;
2079                 nvme_delete_ctrl(&ctrl->ctrl);
2080         }
2081         mutex_unlock(&nvme_rdma_ctrl_mutex);
2082
2083         flush_workqueue(nvme_delete_wq);
2084 }
2085
2086 static struct ib_client nvme_rdma_ib_client = {
2087         .name   = "nvme_rdma",
2088         .remove = nvme_rdma_remove_one
2089 };
2090
2091 static int __init nvme_rdma_init_module(void)
2092 {
2093         int ret;
2094
2095         ret = ib_register_client(&nvme_rdma_ib_client);
2096         if (ret)
2097                 return ret;
2098
2099         ret = nvmf_register_transport(&nvme_rdma_transport);
2100         if (ret)
2101                 goto err_unreg_client;
2102
2103         return 0;
2104
2105 err_unreg_client:
2106         ib_unregister_client(&nvme_rdma_ib_client);
2107         return ret;
2108 }
2109
2110 static void __exit nvme_rdma_cleanup_module(void)
2111 {
2112         nvmf_unregister_transport(&nvme_rdma_transport);
2113         ib_unregister_client(&nvme_rdma_ib_client);
2114 }
2115
2116 module_init(nvme_rdma_init_module);
2117 module_exit(nvme_rdma_cleanup_module);
2118
2119 MODULE_LICENSE("GPL v2");