drivers/nvme/host/rdma.c

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