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