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