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         blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
 757
 758         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 759         if (ret)
 760                 goto stop_admin_q;
 761
 762         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
 763
 764         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 765         if (ret)
 766                 goto stop_admin_q;
 767
 768         nvme_start_keep_alive(&ctrl->ctrl);
 769
 770         if (ctrl->queue_count > 1) {
 771                 ret = nvme_rdma_init_io_queues(ctrl);
 772                 if (ret)
 773                         goto stop_admin_q;
 774
 775                 ret = nvme_rdma_connect_io_queues(ctrl);
 776                 if (ret)
 777                         goto stop_admin_q;
 778         }
 779
 780         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 781         WARN_ON_ONCE(!changed);
 782         ctrl->ctrl.opts->nr_reconnects = 0;
 783
 784         if (ctrl->queue_count > 1) {
 785                 nvme_start_queues(&ctrl->ctrl);
 786                 nvme_queue_scan(&ctrl->ctrl);
 787                 nvme_queue_async_events(&ctrl->ctrl);
 788         }
 789
 790         dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
 791
 792         return;
 793
 794 stop_admin_q:
 795         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 796 requeue:
 797         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
 798                         ctrl->ctrl.opts->nr_reconnects);
 799         nvme_rdma_reconnect_or_remove(ctrl);
 800 }
 801
 802 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 803 {
 804         struct nvme_rdma_ctrl *ctrl = container_of(work,
 805                         struct nvme_rdma_ctrl, err_work);
 806         int i;
 807
 808         nvme_stop_keep_alive(&ctrl->ctrl);
 809
 810         for (i = 0; i < ctrl->queue_count; i++) {
 811                 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
 812                 clear_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 813         }
 814
 815         if (ctrl->queue_count > 1)
 816                 nvme_stop_queues(&ctrl->ctrl);
 817         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 818
 819         /* We must take care of fastfail/requeue all our inflight requests */
 820         if (ctrl->queue_count > 1)
 821                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 822                                         nvme_cancel_request, &ctrl->ctrl);
 823         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 824                                 nvme_cancel_request, &ctrl->ctrl);
 825
 826         nvme_rdma_reconnect_or_remove(ctrl);
 827 }
 828
 829 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 830 {
 831         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 832                 return;
 833
 834         queue_work(nvme_rdma_wq, &ctrl->err_work);
 835 }
 836
 837 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 838                 const char *op)
 839 {
 840         struct nvme_rdma_queue *queue = cq->cq_context;
 841         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 842
 843         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 844                 dev_info(ctrl->ctrl.device,
 845                              "%s for CQE 0x%p failed with status %s (%d)\n",
 846                              op, wc->wr_cqe,
 847                              ib_wc_status_msg(wc->status), wc->status);
 848         nvme_rdma_error_recovery(ctrl);
 849 }
 850
 851 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
 852 {
 853         if (unlikely(wc->status != IB_WC_SUCCESS))
 854                 nvme_rdma_wr_error(cq, wc, "MEMREG");
 855 }
 856
 857 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 858 {
 859         if (unlikely(wc->status != IB_WC_SUCCESS))
 860                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
 861 }
 862
 863 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
 864                 struct nvme_rdma_request *req)
 865 {
 866         struct ib_send_wr *bad_wr;
 867         struct ib_send_wr wr = {
 868                 .opcode             = IB_WR_LOCAL_INV,
 869                 .next               = NULL,
 870                 .num_sge            = 0,
 871                 .send_flags         = 0,
 872                 .ex.invalidate_rkey = req->mr->rkey,
 873         };
 874
 875         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
 876         wr.wr_cqe = &req->reg_cqe;
 877
 878         return ib_post_send(queue->qp, &wr, &bad_wr);
 879 }
 880
 881 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 882                 struct request *rq)
 883 {
 884         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 885         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 886         struct nvme_rdma_device *dev = queue->device;
 887         struct ib_device *ibdev = dev->dev;
 888         int res;
 889
 890         if (!blk_rq_bytes(rq))
 891                 return;
 892
 893         if (req->mr->need_inval) {
 894                 res = nvme_rdma_inv_rkey(queue, req);
 895                 if (res < 0) {
 896                         dev_err(ctrl->ctrl.device,
 897                                 "Queueing INV WR for rkey %#x failed (%d)\n",
 898                                 req->mr->rkey, res);
 899                         nvme_rdma_error_recovery(queue->ctrl);
 900                 }
 901         }
 902
 903         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 904                         req->nents, rq_data_dir(rq) ==
 905                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 906
 907         nvme_cleanup_cmd(rq);
 908         sg_free_table_chained(&req->sg_table, true);
 909 }
 910
 911 static int nvme_rdma_set_sg_null(struct nvme_command *c)
 912 {
 913         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 914
 915         sg->addr = 0;
 916         put_unaligned_le24(0, sg->length);
 917         put_unaligned_le32(0, sg->key);
 918         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 919         return 0;
 920 }
 921
 922 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
 923                 struct nvme_rdma_request *req, struct nvme_command *c)
 924 {
 925         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
 926
 927         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
 928         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
 929         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
 930
 931         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
 932         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
 933         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
 934
 935         req->inline_data = true;
 936         req->num_sge++;
 937         return 0;
 938 }
 939
 940 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
 941                 struct nvme_rdma_request *req, struct nvme_command *c)
 942 {
 943         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 944
 945         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
 946         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
 947         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
 948         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 949         return 0;
 950 }
 951
 952 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 953                 struct nvme_rdma_request *req, struct nvme_command *c,
 954                 int count)
 955 {
 956         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 957         int nr;
 958
 959         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
 960         if (nr < count) {
 961                 if (nr < 0)
 962                         return nr;
 963                 return -EINVAL;
 964         }
 965
 966         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
 967
 968         req->reg_cqe.done = nvme_rdma_memreg_done;
 969         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
 970         req->reg_wr.wr.opcode = IB_WR_REG_MR;
 971         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
 972         req->reg_wr.wr.num_sge = 0;
 973         req->reg_wr.mr = req->mr;
 974         req->reg_wr.key = req->mr->rkey;
 975         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 976                              IB_ACCESS_REMOTE_READ |
 977                              IB_ACCESS_REMOTE_WRITE;
 978
 979         req->mr->need_inval = true;
 980
 981         sg->addr = cpu_to_le64(req->mr->iova);
 982         put_unaligned_le24(req->mr->length, sg->length);
 983         put_unaligned_le32(req->mr->rkey, sg->key);
 984         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
 985                         NVME_SGL_FMT_INVALIDATE;
 986
 987         return 0;
 988 }
 989
 990 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 991                 struct request *rq, struct nvme_command *c)
 992 {
 993         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 994         struct nvme_rdma_device *dev = queue->device;
 995         struct ib_device *ibdev = dev->dev;
 996         int count, ret;
 997
 998         req->num_sge = 1;
 999         req->inline_data = false;
1000         req->mr->need_inval = false;
1001
1002         c->common.flags |= NVME_CMD_SGL_METABUF;
1003
1004         if (!blk_rq_bytes(rq))
1005                 return nvme_rdma_set_sg_null(c);
1006
1007         req->sg_table.sgl = req->first_sgl;
1008         ret = sg_alloc_table_chained(&req->sg_table,
1009                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1010         if (ret)
1011                 return -ENOMEM;
1012
1013         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1014
1015         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1016                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1017         if (unlikely(count <= 0)) {
1018                 sg_free_table_chained(&req->sg_table, true);
1019                 return -EIO;
1020         }
1021
1022         if (count == 1) {
1023                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1024                     blk_rq_payload_bytes(rq) <=
1025                                 nvme_rdma_inline_data_size(queue))
1026                         return nvme_rdma_map_sg_inline(queue, req, c);
1027
1028                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1029                         return nvme_rdma_map_sg_single(queue, req, c);
1030         }
1031
1032         return nvme_rdma_map_sg_fr(queue, req, c, count);
1033 }
1034
1035 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1036 {
1037         if (unlikely(wc->status != IB_WC_SUCCESS))
1038                 nvme_rdma_wr_error(cq, wc, "SEND");
1039 }
1040
1041 static inline int nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1042 {
1043         int sig_limit;
1044
1045         /*
1046          * We signal completion every queue depth/2 and also handle the
1047          * degenerated case of a  device with queue_depth=1, where we
1048          * would need to signal every message.
1049          */
1050         sig_limit = max(queue->queue_size / 2, 1);
1051         return (++queue->sig_count % sig_limit) == 0;
1052 }
1053
1054 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1055                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1056                 struct ib_send_wr *first, bool flush)
1057 {
1058         struct ib_send_wr wr, *bad_wr;
1059         int ret;
1060
1061         sge->addr   = qe->dma;
1062         sge->length = sizeof(struct nvme_command),
1063         sge->lkey   = queue->device->pd->local_dma_lkey;
1064
1065         qe->cqe.done = nvme_rdma_send_done;
1066
1067         wr.next       = NULL;
1068         wr.wr_cqe     = &qe->cqe;
1069         wr.sg_list    = sge;
1070         wr.num_sge    = num_sge;
1071         wr.opcode     = IB_WR_SEND;
1072         wr.send_flags = 0;
1073
1074         /*
1075          * Unsignalled send completions are another giant desaster in the
1076          * IB Verbs spec:  If we don't regularly post signalled sends
1077          * the send queue will fill up and only a QP reset will rescue us.
1078          * Would have been way to obvious to handle this in hardware or
1079          * at least the RDMA stack..
1080          *
1081          * Always signal the flushes. The magic request used for the flush
1082          * sequencer is not allocated in our driver's tagset and it's
1083          * triggered to be freed by blk_cleanup_queue(). So we need to
1084          * always mark it as signaled to ensure that the "wr_cqe", which is
1085          * embedded in request's payload, is not freed when __ib_process_cq()
1086          * calls wr_cqe->done().
1087          */
1088         if (nvme_rdma_queue_sig_limit(queue) || flush)
1089                 wr.send_flags |= IB_SEND_SIGNALED;
1090
1091         if (first)
1092                 first->next = &wr;
1093         else
1094                 first = &wr;
1095
1096         ret = ib_post_send(queue->qp, first, &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 int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1105                 struct nvme_rdma_qe *qe)
1106 {
1107         struct ib_recv_wr wr, *bad_wr;
1108         struct ib_sge list;
1109         int ret;
1110
1111         list.addr   = qe->dma;
1112         list.length = sizeof(struct nvme_completion);
1113         list.lkey   = queue->device->pd->local_dma_lkey;
1114
1115         qe->cqe.done = nvme_rdma_recv_done;
1116
1117         wr.next     = NULL;
1118         wr.wr_cqe   = &qe->cqe;
1119         wr.sg_list  = &list;
1120         wr.num_sge  = 1;
1121
1122         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1123         if (ret) {
1124                 dev_err(queue->ctrl->ctrl.device,
1125                         "%s failed with error code %d\n", __func__, ret);
1126         }
1127         return ret;
1128 }
1129
1130 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1131 {
1132         u32 queue_idx = nvme_rdma_queue_idx(queue);
1133
1134         if (queue_idx == 0)
1135                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1136         return queue->ctrl->tag_set.tags[queue_idx - 1];
1137 }
1138
1139 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1140 {
1141         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1142         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1143         struct ib_device *dev = queue->device->dev;
1144         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1145         struct nvme_command *cmd = sqe->data;
1146         struct ib_sge sge;
1147         int ret;
1148
1149         if (WARN_ON_ONCE(aer_idx != 0))
1150                 return;
1151
1152         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1153
1154         memset(cmd, 0, sizeof(*cmd));
1155         cmd->common.opcode = nvme_admin_async_event;
1156         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1157         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1158         nvme_rdma_set_sg_null(cmd);
1159
1160         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1161                         DMA_TO_DEVICE);
1162
1163         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1164         WARN_ON_ONCE(ret);
1165 }
1166
1167 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1168                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1169 {
1170         struct request *rq;
1171         struct nvme_rdma_request *req;
1172         int ret = 0;
1173
1174         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1175         if (!rq) {
1176                 dev_err(queue->ctrl->ctrl.device,
1177                         "tag 0x%x on QP %#x not found\n",
1178                         cqe->command_id, queue->qp->qp_num);
1179                 nvme_rdma_error_recovery(queue->ctrl);
1180                 return ret;
1181         }
1182         req = blk_mq_rq_to_pdu(rq);
1183
1184         if (rq->tag == tag)
1185                 ret = 1;
1186
1187         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1188             wc->ex.invalidate_rkey == req->mr->rkey)
1189                 req->mr->need_inval = false;
1190
1191         nvme_end_request(rq, cqe->status, cqe->result);
1192         return ret;
1193 }
1194
1195 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1196 {
1197         struct nvme_rdma_qe *qe =
1198                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1199         struct nvme_rdma_queue *queue = cq->cq_context;
1200         struct ib_device *ibdev = queue->device->dev;
1201         struct nvme_completion *cqe = qe->data;
1202         const size_t len = sizeof(struct nvme_completion);
1203         int ret = 0;
1204
1205         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1206                 nvme_rdma_wr_error(cq, wc, "RECV");
1207                 return 0;
1208         }
1209
1210         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1211         /*
1212          * AEN requests are special as they don't time out and can
1213          * survive any kind of queue freeze and often don't respond to
1214          * aborts.  We don't even bother to allocate a struct request
1215          * for them but rather special case them here.
1216          */
1217         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1218                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1219                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1220                                 &cqe->result);
1221         else
1222                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1223         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1224
1225         nvme_rdma_post_recv(queue, qe);
1226         return ret;
1227 }
1228
1229 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1230 {
1231         __nvme_rdma_recv_done(cq, wc, -1);
1232 }
1233
1234 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1235 {
1236         int ret, i;
1237
1238         for (i = 0; i < queue->queue_size; i++) {
1239                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1240                 if (ret)
1241                         goto out_destroy_queue_ib;
1242         }
1243
1244         return 0;
1245
1246 out_destroy_queue_ib:
1247         nvme_rdma_destroy_queue_ib(queue);
1248         return ret;
1249 }
1250
1251 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1252                 struct rdma_cm_event *ev)
1253 {
1254         struct rdma_cm_id *cm_id = queue->cm_id;
1255         int status = ev->status;
1256         const char *rej_msg;
1257         const struct nvme_rdma_cm_rej *rej_data;
1258         u8 rej_data_len;
1259
1260         rej_msg = rdma_reject_msg(cm_id, status);
1261         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1262
1263         if (rej_data && rej_data_len >= sizeof(u16)) {
1264                 u16 sts = le16_to_cpu(rej_data->sts);
1265
1266                 dev_err(queue->ctrl->ctrl.device,
1267                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1268                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1269         } else {
1270                 dev_err(queue->ctrl->ctrl.device,
1271                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1272         }
1273
1274         return -ECONNRESET;
1275 }
1276
1277 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1278 {
1279         struct nvme_rdma_device *dev;
1280         int ret;
1281
1282         dev = nvme_rdma_find_get_device(queue->cm_id);
1283         if (!dev) {
1284                 dev_err(queue->cm_id->device->dev.parent,
1285                         "no client data found!\n");
1286                 return -ECONNREFUSED;
1287         }
1288
1289         ret = nvme_rdma_create_queue_ib(queue, dev);
1290         if (ret) {
1291                 nvme_rdma_dev_put(dev);
1292                 goto out;
1293         }
1294
1295         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1296         if (ret) {
1297                 dev_err(queue->ctrl->ctrl.device,
1298                         "rdma_resolve_route failed (%d).\n",
1299                         queue->cm_error);
1300                 goto out_destroy_queue;
1301         }
1302
1303         return 0;
1304
1305 out_destroy_queue:
1306         nvme_rdma_destroy_queue_ib(queue);
1307 out:
1308         return ret;
1309 }
1310
1311 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1312 {
1313         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1314         struct rdma_conn_param param = { };
1315         struct nvme_rdma_cm_req priv = { };
1316         int ret;
1317
1318         param.qp_num = queue->qp->qp_num;
1319         param.flow_control = 1;
1320
1321         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1322         /* maximum retry count */
1323         param.retry_count = 7;
1324         param.rnr_retry_count = 7;
1325         param.private_data = &priv;
1326         param.private_data_len = sizeof(priv);
1327
1328         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1329         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1330         /*
1331          * set the admin queue depth to the minimum size
1332          * specified by the Fabrics standard.
1333          */
1334         if (priv.qid == 0) {
1335                 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1336                 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1337         } else {
1338                 /*
1339                  * current interpretation of the fabrics spec
1340                  * is at minimum you make hrqsize sqsize+1, or a
1341                  * 1's based representation of sqsize.
1342                  */
1343                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1344                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1345         }
1346
1347         ret = rdma_connect(queue->cm_id, &param);
1348         if (ret) {
1349                 dev_err(ctrl->ctrl.device,
1350                         "rdma_connect failed (%d).\n", ret);
1351                 goto out_destroy_queue_ib;
1352         }
1353
1354         return 0;
1355
1356 out_destroy_queue_ib:
1357         nvme_rdma_destroy_queue_ib(queue);
1358         return ret;
1359 }
1360
1361 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1362                 struct rdma_cm_event *ev)
1363 {
1364         struct nvme_rdma_queue *queue = cm_id->context;
1365         int cm_error = 0;
1366
1367         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1368                 rdma_event_msg(ev->event), ev->event,
1369                 ev->status, cm_id);
1370
1371         switch (ev->event) {
1372         case RDMA_CM_EVENT_ADDR_RESOLVED:
1373                 cm_error = nvme_rdma_addr_resolved(queue);
1374                 break;
1375         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1376                 cm_error = nvme_rdma_route_resolved(queue);
1377                 break;
1378         case RDMA_CM_EVENT_ESTABLISHED:
1379                 queue->cm_error = nvme_rdma_conn_established(queue);
1380                 /* complete cm_done regardless of success/failure */
1381                 complete(&queue->cm_done);
1382                 return 0;
1383         case RDMA_CM_EVENT_REJECTED:
1384                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1385                 break;
1386         case RDMA_CM_EVENT_ADDR_ERROR:
1387         case RDMA_CM_EVENT_ROUTE_ERROR:
1388         case RDMA_CM_EVENT_CONNECT_ERROR:
1389         case RDMA_CM_EVENT_UNREACHABLE:
1390                 dev_dbg(queue->ctrl->ctrl.device,
1391                         "CM error event %d\n", ev->event);
1392                 cm_error = -ECONNRESET;
1393                 break;
1394         case RDMA_CM_EVENT_DISCONNECTED:
1395         case RDMA_CM_EVENT_ADDR_CHANGE:
1396         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1397                 dev_dbg(queue->ctrl->ctrl.device,
1398                         "disconnect received - connection closed\n");
1399                 nvme_rdma_error_recovery(queue->ctrl);
1400                 break;
1401         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1402                 /* device removal is handled via the ib_client API */
1403                 break;
1404         default:
1405                 dev_err(queue->ctrl->ctrl.device,
1406                         "Unexpected RDMA CM event (%d)\n", ev->event);
1407                 nvme_rdma_error_recovery(queue->ctrl);
1408                 break;
1409         }
1410
1411         if (cm_error) {
1412                 queue->cm_error = cm_error;
1413                 complete(&queue->cm_done);
1414         }
1415
1416         return 0;
1417 }
1418
1419 static enum blk_eh_timer_return
1420 nvme_rdma_timeout(struct request *rq, bool reserved)
1421 {
1422         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1423
1424         /* queue error recovery */
1425         nvme_rdma_error_recovery(req->queue->ctrl);
1426
1427         /* fail with DNR on cmd timeout */
1428         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1429
1430         return BLK_EH_HANDLED;
1431 }
1432
1433 /*
1434  * We cannot accept any other command until the Connect command has completed.
1435  */
1436 static inline bool nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue,
1437                 struct request *rq)
1438 {
1439         if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1440                 struct nvme_command *cmd = nvme_req(rq)->cmd;
1441
1442                 if (!blk_rq_is_passthrough(rq) ||
1443                     cmd->common.opcode != nvme_fabrics_command ||
1444                     cmd->fabrics.fctype != nvme_fabrics_type_connect)
1445                         return false;
1446         }
1447
1448         return true;
1449 }
1450
1451 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1452                 const struct blk_mq_queue_data *bd)
1453 {
1454         struct nvme_ns *ns = hctx->queue->queuedata;
1455         struct nvme_rdma_queue *queue = hctx->driver_data;
1456         struct request *rq = bd->rq;
1457         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1458         struct nvme_rdma_qe *sqe = &req->sqe;
1459         struct nvme_command *c = sqe->data;
1460         bool flush = false;
1461         struct ib_device *dev;
1462         int ret;
1463
1464         WARN_ON_ONCE(rq->tag < 0);
1465
1466         if (!nvme_rdma_queue_is_ready(queue, rq))
1467                 return BLK_MQ_RQ_QUEUE_BUSY;
1468
1469         dev = queue->device->dev;
1470         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1471                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1472
1473         ret = nvme_setup_cmd(ns, rq, c);
1474         if (ret != BLK_MQ_RQ_QUEUE_OK)
1475                 return ret;
1476
1477         blk_mq_start_request(rq);
1478
1479         ret = nvme_rdma_map_data(queue, rq, c);
1480         if (ret < 0) {
1481                 dev_err(queue->ctrl->ctrl.device,
1482                              "Failed to map data (%d)\n", ret);
1483                 nvme_cleanup_cmd(rq);
1484                 goto err;
1485         }
1486
1487         ib_dma_sync_single_for_device(dev, sqe->dma,
1488                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1489
1490         if (req_op(rq) == REQ_OP_FLUSH)
1491                 flush = true;
1492         ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1493                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1494         if (ret) {
1495                 nvme_rdma_unmap_data(queue, rq);
1496                 goto err;
1497         }
1498
1499         return BLK_MQ_RQ_QUEUE_OK;
1500 err:
1501         return (ret == -ENOMEM || ret == -EAGAIN) ?
1502                 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1503 }
1504
1505 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1506 {
1507         struct nvme_rdma_queue *queue = hctx->driver_data;
1508         struct ib_cq *cq = queue->ib_cq;
1509         struct ib_wc wc;
1510         int found = 0;
1511
1512         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1513         while (ib_poll_cq(cq, 1, &wc) > 0) {
1514                 struct ib_cqe *cqe = wc.wr_cqe;
1515
1516                 if (cqe) {
1517                         if (cqe->done == nvme_rdma_recv_done)
1518                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1519                         else
1520                                 cqe->done(cq, &wc);
1521                 }
1522         }
1523
1524         return found;
1525 }
1526
1527 static void nvme_rdma_complete_rq(struct request *rq)
1528 {
1529         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1530
1531         nvme_rdma_unmap_data(req->queue, rq);
1532         nvme_complete_rq(rq);
1533 }
1534
1535 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1536         .queue_rq       = nvme_rdma_queue_rq,
1537         .complete       = nvme_rdma_complete_rq,
1538         .init_request   = nvme_rdma_init_request,
1539         .exit_request   = nvme_rdma_exit_request,
1540         .reinit_request = nvme_rdma_reinit_request,
1541         .init_hctx      = nvme_rdma_init_hctx,
1542         .poll           = nvme_rdma_poll,
1543         .timeout        = nvme_rdma_timeout,
1544 };
1545
1546 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1547         .queue_rq       = nvme_rdma_queue_rq,
1548         .complete       = nvme_rdma_complete_rq,
1549         .init_request   = nvme_rdma_init_admin_request,
1550         .exit_request   = nvme_rdma_exit_admin_request,
1551         .reinit_request = nvme_rdma_reinit_request,
1552         .init_hctx      = nvme_rdma_init_admin_hctx,
1553         .timeout        = nvme_rdma_timeout,
1554 };
1555
1556 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1557 {
1558         int error;
1559
1560         error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1561         if (error)
1562                 return error;
1563
1564         ctrl->device = ctrl->queues[0].device;
1565
1566         /*
1567          * We need a reference on the device as long as the tag_set is alive,
1568          * as the MRs in the request structures need a valid ib_device.
1569          */
1570         error = -EINVAL;
1571         if (!nvme_rdma_dev_get(ctrl->device))
1572                 goto out_free_queue;
1573
1574         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1575                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1576
1577         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1578         ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1579         ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1580         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1581         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1582         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1583                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1584         ctrl->admin_tag_set.driver_data = ctrl;
1585         ctrl->admin_tag_set.nr_hw_queues = 1;
1586         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1587
1588         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1589         if (error)
1590                 goto out_put_dev;
1591
1592         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1593         if (IS_ERR(ctrl->ctrl.admin_q)) {
1594                 error = PTR_ERR(ctrl->ctrl.admin_q);
1595                 goto out_free_tagset;
1596         }
1597
1598         error = nvmf_connect_admin_queue(&ctrl->ctrl);
1599         if (error)
1600                 goto out_cleanup_queue;
1601
1602         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
1603
1604         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1605         if (error) {
1606                 dev_err(ctrl->ctrl.device,
1607                         "prop_get NVME_REG_CAP failed\n");
1608                 goto out_cleanup_queue;
1609         }
1610
1611         ctrl->ctrl.sqsize =
1612                 min_t(int, NVME_CAP_MQES(ctrl->cap), ctrl->ctrl.sqsize);
1613
1614         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1615         if (error)
1616                 goto out_cleanup_queue;
1617
1618         ctrl->ctrl.max_hw_sectors =
1619                 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1620
1621         error = nvme_init_identify(&ctrl->ctrl);
1622         if (error)
1623                 goto out_cleanup_queue;
1624
1625         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1626                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
1627                         DMA_TO_DEVICE);
1628         if (error)
1629                 goto out_cleanup_queue;
1630
1631         nvme_start_keep_alive(&ctrl->ctrl);
1632
1633         return 0;
1634
1635 out_cleanup_queue:
1636         blk_cleanup_queue(ctrl->ctrl.admin_q);
1637 out_free_tagset:
1638         /* disconnect and drain the queue before freeing the tagset */
1639         nvme_rdma_stop_queue(&ctrl->queues[0]);
1640         blk_mq_free_tag_set(&ctrl->admin_tag_set);
1641 out_put_dev:
1642         nvme_rdma_dev_put(ctrl->device);
1643 out_free_queue:
1644         nvme_rdma_free_queue(&ctrl->queues[0]);
1645         return error;
1646 }
1647
1648 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1649 {
1650         nvme_stop_keep_alive(&ctrl->ctrl);
1651         cancel_work_sync(&ctrl->err_work);
1652         cancel_delayed_work_sync(&ctrl->reconnect_work);
1653
1654         if (ctrl->queue_count > 1) {
1655                 nvme_stop_queues(&ctrl->ctrl);
1656                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1657                                         nvme_cancel_request, &ctrl->ctrl);
1658                 nvme_rdma_free_io_queues(ctrl);
1659         }
1660
1661         if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1662                 nvme_shutdown_ctrl(&ctrl->ctrl);
1663
1664         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1665         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1666                                 nvme_cancel_request, &ctrl->ctrl);
1667         nvme_rdma_destroy_admin_queue(ctrl);
1668 }
1669
1670 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1671 {
1672         nvme_uninit_ctrl(&ctrl->ctrl);
1673         if (shutdown)
1674                 nvme_rdma_shutdown_ctrl(ctrl);
1675
1676         if (ctrl->ctrl.tagset) {
1677                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1678                 blk_mq_free_tag_set(&ctrl->tag_set);
1679                 nvme_rdma_dev_put(ctrl->device);
1680         }
1681
1682         nvme_put_ctrl(&ctrl->ctrl);
1683 }
1684
1685 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1686 {
1687         struct nvme_rdma_ctrl *ctrl = container_of(work,
1688                                 struct nvme_rdma_ctrl, delete_work);
1689
1690         __nvme_rdma_remove_ctrl(ctrl, true);
1691 }
1692
1693 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1694 {
1695         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1696                 return -EBUSY;
1697
1698         if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1699                 return -EBUSY;
1700
1701         return 0;
1702 }
1703
1704 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1705 {
1706         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1707         int ret = 0;
1708
1709         /*
1710          * Keep a reference until all work is flushed since
1711          * __nvme_rdma_del_ctrl can free the ctrl mem
1712          */
1713         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1714                 return -EBUSY;
1715         ret = __nvme_rdma_del_ctrl(ctrl);
1716         if (!ret)
1717                 flush_work(&ctrl->delete_work);
1718         nvme_put_ctrl(&ctrl->ctrl);
1719         return ret;
1720 }
1721
1722 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1723 {
1724         struct nvme_rdma_ctrl *ctrl = container_of(work,
1725                                 struct nvme_rdma_ctrl, delete_work);
1726
1727         __nvme_rdma_remove_ctrl(ctrl, false);
1728 }
1729
1730 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1731 {
1732         struct nvme_rdma_ctrl *ctrl = container_of(work,
1733                                         struct nvme_rdma_ctrl, reset_work);
1734         int ret;
1735         bool changed;
1736
1737         nvme_rdma_shutdown_ctrl(ctrl);
1738
1739         ret = nvme_rdma_configure_admin_queue(ctrl);
1740         if (ret) {
1741                 /* ctrl is already shutdown, just remove the ctrl */
1742                 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1743                 goto del_dead_ctrl;
1744         }
1745
1746         if (ctrl->queue_count > 1) {
1747                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1748                 if (ret)
1749                         goto del_dead_ctrl;
1750
1751                 ret = nvme_rdma_init_io_queues(ctrl);
1752                 if (ret)
1753                         goto del_dead_ctrl;
1754
1755                 ret = nvme_rdma_connect_io_queues(ctrl);
1756                 if (ret)
1757                         goto del_dead_ctrl;
1758         }
1759
1760         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1761         WARN_ON_ONCE(!changed);
1762
1763         if (ctrl->queue_count > 1) {
1764                 nvme_start_queues(&ctrl->ctrl);
1765                 nvme_queue_scan(&ctrl->ctrl);
1766                 nvme_queue_async_events(&ctrl->ctrl);
1767         }
1768
1769         return;
1770
1771 del_dead_ctrl:
1772         /* Deleting this dead controller... */
1773         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1774         WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1775 }
1776
1777 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1778 {
1779         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1780
1781         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1782                 return -EBUSY;
1783
1784         if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1785                 return -EBUSY;
1786
1787         flush_work(&ctrl->reset_work);
1788
1789         return 0;
1790 }
1791
1792 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1793         .name                   = "rdma",
1794         .module                 = THIS_MODULE,
1795         .flags                  = NVME_F_FABRICS,
1796         .reg_read32             = nvmf_reg_read32,
1797         .reg_read64             = nvmf_reg_read64,
1798         .reg_write32            = nvmf_reg_write32,
1799         .reset_ctrl             = nvme_rdma_reset_ctrl,
1800         .free_ctrl              = nvme_rdma_free_ctrl,
1801         .submit_async_event     = nvme_rdma_submit_async_event,
1802         .delete_ctrl            = nvme_rdma_del_ctrl,
1803         .get_subsysnqn          = nvmf_get_subsysnqn,
1804         .get_address            = nvmf_get_address,
1805 };
1806
1807 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1808 {
1809         int ret;
1810
1811         ret = nvme_rdma_init_io_queues(ctrl);
1812         if (ret)
1813                 return ret;
1814
1815         /*
1816          * We need a reference on the device as long as the tag_set is alive,
1817          * as the MRs in the request structures need a valid ib_device.
1818          */
1819         ret = -EINVAL;
1820         if (!nvme_rdma_dev_get(ctrl->device))
1821                 goto out_free_io_queues;
1822
1823         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1824         ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1825         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1826         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1827         ctrl->tag_set.numa_node = NUMA_NO_NODE;
1828         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1829         ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1830                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1831         ctrl->tag_set.driver_data = ctrl;
1832         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1833         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1834
1835         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1836         if (ret)
1837                 goto out_put_dev;
1838         ctrl->ctrl.tagset = &ctrl->tag_set;
1839
1840         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1841         if (IS_ERR(ctrl->ctrl.connect_q)) {
1842                 ret = PTR_ERR(ctrl->ctrl.connect_q);
1843                 goto out_free_tag_set;
1844         }
1845
1846         ret = nvme_rdma_connect_io_queues(ctrl);
1847         if (ret)
1848                 goto out_cleanup_connect_q;
1849
1850         return 0;
1851
1852 out_cleanup_connect_q:
1853         blk_cleanup_queue(ctrl->ctrl.connect_q);
1854 out_free_tag_set:
1855         blk_mq_free_tag_set(&ctrl->tag_set);
1856 out_put_dev:
1857         nvme_rdma_dev_put(ctrl->device);
1858 out_free_io_queues:
1859         nvme_rdma_free_io_queues(ctrl);
1860         return ret;
1861 }
1862
1863 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1864                 struct nvmf_ctrl_options *opts)
1865 {
1866         struct nvme_rdma_ctrl *ctrl;
1867         int ret;
1868         bool changed;
1869         char *port;
1870
1871         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1872         if (!ctrl)
1873                 return ERR_PTR(-ENOMEM);
1874         ctrl->ctrl.opts = opts;
1875         INIT_LIST_HEAD(&ctrl->list);
1876
1877         if (opts->mask & NVMF_OPT_TRSVCID)
1878                 port = opts->trsvcid;
1879         else
1880                 port = __stringify(NVME_RDMA_IP_PORT);
1881
1882         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1883                         opts->traddr, port, &ctrl->addr);
1884         if (ret) {
1885                 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1886                 goto out_free_ctrl;
1887         }
1888
1889         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1890                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1891                         opts->host_traddr, NULL, &ctrl->src_addr);
1892                 if (ret) {
1893                         pr_err("malformed src address passed: %s\n",
1894                                opts->host_traddr);
1895                         goto out_free_ctrl;
1896                 }
1897         }
1898
1899         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1900                                 0 /* no quirks, we're perfect! */);
1901         if (ret)
1902                 goto out_free_ctrl;
1903
1904         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1905                         nvme_rdma_reconnect_ctrl_work);
1906         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1907         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1908         INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1909         spin_lock_init(&ctrl->lock);
1910
1911         ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1912         ctrl->ctrl.sqsize = opts->queue_size - 1;
1913         ctrl->ctrl.kato = opts->kato;
1914
1915         ret = -ENOMEM;
1916         ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1917                                 GFP_KERNEL);
1918         if (!ctrl->queues)
1919                 goto out_uninit_ctrl;
1920
1921         ret = nvme_rdma_configure_admin_queue(ctrl);
1922         if (ret)
1923                 goto out_kfree_queues;
1924
1925         /* sanity check icdoff */
1926         if (ctrl->ctrl.icdoff) {
1927                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1928                 goto out_remove_admin_queue;
1929         }
1930
1931         /* sanity check keyed sgls */
1932         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1933                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1934                 goto out_remove_admin_queue;
1935         }
1936
1937         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1938                 /* warn if maxcmd is lower than queue_size */
1939                 dev_warn(ctrl->ctrl.device,
1940                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1941                         opts->queue_size, ctrl->ctrl.maxcmd);
1942                 opts->queue_size = ctrl->ctrl.maxcmd;
1943         }
1944
1945         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1946                 /* warn if sqsize is lower than queue_size */
1947                 dev_warn(ctrl->ctrl.device,
1948                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1949                         opts->queue_size, ctrl->ctrl.sqsize + 1);
1950                 opts->queue_size = ctrl->ctrl.sqsize + 1;
1951         }
1952
1953         if (opts->nr_io_queues) {
1954                 ret = nvme_rdma_create_io_queues(ctrl);
1955                 if (ret)
1956                         goto out_remove_admin_queue;
1957         }
1958
1959         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1960         WARN_ON_ONCE(!changed);
1961
1962         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1963                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1964
1965         kref_get(&ctrl->ctrl.kref);
1966
1967         mutex_lock(&nvme_rdma_ctrl_mutex);
1968         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1969         mutex_unlock(&nvme_rdma_ctrl_mutex);
1970
1971         if (opts->nr_io_queues) {
1972                 nvme_queue_scan(&ctrl->ctrl);
1973                 nvme_queue_async_events(&ctrl->ctrl);
1974         }
1975
1976         return &ctrl->ctrl;
1977
1978 out_remove_admin_queue:
1979         nvme_stop_keep_alive(&ctrl->ctrl);
1980         nvme_rdma_destroy_admin_queue(ctrl);
1981 out_kfree_queues:
1982         kfree(ctrl->queues);
1983 out_uninit_ctrl:
1984         nvme_uninit_ctrl(&ctrl->ctrl);
1985         nvme_put_ctrl(&ctrl->ctrl);
1986         if (ret > 0)
1987                 ret = -EIO;
1988         return ERR_PTR(ret);
1989 out_free_ctrl:
1990         kfree(ctrl);
1991         return ERR_PTR(ret);
1992 }
1993
1994 static struct nvmf_transport_ops nvme_rdma_transport = {
1995         .name           = "rdma",
1996         .required_opts  = NVMF_OPT_TRADDR,
1997         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1998                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1999         .create_ctrl    = nvme_rdma_create_ctrl,
2000 };
2001
2002 static void nvme_rdma_add_one(struct ib_device *ib_device)
2003 {
2004 }
2005
2006 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2007 {
2008         struct nvme_rdma_ctrl *ctrl;
2009
2010         /* Delete all controllers using this device */
2011         mutex_lock(&nvme_rdma_ctrl_mutex);
2012         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2013                 if (ctrl->device->dev != ib_device)
2014                         continue;
2015                 dev_info(ctrl->ctrl.device,
2016                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
2017                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2018                 __nvme_rdma_del_ctrl(ctrl);
2019         }
2020         mutex_unlock(&nvme_rdma_ctrl_mutex);
2021
2022         flush_workqueue(nvme_rdma_wq);
2023 }
2024
2025 static struct ib_client nvme_rdma_ib_client = {
2026         .name   = "nvme_rdma",
2027         .add = nvme_rdma_add_one,
2028         .remove = nvme_rdma_remove_one
2029 };
2030
2031 static int __init nvme_rdma_init_module(void)
2032 {
2033         int ret;
2034
2035         nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2036         if (!nvme_rdma_wq)
2037                 return -ENOMEM;
2038
2039         ret = ib_register_client(&nvme_rdma_ib_client);
2040         if (ret)
2041                 goto err_destroy_wq;
2042
2043         ret = nvmf_register_transport(&nvme_rdma_transport);
2044         if (ret)
2045                 goto err_unreg_client;
2046
2047         return 0;
2048
2049 err_unreg_client:
2050         ib_unregister_client(&nvme_rdma_ib_client);
2051 err_destroy_wq:
2052         destroy_workqueue(nvme_rdma_wq);
2053         return ret;
2054 }
2055
2056 static void __exit nvme_rdma_cleanup_module(void)
2057 {
2058         nvmf_unregister_transport(&nvme_rdma_transport);
2059         ib_unregister_client(&nvme_rdma_ib_client);
2060         destroy_workqueue(nvme_rdma_wq);
2061 }
2062
2063 module_init(nvme_rdma_init_module);
2064 module_exit(nvme_rdma_cleanup_module);
2065
2066 MODULE_LICENSE("GPL v2");