drivers/nvme/host/rdma.c

   1 /*
   2  * NVMe over Fabrics RDMA host code.
   3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15 #include <linux/module.h>
  16 #include <linux/init.h>
  17 #include <linux/slab.h>
  18 #include <rdma/mr_pool.h>
  19 #include <linux/err.h>
  20 #include <linux/string.h>
  21 #include <linux/atomic.h>
  22 #include <linux/blk-mq.h>
  23 #include <linux/blk-mq-rdma.h>
  24 #include <linux/types.h>
  25 #include <linux/list.h>
  26 #include <linux/mutex.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/nvme.h>
  29 #include <asm/unaligned.h>
  30
  31 #include <rdma/ib_verbs.h>
  32 #include <rdma/rdma_cm.h>
  33 #include <linux/nvme-rdma.h>
  34
  35 #include "nvme.h"
  36 #include "fabrics.h"
  37
  38
  39 #define NVME_RDMA_CONNECT_TIMEOUT_MS    3000            /* 3 second */
  40
  41 #define NVME_RDMA_MAX_SEGMENTS          256
  42
  43 #define NVME_RDMA_MAX_INLINE_SEGMENTS   4
  44
  45 struct nvme_rdma_device {
  46         struct ib_device        *dev;
  47         struct ib_pd            *pd;
  48         struct kref             ref;
  49         struct list_head        entry;
  50         unsigned int            num_inline_segments;
  51 };
  52
  53 struct nvme_rdma_qe {
  54         struct ib_cqe           cqe;
  55         void                    *data;
  56         u64                     dma;
  57 };
  58
  59 struct nvme_rdma_queue;
  60 struct nvme_rdma_request {
  61         struct nvme_request     req;
  62         struct ib_mr            *mr;
  63         struct nvme_rdma_qe     sqe;
  64         union nvme_result       result;
  65         __le16                  status;
  66         refcount_t              ref;
  67         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  68         u32                     num_sge;
  69         int                     nents;
  70         struct ib_reg_wr        reg_wr;
  71         struct ib_cqe           reg_cqe;
  72         struct nvme_rdma_queue  *queue;
  73         struct sg_table         sg_table;
  74         struct scatterlist      first_sgl[];
  75 };
  76
  77 enum nvme_rdma_queue_flags {
  78         NVME_RDMA_Q_ALLOCATED           = 0,
  79         NVME_RDMA_Q_LIVE                = 1,
  80         NVME_RDMA_Q_TR_READY            = 2,
  81 };
  82
  83 struct nvme_rdma_queue {
  84         struct nvme_rdma_qe     *rsp_ring;
  85         int                     queue_size;
  86         size_t                  cmnd_capsule_len;
  87         struct nvme_rdma_ctrl   *ctrl;
  88         struct nvme_rdma_device *device;
  89         struct ib_cq            *ib_cq;
  90         struct ib_qp            *qp;
  91
  92         unsigned long           flags;
  93         struct rdma_cm_id       *cm_id;
  94         int                     cm_error;
  95         struct completion       cm_done;
  96 };
  97
  98 struct nvme_rdma_ctrl {
  99         /* read only in the hot path */
 100         struct nvme_rdma_queue  *queues;
 101
 102         /* other member variables */
 103         struct blk_mq_tag_set   tag_set;
 104         struct work_struct      err_work;
 105
 106         struct nvme_rdma_qe     async_event_sqe;
 107
 108         struct delayed_work     reconnect_work;
 109
 110         struct list_head        list;
 111
 112         struct blk_mq_tag_set   admin_tag_set;
 113         struct nvme_rdma_device *device;
 114
 115         u32                     max_fr_pages;
 116
 117         struct sockaddr_storage addr;
 118         struct sockaddr_storage src_addr;
 119
 120         struct nvme_ctrl        ctrl;
 121         bool                    use_inline_data;
 122 };
 123
 124 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 125 {
 126         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 127 }
 128
 129 static LIST_HEAD(device_list);
 130 static DEFINE_MUTEX(device_list_mutex);
 131
 132 static LIST_HEAD(nvme_rdma_ctrl_list);
 133 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 134
 135 /*
 136  * Disabling this option makes small I/O goes faster, but is fundamentally
 137  * unsafe.  With it turned off we will have to register a global rkey that
 138  * allows read and write access to all physical memory.
 139  */
 140 static bool register_always = true;
 141 module_param(register_always, bool, 0444);
 142 MODULE_PARM_DESC(register_always,
 143          "Use memory registration even for contiguous memory regions");
 144
 145 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 146                 struct rdma_cm_event *event);
 147 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 148
 149 static const struct blk_mq_ops nvme_rdma_mq_ops;
 150 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 151
 152 /* XXX: really should move to a generic header sooner or later.. */
 153 static inline void put_unaligned_le24(u32 val, u8 *p)
 154 {
 155         *p++ = val;
 156         *p++ = val >> 8;
 157         *p++ = val >> 16;
 158 }
 159
 160 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 161 {
 162         return queue - queue->ctrl->queues;
 163 }
 164
 165 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 166 {
 167         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 168 }
 169
 170 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 171                 size_t capsule_size, enum dma_data_direction dir)
 172 {
 173         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 174         kfree(qe->data);
 175 }
 176
 177 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 178                 size_t capsule_size, enum dma_data_direction dir)
 179 {
 180         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 181         if (!qe->data)
 182                 return -ENOMEM;
 183
 184         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 185         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 186                 kfree(qe->data);
 187                 qe->data = NULL;
 188                 return -ENOMEM;
 189         }
 190
 191         return 0;
 192 }
 193
 194 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 195                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 196                 size_t capsule_size, enum dma_data_direction dir)
 197 {
 198         int i;
 199
 200         for (i = 0; i < ib_queue_size; i++)
 201                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 202         kfree(ring);
 203 }
 204
 205 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 206                 size_t ib_queue_size, size_t capsule_size,
 207                 enum dma_data_direction dir)
 208 {
 209         struct nvme_rdma_qe *ring;
 210         int i;
 211
 212         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 213         if (!ring)
 214                 return NULL;
 215
 216         for (i = 0; i < ib_queue_size; i++) {
 217                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 218                         goto out_free_ring;
 219         }
 220
 221         return ring;
 222
 223 out_free_ring:
 224         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 225         return NULL;
 226 }
 227
 228 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 229 {
 230         pr_debug("QP event %s (%d)\n",
 231                  ib_event_msg(event->event), event->event);
 232
 233 }
 234
 235 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 236 {
 237         int ret;
 238
 239         ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
 240                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 241         if (ret < 0)
 242                 return ret;
 243         if (ret == 0)
 244                 return -ETIMEDOUT;
 245         WARN_ON_ONCE(queue->cm_error > 0);
 246         return queue->cm_error;
 247 }
 248
 249 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 250 {
 251         struct nvme_rdma_device *dev = queue->device;
 252         struct ib_qp_init_attr init_attr;
 253         int ret;
 254
 255         memset(&init_attr, 0, sizeof(init_attr));
 256         init_attr.event_handler = nvme_rdma_qp_event;
 257         /* +1 for drain */
 258         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 259         /* +1 for drain */
 260         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 261         init_attr.cap.max_recv_sge = 1;
 262         init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
 263         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 264         init_attr.qp_type = IB_QPT_RC;
 265         init_attr.send_cq = queue->ib_cq;
 266         init_attr.recv_cq = queue->ib_cq;
 267
 268         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 269
 270         queue->qp = queue->cm_id->qp;
 271         return ret;
 272 }
 273
 274 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 275                 struct request *rq, unsigned int hctx_idx)
 276 {
 277         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 278         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 279         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 280         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 281         struct nvme_rdma_device *dev = queue->device;
 282
 283         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 284                         DMA_TO_DEVICE);
 285 }
 286
 287 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 288                 struct request *rq, unsigned int hctx_idx,
 289                 unsigned int numa_node)
 290 {
 291         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 292         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 293         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 294         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 295         struct nvme_rdma_device *dev = queue->device;
 296         struct ib_device *ibdev = dev->dev;
 297         int ret;
 298
 299         nvme_req(rq)->ctrl = &ctrl->ctrl;
 300         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 301                         DMA_TO_DEVICE);
 302         if (ret)
 303                 return ret;
 304
 305         req->queue = queue;
 306
 307         return 0;
 308 }
 309
 310 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 311                 unsigned int hctx_idx)
 312 {
 313         struct nvme_rdma_ctrl *ctrl = data;
 314         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 315
 316         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 317
 318         hctx->driver_data = queue;
 319         return 0;
 320 }
 321
 322 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 323                 unsigned int hctx_idx)
 324 {
 325         struct nvme_rdma_ctrl *ctrl = data;
 326         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 327
 328         BUG_ON(hctx_idx != 0);
 329
 330         hctx->driver_data = queue;
 331         return 0;
 332 }
 333
 334 static void nvme_rdma_free_dev(struct kref *ref)
 335 {
 336         struct nvme_rdma_device *ndev =
 337                 container_of(ref, struct nvme_rdma_device, ref);
 338
 339         mutex_lock(&device_list_mutex);
 340         list_del(&ndev->entry);
 341         mutex_unlock(&device_list_mutex);
 342
 343         ib_dealloc_pd(ndev->pd);
 344         kfree(ndev);
 345 }
 346
 347 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 348 {
 349         kref_put(&dev->ref, nvme_rdma_free_dev);
 350 }
 351
 352 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 353 {
 354         return kref_get_unless_zero(&dev->ref);
 355 }
 356
 357 static struct nvme_rdma_device *
 358 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 359 {
 360         struct nvme_rdma_device *ndev;
 361
 362         mutex_lock(&device_list_mutex);
 363         list_for_each_entry(ndev, &device_list, entry) {
 364                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 365                     nvme_rdma_dev_get(ndev))
 366                         goto out_unlock;
 367         }
 368
 369         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 370         if (!ndev)
 371                 goto out_err;
 372
 373         ndev->dev = cm_id->device;
 374         kref_init(&ndev->ref);
 375
 376         ndev->pd = ib_alloc_pd(ndev->dev,
 377                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 378         if (IS_ERR(ndev->pd))
 379                 goto out_free_dev;
 380
 381         if (!(ndev->dev->attrs.device_cap_flags &
 382               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 383                 dev_err(&ndev->dev->dev,
 384                         "Memory registrations not supported.\n");
 385                 goto out_free_pd;
 386         }
 387
 388         ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
 389                                         ndev->dev->attrs.max_send_sge - 1);
 390         list_add(&ndev->entry, &device_list);
 391 out_unlock:
 392         mutex_unlock(&device_list_mutex);
 393         return ndev;
 394
 395 out_free_pd:
 396         ib_dealloc_pd(ndev->pd);
 397 out_free_dev:
 398         kfree(ndev);
 399 out_err:
 400         mutex_unlock(&device_list_mutex);
 401         return NULL;
 402 }
 403
 404 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 405 {
 406         struct nvme_rdma_device *dev;
 407         struct ib_device *ibdev;
 408
 409         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 410                 return;
 411
 412         dev = queue->device;
 413         ibdev = dev->dev;
 414
 415         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 416
 417         /*
 418          * The cm_id object might have been destroyed during RDMA connection
 419          * establishment error flow to avoid getting other cma events, thus
 420          * the destruction of the QP shouldn't use rdma_cm API.
 421          */
 422         ib_destroy_qp(queue->qp);
 423         ib_free_cq(queue->ib_cq);
 424
 425         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 426                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 427
 428         nvme_rdma_dev_put(dev);
 429 }
 430
 431 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
 432 {
 433         return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
 434                      ibdev->attrs.max_fast_reg_page_list_len);
 435 }
 436
 437 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 438 {
 439         struct ib_device *ibdev;
 440         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 441         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 442         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 443         int ret;
 444
 445         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 446         if (!queue->device) {
 447                 dev_err(queue->cm_id->device->dev.parent,
 448                         "no client data found!\n");
 449                 return -ECONNREFUSED;
 450         }
 451         ibdev = queue->device->dev;
 452
 453         /*
 454          * Spread I/O queues completion vectors according their queue index.
 455          * Admin queues can always go on completion vector 0.
 456          */
 457         comp_vector = idx == 0 ? idx : idx - 1;
 458
 459         /* +1 for ib_stop_cq */
 460         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 461                                 cq_factor * queue->queue_size + 1,
 462                                 comp_vector, IB_POLL_SOFTIRQ);
 463         if (IS_ERR(queue->ib_cq)) {
 464                 ret = PTR_ERR(queue->ib_cq);
 465                 goto out_put_dev;
 466         }
 467
 468         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 469         if (ret)
 470                 goto out_destroy_ib_cq;
 471
 472         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 473                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 474         if (!queue->rsp_ring) {
 475                 ret = -ENOMEM;
 476                 goto out_destroy_qp;
 477         }
 478
 479         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 480                               queue->queue_size,
 481                               IB_MR_TYPE_MEM_REG,
 482                               nvme_rdma_get_max_fr_pages(ibdev));
 483         if (ret) {
 484                 dev_err(queue->ctrl->ctrl.device,
 485                         "failed to initialize MR pool sized %d for QID %d\n",
 486                         queue->queue_size, idx);
 487                 goto out_destroy_ring;
 488         }
 489
 490         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 491
 492         return 0;
 493
 494 out_destroy_ring:
 495         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 496                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 497 out_destroy_qp:
 498         rdma_destroy_qp(queue->cm_id);
 499 out_destroy_ib_cq:
 500         ib_free_cq(queue->ib_cq);
 501 out_put_dev:
 502         nvme_rdma_dev_put(queue->device);
 503         return ret;
 504 }
 505
 506 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 507                 int idx, size_t queue_size)
 508 {
 509         struct nvme_rdma_queue *queue;
 510         struct sockaddr *src_addr = NULL;
 511         int ret;
 512
 513         queue = &ctrl->queues[idx];
 514         queue->ctrl = ctrl;
 515         init_completion(&queue->cm_done);
 516
 517         if (idx > 0)
 518                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 519         else
 520                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 521
 522         queue->queue_size = queue_size;
 523
 524         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 525                         RDMA_PS_TCP, IB_QPT_RC);
 526         if (IS_ERR(queue->cm_id)) {
 527                 dev_info(ctrl->ctrl.device,
 528                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 529                 return PTR_ERR(queue->cm_id);
 530         }
 531
 532         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 533                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 534
 535         queue->cm_error = -ETIMEDOUT;
 536         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 537                         (struct sockaddr *)&ctrl->addr,
 538                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 539         if (ret) {
 540                 dev_info(ctrl->ctrl.device,
 541                         "rdma_resolve_addr failed (%d).\n", ret);
 542                 goto out_destroy_cm_id;
 543         }
 544
 545         ret = nvme_rdma_wait_for_cm(queue);
 546         if (ret) {
 547                 dev_info(ctrl->ctrl.device,
 548                         "rdma connection establishment failed (%d)\n", ret);
 549                 goto out_destroy_cm_id;
 550         }
 551
 552         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 553
 554         return 0;
 555
 556 out_destroy_cm_id:
 557         rdma_destroy_id(queue->cm_id);
 558         nvme_rdma_destroy_queue_ib(queue);
 559         return ret;
 560 }
 561
 562 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 563 {
 564         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 565                 return;
 566
 567         rdma_disconnect(queue->cm_id);
 568         ib_drain_qp(queue->qp);
 569 }
 570
 571 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 572 {
 573         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 574                 return;
 575
 576         nvme_rdma_destroy_queue_ib(queue);
 577         rdma_destroy_id(queue->cm_id);
 578 }
 579
 580 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 581 {
 582         int i;
 583
 584         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 585                 nvme_rdma_free_queue(&ctrl->queues[i]);
 586 }
 587
 588 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 589 {
 590         int i;
 591
 592         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 593                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 594 }
 595
 596 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 597 {
 598         int ret;
 599
 600         if (idx)
 601                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 602         else
 603                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 604
 605         if (!ret)
 606                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
 607         else
 608                 dev_info(ctrl->ctrl.device,
 609                         "failed to connect queue: %d ret=%d\n", idx, ret);
 610         return ret;
 611 }
 612
 613 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 614 {
 615         int i, ret = 0;
 616
 617         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 618                 ret = nvme_rdma_start_queue(ctrl, i);
 619                 if (ret)
 620                         goto out_stop_queues;
 621         }
 622
 623         return 0;
 624
 625 out_stop_queues:
 626         for (i--; i >= 1; i--)
 627                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 628         return ret;
 629 }
 630
 631 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 632 {
 633         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 634         struct ib_device *ibdev = ctrl->device->dev;
 635         unsigned int nr_io_queues;
 636         int i, ret;
 637
 638         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 639
 640         /*
 641          * we map queues according to the device irq vectors for
 642          * optimal locality so we don't need more queues than
 643          * completion vectors.
 644          */
 645         nr_io_queues = min_t(unsigned int, nr_io_queues,
 646                                 ibdev->num_comp_vectors);
 647
 648         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 649         if (ret)
 650                 return ret;
 651
 652         ctrl->ctrl.queue_count = nr_io_queues + 1;
 653         if (ctrl->ctrl.queue_count < 2)
 654                 return 0;
 655
 656         dev_info(ctrl->ctrl.device,
 657                 "creating %d I/O queues.\n", nr_io_queues);
 658
 659         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 660                 ret = nvme_rdma_alloc_queue(ctrl, i,
 661                                 ctrl->ctrl.sqsize + 1);
 662                 if (ret)
 663                         goto out_free_queues;
 664         }
 665
 666         return 0;
 667
 668 out_free_queues:
 669         for (i--; i >= 1; i--)
 670                 nvme_rdma_free_queue(&ctrl->queues[i]);
 671
 672         return ret;
 673 }
 674
 675 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
 676                 struct blk_mq_tag_set *set)
 677 {
 678         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 679
 680         blk_mq_free_tag_set(set);
 681         nvme_rdma_dev_put(ctrl->device);
 682 }
 683
 684 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 685                 bool admin)
 686 {
 687         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 688         struct blk_mq_tag_set *set;
 689         int ret;
 690
 691         if (admin) {
 692                 set = &ctrl->admin_tag_set;
 693                 memset(set, 0, sizeof(*set));
 694                 set->ops = &nvme_rdma_admin_mq_ops;
 695                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 696                 set->reserved_tags = 2; /* connect + keep-alive */
 697                 set->numa_node = nctrl->numa_node;
 698                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 699                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 700                 set->driver_data = ctrl;
 701                 set->nr_hw_queues = 1;
 702                 set->timeout = ADMIN_TIMEOUT;
 703                 set->flags = BLK_MQ_F_NO_SCHED;
 704         } else {
 705                 set = &ctrl->tag_set;
 706                 memset(set, 0, sizeof(*set));
 707                 set->ops = &nvme_rdma_mq_ops;
 708                 set->queue_depth = nctrl->sqsize + 1;
 709                 set->reserved_tags = 1; /* fabric connect */
 710                 set->numa_node = nctrl->numa_node;
 711                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 712                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 713                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 714                 set->driver_data = ctrl;
 715                 set->nr_hw_queues = nctrl->queue_count - 1;
 716                 set->timeout = NVME_IO_TIMEOUT;
 717         }
 718
 719         ret = blk_mq_alloc_tag_set(set);
 720         if (ret)
 721                 goto out;
 722
 723         /*
 724          * We need a reference on the device as long as the tag_set is alive,
 725          * as the MRs in the request structures need a valid ib_device.
 726          */
 727         ret = nvme_rdma_dev_get(ctrl->device);
 728         if (!ret) {
 729                 ret = -EINVAL;
 730                 goto out_free_tagset;
 731         }
 732
 733         return set;
 734
 735 out_free_tagset:
 736         blk_mq_free_tag_set(set);
 737 out:
 738         return ERR_PTR(ret);
 739 }
 740
 741 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 742                 bool remove)
 743 {
 744         if (remove) {
 745                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 746                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 747         }
 748         if (ctrl->async_event_sqe.data) {
 749                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 750                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 751                 ctrl->async_event_sqe.data = NULL;
 752         }
 753         nvme_rdma_free_queue(&ctrl->queues[0]);
 754 }
 755
 756 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 757                 bool new)
 758 {
 759         int error;
 760
 761         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 762         if (error)
 763                 return error;
 764
 765         ctrl->device = ctrl->queues[0].device;
 766         ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
 767
 768         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
 769
 770         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 771                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 772         if (error)
 773                 goto out_free_queue;
 774
 775         if (new) {
 776                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 777                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 778                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 779                         goto out_free_async_qe;
 780                 }
 781
 782                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 783                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 784                         error = PTR_ERR(ctrl->ctrl.admin_q);
 785                         goto out_free_tagset;
 786                 }
 787         }
 788
 789         error = nvme_rdma_start_queue(ctrl, 0);
 790         if (error)
 791                 goto out_cleanup_queue;
 792
 793         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 794                         &ctrl->ctrl.cap);
 795         if (error) {
 796                 dev_err(ctrl->ctrl.device,
 797                         "prop_get NVME_REG_CAP failed\n");
 798                 goto out_stop_queue;
 799         }
 800
 801         ctrl->ctrl.sqsize =
 802                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 803
 804         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 805         if (error)
 806                 goto out_stop_queue;
 807
 808         ctrl->ctrl.max_hw_sectors =
 809                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 810
 811         error = nvme_init_identify(&ctrl->ctrl);
 812         if (error)
 813                 goto out_stop_queue;
 814
 815         return 0;
 816
 817 out_stop_queue:
 818         nvme_rdma_stop_queue(&ctrl->queues[0]);
 819 out_cleanup_queue:
 820         if (new)
 821                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 822 out_free_tagset:
 823         if (new)
 824                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 825 out_free_async_qe:
 826         nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 827                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 828         ctrl->async_event_sqe.data = NULL;
 829 out_free_queue:
 830         nvme_rdma_free_queue(&ctrl->queues[0]);
 831         return error;
 832 }
 833
 834 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 835                 bool remove)
 836 {
 837         if (remove) {
 838                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 839                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 840         }
 841         nvme_rdma_free_io_queues(ctrl);
 842 }
 843
 844 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 845 {
 846         int ret;
 847
 848         ret = nvme_rdma_alloc_io_queues(ctrl);
 849         if (ret)
 850                 return ret;
 851
 852         if (new) {
 853                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 854                 if (IS_ERR(ctrl->ctrl.tagset)) {
 855                         ret = PTR_ERR(ctrl->ctrl.tagset);
 856                         goto out_free_io_queues;
 857                 }
 858
 859                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 860                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 861                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 862                         goto out_free_tag_set;
 863                 }
 864         } else {
 865                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 866                         ctrl->ctrl.queue_count - 1);
 867         }
 868
 869         ret = nvme_rdma_start_io_queues(ctrl);
 870         if (ret)
 871                 goto out_cleanup_connect_q;
 872
 873         return 0;
 874
 875 out_cleanup_connect_q:
 876         if (new)
 877                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 878 out_free_tag_set:
 879         if (new)
 880                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 881 out_free_io_queues:
 882         nvme_rdma_free_io_queues(ctrl);
 883         return ret;
 884 }
 885
 886 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
 887                 bool remove)
 888 {
 889         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 890         nvme_rdma_stop_queue(&ctrl->queues[0]);
 891         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
 892                         &ctrl->ctrl);
 893         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 894         nvme_rdma_destroy_admin_queue(ctrl, remove);
 895 }
 896
 897 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
 898                 bool remove)
 899 {
 900         if (ctrl->ctrl.queue_count > 1) {
 901                 nvme_stop_queues(&ctrl->ctrl);
 902                 nvme_rdma_stop_io_queues(ctrl);
 903                 blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
 904                                 &ctrl->ctrl);
 905                 if (remove)
 906                         nvme_start_queues(&ctrl->ctrl);
 907                 nvme_rdma_destroy_io_queues(ctrl, remove);
 908         }
 909 }
 910
 911 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
 912 {
 913         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 914
 915         cancel_work_sync(&ctrl->err_work);
 916         cancel_delayed_work_sync(&ctrl->reconnect_work);
 917 }
 918
 919 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 920 {
 921         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 922
 923         if (list_empty(&ctrl->list))
 924                 goto free_ctrl;
 925
 926         mutex_lock(&nvme_rdma_ctrl_mutex);
 927         list_del(&ctrl->list);
 928         mutex_unlock(&nvme_rdma_ctrl_mutex);
 929
 930         nvmf_free_options(nctrl->opts);
 931 free_ctrl:
 932         kfree(ctrl->queues);
 933         kfree(ctrl);
 934 }
 935
 936 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 937 {
 938         /* If we are resetting/deleting then do nothing */
 939         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 940                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 941                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 942                 return;
 943         }
 944
 945         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 946                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 947                         ctrl->ctrl.opts->reconnect_delay);
 948                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 949                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 950         } else {
 951                 nvme_delete_ctrl(&ctrl->ctrl);
 952         }
 953 }
 954
 955 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
 956 {
 957         int ret = -EINVAL;
 958         bool changed;
 959
 960         ret = nvme_rdma_configure_admin_queue(ctrl, new);
 961         if (ret)
 962                 return ret;
 963
 964         if (ctrl->ctrl.icdoff) {
 965                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
 966                 goto destroy_admin;
 967         }
 968
 969         if (!(ctrl->ctrl.sgls & (1 << 2))) {
 970                 dev_err(ctrl->ctrl.device,
 971                         "Mandatory keyed sgls are not supported!\n");
 972                 goto destroy_admin;
 973         }
 974
 975         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
 976                 dev_warn(ctrl->ctrl.device,
 977                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
 978                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
 979         }
 980
 981         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
 982                 dev_warn(ctrl->ctrl.device,
 983                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
 984                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
 985                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
 986         }
 987
 988         if (ctrl->ctrl.sgls & (1 << 20))
 989                 ctrl->use_inline_data = true;
 990
 991         if (ctrl->ctrl.queue_count > 1) {
 992                 ret = nvme_rdma_configure_io_queues(ctrl, new);
 993                 if (ret)
 994                         goto destroy_admin;
 995         }
 996
 997         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 998         if (!changed) {
 999                 /* state change failure is ok if we're in DELETING state */
1000                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1001                 ret = -EINVAL;
1002                 goto destroy_io;
1003         }
1004
1005         nvme_start_ctrl(&ctrl->ctrl);
1006         return 0;
1007
1008 destroy_io:
1009         if (ctrl->ctrl.queue_count > 1)
1010                 nvme_rdma_destroy_io_queues(ctrl, new);
1011 destroy_admin:
1012         nvme_rdma_stop_queue(&ctrl->queues[0]);
1013         nvme_rdma_destroy_admin_queue(ctrl, new);
1014         return ret;
1015 }
1016
1017 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1018 {
1019         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1020                         struct nvme_rdma_ctrl, reconnect_work);
1021
1022         ++ctrl->ctrl.nr_reconnects;
1023
1024         if (nvme_rdma_setup_ctrl(ctrl, false))
1025                 goto requeue;
1026
1027         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1028                         ctrl->ctrl.nr_reconnects);
1029
1030         ctrl->ctrl.nr_reconnects = 0;
1031
1032         return;
1033
1034 requeue:
1035         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1036                         ctrl->ctrl.nr_reconnects);
1037         nvme_rdma_reconnect_or_remove(ctrl);
1038 }
1039
1040 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1041 {
1042         struct nvme_rdma_ctrl *ctrl = container_of(work,
1043                         struct nvme_rdma_ctrl, err_work);
1044
1045         nvme_stop_keep_alive(&ctrl->ctrl);
1046         nvme_rdma_teardown_io_queues(ctrl, false);
1047         nvme_start_queues(&ctrl->ctrl);
1048         nvme_rdma_teardown_admin_queue(ctrl, false);
1049
1050         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1051                 /* state change failure is ok if we're in DELETING state */
1052                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1053                 return;
1054         }
1055
1056         nvme_rdma_reconnect_or_remove(ctrl);
1057 }
1058
1059 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1060 {
1061         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1062                 return;
1063
1064         queue_work(nvme_wq, &ctrl->err_work);
1065 }
1066
1067 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1068                 const char *op)
1069 {
1070         struct nvme_rdma_queue *queue = cq->cq_context;
1071         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1072
1073         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1074                 dev_info(ctrl->ctrl.device,
1075                              "%s for CQE 0x%p failed with status %s (%d)\n",
1076                              op, wc->wr_cqe,
1077                              ib_wc_status_msg(wc->status), wc->status);
1078         nvme_rdma_error_recovery(ctrl);
1079 }
1080
1081 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1082 {
1083         if (unlikely(wc->status != IB_WC_SUCCESS))
1084                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1085 }
1086
1087 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1088 {
1089         struct nvme_rdma_request *req =
1090                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1091         struct request *rq = blk_mq_rq_from_pdu(req);
1092
1093         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1094                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1095                 return;
1096         }
1097
1098         if (refcount_dec_and_test(&req->ref))
1099                 nvme_end_request(rq, req->status, req->result);
1100
1101 }
1102
1103 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1104                 struct nvme_rdma_request *req)
1105 {
1106         struct ib_send_wr wr = {
1107                 .opcode             = IB_WR_LOCAL_INV,
1108                 .next               = NULL,
1109                 .num_sge            = 0,
1110                 .send_flags         = IB_SEND_SIGNALED,
1111                 .ex.invalidate_rkey = req->mr->rkey,
1112         };
1113
1114         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1115         wr.wr_cqe = &req->reg_cqe;
1116
1117         return ib_post_send(queue->qp, &wr, NULL);
1118 }
1119
1120 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1121                 struct request *rq)
1122 {
1123         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1124         struct nvme_rdma_device *dev = queue->device;
1125         struct ib_device *ibdev = dev->dev;
1126
1127         if (!blk_rq_payload_bytes(rq))
1128                 return;
1129
1130         if (req->mr) {
1131                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1132                 req->mr = NULL;
1133         }
1134
1135         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1136                         req->nents, rq_data_dir(rq) ==
1137                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1138
1139         nvme_cleanup_cmd(rq);
1140         sg_free_table_chained(&req->sg_table, true);
1141 }
1142
1143 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1144 {
1145         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1146
1147         sg->addr = 0;
1148         put_unaligned_le24(0, sg->length);
1149         put_unaligned_le32(0, sg->key);
1150         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1151         return 0;
1152 }
1153
1154 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1155                 struct nvme_rdma_request *req, struct nvme_command *c,
1156                 int count)
1157 {
1158         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1159         struct scatterlist *sgl = req->sg_table.sgl;
1160         struct ib_sge *sge = &req->sge[1];
1161         u32 len = 0;
1162         int i;
1163
1164         for (i = 0; i < count; i++, sgl++, sge++) {
1165                 sge->addr = sg_dma_address(sgl);
1166                 sge->length = sg_dma_len(sgl);
1167                 sge->lkey = queue->device->pd->local_dma_lkey;
1168                 len += sge->length;
1169         }
1170
1171         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1172         sg->length = cpu_to_le32(len);
1173         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1174
1175         req->num_sge += count;
1176         return 0;
1177 }
1178
1179 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1180                 struct nvme_rdma_request *req, struct nvme_command *c)
1181 {
1182         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1183
1184         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1185         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1186         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1187         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1188         return 0;
1189 }
1190
1191 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1192                 struct nvme_rdma_request *req, struct nvme_command *c,
1193                 int count)
1194 {
1195         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1196         int nr;
1197
1198         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1199         if (WARN_ON_ONCE(!req->mr))
1200                 return -EAGAIN;
1201
1202         /*
1203          * Align the MR to a 4K page size to match the ctrl page size and
1204          * the block virtual boundary.
1205          */
1206         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1207         if (unlikely(nr < count)) {
1208                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1209                 req->mr = NULL;
1210                 if (nr < 0)
1211                         return nr;
1212                 return -EINVAL;
1213         }
1214
1215         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1216
1217         req->reg_cqe.done = nvme_rdma_memreg_done;
1218         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1219         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1220         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1221         req->reg_wr.wr.num_sge = 0;
1222         req->reg_wr.mr = req->mr;
1223         req->reg_wr.key = req->mr->rkey;
1224         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1225                              IB_ACCESS_REMOTE_READ |
1226                              IB_ACCESS_REMOTE_WRITE;
1227
1228         sg->addr = cpu_to_le64(req->mr->iova);
1229         put_unaligned_le24(req->mr->length, sg->length);
1230         put_unaligned_le32(req->mr->rkey, sg->key);
1231         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1232                         NVME_SGL_FMT_INVALIDATE;
1233
1234         return 0;
1235 }
1236
1237 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1238                 struct request *rq, struct nvme_command *c)
1239 {
1240         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1241         struct nvme_rdma_device *dev = queue->device;
1242         struct ib_device *ibdev = dev->dev;
1243         int count, ret;
1244
1245         req->num_sge = 1;
1246         refcount_set(&req->ref, 2); /* send and recv completions */
1247
1248         c->common.flags |= NVME_CMD_SGL_METABUF;
1249
1250         if (!blk_rq_payload_bytes(rq))
1251                 return nvme_rdma_set_sg_null(c);
1252
1253         req->sg_table.sgl = req->first_sgl;
1254         ret = sg_alloc_table_chained(&req->sg_table,
1255                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1256         if (ret)
1257                 return -ENOMEM;
1258
1259         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1260
1261         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1262                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1263         if (unlikely(count <= 0)) {
1264                 ret = -EIO;
1265                 goto out_free_table;
1266         }
1267
1268         if (count <= dev->num_inline_segments) {
1269                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1270                     queue->ctrl->use_inline_data &&
1271                     blk_rq_payload_bytes(rq) <=
1272                                 nvme_rdma_inline_data_size(queue)) {
1273                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1274                         goto out;
1275                 }
1276
1277                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1278                         ret = nvme_rdma_map_sg_single(queue, req, c);
1279                         goto out;
1280                 }
1281         }
1282
1283         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1284 out:
1285         if (unlikely(ret))
1286                 goto out_unmap_sg;
1287
1288         return 0;
1289
1290 out_unmap_sg:
1291         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1292                         req->nents, rq_data_dir(rq) ==
1293                         WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1294 out_free_table:
1295         sg_free_table_chained(&req->sg_table, true);
1296         return ret;
1297 }
1298
1299 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1300 {
1301         struct nvme_rdma_qe *qe =
1302                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1303         struct nvme_rdma_request *req =
1304                 container_of(qe, struct nvme_rdma_request, sqe);
1305         struct request *rq = blk_mq_rq_from_pdu(req);
1306
1307         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1308                 nvme_rdma_wr_error(cq, wc, "SEND");
1309                 return;
1310         }
1311
1312         if (refcount_dec_and_test(&req->ref))
1313                 nvme_end_request(rq, req->status, req->result);
1314 }
1315
1316 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1317                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1318                 struct ib_send_wr *first)
1319 {
1320         struct ib_send_wr wr;
1321         int ret;
1322
1323         sge->addr   = qe->dma;
1324         sge->length = sizeof(struct nvme_command),
1325         sge->lkey   = queue->device->pd->local_dma_lkey;
1326
1327         wr.next       = NULL;
1328         wr.wr_cqe     = &qe->cqe;
1329         wr.sg_list    = sge;
1330         wr.num_sge    = num_sge;
1331         wr.opcode     = IB_WR_SEND;
1332         wr.send_flags = IB_SEND_SIGNALED;
1333
1334         if (first)
1335                 first->next = &wr;
1336         else
1337                 first = &wr;
1338
1339         ret = ib_post_send(queue->qp, first, NULL);
1340         if (unlikely(ret)) {
1341                 dev_err(queue->ctrl->ctrl.device,
1342                              "%s failed with error code %d\n", __func__, ret);
1343         }
1344         return ret;
1345 }
1346
1347 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1348                 struct nvme_rdma_qe *qe)
1349 {
1350         struct ib_recv_wr wr;
1351         struct ib_sge list;
1352         int ret;
1353
1354         list.addr   = qe->dma;
1355         list.length = sizeof(struct nvme_completion);
1356         list.lkey   = queue->device->pd->local_dma_lkey;
1357
1358         qe->cqe.done = nvme_rdma_recv_done;
1359
1360         wr.next     = NULL;
1361         wr.wr_cqe   = &qe->cqe;
1362         wr.sg_list  = &list;
1363         wr.num_sge  = 1;
1364
1365         ret = ib_post_recv(queue->qp, &wr, NULL);
1366         if (unlikely(ret)) {
1367                 dev_err(queue->ctrl->ctrl.device,
1368                         "%s failed with error code %d\n", __func__, ret);
1369         }
1370         return ret;
1371 }
1372
1373 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1374 {
1375         u32 queue_idx = nvme_rdma_queue_idx(queue);
1376
1377         if (queue_idx == 0)
1378                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1379         return queue->ctrl->tag_set.tags[queue_idx - 1];
1380 }
1381
1382 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1383 {
1384         if (unlikely(wc->status != IB_WC_SUCCESS))
1385                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1386 }
1387
1388 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1389 {
1390         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1391         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1392         struct ib_device *dev = queue->device->dev;
1393         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1394         struct nvme_command *cmd = sqe->data;
1395         struct ib_sge sge;
1396         int ret;
1397
1398         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1399
1400         memset(cmd, 0, sizeof(*cmd));
1401         cmd->common.opcode = nvme_admin_async_event;
1402         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1403         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1404         nvme_rdma_set_sg_null(cmd);
1405
1406         sqe->cqe.done = nvme_rdma_async_done;
1407
1408         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1409                         DMA_TO_DEVICE);
1410
1411         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1412         WARN_ON_ONCE(ret);
1413 }
1414
1415 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1416                 struct nvme_completion *cqe, struct ib_wc *wc)
1417 {
1418         struct request *rq;
1419         struct nvme_rdma_request *req;
1420
1421         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1422         if (!rq) {
1423                 dev_err(queue->ctrl->ctrl.device,
1424                         "tag 0x%x on QP %#x not found\n",
1425                         cqe->command_id, queue->qp->qp_num);
1426                 nvme_rdma_error_recovery(queue->ctrl);
1427                 return;
1428         }
1429         req = blk_mq_rq_to_pdu(rq);
1430
1431         req->status = cqe->status;
1432         req->result = cqe->result;
1433
1434         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1435                 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1436                         dev_err(queue->ctrl->ctrl.device,
1437                                 "Bogus remote invalidation for rkey %#x\n",
1438                                 req->mr->rkey);
1439                         nvme_rdma_error_recovery(queue->ctrl);
1440                 }
1441         } else if (req->mr) {
1442                 int ret;
1443
1444                 ret = nvme_rdma_inv_rkey(queue, req);
1445                 if (unlikely(ret < 0)) {
1446                         dev_err(queue->ctrl->ctrl.device,
1447                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1448                                 req->mr->rkey, ret);
1449                         nvme_rdma_error_recovery(queue->ctrl);
1450                 }
1451                 /* the local invalidation completion will end the request */
1452                 return;
1453         }
1454
1455         if (refcount_dec_and_test(&req->ref))
1456                 nvme_end_request(rq, req->status, req->result);
1457 }
1458
1459 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1460 {
1461         struct nvme_rdma_qe *qe =
1462                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1463         struct nvme_rdma_queue *queue = cq->cq_context;
1464         struct ib_device *ibdev = queue->device->dev;
1465         struct nvme_completion *cqe = qe->data;
1466         const size_t len = sizeof(struct nvme_completion);
1467
1468         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1469                 nvme_rdma_wr_error(cq, wc, "RECV");
1470                 return;
1471         }
1472
1473         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1474         /*
1475          * AEN requests are special as they don't time out and can
1476          * survive any kind of queue freeze and often don't respond to
1477          * aborts.  We don't even bother to allocate a struct request
1478          * for them but rather special case them here.
1479          */
1480         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1481                         cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1482                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1483                                 &cqe->result);
1484         else
1485                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1486         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1487
1488         nvme_rdma_post_recv(queue, qe);
1489 }
1490
1491 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1492 {
1493         int ret, i;
1494
1495         for (i = 0; i < queue->queue_size; i++) {
1496                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1497                 if (ret)
1498                         goto out_destroy_queue_ib;
1499         }
1500
1501         return 0;
1502
1503 out_destroy_queue_ib:
1504         nvme_rdma_destroy_queue_ib(queue);
1505         return ret;
1506 }
1507
1508 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1509                 struct rdma_cm_event *ev)
1510 {
1511         struct rdma_cm_id *cm_id = queue->cm_id;
1512         int status = ev->status;
1513         const char *rej_msg;
1514         const struct nvme_rdma_cm_rej *rej_data;
1515         u8 rej_data_len;
1516
1517         rej_msg = rdma_reject_msg(cm_id, status);
1518         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1519
1520         if (rej_data && rej_data_len >= sizeof(u16)) {
1521                 u16 sts = le16_to_cpu(rej_data->sts);
1522
1523                 dev_err(queue->ctrl->ctrl.device,
1524                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1525                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1526         } else {
1527                 dev_err(queue->ctrl->ctrl.device,
1528                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1529         }
1530
1531         return -ECONNRESET;
1532 }
1533
1534 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1535 {
1536         int ret;
1537
1538         ret = nvme_rdma_create_queue_ib(queue);
1539         if (ret)
1540                 return ret;
1541
1542         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1543         if (ret) {
1544                 dev_err(queue->ctrl->ctrl.device,
1545                         "rdma_resolve_route failed (%d).\n",
1546                         queue->cm_error);
1547                 goto out_destroy_queue;
1548         }
1549
1550         return 0;
1551
1552 out_destroy_queue:
1553         nvme_rdma_destroy_queue_ib(queue);
1554         return ret;
1555 }
1556
1557 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1558 {
1559         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1560         struct rdma_conn_param param = { };
1561         struct nvme_rdma_cm_req priv = { };
1562         int ret;
1563
1564         param.qp_num = queue->qp->qp_num;
1565         param.flow_control = 1;
1566
1567         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1568         /* maximum retry count */
1569         param.retry_count = 7;
1570         param.rnr_retry_count = 7;
1571         param.private_data = &priv;
1572         param.private_data_len = sizeof(priv);
1573
1574         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1575         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1576         /*
1577          * set the admin queue depth to the minimum size
1578          * specified by the Fabrics standard.
1579          */
1580         if (priv.qid == 0) {
1581                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1582                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1583         } else {
1584                 /*
1585                  * current interpretation of the fabrics spec
1586                  * is at minimum you make hrqsize sqsize+1, or a
1587                  * 1's based representation of sqsize.
1588                  */
1589                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1590                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1591         }
1592
1593         ret = rdma_connect(queue->cm_id, &param);
1594         if (ret) {
1595                 dev_err(ctrl->ctrl.device,
1596                         "rdma_connect failed (%d).\n", ret);
1597                 goto out_destroy_queue_ib;
1598         }
1599
1600         return 0;
1601
1602 out_destroy_queue_ib:
1603         nvme_rdma_destroy_queue_ib(queue);
1604         return ret;
1605 }
1606
1607 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1608                 struct rdma_cm_event *ev)
1609 {
1610         struct nvme_rdma_queue *queue = cm_id->context;
1611         int cm_error = 0;
1612
1613         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1614                 rdma_event_msg(ev->event), ev->event,
1615                 ev->status, cm_id);
1616
1617         switch (ev->event) {
1618         case RDMA_CM_EVENT_ADDR_RESOLVED:
1619                 cm_error = nvme_rdma_addr_resolved(queue);
1620                 break;
1621         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1622                 cm_error = nvme_rdma_route_resolved(queue);
1623                 break;
1624         case RDMA_CM_EVENT_ESTABLISHED:
1625                 queue->cm_error = nvme_rdma_conn_established(queue);
1626                 /* complete cm_done regardless of success/failure */
1627                 complete(&queue->cm_done);
1628                 return 0;
1629         case RDMA_CM_EVENT_REJECTED:
1630                 nvme_rdma_destroy_queue_ib(queue);
1631                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1632                 break;
1633         case RDMA_CM_EVENT_ROUTE_ERROR:
1634         case RDMA_CM_EVENT_CONNECT_ERROR:
1635         case RDMA_CM_EVENT_UNREACHABLE:
1636                 nvme_rdma_destroy_queue_ib(queue);
1637                 /* fall through */
1638         case RDMA_CM_EVENT_ADDR_ERROR:
1639                 dev_dbg(queue->ctrl->ctrl.device,
1640                         "CM error event %d\n", ev->event);
1641                 cm_error = -ECONNRESET;
1642                 break;
1643         case RDMA_CM_EVENT_DISCONNECTED:
1644         case RDMA_CM_EVENT_ADDR_CHANGE:
1645         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1646                 dev_dbg(queue->ctrl->ctrl.device,
1647                         "disconnect received - connection closed\n");
1648                 nvme_rdma_error_recovery(queue->ctrl);
1649                 break;
1650         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1651                 /* device removal is handled via the ib_client API */
1652                 break;
1653         default:
1654                 dev_err(queue->ctrl->ctrl.device,
1655                         "Unexpected RDMA CM event (%d)\n", ev->event);
1656                 nvme_rdma_error_recovery(queue->ctrl);
1657                 break;
1658         }
1659
1660         if (cm_error) {
1661                 queue->cm_error = cm_error;
1662                 complete(&queue->cm_done);
1663         }
1664
1665         return 0;
1666 }
1667
1668 static enum blk_eh_timer_return
1669 nvme_rdma_timeout(struct request *rq, bool reserved)
1670 {
1671         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1672
1673         dev_warn(req->queue->ctrl->ctrl.device,
1674                  "I/O %d QID %d timeout, reset controller\n",
1675                  rq->tag, nvme_rdma_queue_idx(req->queue));
1676
1677         /* queue error recovery */
1678         nvme_rdma_error_recovery(req->queue->ctrl);
1679
1680         /* fail with DNR on cmd timeout */
1681         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1682
1683         return BLK_EH_DONE;
1684 }
1685
1686 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1687                 const struct blk_mq_queue_data *bd)
1688 {
1689         struct nvme_ns *ns = hctx->queue->queuedata;
1690         struct nvme_rdma_queue *queue = hctx->driver_data;
1691         struct request *rq = bd->rq;
1692         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1693         struct nvme_rdma_qe *sqe = &req->sqe;
1694         struct nvme_command *c = sqe->data;
1695         struct ib_device *dev;
1696         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1697         blk_status_t ret;
1698         int err;
1699
1700         WARN_ON_ONCE(rq->tag < 0);
1701
1702         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1703                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1704
1705         dev = queue->device->dev;
1706         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1707                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1708
1709         ret = nvme_setup_cmd(ns, rq, c);
1710         if (ret)
1711                 return ret;
1712
1713         blk_mq_start_request(rq);
1714
1715         err = nvme_rdma_map_data(queue, rq, c);
1716         if (unlikely(err < 0)) {
1717                 dev_err(queue->ctrl->ctrl.device,
1718                              "Failed to map data (%d)\n", err);
1719                 nvme_cleanup_cmd(rq);
1720                 goto err;
1721         }
1722
1723         sqe->cqe.done = nvme_rdma_send_done;
1724
1725         ib_dma_sync_single_for_device(dev, sqe->dma,
1726                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1727
1728         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1729                         req->mr ? &req->reg_wr.wr : NULL);
1730         if (unlikely(err)) {
1731                 nvme_rdma_unmap_data(queue, rq);
1732                 goto err;
1733         }
1734
1735         return BLK_STS_OK;
1736 err:
1737         if (err == -ENOMEM || err == -EAGAIN)
1738                 return BLK_STS_RESOURCE;
1739         return BLK_STS_IOERR;
1740 }
1741
1742 static void nvme_rdma_complete_rq(struct request *rq)
1743 {
1744         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1745
1746         nvme_rdma_unmap_data(req->queue, rq);
1747         nvme_complete_rq(rq);
1748 }
1749
1750 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1751 {
1752         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1753
1754         return blk_mq_rdma_map_queues(&set->map[0], ctrl->device->dev, 0);
1755 }
1756
1757 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1758         .queue_rq       = nvme_rdma_queue_rq,
1759         .complete       = nvme_rdma_complete_rq,
1760         .init_request   = nvme_rdma_init_request,
1761         .exit_request   = nvme_rdma_exit_request,
1762         .init_hctx      = nvme_rdma_init_hctx,
1763         .timeout        = nvme_rdma_timeout,
1764         .map_queues     = nvme_rdma_map_queues,
1765 };
1766
1767 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1768         .queue_rq       = nvme_rdma_queue_rq,
1769         .complete       = nvme_rdma_complete_rq,
1770         .init_request   = nvme_rdma_init_request,
1771         .exit_request   = nvme_rdma_exit_request,
1772         .init_hctx      = nvme_rdma_init_admin_hctx,
1773         .timeout        = nvme_rdma_timeout,
1774 };
1775
1776 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1777 {
1778         nvme_rdma_teardown_io_queues(ctrl, shutdown);
1779         if (shutdown)
1780                 nvme_shutdown_ctrl(&ctrl->ctrl);
1781         else
1782                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1783         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1784 }
1785
1786 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1787 {
1788         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1789 }
1790
1791 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1792 {
1793         struct nvme_rdma_ctrl *ctrl =
1794                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1795
1796         nvme_stop_ctrl(&ctrl->ctrl);
1797         nvme_rdma_shutdown_ctrl(ctrl, false);
1798
1799         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1800                 /* state change failure should never happen */
1801                 WARN_ON_ONCE(1);
1802                 return;
1803         }
1804
1805         if (nvme_rdma_setup_ctrl(ctrl, false))
1806                 goto out_fail;
1807
1808         return;
1809
1810 out_fail:
1811         ++ctrl->ctrl.nr_reconnects;
1812         nvme_rdma_reconnect_or_remove(ctrl);
1813 }
1814
1815 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1816         .name                   = "rdma",
1817         .module                 = THIS_MODULE,
1818         .flags                  = NVME_F_FABRICS,
1819         .reg_read32             = nvmf_reg_read32,
1820         .reg_read64             = nvmf_reg_read64,
1821         .reg_write32            = nvmf_reg_write32,
1822         .free_ctrl              = nvme_rdma_free_ctrl,
1823         .submit_async_event     = nvme_rdma_submit_async_event,
1824         .delete_ctrl            = nvme_rdma_delete_ctrl,
1825         .get_address            = nvmf_get_address,
1826         .stop_ctrl              = nvme_rdma_stop_ctrl,
1827 };
1828
1829 /*
1830  * Fails a connection request if it matches an existing controller
1831  * (association) with the same tuple:
1832  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1833  *
1834  * if local address is not specified in the request, it will match an
1835  * existing controller with all the other parameters the same and no
1836  * local port address specified as well.
1837  *
1838  * The ports don't need to be compared as they are intrinsically
1839  * already matched by the port pointers supplied.
1840  */
1841 static bool
1842 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1843 {
1844         struct nvme_rdma_ctrl *ctrl;
1845         bool found = false;
1846
1847         mutex_lock(&nvme_rdma_ctrl_mutex);
1848         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1849                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1850                 if (found)
1851                         break;
1852         }
1853         mutex_unlock(&nvme_rdma_ctrl_mutex);
1854
1855         return found;
1856 }
1857
1858 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1859                 struct nvmf_ctrl_options *opts)
1860 {
1861         struct nvme_rdma_ctrl *ctrl;
1862         int ret;
1863         bool changed;
1864
1865         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1866         if (!ctrl)
1867                 return ERR_PTR(-ENOMEM);
1868         ctrl->ctrl.opts = opts;
1869         INIT_LIST_HEAD(&ctrl->list);
1870
1871         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1872                 opts->trsvcid =
1873                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1874                 if (!opts->trsvcid) {
1875                         ret = -ENOMEM;
1876                         goto out_free_ctrl;
1877                 }
1878                 opts->mask |= NVMF_OPT_TRSVCID;
1879         }
1880
1881         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1882                         opts->traddr, opts->trsvcid, &ctrl->addr);
1883         if (ret) {
1884                 pr_err("malformed address passed: %s:%s\n",
1885                         opts->traddr, opts->trsvcid);
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         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1900                 ret = -EALREADY;
1901                 goto out_free_ctrl;
1902         }
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->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1908
1909         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1910         ctrl->ctrl.sqsize = opts->queue_size - 1;
1911         ctrl->ctrl.kato = opts->kato;
1912
1913         ret = -ENOMEM;
1914         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1915                                 GFP_KERNEL);
1916         if (!ctrl->queues)
1917                 goto out_free_ctrl;
1918
1919         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1920                                 0 /* no quirks, we're perfect! */);
1921         if (ret)
1922                 goto out_kfree_queues;
1923
1924         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1925         WARN_ON_ONCE(!changed);
1926
1927         ret = nvme_rdma_setup_ctrl(ctrl, true);
1928         if (ret)
1929                 goto out_uninit_ctrl;
1930
1931         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1932                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1933
1934         nvme_get_ctrl(&ctrl->ctrl);
1935
1936         mutex_lock(&nvme_rdma_ctrl_mutex);
1937         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1938         mutex_unlock(&nvme_rdma_ctrl_mutex);
1939
1940         return &ctrl->ctrl;
1941
1942 out_uninit_ctrl:
1943         nvme_uninit_ctrl(&ctrl->ctrl);
1944         nvme_put_ctrl(&ctrl->ctrl);
1945         if (ret > 0)
1946                 ret = -EIO;
1947         return ERR_PTR(ret);
1948 out_kfree_queues:
1949         kfree(ctrl->queues);
1950 out_free_ctrl:
1951         kfree(ctrl);
1952         return ERR_PTR(ret);
1953 }
1954
1955 static struct nvmf_transport_ops nvme_rdma_transport = {
1956         .name           = "rdma",
1957         .module         = THIS_MODULE,
1958         .required_opts  = NVMF_OPT_TRADDR,
1959         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1960                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1961         .create_ctrl    = nvme_rdma_create_ctrl,
1962 };
1963
1964 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1965 {
1966         struct nvme_rdma_ctrl *ctrl;
1967         struct nvme_rdma_device *ndev;
1968         bool found = false;
1969
1970         mutex_lock(&device_list_mutex);
1971         list_for_each_entry(ndev, &device_list, entry) {
1972                 if (ndev->dev == ib_device) {
1973                         found = true;
1974                         break;
1975                 }
1976         }
1977         mutex_unlock(&device_list_mutex);
1978
1979         if (!found)
1980                 return;
1981
1982         /* Delete all controllers using this device */
1983         mutex_lock(&nvme_rdma_ctrl_mutex);
1984         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1985                 if (ctrl->device->dev != ib_device)
1986                         continue;
1987                 nvme_delete_ctrl(&ctrl->ctrl);
1988         }
1989         mutex_unlock(&nvme_rdma_ctrl_mutex);
1990
1991         flush_workqueue(nvme_delete_wq);
1992 }
1993
1994 static struct ib_client nvme_rdma_ib_client = {
1995         .name   = "nvme_rdma",
1996         .remove = nvme_rdma_remove_one
1997 };
1998
1999 static int __init nvme_rdma_init_module(void)
2000 {
2001         int ret;
2002
2003         ret = ib_register_client(&nvme_rdma_ib_client);
2004         if (ret)
2005                 return ret;
2006
2007         ret = nvmf_register_transport(&nvme_rdma_transport);
2008         if (ret)
2009                 goto err_unreg_client;
2010
2011         return 0;
2012
2013 err_unreg_client:
2014         ib_unregister_client(&nvme_rdma_ib_client);
2015         return ret;
2016 }
2017
2018 static void __exit nvme_rdma_cleanup_module(void)
2019 {
2020         nvmf_unregister_transport(&nvme_rdma_transport);
2021         ib_unregister_client(&nvme_rdma_ib_client);
2022 }
2023
2024 module_init(nvme_rdma_init_module);
2025 module_exit(nvme_rdma_cleanup_module);
2026
2027 MODULE_LICENSE("GPL v2");