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 = NUMA_NO_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 = NUMA_NO_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
 767         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
 768
 769         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 770                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 771         if (error)
 772                 goto out_free_queue;
 773
 774         if (new) {
 775                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 776                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 777                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 778                         goto out_free_async_qe;
 779                 }
 780
 781                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 782                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 783                         error = PTR_ERR(ctrl->ctrl.admin_q);
 784                         goto out_free_tagset;
 785                 }
 786         }
 787
 788         error = nvme_rdma_start_queue(ctrl, 0);
 789         if (error)
 790                 goto out_cleanup_queue;
 791
 792         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 793                         &ctrl->ctrl.cap);
 794         if (error) {
 795                 dev_err(ctrl->ctrl.device,
 796                         "prop_get NVME_REG_CAP failed\n");
 797                 goto out_stop_queue;
 798         }
 799
 800         ctrl->ctrl.sqsize =
 801                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 802
 803         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 804         if (error)
 805                 goto out_stop_queue;
 806
 807         ctrl->ctrl.max_hw_sectors =
 808                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 809
 810         error = nvme_init_identify(&ctrl->ctrl);
 811         if (error)
 812                 goto out_stop_queue;
 813
 814         return 0;
 815
 816 out_stop_queue:
 817         nvme_rdma_stop_queue(&ctrl->queues[0]);
 818 out_cleanup_queue:
 819         if (new)
 820                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 821 out_free_tagset:
 822         if (new)
 823                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 824 out_free_async_qe:
 825         nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 826                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 827         ctrl->async_event_sqe.data = NULL;
 828 out_free_queue:
 829         nvme_rdma_free_queue(&ctrl->queues[0]);
 830         return error;
 831 }
 832
 833 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 834                 bool remove)
 835 {
 836         if (remove) {
 837                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 838                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 839         }
 840         nvme_rdma_free_io_queues(ctrl);
 841 }
 842
 843 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 844 {
 845         int ret;
 846
 847         ret = nvme_rdma_alloc_io_queues(ctrl);
 848         if (ret)
 849                 return ret;
 850
 851         if (new) {
 852                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 853                 if (IS_ERR(ctrl->ctrl.tagset)) {
 854                         ret = PTR_ERR(ctrl->ctrl.tagset);
 855                         goto out_free_io_queues;
 856                 }
 857
 858                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 859                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 860                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 861                         goto out_free_tag_set;
 862                 }
 863         } else {
 864                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 865                         ctrl->ctrl.queue_count - 1);
 866         }
 867
 868         ret = nvme_rdma_start_io_queues(ctrl);
 869         if (ret)
 870                 goto out_cleanup_connect_q;
 871
 872         return 0;
 873
 874 out_cleanup_connect_q:
 875         if (new)
 876                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 877 out_free_tag_set:
 878         if (new)
 879                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 880 out_free_io_queues:
 881         nvme_rdma_free_io_queues(ctrl);
 882         return ret;
 883 }
 884
 885 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
 886                 bool remove)
 887 {
 888         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 889         nvme_rdma_stop_queue(&ctrl->queues[0]);
 890         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
 891                         &ctrl->ctrl);
 892         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 893         nvme_rdma_destroy_admin_queue(ctrl, remove);
 894 }
 895
 896 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
 897                 bool remove)
 898 {
 899         if (ctrl->ctrl.queue_count > 1) {
 900                 nvme_stop_queues(&ctrl->ctrl);
 901                 nvme_rdma_stop_io_queues(ctrl);
 902                 blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
 903                                 &ctrl->ctrl);
 904                 if (remove)
 905                         nvme_start_queues(&ctrl->ctrl);
 906                 nvme_rdma_destroy_io_queues(ctrl, remove);
 907         }
 908 }
 909
 910 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
 911 {
 912         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 913
 914         cancel_work_sync(&ctrl->err_work);
 915         cancel_delayed_work_sync(&ctrl->reconnect_work);
 916 }
 917
 918 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 919 {
 920         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 921
 922         if (list_empty(&ctrl->list))
 923                 goto free_ctrl;
 924
 925         mutex_lock(&nvme_rdma_ctrl_mutex);
 926         list_del(&ctrl->list);
 927         mutex_unlock(&nvme_rdma_ctrl_mutex);
 928
 929         nvmf_free_options(nctrl->opts);
 930 free_ctrl:
 931         kfree(ctrl->queues);
 932         kfree(ctrl);
 933 }
 934
 935 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 936 {
 937         /* If we are resetting/deleting then do nothing */
 938         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 939                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 940                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 941                 return;
 942         }
 943
 944         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 945                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 946                         ctrl->ctrl.opts->reconnect_delay);
 947                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 948                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 949         } else {
 950                 nvme_delete_ctrl(&ctrl->ctrl);
 951         }
 952 }
 953
 954 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
 955 {
 956         int ret = -EINVAL;
 957         bool changed;
 958
 959         ret = nvme_rdma_configure_admin_queue(ctrl, new);
 960         if (ret)
 961                 return ret;
 962
 963         if (ctrl->ctrl.icdoff) {
 964                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
 965                 goto destroy_admin;
 966         }
 967
 968         if (!(ctrl->ctrl.sgls & (1 << 2))) {
 969                 dev_err(ctrl->ctrl.device,
 970                         "Mandatory keyed sgls are not supported!\n");
 971                 goto destroy_admin;
 972         }
 973
 974         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
 975                 dev_warn(ctrl->ctrl.device,
 976                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
 977                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
 978         }
 979
 980         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
 981                 dev_warn(ctrl->ctrl.device,
 982                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
 983                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
 984                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
 985         }
 986
 987         if (ctrl->ctrl.sgls & (1 << 20))
 988                 ctrl->use_inline_data = true;
 989
 990         if (ctrl->ctrl.queue_count > 1) {
 991                 ret = nvme_rdma_configure_io_queues(ctrl, new);
 992                 if (ret)
 993                         goto destroy_admin;
 994         }
 995
 996         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 997         if (!changed) {
 998                 /* state change failure is ok if we're in DELETING state */
 999                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1000                 ret = -EINVAL;
1001                 goto destroy_io;
1002         }
1003
1004         nvme_start_ctrl(&ctrl->ctrl);
1005         return 0;
1006
1007 destroy_io:
1008         if (ctrl->ctrl.queue_count > 1)
1009                 nvme_rdma_destroy_io_queues(ctrl, new);
1010 destroy_admin:
1011         nvme_rdma_stop_queue(&ctrl->queues[0]);
1012         nvme_rdma_destroy_admin_queue(ctrl, new);
1013         return ret;
1014 }
1015
1016 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1017 {
1018         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1019                         struct nvme_rdma_ctrl, reconnect_work);
1020
1021         ++ctrl->ctrl.nr_reconnects;
1022
1023         if (nvme_rdma_setup_ctrl(ctrl, false))
1024                 goto requeue;
1025
1026         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1027                         ctrl->ctrl.nr_reconnects);
1028
1029         ctrl->ctrl.nr_reconnects = 0;
1030
1031         return;
1032
1033 requeue:
1034         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1035                         ctrl->ctrl.nr_reconnects);
1036         nvme_rdma_reconnect_or_remove(ctrl);
1037 }
1038
1039 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1040 {
1041         struct nvme_rdma_ctrl *ctrl = container_of(work,
1042                         struct nvme_rdma_ctrl, err_work);
1043
1044         nvme_stop_keep_alive(&ctrl->ctrl);
1045         nvme_rdma_teardown_io_queues(ctrl, false);
1046         nvme_start_queues(&ctrl->ctrl);
1047         nvme_rdma_teardown_admin_queue(ctrl, false);
1048
1049         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1050                 /* state change failure is ok if we're in DELETING state */
1051                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1052                 return;
1053         }
1054
1055         nvme_rdma_reconnect_or_remove(ctrl);
1056 }
1057
1058 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1059 {
1060         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1061                 return;
1062
1063         queue_work(nvme_wq, &ctrl->err_work);
1064 }
1065
1066 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1067                 const char *op)
1068 {
1069         struct nvme_rdma_queue *queue = cq->cq_context;
1070         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1071
1072         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1073                 dev_info(ctrl->ctrl.device,
1074                              "%s for CQE 0x%p failed with status %s (%d)\n",
1075                              op, wc->wr_cqe,
1076                              ib_wc_status_msg(wc->status), wc->status);
1077         nvme_rdma_error_recovery(ctrl);
1078 }
1079
1080 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1081 {
1082         if (unlikely(wc->status != IB_WC_SUCCESS))
1083                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1084 }
1085
1086 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1087 {
1088         struct nvme_rdma_request *req =
1089                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1090         struct request *rq = blk_mq_rq_from_pdu(req);
1091
1092         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1093                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1094                 return;
1095         }
1096
1097         if (refcount_dec_and_test(&req->ref))
1098                 nvme_end_request(rq, req->status, req->result);
1099
1100 }
1101
1102 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1103                 struct nvme_rdma_request *req)
1104 {
1105         struct ib_send_wr wr = {
1106                 .opcode             = IB_WR_LOCAL_INV,
1107                 .next               = NULL,
1108                 .num_sge            = 0,
1109                 .send_flags         = IB_SEND_SIGNALED,
1110                 .ex.invalidate_rkey = req->mr->rkey,
1111         };
1112
1113         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1114         wr.wr_cqe = &req->reg_cqe;
1115
1116         return ib_post_send(queue->qp, &wr, NULL);
1117 }
1118
1119 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1120                 struct request *rq)
1121 {
1122         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1123         struct nvme_rdma_device *dev = queue->device;
1124         struct ib_device *ibdev = dev->dev;
1125
1126         if (!blk_rq_payload_bytes(rq))
1127                 return;
1128
1129         if (req->mr) {
1130                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1131                 req->mr = NULL;
1132         }
1133
1134         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1135                         req->nents, rq_data_dir(rq) ==
1136                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1137
1138         nvme_cleanup_cmd(rq);
1139         sg_free_table_chained(&req->sg_table, true);
1140 }
1141
1142 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1143 {
1144         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1145
1146         sg->addr = 0;
1147         put_unaligned_le24(0, sg->length);
1148         put_unaligned_le32(0, sg->key);
1149         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1150         return 0;
1151 }
1152
1153 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1154                 struct nvme_rdma_request *req, struct nvme_command *c,
1155                 int count)
1156 {
1157         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1158         struct scatterlist *sgl = req->sg_table.sgl;
1159         struct ib_sge *sge = &req->sge[1];
1160         u32 len = 0;
1161         int i;
1162
1163         for (i = 0; i < count; i++, sgl++, sge++) {
1164                 sge->addr = sg_dma_address(sgl);
1165                 sge->length = sg_dma_len(sgl);
1166                 sge->lkey = queue->device->pd->local_dma_lkey;
1167                 len += sge->length;
1168         }
1169
1170         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1171         sg->length = cpu_to_le32(len);
1172         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1173
1174         req->num_sge += count;
1175         return 0;
1176 }
1177
1178 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1179                 struct nvme_rdma_request *req, struct nvme_command *c)
1180 {
1181         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1182
1183         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1184         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1185         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1186         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1187         return 0;
1188 }
1189
1190 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1191                 struct nvme_rdma_request *req, struct nvme_command *c,
1192                 int count)
1193 {
1194         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1195         int nr;
1196
1197         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1198         if (WARN_ON_ONCE(!req->mr))
1199                 return -EAGAIN;
1200
1201         /*
1202          * Align the MR to a 4K page size to match the ctrl page size and
1203          * the block virtual boundary.
1204          */
1205         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1206         if (unlikely(nr < count)) {
1207                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1208                 req->mr = NULL;
1209                 if (nr < 0)
1210                         return nr;
1211                 return -EINVAL;
1212         }
1213
1214         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1215
1216         req->reg_cqe.done = nvme_rdma_memreg_done;
1217         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1218         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1219         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1220         req->reg_wr.wr.num_sge = 0;
1221         req->reg_wr.mr = req->mr;
1222         req->reg_wr.key = req->mr->rkey;
1223         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1224                              IB_ACCESS_REMOTE_READ |
1225                              IB_ACCESS_REMOTE_WRITE;
1226
1227         sg->addr = cpu_to_le64(req->mr->iova);
1228         put_unaligned_le24(req->mr->length, sg->length);
1229         put_unaligned_le32(req->mr->rkey, sg->key);
1230         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1231                         NVME_SGL_FMT_INVALIDATE;
1232
1233         return 0;
1234 }
1235
1236 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1237                 struct request *rq, struct nvme_command *c)
1238 {
1239         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1240         struct nvme_rdma_device *dev = queue->device;
1241         struct ib_device *ibdev = dev->dev;
1242         int count, ret;
1243
1244         req->num_sge = 1;
1245         refcount_set(&req->ref, 2); /* send and recv completions */
1246
1247         c->common.flags |= NVME_CMD_SGL_METABUF;
1248
1249         if (!blk_rq_payload_bytes(rq))
1250                 return nvme_rdma_set_sg_null(c);
1251
1252         req->sg_table.sgl = req->first_sgl;
1253         ret = sg_alloc_table_chained(&req->sg_table,
1254                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1255         if (ret)
1256                 return -ENOMEM;
1257
1258         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1259
1260         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1261                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1262         if (unlikely(count <= 0)) {
1263                 ret = -EIO;
1264                 goto out_free_table;
1265         }
1266
1267         if (count <= dev->num_inline_segments) {
1268                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1269                     queue->ctrl->use_inline_data &&
1270                     blk_rq_payload_bytes(rq) <=
1271                                 nvme_rdma_inline_data_size(queue)) {
1272                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1273                         goto out;
1274                 }
1275
1276                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1277                         ret = nvme_rdma_map_sg_single(queue, req, c);
1278                         goto out;
1279                 }
1280         }
1281
1282         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1283 out:
1284         if (unlikely(ret))
1285                 goto out_unmap_sg;
1286
1287         return 0;
1288
1289 out_unmap_sg:
1290         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1291                         req->nents, rq_data_dir(rq) ==
1292                         WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1293 out_free_table:
1294         sg_free_table_chained(&req->sg_table, true);
1295         return ret;
1296 }
1297
1298 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1299 {
1300         struct nvme_rdma_qe *qe =
1301                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1302         struct nvme_rdma_request *req =
1303                 container_of(qe, struct nvme_rdma_request, sqe);
1304         struct request *rq = blk_mq_rq_from_pdu(req);
1305
1306         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1307                 nvme_rdma_wr_error(cq, wc, "SEND");
1308                 return;
1309         }
1310
1311         if (refcount_dec_and_test(&req->ref))
1312                 nvme_end_request(rq, req->status, req->result);
1313 }
1314
1315 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1316                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1317                 struct ib_send_wr *first)
1318 {
1319         struct ib_send_wr wr;
1320         int ret;
1321
1322         sge->addr   = qe->dma;
1323         sge->length = sizeof(struct nvme_command),
1324         sge->lkey   = queue->device->pd->local_dma_lkey;
1325
1326         wr.next       = NULL;
1327         wr.wr_cqe     = &qe->cqe;
1328         wr.sg_list    = sge;
1329         wr.num_sge    = num_sge;
1330         wr.opcode     = IB_WR_SEND;
1331         wr.send_flags = IB_SEND_SIGNALED;
1332
1333         if (first)
1334                 first->next = &wr;
1335         else
1336                 first = &wr;
1337
1338         ret = ib_post_send(queue->qp, first, NULL);
1339         if (unlikely(ret)) {
1340                 dev_err(queue->ctrl->ctrl.device,
1341                              "%s failed with error code %d\n", __func__, ret);
1342         }
1343         return ret;
1344 }
1345
1346 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1347                 struct nvme_rdma_qe *qe)
1348 {
1349         struct ib_recv_wr wr;
1350         struct ib_sge list;
1351         int ret;
1352
1353         list.addr   = qe->dma;
1354         list.length = sizeof(struct nvme_completion);
1355         list.lkey   = queue->device->pd->local_dma_lkey;
1356
1357         qe->cqe.done = nvme_rdma_recv_done;
1358
1359         wr.next     = NULL;
1360         wr.wr_cqe   = &qe->cqe;
1361         wr.sg_list  = &list;
1362         wr.num_sge  = 1;
1363
1364         ret = ib_post_recv(queue->qp, &wr, NULL);
1365         if (unlikely(ret)) {
1366                 dev_err(queue->ctrl->ctrl.device,
1367                         "%s failed with error code %d\n", __func__, ret);
1368         }
1369         return ret;
1370 }
1371
1372 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1373 {
1374         u32 queue_idx = nvme_rdma_queue_idx(queue);
1375
1376         if (queue_idx == 0)
1377                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1378         return queue->ctrl->tag_set.tags[queue_idx - 1];
1379 }
1380
1381 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1382 {
1383         if (unlikely(wc->status != IB_WC_SUCCESS))
1384                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1385 }
1386
1387 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1388 {
1389         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1390         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1391         struct ib_device *dev = queue->device->dev;
1392         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1393         struct nvme_command *cmd = sqe->data;
1394         struct ib_sge sge;
1395         int ret;
1396
1397         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1398
1399         memset(cmd, 0, sizeof(*cmd));
1400         cmd->common.opcode = nvme_admin_async_event;
1401         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1402         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1403         nvme_rdma_set_sg_null(cmd);
1404
1405         sqe->cqe.done = nvme_rdma_async_done;
1406
1407         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1408                         DMA_TO_DEVICE);
1409
1410         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1411         WARN_ON_ONCE(ret);
1412 }
1413
1414 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1415                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1416 {
1417         struct request *rq;
1418         struct nvme_rdma_request *req;
1419         int ret = 0;
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 ret;
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                 ret = nvme_rdma_inv_rkey(queue, req);
1443                 if (unlikely(ret < 0)) {
1444                         dev_err(queue->ctrl->ctrl.device,
1445                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1446                                 req->mr->rkey, ret);
1447                         nvme_rdma_error_recovery(queue->ctrl);
1448                 }
1449                 /* the local invalidation completion will end the request */
1450                 return 0;
1451         }
1452
1453         if (refcount_dec_and_test(&req->ref)) {
1454                 if (rq->tag == tag)
1455                         ret = 1;
1456                 nvme_end_request(rq, req->status, req->result);
1457         }
1458
1459         return ret;
1460 }
1461
1462 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1463 {
1464         struct nvme_rdma_qe *qe =
1465                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1466         struct nvme_rdma_queue *queue = cq->cq_context;
1467         struct ib_device *ibdev = queue->device->dev;
1468         struct nvme_completion *cqe = qe->data;
1469         const size_t len = sizeof(struct nvme_completion);
1470         int ret = 0;
1471
1472         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1473                 nvme_rdma_wr_error(cq, wc, "RECV");
1474                 return 0;
1475         }
1476
1477         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1478         /*
1479          * AEN requests are special as they don't time out and can
1480          * survive any kind of queue freeze and often don't respond to
1481          * aborts.  We don't even bother to allocate a struct request
1482          * for them but rather special case them here.
1483          */
1484         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1485                         cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1486                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1487                                 &cqe->result);
1488         else
1489                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1490         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1491
1492         nvme_rdma_post_recv(queue, qe);
1493         return ret;
1494 }
1495
1496 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1497 {
1498         __nvme_rdma_recv_done(cq, wc, -1);
1499 }
1500
1501 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1502 {
1503         int ret, i;
1504
1505         for (i = 0; i < queue->queue_size; i++) {
1506                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1507                 if (ret)
1508                         goto out_destroy_queue_ib;
1509         }
1510
1511         return 0;
1512
1513 out_destroy_queue_ib:
1514         nvme_rdma_destroy_queue_ib(queue);
1515         return ret;
1516 }
1517
1518 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1519                 struct rdma_cm_event *ev)
1520 {
1521         struct rdma_cm_id *cm_id = queue->cm_id;
1522         int status = ev->status;
1523         const char *rej_msg;
1524         const struct nvme_rdma_cm_rej *rej_data;
1525         u8 rej_data_len;
1526
1527         rej_msg = rdma_reject_msg(cm_id, status);
1528         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1529
1530         if (rej_data && rej_data_len >= sizeof(u16)) {
1531                 u16 sts = le16_to_cpu(rej_data->sts);
1532
1533                 dev_err(queue->ctrl->ctrl.device,
1534                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1535                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1536         } else {
1537                 dev_err(queue->ctrl->ctrl.device,
1538                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1539         }
1540
1541         return -ECONNRESET;
1542 }
1543
1544 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1545 {
1546         int ret;
1547
1548         ret = nvme_rdma_create_queue_ib(queue);
1549         if (ret)
1550                 return ret;
1551
1552         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1553         if (ret) {
1554                 dev_err(queue->ctrl->ctrl.device,
1555                         "rdma_resolve_route failed (%d).\n",
1556                         queue->cm_error);
1557                 goto out_destroy_queue;
1558         }
1559
1560         return 0;
1561
1562 out_destroy_queue:
1563         nvme_rdma_destroy_queue_ib(queue);
1564         return ret;
1565 }
1566
1567 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1568 {
1569         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1570         struct rdma_conn_param param = { };
1571         struct nvme_rdma_cm_req priv = { };
1572         int ret;
1573
1574         param.qp_num = queue->qp->qp_num;
1575         param.flow_control = 1;
1576
1577         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1578         /* maximum retry count */
1579         param.retry_count = 7;
1580         param.rnr_retry_count = 7;
1581         param.private_data = &priv;
1582         param.private_data_len = sizeof(priv);
1583
1584         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1585         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1586         /*
1587          * set the admin queue depth to the minimum size
1588          * specified by the Fabrics standard.
1589          */
1590         if (priv.qid == 0) {
1591                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1592                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1593         } else {
1594                 /*
1595                  * current interpretation of the fabrics spec
1596                  * is at minimum you make hrqsize sqsize+1, or a
1597                  * 1's based representation of sqsize.
1598                  */
1599                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1600                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1601         }
1602
1603         ret = rdma_connect(queue->cm_id, &param);
1604         if (ret) {
1605                 dev_err(ctrl->ctrl.device,
1606                         "rdma_connect failed (%d).\n", ret);
1607                 goto out_destroy_queue_ib;
1608         }
1609
1610         return 0;
1611
1612 out_destroy_queue_ib:
1613         nvme_rdma_destroy_queue_ib(queue);
1614         return ret;
1615 }
1616
1617 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1618                 struct rdma_cm_event *ev)
1619 {
1620         struct nvme_rdma_queue *queue = cm_id->context;
1621         int cm_error = 0;
1622
1623         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1624                 rdma_event_msg(ev->event), ev->event,
1625                 ev->status, cm_id);
1626
1627         switch (ev->event) {
1628         case RDMA_CM_EVENT_ADDR_RESOLVED:
1629                 cm_error = nvme_rdma_addr_resolved(queue);
1630                 break;
1631         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1632                 cm_error = nvme_rdma_route_resolved(queue);
1633                 break;
1634         case RDMA_CM_EVENT_ESTABLISHED:
1635                 queue->cm_error = nvme_rdma_conn_established(queue);
1636                 /* complete cm_done regardless of success/failure */
1637                 complete(&queue->cm_done);
1638                 return 0;
1639         case RDMA_CM_EVENT_REJECTED:
1640                 nvme_rdma_destroy_queue_ib(queue);
1641                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1642                 break;
1643         case RDMA_CM_EVENT_ROUTE_ERROR:
1644         case RDMA_CM_EVENT_CONNECT_ERROR:
1645         case RDMA_CM_EVENT_UNREACHABLE:
1646                 nvme_rdma_destroy_queue_ib(queue);
1647                 /* fall through */
1648         case RDMA_CM_EVENT_ADDR_ERROR:
1649                 dev_dbg(queue->ctrl->ctrl.device,
1650                         "CM error event %d\n", ev->event);
1651                 cm_error = -ECONNRESET;
1652                 break;
1653         case RDMA_CM_EVENT_DISCONNECTED:
1654         case RDMA_CM_EVENT_ADDR_CHANGE:
1655         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1656                 dev_dbg(queue->ctrl->ctrl.device,
1657                         "disconnect received - connection closed\n");
1658                 nvme_rdma_error_recovery(queue->ctrl);
1659                 break;
1660         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1661                 /* device removal is handled via the ib_client API */
1662                 break;
1663         default:
1664                 dev_err(queue->ctrl->ctrl.device,
1665                         "Unexpected RDMA CM event (%d)\n", ev->event);
1666                 nvme_rdma_error_recovery(queue->ctrl);
1667                 break;
1668         }
1669
1670         if (cm_error) {
1671                 queue->cm_error = cm_error;
1672                 complete(&queue->cm_done);
1673         }
1674
1675         return 0;
1676 }
1677
1678 static enum blk_eh_timer_return
1679 nvme_rdma_timeout(struct request *rq, bool reserved)
1680 {
1681         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1682
1683         dev_warn(req->queue->ctrl->ctrl.device,
1684                  "I/O %d QID %d timeout, reset controller\n",
1685                  rq->tag, nvme_rdma_queue_idx(req->queue));
1686
1687         /* queue error recovery */
1688         nvme_rdma_error_recovery(req->queue->ctrl);
1689
1690         /* fail with DNR on cmd timeout */
1691         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1692
1693         return BLK_EH_DONE;
1694 }
1695
1696 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1697                 const struct blk_mq_queue_data *bd)
1698 {
1699         struct nvme_ns *ns = hctx->queue->queuedata;
1700         struct nvme_rdma_queue *queue = hctx->driver_data;
1701         struct request *rq = bd->rq;
1702         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1703         struct nvme_rdma_qe *sqe = &req->sqe;
1704         struct nvme_command *c = sqe->data;
1705         struct ib_device *dev;
1706         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1707         blk_status_t ret;
1708         int err;
1709
1710         WARN_ON_ONCE(rq->tag < 0);
1711
1712         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1713                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1714
1715         dev = queue->device->dev;
1716         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1717                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1718
1719         ret = nvme_setup_cmd(ns, rq, c);
1720         if (ret)
1721                 return ret;
1722
1723         blk_mq_start_request(rq);
1724
1725         err = nvme_rdma_map_data(queue, rq, c);
1726         if (unlikely(err < 0)) {
1727                 dev_err(queue->ctrl->ctrl.device,
1728                              "Failed to map data (%d)\n", err);
1729                 nvme_cleanup_cmd(rq);
1730                 goto err;
1731         }
1732
1733         sqe->cqe.done = nvme_rdma_send_done;
1734
1735         ib_dma_sync_single_for_device(dev, sqe->dma,
1736                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1737
1738         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1739                         req->mr ? &req->reg_wr.wr : NULL);
1740         if (unlikely(err)) {
1741                 nvme_rdma_unmap_data(queue, rq);
1742                 goto err;
1743         }
1744
1745         return BLK_STS_OK;
1746 err:
1747         if (err == -ENOMEM || err == -EAGAIN)
1748                 return BLK_STS_RESOURCE;
1749         return BLK_STS_IOERR;
1750 }
1751
1752 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1753 {
1754         struct nvme_rdma_queue *queue = hctx->driver_data;
1755         struct ib_cq *cq = queue->ib_cq;
1756         struct ib_wc wc;
1757         int found = 0;
1758
1759         while (ib_poll_cq(cq, 1, &wc) > 0) {
1760                 struct ib_cqe *cqe = wc.wr_cqe;
1761
1762                 if (cqe) {
1763                         if (cqe->done == nvme_rdma_recv_done)
1764                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1765                         else
1766                                 cqe->done(cq, &wc);
1767                 }
1768         }
1769
1770         return found;
1771 }
1772
1773 static void nvme_rdma_complete_rq(struct request *rq)
1774 {
1775         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1776
1777         nvme_rdma_unmap_data(req->queue, rq);
1778         nvme_complete_rq(rq);
1779 }
1780
1781 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1782 {
1783         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1784
1785         return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1786 }
1787
1788 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1789         .queue_rq       = nvme_rdma_queue_rq,
1790         .complete       = nvme_rdma_complete_rq,
1791         .init_request   = nvme_rdma_init_request,
1792         .exit_request   = nvme_rdma_exit_request,
1793         .init_hctx      = nvme_rdma_init_hctx,
1794         .poll           = nvme_rdma_poll,
1795         .timeout        = nvme_rdma_timeout,
1796         .map_queues     = nvme_rdma_map_queues,
1797 };
1798
1799 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1800         .queue_rq       = nvme_rdma_queue_rq,
1801         .complete       = nvme_rdma_complete_rq,
1802         .init_request   = nvme_rdma_init_request,
1803         .exit_request   = nvme_rdma_exit_request,
1804         .init_hctx      = nvme_rdma_init_admin_hctx,
1805         .timeout        = nvme_rdma_timeout,
1806 };
1807
1808 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1809 {
1810         nvme_rdma_teardown_io_queues(ctrl, shutdown);
1811         if (shutdown)
1812                 nvme_shutdown_ctrl(&ctrl->ctrl);
1813         else
1814                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1815         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1816 }
1817
1818 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1819 {
1820         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1821 }
1822
1823 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1824 {
1825         struct nvme_rdma_ctrl *ctrl =
1826                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1827
1828         nvme_stop_ctrl(&ctrl->ctrl);
1829         nvme_rdma_shutdown_ctrl(ctrl, false);
1830
1831         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1832                 /* state change failure should never happen */
1833                 WARN_ON_ONCE(1);
1834                 return;
1835         }
1836
1837         if (nvme_rdma_setup_ctrl(ctrl, false))
1838                 goto out_fail;
1839
1840         return;
1841
1842 out_fail:
1843         ++ctrl->ctrl.nr_reconnects;
1844         nvme_rdma_reconnect_or_remove(ctrl);
1845 }
1846
1847 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1848         .name                   = "rdma",
1849         .module                 = THIS_MODULE,
1850         .flags                  = NVME_F_FABRICS,
1851         .reg_read32             = nvmf_reg_read32,
1852         .reg_read64             = nvmf_reg_read64,
1853         .reg_write32            = nvmf_reg_write32,
1854         .free_ctrl              = nvme_rdma_free_ctrl,
1855         .submit_async_event     = nvme_rdma_submit_async_event,
1856         .delete_ctrl            = nvme_rdma_delete_ctrl,
1857         .get_address            = nvmf_get_address,
1858         .stop_ctrl              = nvme_rdma_stop_ctrl,
1859 };
1860
1861 /*
1862  * Fails a connection request if it matches an existing controller
1863  * (association) with the same tuple:
1864  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1865  *
1866  * if local address is not specified in the request, it will match an
1867  * existing controller with all the other parameters the same and no
1868  * local port address specified as well.
1869  *
1870  * The ports don't need to be compared as they are intrinsically
1871  * already matched by the port pointers supplied.
1872  */
1873 static bool
1874 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1875 {
1876         struct nvme_rdma_ctrl *ctrl;
1877         bool found = false;
1878
1879         mutex_lock(&nvme_rdma_ctrl_mutex);
1880         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1881                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1882                 if (found)
1883                         break;
1884         }
1885         mutex_unlock(&nvme_rdma_ctrl_mutex);
1886
1887         return found;
1888 }
1889
1890 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1891                 struct nvmf_ctrl_options *opts)
1892 {
1893         struct nvme_rdma_ctrl *ctrl;
1894         int ret;
1895         bool changed;
1896
1897         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1898         if (!ctrl)
1899                 return ERR_PTR(-ENOMEM);
1900         ctrl->ctrl.opts = opts;
1901         INIT_LIST_HEAD(&ctrl->list);
1902
1903         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1904                 opts->trsvcid =
1905                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1906                 if (!opts->trsvcid) {
1907                         ret = -ENOMEM;
1908                         goto out_free_ctrl;
1909                 }
1910                 opts->mask |= NVMF_OPT_TRSVCID;
1911         }
1912
1913         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1914                         opts->traddr, opts->trsvcid, &ctrl->addr);
1915         if (ret) {
1916                 pr_err("malformed address passed: %s:%s\n",
1917                         opts->traddr, opts->trsvcid);
1918                 goto out_free_ctrl;
1919         }
1920
1921         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1922                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1923                         opts->host_traddr, NULL, &ctrl->src_addr);
1924                 if (ret) {
1925                         pr_err("malformed src address passed: %s\n",
1926                                opts->host_traddr);
1927                         goto out_free_ctrl;
1928                 }
1929         }
1930
1931         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1932                 ret = -EALREADY;
1933                 goto out_free_ctrl;
1934         }
1935
1936         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1937                         nvme_rdma_reconnect_ctrl_work);
1938         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1939         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1940
1941         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1942         ctrl->ctrl.sqsize = opts->queue_size - 1;
1943         ctrl->ctrl.kato = opts->kato;
1944
1945         ret = -ENOMEM;
1946         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1947                                 GFP_KERNEL);
1948         if (!ctrl->queues)
1949                 goto out_free_ctrl;
1950
1951         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1952                                 0 /* no quirks, we're perfect! */);
1953         if (ret)
1954                 goto out_kfree_queues;
1955
1956         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1957         WARN_ON_ONCE(!changed);
1958
1959         ret = nvme_rdma_setup_ctrl(ctrl, true);
1960         if (ret)
1961                 goto out_uninit_ctrl;
1962
1963         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1964                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1965
1966         nvme_get_ctrl(&ctrl->ctrl);
1967
1968         mutex_lock(&nvme_rdma_ctrl_mutex);
1969         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1970         mutex_unlock(&nvme_rdma_ctrl_mutex);
1971
1972         return &ctrl->ctrl;
1973
1974 out_uninit_ctrl:
1975         nvme_uninit_ctrl(&ctrl->ctrl);
1976         nvme_put_ctrl(&ctrl->ctrl);
1977         if (ret > 0)
1978                 ret = -EIO;
1979         return ERR_PTR(ret);
1980 out_kfree_queues:
1981         kfree(ctrl->queues);
1982 out_free_ctrl:
1983         kfree(ctrl);
1984         return ERR_PTR(ret);
1985 }
1986
1987 static struct nvmf_transport_ops nvme_rdma_transport = {
1988         .name           = "rdma",
1989         .module         = THIS_MODULE,
1990         .required_opts  = NVMF_OPT_TRADDR,
1991         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1992                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1993         .create_ctrl    = nvme_rdma_create_ctrl,
1994 };
1995
1996 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1997 {
1998         struct nvme_rdma_ctrl *ctrl;
1999         struct nvme_rdma_device *ndev;
2000         bool found = false;
2001
2002         mutex_lock(&device_list_mutex);
2003         list_for_each_entry(ndev, &device_list, entry) {
2004                 if (ndev->dev == ib_device) {
2005                         found = true;
2006                         break;
2007                 }
2008         }
2009         mutex_unlock(&device_list_mutex);
2010
2011         if (!found)
2012                 return;
2013
2014         /* Delete all controllers using this device */
2015         mutex_lock(&nvme_rdma_ctrl_mutex);
2016         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2017                 if (ctrl->device->dev != ib_device)
2018                         continue;
2019                 nvme_delete_ctrl(&ctrl->ctrl);
2020         }
2021         mutex_unlock(&nvme_rdma_ctrl_mutex);
2022
2023         flush_workqueue(nvme_delete_wq);
2024 }
2025
2026 static struct ib_client nvme_rdma_ib_client = {
2027         .name   = "nvme_rdma",
2028         .remove = nvme_rdma_remove_one
2029 };
2030
2031 static int __init nvme_rdma_init_module(void)
2032 {
2033         int ret;
2034
2035         ret = ib_register_client(&nvme_rdma_ib_client);
2036         if (ret)
2037                 return ret;
2038
2039         ret = nvmf_register_transport(&nvme_rdma_transport);
2040         if (ret)
2041                 goto err_unreg_client;
2042
2043         return 0;
2044
2045 err_unreg_client:
2046         ib_unregister_client(&nvme_rdma_ib_client);
2047         return ret;
2048 }
2049
2050 static void __exit nvme_rdma_cleanup_module(void)
2051 {
2052         nvmf_unregister_transport(&nvme_rdma_transport);
2053         ib_unregister_client(&nvme_rdma_ib_client);
2054 }
2055
2056 module_init(nvme_rdma_init_module);
2057 module_exit(nvme_rdma_cleanup_module);
2058
2059 MODULE_LICENSE("GPL v2");