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   1
  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 };
  51
  52 struct nvme_rdma_qe {
  53         struct ib_cqe           cqe;
  54         void                    *data;
  55         u64                     dma;
  56 };
  57
  58 struct nvme_rdma_queue;
  59 struct nvme_rdma_request {
  60         struct nvme_request     req;
  61         struct ib_mr            *mr;
  62         struct nvme_rdma_qe     sqe;
  63         union nvme_result       result;
  64         __le16                  status;
  65         refcount_t              ref;
  66         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  67         u32                     num_sge;
  68         int                     nents;
  69         struct ib_reg_wr        reg_wr;
  70         struct ib_cqe           reg_cqe;
  71         struct nvme_rdma_queue  *queue;
  72         struct sg_table         sg_table;
  73         struct scatterlist      first_sgl[];
  74 };
  75
  76 enum nvme_rdma_queue_flags {
  77         NVME_RDMA_Q_ALLOCATED           = 0,
  78         NVME_RDMA_Q_LIVE                = 1,
  79         NVME_RDMA_Q_TR_READY            = 2,
  80 };
  81
  82 struct nvme_rdma_queue {
  83         struct nvme_rdma_qe     *rsp_ring;
  84         int                     queue_size;
  85         size_t                  cmnd_capsule_len;
  86         struct nvme_rdma_ctrl   *ctrl;
  87         struct nvme_rdma_device *device;
  88         struct ib_cq            *ib_cq;
  89         struct ib_qp            *qp;
  90
  91         unsigned long           flags;
  92         struct rdma_cm_id       *cm_id;
  93         int                     cm_error;
  94         struct completion       cm_done;
  95 };
  96
  97 struct nvme_rdma_ctrl {
  98         /* read only in the hot path */
  99         struct nvme_rdma_queue  *queues;
 100
 101         /* other member variables */
 102         struct blk_mq_tag_set   tag_set;
 103         struct work_struct      err_work;
 104
 105         struct nvme_rdma_qe     async_event_sqe;
 106
 107         struct delayed_work     reconnect_work;
 108
 109         struct list_head        list;
 110
 111         struct blk_mq_tag_set   admin_tag_set;
 112         struct nvme_rdma_device *device;
 113
 114         u32                     max_fr_pages;
 115
 116         struct sockaddr_storage addr;
 117         struct sockaddr_storage src_addr;
 118
 119         struct nvme_ctrl        ctrl;
 120 };
 121
 122 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 123 {
 124         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 125 }
 126
 127 static LIST_HEAD(device_list);
 128 static DEFINE_MUTEX(device_list_mutex);
 129
 130 static LIST_HEAD(nvme_rdma_ctrl_list);
 131 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 132
 133 /*
 134  * Disabling this option makes small I/O goes faster, but is fundamentally
 135  * unsafe.  With it turned off we will have to register a global rkey that
 136  * allows read and write access to all physical memory.
 137  */
 138 static bool register_always = true;
 139 module_param(register_always, bool, 0444);
 140 MODULE_PARM_DESC(register_always,
 141          "Use memory registration even for contiguous memory regions");
 142
 143 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 144                 struct rdma_cm_event *event);
 145 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 146
 147 static const struct blk_mq_ops nvme_rdma_mq_ops;
 148 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 149
 150 /* XXX: really should move to a generic header sooner or later.. */
 151 static inline void put_unaligned_le24(u32 val, u8 *p)
 152 {
 153         *p++ = val;
 154         *p++ = val >> 8;
 155         *p++ = val >> 16;
 156 }
 157
 158 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 159 {
 160         return queue - queue->ctrl->queues;
 161 }
 162
 163 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 164 {
 165         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 166 }
 167
 168 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 169                 size_t capsule_size, enum dma_data_direction dir)
 170 {
 171         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 172         kfree(qe->data);
 173 }
 174
 175 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 176                 size_t capsule_size, enum dma_data_direction dir)
 177 {
 178         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 179         if (!qe->data)
 180                 return -ENOMEM;
 181
 182         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 183         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 184                 kfree(qe->data);
 185                 return -ENOMEM;
 186         }
 187
 188         return 0;
 189 }
 190
 191 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 192                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 193                 size_t capsule_size, enum dma_data_direction dir)
 194 {
 195         int i;
 196
 197         for (i = 0; i < ib_queue_size; i++)
 198                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 199         kfree(ring);
 200 }
 201
 202 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 203                 size_t ib_queue_size, size_t capsule_size,
 204                 enum dma_data_direction dir)
 205 {
 206         struct nvme_rdma_qe *ring;
 207         int i;
 208
 209         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 210         if (!ring)
 211                 return NULL;
 212
 213         for (i = 0; i < ib_queue_size; i++) {
 214                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 215                         goto out_free_ring;
 216         }
 217
 218         return ring;
 219
 220 out_free_ring:
 221         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 222         return NULL;
 223 }
 224
 225 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 226 {
 227         pr_debug("QP event %s (%d)\n",
 228                  ib_event_msg(event->event), event->event);
 229
 230 }
 231
 232 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 233 {
 234         wait_for_completion_interruptible_timeout(&queue->cm_done,
 235                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 236         return queue->cm_error;
 237 }
 238
 239 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 240 {
 241         struct nvme_rdma_device *dev = queue->device;
 242         struct ib_qp_init_attr init_attr;
 243         int ret;
 244
 245         memset(&init_attr, 0, sizeof(init_attr));
 246         init_attr.event_handler = nvme_rdma_qp_event;
 247         /* +1 for drain */
 248         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 249         /* +1 for drain */
 250         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 251         init_attr.cap.max_recv_sge = 1;
 252         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 253         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 254         init_attr.qp_type = IB_QPT_RC;
 255         init_attr.send_cq = queue->ib_cq;
 256         init_attr.recv_cq = queue->ib_cq;
 257
 258         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 259
 260         queue->qp = queue->cm_id->qp;
 261         return ret;
 262 }
 263
 264 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 265                 struct request *rq, unsigned int hctx_idx)
 266 {
 267         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 268         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 269         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 270         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 271         struct nvme_rdma_device *dev = queue->device;
 272
 273         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 274                         DMA_TO_DEVICE);
 275 }
 276
 277 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 278                 struct request *rq, unsigned int hctx_idx,
 279                 unsigned int numa_node)
 280 {
 281         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 282         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 283         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 284         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 285         struct nvme_rdma_device *dev = queue->device;
 286         struct ib_device *ibdev = dev->dev;
 287         int ret;
 288
 289         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 290                         DMA_TO_DEVICE);
 291         if (ret)
 292                 return ret;
 293
 294         req->queue = queue;
 295
 296         return 0;
 297 }
 298
 299 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 300                 unsigned int hctx_idx)
 301 {
 302         struct nvme_rdma_ctrl *ctrl = data;
 303         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 304
 305         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 306
 307         hctx->driver_data = queue;
 308         return 0;
 309 }
 310
 311 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 312                 unsigned int hctx_idx)
 313 {
 314         struct nvme_rdma_ctrl *ctrl = data;
 315         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 316
 317         BUG_ON(hctx_idx != 0);
 318
 319         hctx->driver_data = queue;
 320         return 0;
 321 }
 322
 323 static void nvme_rdma_free_dev(struct kref *ref)
 324 {
 325         struct nvme_rdma_device *ndev =
 326                 container_of(ref, struct nvme_rdma_device, ref);
 327
 328         mutex_lock(&device_list_mutex);
 329         list_del(&ndev->entry);
 330         mutex_unlock(&device_list_mutex);
 331
 332         ib_dealloc_pd(ndev->pd);
 333         kfree(ndev);
 334 }
 335
 336 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 337 {
 338         kref_put(&dev->ref, nvme_rdma_free_dev);
 339 }
 340
 341 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 342 {
 343         return kref_get_unless_zero(&dev->ref);
 344 }
 345
 346 static struct nvme_rdma_device *
 347 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 348 {
 349         struct nvme_rdma_device *ndev;
 350
 351         mutex_lock(&device_list_mutex);
 352         list_for_each_entry(ndev, &device_list, entry) {
 353                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 354                     nvme_rdma_dev_get(ndev))
 355                         goto out_unlock;
 356         }
 357
 358         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 359         if (!ndev)
 360                 goto out_err;
 361
 362         ndev->dev = cm_id->device;
 363         kref_init(&ndev->ref);
 364
 365         ndev->pd = ib_alloc_pd(ndev->dev,
 366                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 367         if (IS_ERR(ndev->pd))
 368                 goto out_free_dev;
 369
 370         if (!(ndev->dev->attrs.device_cap_flags &
 371               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 372                 dev_err(&ndev->dev->dev,
 373                         "Memory registrations not supported.\n");
 374                 goto out_free_pd;
 375         }
 376
 377         list_add(&ndev->entry, &device_list);
 378 out_unlock:
 379         mutex_unlock(&device_list_mutex);
 380         return ndev;
 381
 382 out_free_pd:
 383         ib_dealloc_pd(ndev->pd);
 384 out_free_dev:
 385         kfree(ndev);
 386 out_err:
 387         mutex_unlock(&device_list_mutex);
 388         return NULL;
 389 }
 390
 391 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 392 {
 393         struct nvme_rdma_device *dev;
 394         struct ib_device *ibdev;
 395
 396         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 397                 return;
 398
 399         dev = queue->device;
 400         ibdev = dev->dev;
 401
 402         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 403
 404         /*
 405          * The cm_id object might have been destroyed during RDMA connection
 406          * establishment error flow to avoid getting other cma events, thus
 407          * the destruction of the QP shouldn't use rdma_cm API.
 408          */
 409         ib_destroy_qp(queue->qp);
 410         ib_free_cq(queue->ib_cq);
 411
 412         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 413                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 414
 415         nvme_rdma_dev_put(dev);
 416 }
 417
 418 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
 419 {
 420         return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
 421                      ibdev->attrs.max_fast_reg_page_list_len);
 422 }
 423
 424 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 425 {
 426         struct ib_device *ibdev;
 427         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 428         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 429         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 430         int ret;
 431
 432         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 433         if (!queue->device) {
 434                 dev_err(queue->cm_id->device->dev.parent,
 435                         "no client data found!\n");
 436                 return -ECONNREFUSED;
 437         }
 438         ibdev = queue->device->dev;
 439
 440         /*
 441          * Spread I/O queues completion vectors according their queue index.
 442          * Admin queues can always go on completion vector 0.
 443          */
 444         comp_vector = idx == 0 ? idx : idx - 1;
 445
 446         /* +1 for ib_stop_cq */
 447         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 448                                 cq_factor * queue->queue_size + 1,
 449                                 comp_vector, IB_POLL_SOFTIRQ);
 450         if (IS_ERR(queue->ib_cq)) {
 451                 ret = PTR_ERR(queue->ib_cq);
 452                 goto out_put_dev;
 453         }
 454
 455         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 456         if (ret)
 457                 goto out_destroy_ib_cq;
 458
 459         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 460                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 461         if (!queue->rsp_ring) {
 462                 ret = -ENOMEM;
 463                 goto out_destroy_qp;
 464         }
 465
 466         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 467                               queue->queue_size,
 468                               IB_MR_TYPE_MEM_REG,
 469                               nvme_rdma_get_max_fr_pages(ibdev));
 470         if (ret) {
 471                 dev_err(queue->ctrl->ctrl.device,
 472                         "failed to initialize MR pool sized %d for QID %d\n",
 473                         queue->queue_size, idx);
 474                 goto out_destroy_ring;
 475         }
 476
 477         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 478
 479         return 0;
 480
 481 out_destroy_ring:
 482         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 483                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 484 out_destroy_qp:
 485         rdma_destroy_qp(queue->cm_id);
 486 out_destroy_ib_cq:
 487         ib_free_cq(queue->ib_cq);
 488 out_put_dev:
 489         nvme_rdma_dev_put(queue->device);
 490         return ret;
 491 }
 492
 493 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 494                 int idx, size_t queue_size)
 495 {
 496         struct nvme_rdma_queue *queue;
 497         struct sockaddr *src_addr = NULL;
 498         int ret;
 499
 500         queue = &ctrl->queues[idx];
 501         queue->ctrl = ctrl;
 502         init_completion(&queue->cm_done);
 503
 504         if (idx > 0)
 505                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 506         else
 507                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 508
 509         queue->queue_size = queue_size;
 510
 511         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 512                         RDMA_PS_TCP, IB_QPT_RC);
 513         if (IS_ERR(queue->cm_id)) {
 514                 dev_info(ctrl->ctrl.device,
 515                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 516                 return PTR_ERR(queue->cm_id);
 517         }
 518
 519         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 520                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 521
 522         queue->cm_error = -ETIMEDOUT;
 523         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 524                         (struct sockaddr *)&ctrl->addr,
 525                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 526         if (ret) {
 527                 dev_info(ctrl->ctrl.device,
 528                         "rdma_resolve_addr failed (%d).\n", ret);
 529                 goto out_destroy_cm_id;
 530         }
 531
 532         ret = nvme_rdma_wait_for_cm(queue);
 533         if (ret) {
 534                 dev_info(ctrl->ctrl.device,
 535                         "rdma connection establishment failed (%d)\n", ret);
 536                 goto out_destroy_cm_id;
 537         }
 538
 539         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 540
 541         return 0;
 542
 543 out_destroy_cm_id:
 544         rdma_destroy_id(queue->cm_id);
 545         nvme_rdma_destroy_queue_ib(queue);
 546         return ret;
 547 }
 548
 549 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 550 {
 551         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 552                 return;
 553
 554         rdma_disconnect(queue->cm_id);
 555         ib_drain_qp(queue->qp);
 556 }
 557
 558 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 559 {
 560         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 561                 return;
 562
 563         if (nvme_rdma_queue_idx(queue) == 0) {
 564                 nvme_rdma_free_qe(queue->device->dev,
 565                         &queue->ctrl->async_event_sqe,
 566                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 567         }
 568
 569         nvme_rdma_destroy_queue_ib(queue);
 570         rdma_destroy_id(queue->cm_id);
 571 }
 572
 573 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 574 {
 575         int i;
 576
 577         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 578                 nvme_rdma_free_queue(&ctrl->queues[i]);
 579 }
 580
 581 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 582 {
 583         int i;
 584
 585         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 586                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 587 }
 588
 589 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 590 {
 591         int ret;
 592
 593         if (idx)
 594                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 595         else
 596                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 597
 598         if (!ret)
 599                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
 600         else
 601                 dev_info(ctrl->ctrl.device,
 602                         "failed to connect queue: %d ret=%d\n", idx, ret);
 603         return ret;
 604 }
 605
 606 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 607 {
 608         int i, ret = 0;
 609
 610         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 611                 ret = nvme_rdma_start_queue(ctrl, i);
 612                 if (ret)
 613                         goto out_stop_queues;
 614         }
 615
 616         return 0;
 617
 618 out_stop_queues:
 619         for (i--; i >= 1; i--)
 620                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 621         return ret;
 622 }
 623
 624 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 625 {
 626         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 627         struct ib_device *ibdev = ctrl->device->dev;
 628         unsigned int nr_io_queues;
 629         int i, ret;
 630
 631         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 632
 633         /*
 634          * we map queues according to the device irq vectors for
 635          * optimal locality so we don't need more queues than
 636          * completion vectors.
 637          */
 638         nr_io_queues = min_t(unsigned int, nr_io_queues,
 639                                 ibdev->num_comp_vectors);
 640
 641         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 642         if (ret)
 643                 return ret;
 644
 645         ctrl->ctrl.queue_count = nr_io_queues + 1;
 646         if (ctrl->ctrl.queue_count < 2)
 647                 return 0;
 648
 649         dev_info(ctrl->ctrl.device,
 650                 "creating %d I/O queues.\n", nr_io_queues);
 651
 652         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 653                 ret = nvme_rdma_alloc_queue(ctrl, i,
 654                                 ctrl->ctrl.sqsize + 1);
 655                 if (ret)
 656                         goto out_free_queues;
 657         }
 658
 659         return 0;
 660
 661 out_free_queues:
 662         for (i--; i >= 1; i--)
 663                 nvme_rdma_free_queue(&ctrl->queues[i]);
 664
 665         return ret;
 666 }
 667
 668 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
 669                 struct blk_mq_tag_set *set)
 670 {
 671         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 672
 673         blk_mq_free_tag_set(set);
 674         nvme_rdma_dev_put(ctrl->device);
 675 }
 676
 677 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 678                 bool admin)
 679 {
 680         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 681         struct blk_mq_tag_set *set;
 682         int ret;
 683
 684         if (admin) {
 685                 set = &ctrl->admin_tag_set;
 686                 memset(set, 0, sizeof(*set));
 687                 set->ops = &nvme_rdma_admin_mq_ops;
 688                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 689                 set->reserved_tags = 2; /* connect + keep-alive */
 690                 set->numa_node = NUMA_NO_NODE;
 691                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 692                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 693                 set->driver_data = ctrl;
 694                 set->nr_hw_queues = 1;
 695                 set->timeout = ADMIN_TIMEOUT;
 696                 set->flags = BLK_MQ_F_NO_SCHED;
 697         } else {
 698                 set = &ctrl->tag_set;
 699                 memset(set, 0, sizeof(*set));
 700                 set->ops = &nvme_rdma_mq_ops;
 701                 set->queue_depth = nctrl->opts->queue_size;
 702                 set->reserved_tags = 1; /* fabric connect */
 703                 set->numa_node = NUMA_NO_NODE;
 704                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 705                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 706                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 707                 set->driver_data = ctrl;
 708                 set->nr_hw_queues = nctrl->queue_count - 1;
 709                 set->timeout = NVME_IO_TIMEOUT;
 710         }
 711
 712         ret = blk_mq_alloc_tag_set(set);
 713         if (ret)
 714                 goto out;
 715
 716         /*
 717          * We need a reference on the device as long as the tag_set is alive,
 718          * as the MRs in the request structures need a valid ib_device.
 719          */
 720         ret = nvme_rdma_dev_get(ctrl->device);
 721         if (!ret) {
 722                 ret = -EINVAL;
 723                 goto out_free_tagset;
 724         }
 725
 726         return set;
 727
 728 out_free_tagset:
 729         blk_mq_free_tag_set(set);
 730 out:
 731         return ERR_PTR(ret);
 732 }
 733
 734 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 735                 bool remove)
 736 {
 737         nvme_rdma_stop_queue(&ctrl->queues[0]);
 738         if (remove) {
 739                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 740                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 741         }
 742         nvme_rdma_free_queue(&ctrl->queues[0]);
 743 }
 744
 745 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 746                 bool new)
 747 {
 748         int error;
 749
 750         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 751         if (error)
 752                 return error;
 753
 754         ctrl->device = ctrl->queues[0].device;
 755
 756         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
 757
 758         if (new) {
 759                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 760                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 761                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 762                         goto out_free_queue;
 763                 }
 764
 765                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 766                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 767                         error = PTR_ERR(ctrl->ctrl.admin_q);
 768                         goto out_free_tagset;
 769                 }
 770         }
 771
 772         error = nvme_rdma_start_queue(ctrl, 0);
 773         if (error)
 774                 goto out_cleanup_queue;
 775
 776         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 777                         &ctrl->ctrl.cap);
 778         if (error) {
 779                 dev_err(ctrl->ctrl.device,
 780                         "prop_get NVME_REG_CAP failed\n");
 781                 goto out_cleanup_queue;
 782         }
 783
 784         ctrl->ctrl.sqsize =
 785                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 786
 787         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 788         if (error)
 789                 goto out_cleanup_queue;
 790
 791         ctrl->ctrl.max_hw_sectors =
 792                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 793
 794         error = nvme_init_identify(&ctrl->ctrl);
 795         if (error)
 796                 goto out_cleanup_queue;
 797
 798         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
 799                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
 800                         DMA_TO_DEVICE);
 801         if (error)
 802                 goto out_cleanup_queue;
 803
 804         return 0;
 805
 806 out_cleanup_queue:
 807         if (new)
 808                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 809 out_free_tagset:
 810         if (new)
 811                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 812 out_free_queue:
 813         nvme_rdma_free_queue(&ctrl->queues[0]);
 814         return error;
 815 }
 816
 817 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 818                 bool remove)
 819 {
 820         nvme_rdma_stop_io_queues(ctrl);
 821         if (remove) {
 822                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 823                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 824         }
 825         nvme_rdma_free_io_queues(ctrl);
 826 }
 827
 828 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 829 {
 830         int ret;
 831
 832         ret = nvme_rdma_alloc_io_queues(ctrl);
 833         if (ret)
 834                 return ret;
 835
 836         if (new) {
 837                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 838                 if (IS_ERR(ctrl->ctrl.tagset)) {
 839                         ret = PTR_ERR(ctrl->ctrl.tagset);
 840                         goto out_free_io_queues;
 841                 }
 842
 843                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 844                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 845                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 846                         goto out_free_tag_set;
 847                 }
 848         } else {
 849                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 850                         ctrl->ctrl.queue_count - 1);
 851         }
 852
 853         ret = nvme_rdma_start_io_queues(ctrl);
 854         if (ret)
 855                 goto out_cleanup_connect_q;
 856
 857         return 0;
 858
 859 out_cleanup_connect_q:
 860         if (new)
 861                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 862 out_free_tag_set:
 863         if (new)
 864                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 865 out_free_io_queues:
 866         nvme_rdma_free_io_queues(ctrl);
 867         return ret;
 868 }
 869
 870 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
 871 {
 872         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 873
 874         cancel_work_sync(&ctrl->err_work);
 875         cancel_delayed_work_sync(&ctrl->reconnect_work);
 876 }
 877
 878 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 879 {
 880         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 881
 882         if (list_empty(&ctrl->list))
 883                 goto free_ctrl;
 884
 885         mutex_lock(&nvme_rdma_ctrl_mutex);
 886         list_del(&ctrl->list);
 887         mutex_unlock(&nvme_rdma_ctrl_mutex);
 888
 889         kfree(ctrl->queues);
 890         nvmf_free_options(nctrl->opts);
 891 free_ctrl:
 892         kfree(ctrl);
 893 }
 894
 895 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 896 {
 897         /* If we are resetting/deleting then do nothing */
 898         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 899                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 900                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 901                 return;
 902         }
 903
 904         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 905                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 906                         ctrl->ctrl.opts->reconnect_delay);
 907                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 908                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 909         } else {
 910                 nvme_delete_ctrl(&ctrl->ctrl);
 911         }
 912 }
 913
 914 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 915 {
 916         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 917                         struct nvme_rdma_ctrl, reconnect_work);
 918         bool changed;
 919         int ret;
 920
 921         ++ctrl->ctrl.nr_reconnects;
 922
 923         ret = nvme_rdma_configure_admin_queue(ctrl, false);
 924         if (ret)
 925                 goto requeue;
 926
 927         if (ctrl->ctrl.queue_count > 1) {
 928                 ret = nvme_rdma_configure_io_queues(ctrl, false);
 929                 if (ret)
 930                         goto destroy_admin;
 931         }
 932
 933         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 934         if (!changed) {
 935                 /* state change failure is ok if we're in DELETING state */
 936                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
 937                 return;
 938         }
 939
 940         nvme_start_ctrl(&ctrl->ctrl);
 941
 942         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
 943                         ctrl->ctrl.nr_reconnects);
 944
 945         ctrl->ctrl.nr_reconnects = 0;
 946
 947         return;
 948
 949 destroy_admin:
 950         nvme_rdma_destroy_admin_queue(ctrl, false);
 951 requeue:
 952         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
 953                         ctrl->ctrl.nr_reconnects);
 954         nvme_rdma_reconnect_or_remove(ctrl);
 955 }
 956
 957 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 958 {
 959         struct nvme_rdma_ctrl *ctrl = container_of(work,
 960                         struct nvme_rdma_ctrl, err_work);
 961
 962         nvme_stop_keep_alive(&ctrl->ctrl);
 963
 964         if (ctrl->ctrl.queue_count > 1) {
 965                 nvme_stop_queues(&ctrl->ctrl);
 966                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 967                                         nvme_cancel_request, &ctrl->ctrl);
 968                 nvme_rdma_destroy_io_queues(ctrl, false);
 969         }
 970
 971         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 972         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 973                                 nvme_cancel_request, &ctrl->ctrl);
 974         nvme_rdma_destroy_admin_queue(ctrl, false);
 975
 976         /*
 977          * queues are not a live anymore, so restart the queues to fail fast
 978          * new IO
 979          */
 980         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 981         nvme_start_queues(&ctrl->ctrl);
 982
 983         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
 984                 /* state change failure is ok if we're in DELETING state */
 985                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
 986                 return;
 987         }
 988
 989         nvme_rdma_reconnect_or_remove(ctrl);
 990 }
 991
 992 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 993 {
 994         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
 995                 return;
 996
 997         queue_work(nvme_wq, &ctrl->err_work);
 998 }
 999
1000 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1001                 const char *op)
1002 {
1003         struct nvme_rdma_queue *queue = cq->cq_context;
1004         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1005
1006         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1007                 dev_info(ctrl->ctrl.device,
1008                              "%s for CQE 0x%p failed with status %s (%d)\n",
1009                              op, wc->wr_cqe,
1010                              ib_wc_status_msg(wc->status), wc->status);
1011         nvme_rdma_error_recovery(ctrl);
1012 }
1013
1014 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1015 {
1016         if (unlikely(wc->status != IB_WC_SUCCESS))
1017                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1018 }
1019
1020 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1021 {
1022         struct nvme_rdma_request *req =
1023                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1024         struct request *rq = blk_mq_rq_from_pdu(req);
1025
1026         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1027                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1028                 return;
1029         }
1030
1031         if (refcount_dec_and_test(&req->ref))
1032                 nvme_end_request(rq, req->status, req->result);
1033
1034 }
1035
1036 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1037                 struct nvme_rdma_request *req)
1038 {
1039         struct ib_send_wr *bad_wr;
1040         struct ib_send_wr wr = {
1041                 .opcode             = IB_WR_LOCAL_INV,
1042                 .next               = NULL,
1043                 .num_sge            = 0,
1044                 .send_flags         = IB_SEND_SIGNALED,
1045                 .ex.invalidate_rkey = req->mr->rkey,
1046         };
1047
1048         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1049         wr.wr_cqe = &req->reg_cqe;
1050
1051         return ib_post_send(queue->qp, &wr, &bad_wr);
1052 }
1053
1054 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1055                 struct request *rq)
1056 {
1057         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1058         struct nvme_rdma_device *dev = queue->device;
1059         struct ib_device *ibdev = dev->dev;
1060
1061         if (!blk_rq_payload_bytes(rq))
1062                 return;
1063
1064         if (req->mr) {
1065                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1066                 req->mr = NULL;
1067         }
1068
1069         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1070                         req->nents, rq_data_dir(rq) ==
1071                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1072
1073         nvme_cleanup_cmd(rq);
1074         sg_free_table_chained(&req->sg_table, true);
1075 }
1076
1077 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1078 {
1079         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1080
1081         sg->addr = 0;
1082         put_unaligned_le24(0, sg->length);
1083         put_unaligned_le32(0, sg->key);
1084         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1085         return 0;
1086 }
1087
1088 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1089                 struct nvme_rdma_request *req, struct nvme_command *c)
1090 {
1091         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1092
1093         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1094         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1095         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1096
1097         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1098         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1099         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1100
1101         req->num_sge++;
1102         return 0;
1103 }
1104
1105 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1106                 struct nvme_rdma_request *req, struct nvme_command *c)
1107 {
1108         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1109
1110         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1111         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1112         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1113         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1114         return 0;
1115 }
1116
1117 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1118                 struct nvme_rdma_request *req, struct nvme_command *c,
1119                 int count)
1120 {
1121         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1122         int nr;
1123
1124         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1125         if (WARN_ON_ONCE(!req->mr))
1126                 return -EAGAIN;
1127
1128         /*
1129          * Align the MR to a 4K page size to match the ctrl page size and
1130          * the block virtual boundary.
1131          */
1132         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1133         if (unlikely(nr < count)) {
1134                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1135                 req->mr = NULL;
1136                 if (nr < 0)
1137                         return nr;
1138                 return -EINVAL;
1139         }
1140
1141         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1142
1143         req->reg_cqe.done = nvme_rdma_memreg_done;
1144         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1145         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1146         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1147         req->reg_wr.wr.num_sge = 0;
1148         req->reg_wr.mr = req->mr;
1149         req->reg_wr.key = req->mr->rkey;
1150         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1151                              IB_ACCESS_REMOTE_READ |
1152                              IB_ACCESS_REMOTE_WRITE;
1153
1154         sg->addr = cpu_to_le64(req->mr->iova);
1155         put_unaligned_le24(req->mr->length, sg->length);
1156         put_unaligned_le32(req->mr->rkey, sg->key);
1157         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1158                         NVME_SGL_FMT_INVALIDATE;
1159
1160         return 0;
1161 }
1162
1163 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1164                 struct request *rq, struct nvme_command *c)
1165 {
1166         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1167         struct nvme_rdma_device *dev = queue->device;
1168         struct ib_device *ibdev = dev->dev;
1169         int count, ret;
1170
1171         req->num_sge = 1;
1172         refcount_set(&req->ref, 2); /* send and recv completions */
1173
1174         c->common.flags |= NVME_CMD_SGL_METABUF;
1175
1176         if (!blk_rq_payload_bytes(rq))
1177                 return nvme_rdma_set_sg_null(c);
1178
1179         req->sg_table.sgl = req->first_sgl;
1180         ret = sg_alloc_table_chained(&req->sg_table,
1181                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1182         if (ret)
1183                 return -ENOMEM;
1184
1185         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1186
1187         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1188                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1189         if (unlikely(count <= 0)) {
1190                 sg_free_table_chained(&req->sg_table, true);
1191                 return -EIO;
1192         }
1193
1194         if (count == 1) {
1195                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1196                     blk_rq_payload_bytes(rq) <=
1197                                 nvme_rdma_inline_data_size(queue))
1198                         return nvme_rdma_map_sg_inline(queue, req, c);
1199
1200                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1201                         return nvme_rdma_map_sg_single(queue, req, c);
1202         }
1203
1204         return nvme_rdma_map_sg_fr(queue, req, c, count);
1205 }
1206
1207 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1208 {
1209         struct nvme_rdma_qe *qe =
1210                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1211         struct nvme_rdma_request *req =
1212                 container_of(qe, struct nvme_rdma_request, sqe);
1213         struct request *rq = blk_mq_rq_from_pdu(req);
1214
1215         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1216                 nvme_rdma_wr_error(cq, wc, "SEND");
1217                 return;
1218         }
1219
1220         if (refcount_dec_and_test(&req->ref))
1221                 nvme_end_request(rq, req->status, req->result);
1222 }
1223
1224 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1225                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1226                 struct ib_send_wr *first)
1227 {
1228         struct ib_send_wr wr, *bad_wr;
1229         int ret;
1230
1231         sge->addr   = qe->dma;
1232         sge->length = sizeof(struct nvme_command),
1233         sge->lkey   = queue->device->pd->local_dma_lkey;
1234
1235         wr.next       = NULL;
1236         wr.wr_cqe     = &qe->cqe;
1237         wr.sg_list    = sge;
1238         wr.num_sge    = num_sge;
1239         wr.opcode     = IB_WR_SEND;
1240         wr.send_flags = IB_SEND_SIGNALED;
1241
1242         if (first)
1243                 first->next = &wr;
1244         else
1245                 first = &wr;
1246
1247         ret = ib_post_send(queue->qp, first, &bad_wr);
1248         if (unlikely(ret)) {
1249                 dev_err(queue->ctrl->ctrl.device,
1250                              "%s failed with error code %d\n", __func__, ret);
1251         }
1252         return ret;
1253 }
1254
1255 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1256                 struct nvme_rdma_qe *qe)
1257 {
1258         struct ib_recv_wr wr, *bad_wr;
1259         struct ib_sge list;
1260         int ret;
1261
1262         list.addr   = qe->dma;
1263         list.length = sizeof(struct nvme_completion);
1264         list.lkey   = queue->device->pd->local_dma_lkey;
1265
1266         qe->cqe.done = nvme_rdma_recv_done;
1267
1268         wr.next     = NULL;
1269         wr.wr_cqe   = &qe->cqe;
1270         wr.sg_list  = &list;
1271         wr.num_sge  = 1;
1272
1273         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1274         if (unlikely(ret)) {
1275                 dev_err(queue->ctrl->ctrl.device,
1276                         "%s failed with error code %d\n", __func__, ret);
1277         }
1278         return ret;
1279 }
1280
1281 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1282 {
1283         u32 queue_idx = nvme_rdma_queue_idx(queue);
1284
1285         if (queue_idx == 0)
1286                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1287         return queue->ctrl->tag_set.tags[queue_idx - 1];
1288 }
1289
1290 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1291 {
1292         if (unlikely(wc->status != IB_WC_SUCCESS))
1293                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1294 }
1295
1296 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1297 {
1298         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1299         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1300         struct ib_device *dev = queue->device->dev;
1301         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1302         struct nvme_command *cmd = sqe->data;
1303         struct ib_sge sge;
1304         int ret;
1305
1306         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1307
1308         memset(cmd, 0, sizeof(*cmd));
1309         cmd->common.opcode = nvme_admin_async_event;
1310         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1311         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1312         nvme_rdma_set_sg_null(cmd);
1313
1314         sqe->cqe.done = nvme_rdma_async_done;
1315
1316         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1317                         DMA_TO_DEVICE);
1318
1319         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1320         WARN_ON_ONCE(ret);
1321 }
1322
1323 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1324                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1325 {
1326         struct request *rq;
1327         struct nvme_rdma_request *req;
1328         int ret = 0;
1329
1330         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1331         if (!rq) {
1332                 dev_err(queue->ctrl->ctrl.device,
1333                         "tag 0x%x on QP %#x not found\n",
1334                         cqe->command_id, queue->qp->qp_num);
1335                 nvme_rdma_error_recovery(queue->ctrl);
1336                 return ret;
1337         }
1338         req = blk_mq_rq_to_pdu(rq);
1339
1340         req->status = cqe->status;
1341         req->result = cqe->result;
1342
1343         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1344                 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1345                         dev_err(queue->ctrl->ctrl.device,
1346                                 "Bogus remote invalidation for rkey %#x\n",
1347                                 req->mr->rkey);
1348                         nvme_rdma_error_recovery(queue->ctrl);
1349                 }
1350         } else if (req->mr) {
1351                 ret = nvme_rdma_inv_rkey(queue, req);
1352                 if (unlikely(ret < 0)) {
1353                         dev_err(queue->ctrl->ctrl.device,
1354                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1355                                 req->mr->rkey, ret);
1356                         nvme_rdma_error_recovery(queue->ctrl);
1357                 }
1358                 /* the local invalidation completion will end the request */
1359                 return 0;
1360         }
1361
1362         if (refcount_dec_and_test(&req->ref)) {
1363                 if (rq->tag == tag)
1364                         ret = 1;
1365                 nvme_end_request(rq, req->status, req->result);
1366         }
1367
1368         return ret;
1369 }
1370
1371 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1372 {
1373         struct nvme_rdma_qe *qe =
1374                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1375         struct nvme_rdma_queue *queue = cq->cq_context;
1376         struct ib_device *ibdev = queue->device->dev;
1377         struct nvme_completion *cqe = qe->data;
1378         const size_t len = sizeof(struct nvme_completion);
1379         int ret = 0;
1380
1381         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1382                 nvme_rdma_wr_error(cq, wc, "RECV");
1383                 return 0;
1384         }
1385
1386         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1387         /*
1388          * AEN requests are special as they don't time out and can
1389          * survive any kind of queue freeze and often don't respond to
1390          * aborts.  We don't even bother to allocate a struct request
1391          * for them but rather special case them here.
1392          */
1393         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1394                         cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1395                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1396                                 &cqe->result);
1397         else
1398                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1399         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1400
1401         nvme_rdma_post_recv(queue, qe);
1402         return ret;
1403 }
1404
1405 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1406 {
1407         __nvme_rdma_recv_done(cq, wc, -1);
1408 }
1409
1410 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1411 {
1412         int ret, i;
1413
1414         for (i = 0; i < queue->queue_size; i++) {
1415                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1416                 if (ret)
1417                         goto out_destroy_queue_ib;
1418         }
1419
1420         return 0;
1421
1422 out_destroy_queue_ib:
1423         nvme_rdma_destroy_queue_ib(queue);
1424         return ret;
1425 }
1426
1427 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1428                 struct rdma_cm_event *ev)
1429 {
1430         struct rdma_cm_id *cm_id = queue->cm_id;
1431         int status = ev->status;
1432         const char *rej_msg;
1433         const struct nvme_rdma_cm_rej *rej_data;
1434         u8 rej_data_len;
1435
1436         rej_msg = rdma_reject_msg(cm_id, status);
1437         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1438
1439         if (rej_data && rej_data_len >= sizeof(u16)) {
1440                 u16 sts = le16_to_cpu(rej_data->sts);
1441
1442                 dev_err(queue->ctrl->ctrl.device,
1443                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1444                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1445         } else {
1446                 dev_err(queue->ctrl->ctrl.device,
1447                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1448         }
1449
1450         return -ECONNRESET;
1451 }
1452
1453 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1454 {
1455         int ret;
1456
1457         ret = nvme_rdma_create_queue_ib(queue);
1458         if (ret)
1459                 return ret;
1460
1461         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1462         if (ret) {
1463                 dev_err(queue->ctrl->ctrl.device,
1464                         "rdma_resolve_route failed (%d).\n",
1465                         queue->cm_error);
1466                 goto out_destroy_queue;
1467         }
1468
1469         return 0;
1470
1471 out_destroy_queue:
1472         nvme_rdma_destroy_queue_ib(queue);
1473         return ret;
1474 }
1475
1476 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1477 {
1478         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1479         struct rdma_conn_param param = { };
1480         struct nvme_rdma_cm_req priv = { };
1481         int ret;
1482
1483         param.qp_num = queue->qp->qp_num;
1484         param.flow_control = 1;
1485
1486         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1487         /* maximum retry count */
1488         param.retry_count = 7;
1489         param.rnr_retry_count = 7;
1490         param.private_data = &priv;
1491         param.private_data_len = sizeof(priv);
1492
1493         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1494         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1495         /*
1496          * set the admin queue depth to the minimum size
1497          * specified by the Fabrics standard.
1498          */
1499         if (priv.qid == 0) {
1500                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1501                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1502         } else {
1503                 /*
1504                  * current interpretation of the fabrics spec
1505                  * is at minimum you make hrqsize sqsize+1, or a
1506                  * 1's based representation of sqsize.
1507                  */
1508                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1509                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1510         }
1511
1512         ret = rdma_connect(queue->cm_id, &param);
1513         if (ret) {
1514                 dev_err(ctrl->ctrl.device,
1515                         "rdma_connect failed (%d).\n", ret);
1516                 goto out_destroy_queue_ib;
1517         }
1518
1519         return 0;
1520
1521 out_destroy_queue_ib:
1522         nvme_rdma_destroy_queue_ib(queue);
1523         return ret;
1524 }
1525
1526 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1527                 struct rdma_cm_event *ev)
1528 {
1529         struct nvme_rdma_queue *queue = cm_id->context;
1530         int cm_error = 0;
1531
1532         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1533                 rdma_event_msg(ev->event), ev->event,
1534                 ev->status, cm_id);
1535
1536         switch (ev->event) {
1537         case RDMA_CM_EVENT_ADDR_RESOLVED:
1538                 cm_error = nvme_rdma_addr_resolved(queue);
1539                 break;
1540         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1541                 cm_error = nvme_rdma_route_resolved(queue);
1542                 break;
1543         case RDMA_CM_EVENT_ESTABLISHED:
1544                 queue->cm_error = nvme_rdma_conn_established(queue);
1545                 /* complete cm_done regardless of success/failure */
1546                 complete(&queue->cm_done);
1547                 return 0;
1548         case RDMA_CM_EVENT_REJECTED:
1549                 nvme_rdma_destroy_queue_ib(queue);
1550                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1551                 break;
1552         case RDMA_CM_EVENT_ROUTE_ERROR:
1553         case RDMA_CM_EVENT_CONNECT_ERROR:
1554         case RDMA_CM_EVENT_UNREACHABLE:
1555                 nvme_rdma_destroy_queue_ib(queue);
1556         case RDMA_CM_EVENT_ADDR_ERROR:
1557                 dev_dbg(queue->ctrl->ctrl.device,
1558                         "CM error event %d\n", ev->event);
1559                 cm_error = -ECONNRESET;
1560                 break;
1561         case RDMA_CM_EVENT_DISCONNECTED:
1562         case RDMA_CM_EVENT_ADDR_CHANGE:
1563         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1564                 dev_dbg(queue->ctrl->ctrl.device,
1565                         "disconnect received - connection closed\n");
1566                 nvme_rdma_error_recovery(queue->ctrl);
1567                 break;
1568         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1569                 /* device removal is handled via the ib_client API */
1570                 break;
1571         default:
1572                 dev_err(queue->ctrl->ctrl.device,
1573                         "Unexpected RDMA CM event (%d)\n", ev->event);
1574                 nvme_rdma_error_recovery(queue->ctrl);
1575                 break;
1576         }
1577
1578         if (cm_error) {
1579                 queue->cm_error = cm_error;
1580                 complete(&queue->cm_done);
1581         }
1582
1583         return 0;
1584 }
1585
1586 static enum blk_eh_timer_return
1587 nvme_rdma_timeout(struct request *rq, bool reserved)
1588 {
1589         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1590
1591         dev_warn(req->queue->ctrl->ctrl.device,
1592                  "I/O %d QID %d timeout, reset controller\n",
1593                  rq->tag, nvme_rdma_queue_idx(req->queue));
1594
1595         /* queue error recovery */
1596         nvme_rdma_error_recovery(req->queue->ctrl);
1597
1598         /* fail with DNR on cmd timeout */
1599         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1600
1601         return BLK_EH_HANDLED;
1602 }
1603
1604 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1605                 const struct blk_mq_queue_data *bd)
1606 {
1607         struct nvme_ns *ns = hctx->queue->queuedata;
1608         struct nvme_rdma_queue *queue = hctx->driver_data;
1609         struct request *rq = bd->rq;
1610         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1611         struct nvme_rdma_qe *sqe = &req->sqe;
1612         struct nvme_command *c = sqe->data;
1613         struct ib_device *dev;
1614         blk_status_t ret;
1615         int err;
1616
1617         WARN_ON_ONCE(rq->tag < 0);
1618
1619         ret = nvmf_check_if_ready(&queue->ctrl->ctrl, rq,
1620                 test_bit(NVME_RDMA_Q_LIVE, &queue->flags), true);
1621         if (unlikely(ret))
1622                 return ret;
1623
1624         dev = queue->device->dev;
1625         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1626                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1627
1628         ret = nvme_setup_cmd(ns, rq, c);
1629         if (ret)
1630                 return ret;
1631
1632         blk_mq_start_request(rq);
1633
1634         err = nvme_rdma_map_data(queue, rq, c);
1635         if (unlikely(err < 0)) {
1636                 dev_err(queue->ctrl->ctrl.device,
1637                              "Failed to map data (%d)\n", err);
1638                 nvme_cleanup_cmd(rq);
1639                 goto err;
1640         }
1641
1642         sqe->cqe.done = nvme_rdma_send_done;
1643
1644         ib_dma_sync_single_for_device(dev, sqe->dma,
1645                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1646
1647         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1648                         req->mr ? &req->reg_wr.wr : NULL);
1649         if (unlikely(err)) {
1650                 nvme_rdma_unmap_data(queue, rq);
1651                 goto err;
1652         }
1653
1654         return BLK_STS_OK;
1655 err:
1656         if (err == -ENOMEM || err == -EAGAIN)
1657                 return BLK_STS_RESOURCE;
1658         return BLK_STS_IOERR;
1659 }
1660
1661 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1662 {
1663         struct nvme_rdma_queue *queue = hctx->driver_data;
1664         struct ib_cq *cq = queue->ib_cq;
1665         struct ib_wc wc;
1666         int found = 0;
1667
1668         while (ib_poll_cq(cq, 1, &wc) > 0) {
1669                 struct ib_cqe *cqe = wc.wr_cqe;
1670
1671                 if (cqe) {
1672                         if (cqe->done == nvme_rdma_recv_done)
1673                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1674                         else
1675                                 cqe->done(cq, &wc);
1676                 }
1677         }
1678
1679         return found;
1680 }
1681
1682 static void nvme_rdma_complete_rq(struct request *rq)
1683 {
1684         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1685
1686         nvme_rdma_unmap_data(req->queue, rq);
1687         nvme_complete_rq(rq);
1688 }
1689
1690 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1691 {
1692         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1693
1694         return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1695 }
1696
1697 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1698         .queue_rq       = nvme_rdma_queue_rq,
1699         .complete       = nvme_rdma_complete_rq,
1700         .init_request   = nvme_rdma_init_request,
1701         .exit_request   = nvme_rdma_exit_request,
1702         .init_hctx      = nvme_rdma_init_hctx,
1703         .poll           = nvme_rdma_poll,
1704         .timeout        = nvme_rdma_timeout,
1705         .map_queues     = nvme_rdma_map_queues,
1706 };
1707
1708 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1709         .queue_rq       = nvme_rdma_queue_rq,
1710         .complete       = nvme_rdma_complete_rq,
1711         .init_request   = nvme_rdma_init_request,
1712         .exit_request   = nvme_rdma_exit_request,
1713         .init_hctx      = nvme_rdma_init_admin_hctx,
1714         .timeout        = nvme_rdma_timeout,
1715 };
1716
1717 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1718 {
1719         if (ctrl->ctrl.queue_count > 1) {
1720                 nvme_stop_queues(&ctrl->ctrl);
1721                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1722                                         nvme_cancel_request, &ctrl->ctrl);
1723                 nvme_rdma_destroy_io_queues(ctrl, shutdown);
1724         }
1725
1726         if (shutdown)
1727                 nvme_shutdown_ctrl(&ctrl->ctrl);
1728         else
1729                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1730
1731         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1732         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1733                                 nvme_cancel_request, &ctrl->ctrl);
1734         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1735         nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1736 }
1737
1738 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1739 {
1740         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1741 }
1742
1743 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1744 {
1745         struct nvme_rdma_ctrl *ctrl =
1746                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1747         int ret;
1748         bool changed;
1749
1750         nvme_stop_ctrl(&ctrl->ctrl);
1751         nvme_rdma_shutdown_ctrl(ctrl, false);
1752
1753         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1754                 /* state change failure should never happen */
1755                 WARN_ON_ONCE(1);
1756                 return;
1757         }
1758
1759         ret = nvme_rdma_configure_admin_queue(ctrl, false);
1760         if (ret)
1761                 goto out_fail;
1762
1763         if (ctrl->ctrl.queue_count > 1) {
1764                 ret = nvme_rdma_configure_io_queues(ctrl, false);
1765                 if (ret)
1766                         goto out_fail;
1767         }
1768
1769         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1770         if (!changed) {
1771                 /* state change failure is ok if we're in DELETING state */
1772                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1773                 return;
1774         }
1775
1776         nvme_start_ctrl(&ctrl->ctrl);
1777
1778         return;
1779
1780 out_fail:
1781         ++ctrl->ctrl.nr_reconnects;
1782         nvme_rdma_reconnect_or_remove(ctrl);
1783 }
1784
1785 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1786         .name                   = "rdma",
1787         .module                 = THIS_MODULE,
1788         .flags                  = NVME_F_FABRICS,
1789         .reg_read32             = nvmf_reg_read32,
1790         .reg_read64             = nvmf_reg_read64,
1791         .reg_write32            = nvmf_reg_write32,
1792         .free_ctrl              = nvme_rdma_free_ctrl,
1793         .submit_async_event     = nvme_rdma_submit_async_event,
1794         .delete_ctrl            = nvme_rdma_delete_ctrl,
1795         .get_address            = nvmf_get_address,
1796         .stop_ctrl              = nvme_rdma_stop_ctrl,
1797 };
1798
1799 static inline bool
1800 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1801         struct nvmf_ctrl_options *opts)
1802 {
1803         char *stdport = __stringify(NVME_RDMA_IP_PORT);
1804
1805
1806         if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1807             strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1808                 return false;
1809
1810         if (opts->mask & NVMF_OPT_TRSVCID &&
1811             ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1812                 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1813                         return false;
1814         } else if (opts->mask & NVMF_OPT_TRSVCID) {
1815                 if (strcmp(opts->trsvcid, stdport))
1816                         return false;
1817         } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1818                 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1819                         return false;
1820         }
1821         /* else, it's a match as both have stdport. Fall to next checks */
1822
1823         /*
1824          * checking the local address is rough. In most cases, one
1825          * is not specified and the host port is selected by the stack.
1826          *
1827          * Assume no match if:
1828          *  local address is specified and address is not the same
1829          *  local address is not specified but remote is, or vice versa
1830          *    (admin using specific host_traddr when it matters).
1831          */
1832         if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1833             ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1834                 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1835                         return false;
1836         } else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1837                    ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1838                 return false;
1839         /*
1840          * if neither controller had an host port specified, assume it's
1841          * a match as everything else matched.
1842          */
1843
1844         return true;
1845 }
1846
1847 /*
1848  * Fails a connection request if it matches an existing controller
1849  * (association) with the same tuple:
1850  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1851  *
1852  * if local address is not specified in the request, it will match an
1853  * existing controller with all the other parameters the same and no
1854  * local port address specified as well.
1855  *
1856  * The ports don't need to be compared as they are intrinsically
1857  * already matched by the port pointers supplied.
1858  */
1859 static bool
1860 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1861 {
1862         struct nvme_rdma_ctrl *ctrl;
1863         bool found = false;
1864
1865         mutex_lock(&nvme_rdma_ctrl_mutex);
1866         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1867                 found = __nvme_rdma_options_match(ctrl, opts);
1868                 if (found)
1869                         break;
1870         }
1871         mutex_unlock(&nvme_rdma_ctrl_mutex);
1872
1873         return found;
1874 }
1875
1876 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1877                 struct nvmf_ctrl_options *opts)
1878 {
1879         struct nvme_rdma_ctrl *ctrl;
1880         int ret;
1881         bool changed;
1882         char *port;
1883
1884         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1885         if (!ctrl)
1886                 return ERR_PTR(-ENOMEM);
1887         ctrl->ctrl.opts = opts;
1888         INIT_LIST_HEAD(&ctrl->list);
1889
1890         if (opts->mask & NVMF_OPT_TRSVCID)
1891                 port = opts->trsvcid;
1892         else
1893                 port = __stringify(NVME_RDMA_IP_PORT);
1894
1895         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1896                         opts->traddr, port, &ctrl->addr);
1897         if (ret) {
1898                 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1899                 goto out_free_ctrl;
1900         }
1901
1902         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1903                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1904                         opts->host_traddr, NULL, &ctrl->src_addr);
1905                 if (ret) {
1906                         pr_err("malformed src address passed: %s\n",
1907                                opts->host_traddr);
1908                         goto out_free_ctrl;
1909                 }
1910         }
1911
1912         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1913                 ret = -EALREADY;
1914                 goto out_free_ctrl;
1915         }
1916
1917         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1918                                 0 /* no quirks, we're perfect! */);
1919         if (ret)
1920                 goto out_free_ctrl;
1921
1922         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1923                         nvme_rdma_reconnect_ctrl_work);
1924         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1925         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1926
1927         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1928         ctrl->ctrl.sqsize = opts->queue_size - 1;
1929         ctrl->ctrl.kato = opts->kato;
1930
1931         ret = -ENOMEM;
1932         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1933                                 GFP_KERNEL);
1934         if (!ctrl->queues)
1935                 goto out_uninit_ctrl;
1936
1937         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1938         WARN_ON_ONCE(!changed);
1939
1940         ret = nvme_rdma_configure_admin_queue(ctrl, true);
1941         if (ret)
1942                 goto out_kfree_queues;
1943
1944         /* sanity check icdoff */
1945         if (ctrl->ctrl.icdoff) {
1946                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1947                 ret = -EINVAL;
1948                 goto out_remove_admin_queue;
1949         }
1950
1951         /* sanity check keyed sgls */
1952         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1953                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1954                 ret = -EINVAL;
1955                 goto out_remove_admin_queue;
1956         }
1957
1958         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1959                 /* warn if maxcmd is lower than queue_size */
1960                 dev_warn(ctrl->ctrl.device,
1961                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1962                         opts->queue_size, ctrl->ctrl.maxcmd);
1963                 opts->queue_size = ctrl->ctrl.maxcmd;
1964         }
1965
1966         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1967                 /* warn if sqsize is lower than queue_size */
1968                 dev_warn(ctrl->ctrl.device,
1969                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1970                         opts->queue_size, ctrl->ctrl.sqsize + 1);
1971                 opts->queue_size = ctrl->ctrl.sqsize + 1;
1972         }
1973
1974         if (opts->nr_io_queues) {
1975                 ret = nvme_rdma_configure_io_queues(ctrl, true);
1976                 if (ret)
1977                         goto out_remove_admin_queue;
1978         }
1979
1980         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1981         WARN_ON_ONCE(!changed);
1982
1983         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1984                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1985
1986         nvme_get_ctrl(&ctrl->ctrl);
1987
1988         mutex_lock(&nvme_rdma_ctrl_mutex);
1989         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1990         mutex_unlock(&nvme_rdma_ctrl_mutex);
1991
1992         nvme_start_ctrl(&ctrl->ctrl);
1993
1994         return &ctrl->ctrl;
1995
1996 out_remove_admin_queue:
1997         nvme_rdma_destroy_admin_queue(ctrl, true);
1998 out_kfree_queues:
1999         kfree(ctrl->queues);
2000 out_uninit_ctrl:
2001         nvme_uninit_ctrl(&ctrl->ctrl);
2002         nvme_put_ctrl(&ctrl->ctrl);
2003         if (ret > 0)
2004                 ret = -EIO;
2005         return ERR_PTR(ret);
2006 out_free_ctrl:
2007         kfree(ctrl);
2008         return ERR_PTR(ret);
2009 }
2010
2011 static struct nvmf_transport_ops nvme_rdma_transport = {
2012         .name           = "rdma",
2013         .module         = THIS_MODULE,
2014         .required_opts  = NVMF_OPT_TRADDR,
2015         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2016                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2017         .create_ctrl    = nvme_rdma_create_ctrl,
2018 };
2019
2020 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2021 {
2022         struct nvme_rdma_ctrl *ctrl;
2023         struct nvme_rdma_device *ndev;
2024         bool found = false;
2025
2026         mutex_lock(&device_list_mutex);
2027         list_for_each_entry(ndev, &device_list, entry) {
2028                 if (ndev->dev == ib_device) {
2029                         found = true;
2030                         break;
2031                 }
2032         }
2033         mutex_unlock(&device_list_mutex);
2034
2035         if (!found)
2036                 return;
2037
2038         /* Delete all controllers using this device */
2039         mutex_lock(&nvme_rdma_ctrl_mutex);
2040         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2041                 if (ctrl->device->dev != ib_device)
2042                         continue;
2043                 nvme_delete_ctrl(&ctrl->ctrl);
2044         }
2045         mutex_unlock(&nvme_rdma_ctrl_mutex);
2046
2047         flush_workqueue(nvme_delete_wq);
2048 }
2049
2050 static struct ib_client nvme_rdma_ib_client = {
2051         .name   = "nvme_rdma",
2052         .remove = nvme_rdma_remove_one
2053 };
2054
2055 static int __init nvme_rdma_init_module(void)
2056 {
2057         int ret;
2058
2059         ret = ib_register_client(&nvme_rdma_ib_client);
2060         if (ret)
2061                 return ret;
2062
2063         ret = nvmf_register_transport(&nvme_rdma_transport);
2064         if (ret)
2065                 goto err_unreg_client;
2066
2067         return 0;
2068
2069 err_unreg_client:
2070         ib_unregister_client(&nvme_rdma_ib_client);
2071         return ret;
2072 }
2073
2074 static void __exit nvme_rdma_cleanup_module(void)
2075 {
2076         nvmf_unregister_transport(&nvme_rdma_transport);
2077         ib_unregister_client(&nvme_rdma_ib_client);
2078 }
2079
2080 module_init(nvme_rdma_init_module);
2081 module_exit(nvme_rdma_cleanup_module);
2082
2083 MODULE_LICENSE("GPL v2");