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                 return -ENOMEM;
 188         }
 189
 190         return 0;
 191 }
 192
 193 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 194                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 195                 size_t capsule_size, enum dma_data_direction dir)
 196 {
 197         int i;
 198
 199         for (i = 0; i < ib_queue_size; i++)
 200                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 201         kfree(ring);
 202 }
 203
 204 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 205                 size_t ib_queue_size, size_t capsule_size,
 206                 enum dma_data_direction dir)
 207 {
 208         struct nvme_rdma_qe *ring;
 209         int i;
 210
 211         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 212         if (!ring)
 213                 return NULL;
 214
 215         for (i = 0; i < ib_queue_size; i++) {
 216                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 217                         goto out_free_ring;
 218         }
 219
 220         return ring;
 221
 222 out_free_ring:
 223         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 224         return NULL;
 225 }
 226
 227 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 228 {
 229         pr_debug("QP event %s (%d)\n",
 230                  ib_event_msg(event->event), event->event);
 231
 232 }
 233
 234 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 235 {
 236         wait_for_completion_interruptible_timeout(&queue->cm_done,
 237                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 238         return queue->cm_error;
 239 }
 240
 241 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 242 {
 243         struct nvme_rdma_device *dev = queue->device;
 244         struct ib_qp_init_attr init_attr;
 245         int ret;
 246
 247         memset(&init_attr, 0, sizeof(init_attr));
 248         init_attr.event_handler = nvme_rdma_qp_event;
 249         /* +1 for drain */
 250         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 251         /* +1 for drain */
 252         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 253         init_attr.cap.max_recv_sge = 1;
 254         init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
 255         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 256         init_attr.qp_type = IB_QPT_RC;
 257         init_attr.send_cq = queue->ib_cq;
 258         init_attr.recv_cq = queue->ib_cq;
 259
 260         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 261
 262         queue->qp = queue->cm_id->qp;
 263         return ret;
 264 }
 265
 266 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 267                 struct request *rq, unsigned int hctx_idx)
 268 {
 269         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 270         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 271         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 272         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 273         struct nvme_rdma_device *dev = queue->device;
 274
 275         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 276                         DMA_TO_DEVICE);
 277 }
 278
 279 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 280                 struct request *rq, unsigned int hctx_idx,
 281                 unsigned int numa_node)
 282 {
 283         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 284         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 285         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 286         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 287         struct nvme_rdma_device *dev = queue->device;
 288         struct ib_device *ibdev = dev->dev;
 289         int ret;
 290
 291         nvme_req(rq)->ctrl = &ctrl->ctrl;
 292         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 293                         DMA_TO_DEVICE);
 294         if (ret)
 295                 return ret;
 296
 297         req->queue = queue;
 298
 299         return 0;
 300 }
 301
 302 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 303                 unsigned int hctx_idx)
 304 {
 305         struct nvme_rdma_ctrl *ctrl = data;
 306         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 307
 308         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 309
 310         hctx->driver_data = queue;
 311         return 0;
 312 }
 313
 314 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 315                 unsigned int hctx_idx)
 316 {
 317         struct nvme_rdma_ctrl *ctrl = data;
 318         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 319
 320         BUG_ON(hctx_idx != 0);
 321
 322         hctx->driver_data = queue;
 323         return 0;
 324 }
 325
 326 static void nvme_rdma_free_dev(struct kref *ref)
 327 {
 328         struct nvme_rdma_device *ndev =
 329                 container_of(ref, struct nvme_rdma_device, ref);
 330
 331         mutex_lock(&device_list_mutex);
 332         list_del(&ndev->entry);
 333         mutex_unlock(&device_list_mutex);
 334
 335         ib_dealloc_pd(ndev->pd);
 336         kfree(ndev);
 337 }
 338
 339 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 340 {
 341         kref_put(&dev->ref, nvme_rdma_free_dev);
 342 }
 343
 344 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 345 {
 346         return kref_get_unless_zero(&dev->ref);
 347 }
 348
 349 static struct nvme_rdma_device *
 350 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 351 {
 352         struct nvme_rdma_device *ndev;
 353
 354         mutex_lock(&device_list_mutex);
 355         list_for_each_entry(ndev, &device_list, entry) {
 356                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 357                     nvme_rdma_dev_get(ndev))
 358                         goto out_unlock;
 359         }
 360
 361         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 362         if (!ndev)
 363                 goto out_err;
 364
 365         ndev->dev = cm_id->device;
 366         kref_init(&ndev->ref);
 367
 368         ndev->pd = ib_alloc_pd(ndev->dev,
 369                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 370         if (IS_ERR(ndev->pd))
 371                 goto out_free_dev;
 372
 373         if (!(ndev->dev->attrs.device_cap_flags &
 374               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 375                 dev_err(&ndev->dev->dev,
 376                         "Memory registrations not supported.\n");
 377                 goto out_free_pd;
 378         }
 379
 380         ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
 381                                         ndev->dev->attrs.max_sge - 1);
 382         list_add(&ndev->entry, &device_list);
 383 out_unlock:
 384         mutex_unlock(&device_list_mutex);
 385         return ndev;
 386
 387 out_free_pd:
 388         ib_dealloc_pd(ndev->pd);
 389 out_free_dev:
 390         kfree(ndev);
 391 out_err:
 392         mutex_unlock(&device_list_mutex);
 393         return NULL;
 394 }
 395
 396 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 397 {
 398         struct nvme_rdma_device *dev;
 399         struct ib_device *ibdev;
 400
 401         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 402                 return;
 403
 404         dev = queue->device;
 405         ibdev = dev->dev;
 406
 407         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 408
 409         /*
 410          * The cm_id object might have been destroyed during RDMA connection
 411          * establishment error flow to avoid getting other cma events, thus
 412          * the destruction of the QP shouldn't use rdma_cm API.
 413          */
 414         ib_destroy_qp(queue->qp);
 415         ib_free_cq(queue->ib_cq);
 416
 417         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 418                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 419
 420         nvme_rdma_dev_put(dev);
 421 }
 422
 423 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
 424 {
 425         return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
 426                      ibdev->attrs.max_fast_reg_page_list_len);
 427 }
 428
 429 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 430 {
 431         struct ib_device *ibdev;
 432         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 433         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 434         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 435         int ret;
 436
 437         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 438         if (!queue->device) {
 439                 dev_err(queue->cm_id->device->dev.parent,
 440                         "no client data found!\n");
 441                 return -ECONNREFUSED;
 442         }
 443         ibdev = queue->device->dev;
 444
 445         /*
 446          * Spread I/O queues completion vectors according their queue index.
 447          * Admin queues can always go on completion vector 0.
 448          */
 449         comp_vector = idx == 0 ? idx : idx - 1;
 450
 451         /* +1 for ib_stop_cq */
 452         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 453                                 cq_factor * queue->queue_size + 1,
 454                                 comp_vector, IB_POLL_SOFTIRQ);
 455         if (IS_ERR(queue->ib_cq)) {
 456                 ret = PTR_ERR(queue->ib_cq);
 457                 goto out_put_dev;
 458         }
 459
 460         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 461         if (ret)
 462                 goto out_destroy_ib_cq;
 463
 464         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 465                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 466         if (!queue->rsp_ring) {
 467                 ret = -ENOMEM;
 468                 goto out_destroy_qp;
 469         }
 470
 471         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 472                               queue->queue_size,
 473                               IB_MR_TYPE_MEM_REG,
 474                               nvme_rdma_get_max_fr_pages(ibdev));
 475         if (ret) {
 476                 dev_err(queue->ctrl->ctrl.device,
 477                         "failed to initialize MR pool sized %d for QID %d\n",
 478                         queue->queue_size, idx);
 479                 goto out_destroy_ring;
 480         }
 481
 482         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 483
 484         return 0;
 485
 486 out_destroy_ring:
 487         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 488                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 489 out_destroy_qp:
 490         rdma_destroy_qp(queue->cm_id);
 491 out_destroy_ib_cq:
 492         ib_free_cq(queue->ib_cq);
 493 out_put_dev:
 494         nvme_rdma_dev_put(queue->device);
 495         return ret;
 496 }
 497
 498 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 499                 int idx, size_t queue_size)
 500 {
 501         struct nvme_rdma_queue *queue;
 502         struct sockaddr *src_addr = NULL;
 503         int ret;
 504
 505         queue = &ctrl->queues[idx];
 506         queue->ctrl = ctrl;
 507         init_completion(&queue->cm_done);
 508
 509         if (idx > 0)
 510                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 511         else
 512                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 513
 514         queue->queue_size = queue_size;
 515
 516         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 517                         RDMA_PS_TCP, IB_QPT_RC);
 518         if (IS_ERR(queue->cm_id)) {
 519                 dev_info(ctrl->ctrl.device,
 520                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 521                 return PTR_ERR(queue->cm_id);
 522         }
 523
 524         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 525                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 526
 527         queue->cm_error = -ETIMEDOUT;
 528         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 529                         (struct sockaddr *)&ctrl->addr,
 530                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 531         if (ret) {
 532                 dev_info(ctrl->ctrl.device,
 533                         "rdma_resolve_addr failed (%d).\n", ret);
 534                 goto out_destroy_cm_id;
 535         }
 536
 537         ret = nvme_rdma_wait_for_cm(queue);
 538         if (ret) {
 539                 dev_info(ctrl->ctrl.device,
 540                         "rdma connection establishment failed (%d)\n", ret);
 541                 goto out_destroy_cm_id;
 542         }
 543
 544         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 545
 546         return 0;
 547
 548 out_destroy_cm_id:
 549         rdma_destroy_id(queue->cm_id);
 550         nvme_rdma_destroy_queue_ib(queue);
 551         return ret;
 552 }
 553
 554 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 555 {
 556         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 557                 return;
 558
 559         rdma_disconnect(queue->cm_id);
 560         ib_drain_qp(queue->qp);
 561 }
 562
 563 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 564 {
 565         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 566                 return;
 567
 568         nvme_rdma_destroy_queue_ib(queue);
 569         rdma_destroy_id(queue->cm_id);
 570 }
 571
 572 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 573 {
 574         int i;
 575
 576         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 577                 nvme_rdma_free_queue(&ctrl->queues[i]);
 578 }
 579
 580 static void nvme_rdma_stop_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_stop_queue(&ctrl->queues[i]);
 586 }
 587
 588 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 589 {
 590         int ret;
 591
 592         if (idx)
 593                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 594         else
 595                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 596
 597         if (!ret)
 598                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
 599         else
 600                 dev_info(ctrl->ctrl.device,
 601                         "failed to connect queue: %d ret=%d\n", idx, ret);
 602         return ret;
 603 }
 604
 605 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 606 {
 607         int i, ret = 0;
 608
 609         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 610                 ret = nvme_rdma_start_queue(ctrl, i);
 611                 if (ret)
 612                         goto out_stop_queues;
 613         }
 614
 615         return 0;
 616
 617 out_stop_queues:
 618         for (i--; i >= 1; i--)
 619                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 620         return ret;
 621 }
 622
 623 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 624 {
 625         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 626         struct ib_device *ibdev = ctrl->device->dev;
 627         unsigned int nr_io_queues;
 628         int i, ret;
 629
 630         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 631
 632         /*
 633          * we map queues according to the device irq vectors for
 634          * optimal locality so we don't need more queues than
 635          * completion vectors.
 636          */
 637         nr_io_queues = min_t(unsigned int, nr_io_queues,
 638                                 ibdev->num_comp_vectors);
 639
 640         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 641         if (ret)
 642                 return ret;
 643
 644         ctrl->ctrl.queue_count = nr_io_queues + 1;
 645         if (ctrl->ctrl.queue_count < 2)
 646                 return 0;
 647
 648         dev_info(ctrl->ctrl.device,
 649                 "creating %d I/O queues.\n", nr_io_queues);
 650
 651         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 652                 ret = nvme_rdma_alloc_queue(ctrl, i,
 653                                 ctrl->ctrl.sqsize + 1);
 654                 if (ret)
 655                         goto out_free_queues;
 656         }
 657
 658         return 0;
 659
 660 out_free_queues:
 661         for (i--; i >= 1; i--)
 662                 nvme_rdma_free_queue(&ctrl->queues[i]);
 663
 664         return ret;
 665 }
 666
 667 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
 668                 struct blk_mq_tag_set *set)
 669 {
 670         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 671
 672         blk_mq_free_tag_set(set);
 673         nvme_rdma_dev_put(ctrl->device);
 674 }
 675
 676 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 677                 bool admin)
 678 {
 679         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 680         struct blk_mq_tag_set *set;
 681         int ret;
 682
 683         if (admin) {
 684                 set = &ctrl->admin_tag_set;
 685                 memset(set, 0, sizeof(*set));
 686                 set->ops = &nvme_rdma_admin_mq_ops;
 687                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 688                 set->reserved_tags = 2; /* connect + keep-alive */
 689                 set->numa_node = NUMA_NO_NODE;
 690                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 691                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 692                 set->driver_data = ctrl;
 693                 set->nr_hw_queues = 1;
 694                 set->timeout = ADMIN_TIMEOUT;
 695                 set->flags = BLK_MQ_F_NO_SCHED;
 696         } else {
 697                 set = &ctrl->tag_set;
 698                 memset(set, 0, sizeof(*set));
 699                 set->ops = &nvme_rdma_mq_ops;
 700                 set->queue_depth = nctrl->sqsize + 1;
 701                 set->reserved_tags = 1; /* fabric connect */
 702                 set->numa_node = NUMA_NO_NODE;
 703                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 704                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 705                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 706                 set->driver_data = ctrl;
 707                 set->nr_hw_queues = nctrl->queue_count - 1;
 708                 set->timeout = NVME_IO_TIMEOUT;
 709         }
 710
 711         ret = blk_mq_alloc_tag_set(set);
 712         if (ret)
 713                 goto out;
 714
 715         /*
 716          * We need a reference on the device as long as the tag_set is alive,
 717          * as the MRs in the request structures need a valid ib_device.
 718          */
 719         ret = nvme_rdma_dev_get(ctrl->device);
 720         if (!ret) {
 721                 ret = -EINVAL;
 722                 goto out_free_tagset;
 723         }
 724
 725         return set;
 726
 727 out_free_tagset:
 728         blk_mq_free_tag_set(set);
 729 out:
 730         return ERR_PTR(ret);
 731 }
 732
 733 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 734                 bool remove)
 735 {
 736         if (remove) {
 737                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 738                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 739         }
 740         if (ctrl->async_event_sqe.data) {
 741                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 742                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 743                 ctrl->async_event_sqe.data = NULL;
 744         }
 745         nvme_rdma_free_queue(&ctrl->queues[0]);
 746 }
 747
 748 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 749                 bool new)
 750 {
 751         int error;
 752
 753         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 754         if (error)
 755                 return error;
 756
 757         ctrl->device = ctrl->queues[0].device;
 758
 759         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
 760
 761         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 762                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 763         if (error)
 764                 goto out_free_queue;
 765
 766         if (new) {
 767                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 768                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 769                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 770                         goto out_free_async_qe;
 771                 }
 772
 773                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 774                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 775                         error = PTR_ERR(ctrl->ctrl.admin_q);
 776                         goto out_free_tagset;
 777                 }
 778         }
 779
 780         error = nvme_rdma_start_queue(ctrl, 0);
 781         if (error)
 782                 goto out_cleanup_queue;
 783
 784         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 785                         &ctrl->ctrl.cap);
 786         if (error) {
 787                 dev_err(ctrl->ctrl.device,
 788                         "prop_get NVME_REG_CAP failed\n");
 789                 goto out_stop_queue;
 790         }
 791
 792         ctrl->ctrl.sqsize =
 793                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 794
 795         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 796         if (error)
 797                 goto out_stop_queue;
 798
 799         ctrl->ctrl.max_hw_sectors =
 800                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 801
 802         error = nvme_init_identify(&ctrl->ctrl);
 803         if (error)
 804                 goto out_stop_queue;
 805
 806         return 0;
 807
 808 out_stop_queue:
 809         nvme_rdma_stop_queue(&ctrl->queues[0]);
 810 out_cleanup_queue:
 811         if (new)
 812                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 813 out_free_tagset:
 814         if (new)
 815                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 816 out_free_async_qe:
 817         nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 818                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 819 out_free_queue:
 820         nvme_rdma_free_queue(&ctrl->queues[0]);
 821         return error;
 822 }
 823
 824 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 825                 bool remove)
 826 {
 827         if (remove) {
 828                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 829                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 830         }
 831         nvme_rdma_free_io_queues(ctrl);
 832 }
 833
 834 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 835 {
 836         int ret;
 837
 838         ret = nvme_rdma_alloc_io_queues(ctrl);
 839         if (ret)
 840                 return ret;
 841
 842         if (new) {
 843                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 844                 if (IS_ERR(ctrl->ctrl.tagset)) {
 845                         ret = PTR_ERR(ctrl->ctrl.tagset);
 846                         goto out_free_io_queues;
 847                 }
 848
 849                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 850                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 851                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 852                         goto out_free_tag_set;
 853                 }
 854         } else {
 855                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 856                         ctrl->ctrl.queue_count - 1);
 857         }
 858
 859         ret = nvme_rdma_start_io_queues(ctrl);
 860         if (ret)
 861                 goto out_cleanup_connect_q;
 862
 863         return 0;
 864
 865 out_cleanup_connect_q:
 866         if (new)
 867                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 868 out_free_tag_set:
 869         if (new)
 870                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 871 out_free_io_queues:
 872         nvme_rdma_free_io_queues(ctrl);
 873         return ret;
 874 }
 875
 876 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
 877                 bool remove)
 878 {
 879         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 880         nvme_rdma_stop_queue(&ctrl->queues[0]);
 881         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
 882                         &ctrl->ctrl);
 883         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 884         nvme_rdma_destroy_admin_queue(ctrl, remove);
 885 }
 886
 887 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
 888                 bool remove)
 889 {
 890         if (ctrl->ctrl.queue_count > 1) {
 891                 nvme_stop_queues(&ctrl->ctrl);
 892                 nvme_rdma_stop_io_queues(ctrl);
 893                 blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
 894                                 &ctrl->ctrl);
 895                 if (remove)
 896                         nvme_start_queues(&ctrl->ctrl);
 897                 nvme_rdma_destroy_io_queues(ctrl, remove);
 898         }
 899 }
 900
 901 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
 902 {
 903         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 904
 905         cancel_work_sync(&ctrl->err_work);
 906         cancel_delayed_work_sync(&ctrl->reconnect_work);
 907 }
 908
 909 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 910 {
 911         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 912
 913         if (list_empty(&ctrl->list))
 914                 goto free_ctrl;
 915
 916         mutex_lock(&nvme_rdma_ctrl_mutex);
 917         list_del(&ctrl->list);
 918         mutex_unlock(&nvme_rdma_ctrl_mutex);
 919
 920         nvmf_free_options(nctrl->opts);
 921 free_ctrl:
 922         kfree(ctrl->queues);
 923         kfree(ctrl);
 924 }
 925
 926 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 927 {
 928         /* If we are resetting/deleting then do nothing */
 929         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 930                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 931                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 932                 return;
 933         }
 934
 935         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 936                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 937                         ctrl->ctrl.opts->reconnect_delay);
 938                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 939                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 940         } else {
 941                 nvme_delete_ctrl(&ctrl->ctrl);
 942         }
 943 }
 944
 945 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
 946 {
 947         int ret = -EINVAL;
 948         bool changed;
 949
 950         ret = nvme_rdma_configure_admin_queue(ctrl, new);
 951         if (ret)
 952                 return ret;
 953
 954         if (ctrl->ctrl.icdoff) {
 955                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
 956                 goto destroy_admin;
 957         }
 958
 959         if (!(ctrl->ctrl.sgls & (1 << 2))) {
 960                 dev_err(ctrl->ctrl.device,
 961                         "Mandatory keyed sgls are not supported!\n");
 962                 goto destroy_admin;
 963         }
 964
 965         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
 966                 dev_warn(ctrl->ctrl.device,
 967                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
 968                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
 969         }
 970
 971         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
 972                 dev_warn(ctrl->ctrl.device,
 973                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
 974                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
 975                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
 976         }
 977
 978         if (ctrl->ctrl.sgls & (1 << 20))
 979                 ctrl->use_inline_data = true;
 980
 981         if (ctrl->ctrl.queue_count > 1) {
 982                 ret = nvme_rdma_configure_io_queues(ctrl, new);
 983                 if (ret)
 984                         goto destroy_admin;
 985         }
 986
 987         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 988         if (!changed) {
 989                 /* state change failure is ok if we're in DELETING state */
 990                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
 991                 ret = -EINVAL;
 992                 goto destroy_io;
 993         }
 994
 995         nvme_start_ctrl(&ctrl->ctrl);
 996         return 0;
 997
 998 destroy_io:
 999         if (ctrl->ctrl.queue_count > 1)
1000                 nvme_rdma_destroy_io_queues(ctrl, new);
1001 destroy_admin:
1002         nvme_rdma_stop_queue(&ctrl->queues[0]);
1003         nvme_rdma_destroy_admin_queue(ctrl, new);
1004         return ret;
1005 }
1006
1007 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1008 {
1009         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1010                         struct nvme_rdma_ctrl, reconnect_work);
1011
1012         ++ctrl->ctrl.nr_reconnects;
1013
1014         if (nvme_rdma_setup_ctrl(ctrl, false))
1015                 goto requeue;
1016
1017         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1018                         ctrl->ctrl.nr_reconnects);
1019
1020         ctrl->ctrl.nr_reconnects = 0;
1021
1022         return;
1023
1024 requeue:
1025         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1026                         ctrl->ctrl.nr_reconnects);
1027         nvme_rdma_reconnect_or_remove(ctrl);
1028 }
1029
1030 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1031 {
1032         struct nvme_rdma_ctrl *ctrl = container_of(work,
1033                         struct nvme_rdma_ctrl, err_work);
1034
1035         nvme_stop_keep_alive(&ctrl->ctrl);
1036         nvme_rdma_teardown_io_queues(ctrl, false);
1037         nvme_start_queues(&ctrl->ctrl);
1038         nvme_rdma_teardown_admin_queue(ctrl, false);
1039
1040         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1041                 /* state change failure is ok if we're in DELETING state */
1042                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1043                 return;
1044         }
1045
1046         nvme_rdma_reconnect_or_remove(ctrl);
1047 }
1048
1049 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1050 {
1051         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1052                 return;
1053
1054         queue_work(nvme_wq, &ctrl->err_work);
1055 }
1056
1057 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1058                 const char *op)
1059 {
1060         struct nvme_rdma_queue *queue = cq->cq_context;
1061         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1062
1063         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1064                 dev_info(ctrl->ctrl.device,
1065                              "%s for CQE 0x%p failed with status %s (%d)\n",
1066                              op, wc->wr_cqe,
1067                              ib_wc_status_msg(wc->status), wc->status);
1068         nvme_rdma_error_recovery(ctrl);
1069 }
1070
1071 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1072 {
1073         if (unlikely(wc->status != IB_WC_SUCCESS))
1074                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1075 }
1076
1077 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1078 {
1079         struct nvme_rdma_request *req =
1080                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1081         struct request *rq = blk_mq_rq_from_pdu(req);
1082
1083         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1084                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1085                 return;
1086         }
1087
1088         if (refcount_dec_and_test(&req->ref))
1089                 nvme_end_request(rq, req->status, req->result);
1090
1091 }
1092
1093 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1094                 struct nvme_rdma_request *req)
1095 {
1096         struct ib_send_wr *bad_wr;
1097         struct ib_send_wr wr = {
1098                 .opcode             = IB_WR_LOCAL_INV,
1099                 .next               = NULL,
1100                 .num_sge            = 0,
1101                 .send_flags         = IB_SEND_SIGNALED,
1102                 .ex.invalidate_rkey = req->mr->rkey,
1103         };
1104
1105         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1106         wr.wr_cqe = &req->reg_cqe;
1107
1108         return ib_post_send(queue->qp, &wr, &bad_wr);
1109 }
1110
1111 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1112                 struct request *rq)
1113 {
1114         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1115         struct nvme_rdma_device *dev = queue->device;
1116         struct ib_device *ibdev = dev->dev;
1117
1118         if (!blk_rq_payload_bytes(rq))
1119                 return;
1120
1121         if (req->mr) {
1122                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1123                 req->mr = NULL;
1124         }
1125
1126         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1127                         req->nents, rq_data_dir(rq) ==
1128                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1129
1130         nvme_cleanup_cmd(rq);
1131         sg_free_table_chained(&req->sg_table, true);
1132 }
1133
1134 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1135 {
1136         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1137
1138         sg->addr = 0;
1139         put_unaligned_le24(0, sg->length);
1140         put_unaligned_le32(0, sg->key);
1141         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1142         return 0;
1143 }
1144
1145 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1146                 struct nvme_rdma_request *req, struct nvme_command *c,
1147                 int count)
1148 {
1149         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1150         struct scatterlist *sgl = req->sg_table.sgl;
1151         struct ib_sge *sge = &req->sge[1];
1152         u32 len = 0;
1153         int i;
1154
1155         for (i = 0; i < count; i++, sgl++, sge++) {
1156                 sge->addr = sg_dma_address(sgl);
1157                 sge->length = sg_dma_len(sgl);
1158                 sge->lkey = queue->device->pd->local_dma_lkey;
1159                 len += sge->length;
1160         }
1161
1162         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1163         sg->length = cpu_to_le32(len);
1164         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1165
1166         req->num_sge += count;
1167         return 0;
1168 }
1169
1170 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1171                 struct nvme_rdma_request *req, struct nvme_command *c)
1172 {
1173         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1174
1175         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1176         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1177         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1178         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1179         return 0;
1180 }
1181
1182 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1183                 struct nvme_rdma_request *req, struct nvme_command *c,
1184                 int count)
1185 {
1186         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1187         int nr;
1188
1189         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1190         if (WARN_ON_ONCE(!req->mr))
1191                 return -EAGAIN;
1192
1193         /*
1194          * Align the MR to a 4K page size to match the ctrl page size and
1195          * the block virtual boundary.
1196          */
1197         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1198         if (unlikely(nr < count)) {
1199                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1200                 req->mr = NULL;
1201                 if (nr < 0)
1202                         return nr;
1203                 return -EINVAL;
1204         }
1205
1206         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1207
1208         req->reg_cqe.done = nvme_rdma_memreg_done;
1209         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1210         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1211         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1212         req->reg_wr.wr.num_sge = 0;
1213         req->reg_wr.mr = req->mr;
1214         req->reg_wr.key = req->mr->rkey;
1215         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1216                              IB_ACCESS_REMOTE_READ |
1217                              IB_ACCESS_REMOTE_WRITE;
1218
1219         sg->addr = cpu_to_le64(req->mr->iova);
1220         put_unaligned_le24(req->mr->length, sg->length);
1221         put_unaligned_le32(req->mr->rkey, sg->key);
1222         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1223                         NVME_SGL_FMT_INVALIDATE;
1224
1225         return 0;
1226 }
1227
1228 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1229                 struct request *rq, struct nvme_command *c)
1230 {
1231         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1232         struct nvme_rdma_device *dev = queue->device;
1233         struct ib_device *ibdev = dev->dev;
1234         int count, ret;
1235
1236         req->num_sge = 1;
1237         refcount_set(&req->ref, 2); /* send and recv completions */
1238
1239         c->common.flags |= NVME_CMD_SGL_METABUF;
1240
1241         if (!blk_rq_payload_bytes(rq))
1242                 return nvme_rdma_set_sg_null(c);
1243
1244         req->sg_table.sgl = req->first_sgl;
1245         ret = sg_alloc_table_chained(&req->sg_table,
1246                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1247         if (ret)
1248                 return -ENOMEM;
1249
1250         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1251
1252         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1253                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1254         if (unlikely(count <= 0)) {
1255                 ret = -EIO;
1256                 goto out_free_table;
1257         }
1258
1259         if (count <= dev->num_inline_segments) {
1260                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1261                     queue->ctrl->use_inline_data &&
1262                     blk_rq_payload_bytes(rq) <=
1263                                 nvme_rdma_inline_data_size(queue)) {
1264                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1265                         goto out;
1266                 }
1267
1268                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1269                         ret = nvme_rdma_map_sg_single(queue, req, c);
1270                         goto out;
1271                 }
1272         }
1273
1274         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1275 out:
1276         if (unlikely(ret))
1277                 goto out_unmap_sg;
1278
1279         return 0;
1280
1281 out_unmap_sg:
1282         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1283                         req->nents, rq_data_dir(rq) ==
1284                         WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1285 out_free_table:
1286         sg_free_table_chained(&req->sg_table, true);
1287         return ret;
1288 }
1289
1290 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1291 {
1292         struct nvme_rdma_qe *qe =
1293                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1294         struct nvme_rdma_request *req =
1295                 container_of(qe, struct nvme_rdma_request, sqe);
1296         struct request *rq = blk_mq_rq_from_pdu(req);
1297
1298         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1299                 nvme_rdma_wr_error(cq, wc, "SEND");
1300                 return;
1301         }
1302
1303         if (refcount_dec_and_test(&req->ref))
1304                 nvme_end_request(rq, req->status, req->result);
1305 }
1306
1307 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1308                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1309                 struct ib_send_wr *first)
1310 {
1311         struct ib_send_wr wr, *bad_wr;
1312         int ret;
1313
1314         sge->addr   = qe->dma;
1315         sge->length = sizeof(struct nvme_command),
1316         sge->lkey   = queue->device->pd->local_dma_lkey;
1317
1318         wr.next       = NULL;
1319         wr.wr_cqe     = &qe->cqe;
1320         wr.sg_list    = sge;
1321         wr.num_sge    = num_sge;
1322         wr.opcode     = IB_WR_SEND;
1323         wr.send_flags = IB_SEND_SIGNALED;
1324
1325         if (first)
1326                 first->next = &wr;
1327         else
1328                 first = &wr;
1329
1330         ret = ib_post_send(queue->qp, first, &bad_wr);
1331         if (unlikely(ret)) {
1332                 dev_err(queue->ctrl->ctrl.device,
1333                              "%s failed with error code %d\n", __func__, ret);
1334         }
1335         return ret;
1336 }
1337
1338 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1339                 struct nvme_rdma_qe *qe)
1340 {
1341         struct ib_recv_wr wr, *bad_wr;
1342         struct ib_sge list;
1343         int ret;
1344
1345         list.addr   = qe->dma;
1346         list.length = sizeof(struct nvme_completion);
1347         list.lkey   = queue->device->pd->local_dma_lkey;
1348
1349         qe->cqe.done = nvme_rdma_recv_done;
1350
1351         wr.next     = NULL;
1352         wr.wr_cqe   = &qe->cqe;
1353         wr.sg_list  = &list;
1354         wr.num_sge  = 1;
1355
1356         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1357         if (unlikely(ret)) {
1358                 dev_err(queue->ctrl->ctrl.device,
1359                         "%s failed with error code %d\n", __func__, ret);
1360         }
1361         return ret;
1362 }
1363
1364 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1365 {
1366         u32 queue_idx = nvme_rdma_queue_idx(queue);
1367
1368         if (queue_idx == 0)
1369                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1370         return queue->ctrl->tag_set.tags[queue_idx - 1];
1371 }
1372
1373 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1374 {
1375         if (unlikely(wc->status != IB_WC_SUCCESS))
1376                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1377 }
1378
1379 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1380 {
1381         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1382         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1383         struct ib_device *dev = queue->device->dev;
1384         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1385         struct nvme_command *cmd = sqe->data;
1386         struct ib_sge sge;
1387         int ret;
1388
1389         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1390
1391         memset(cmd, 0, sizeof(*cmd));
1392         cmd->common.opcode = nvme_admin_async_event;
1393         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1394         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1395         nvme_rdma_set_sg_null(cmd);
1396
1397         sqe->cqe.done = nvme_rdma_async_done;
1398
1399         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1400                         DMA_TO_DEVICE);
1401
1402         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1403         WARN_ON_ONCE(ret);
1404 }
1405
1406 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1407                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1408 {
1409         struct request *rq;
1410         struct nvme_rdma_request *req;
1411         int ret = 0;
1412
1413         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1414         if (!rq) {
1415                 dev_err(queue->ctrl->ctrl.device,
1416                         "tag 0x%x on QP %#x not found\n",
1417                         cqe->command_id, queue->qp->qp_num);
1418                 nvme_rdma_error_recovery(queue->ctrl);
1419                 return ret;
1420         }
1421         req = blk_mq_rq_to_pdu(rq);
1422
1423         req->status = cqe->status;
1424         req->result = cqe->result;
1425
1426         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1427                 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1428                         dev_err(queue->ctrl->ctrl.device,
1429                                 "Bogus remote invalidation for rkey %#x\n",
1430                                 req->mr->rkey);
1431                         nvme_rdma_error_recovery(queue->ctrl);
1432                 }
1433         } else if (req->mr) {
1434                 ret = nvme_rdma_inv_rkey(queue, req);
1435                 if (unlikely(ret < 0)) {
1436                         dev_err(queue->ctrl->ctrl.device,
1437                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1438                                 req->mr->rkey, ret);
1439                         nvme_rdma_error_recovery(queue->ctrl);
1440                 }
1441                 /* the local invalidation completion will end the request */
1442                 return 0;
1443         }
1444
1445         if (refcount_dec_and_test(&req->ref)) {
1446                 if (rq->tag == tag)
1447                         ret = 1;
1448                 nvme_end_request(rq, req->status, req->result);
1449         }
1450
1451         return ret;
1452 }
1453
1454 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1455 {
1456         struct nvme_rdma_qe *qe =
1457                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1458         struct nvme_rdma_queue *queue = cq->cq_context;
1459         struct ib_device *ibdev = queue->device->dev;
1460         struct nvme_completion *cqe = qe->data;
1461         const size_t len = sizeof(struct nvme_completion);
1462         int ret = 0;
1463
1464         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1465                 nvme_rdma_wr_error(cq, wc, "RECV");
1466                 return 0;
1467         }
1468
1469         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1470         /*
1471          * AEN requests are special as they don't time out and can
1472          * survive any kind of queue freeze and often don't respond to
1473          * aborts.  We don't even bother to allocate a struct request
1474          * for them but rather special case them here.
1475          */
1476         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1477                         cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1478                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1479                                 &cqe->result);
1480         else
1481                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1482         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1483
1484         nvme_rdma_post_recv(queue, qe);
1485         return ret;
1486 }
1487
1488 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1489 {
1490         __nvme_rdma_recv_done(cq, wc, -1);
1491 }
1492
1493 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1494 {
1495         int ret, i;
1496
1497         for (i = 0; i < queue->queue_size; i++) {
1498                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1499                 if (ret)
1500                         goto out_destroy_queue_ib;
1501         }
1502
1503         return 0;
1504
1505 out_destroy_queue_ib:
1506         nvme_rdma_destroy_queue_ib(queue);
1507         return ret;
1508 }
1509
1510 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1511                 struct rdma_cm_event *ev)
1512 {
1513         struct rdma_cm_id *cm_id = queue->cm_id;
1514         int status = ev->status;
1515         const char *rej_msg;
1516         const struct nvme_rdma_cm_rej *rej_data;
1517         u8 rej_data_len;
1518
1519         rej_msg = rdma_reject_msg(cm_id, status);
1520         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1521
1522         if (rej_data && rej_data_len >= sizeof(u16)) {
1523                 u16 sts = le16_to_cpu(rej_data->sts);
1524
1525                 dev_err(queue->ctrl->ctrl.device,
1526                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1527                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1528         } else {
1529                 dev_err(queue->ctrl->ctrl.device,
1530                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1531         }
1532
1533         return -ECONNRESET;
1534 }
1535
1536 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1537 {
1538         int ret;
1539
1540         ret = nvme_rdma_create_queue_ib(queue);
1541         if (ret)
1542                 return ret;
1543
1544         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1545         if (ret) {
1546                 dev_err(queue->ctrl->ctrl.device,
1547                         "rdma_resolve_route failed (%d).\n",
1548                         queue->cm_error);
1549                 goto out_destroy_queue;
1550         }
1551
1552         return 0;
1553
1554 out_destroy_queue:
1555         nvme_rdma_destroy_queue_ib(queue);
1556         return ret;
1557 }
1558
1559 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1560 {
1561         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1562         struct rdma_conn_param param = { };
1563         struct nvme_rdma_cm_req priv = { };
1564         int ret;
1565
1566         param.qp_num = queue->qp->qp_num;
1567         param.flow_control = 1;
1568
1569         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1570         /* maximum retry count */
1571         param.retry_count = 7;
1572         param.rnr_retry_count = 7;
1573         param.private_data = &priv;
1574         param.private_data_len = sizeof(priv);
1575
1576         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1577         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1578         /*
1579          * set the admin queue depth to the minimum size
1580          * specified by the Fabrics standard.
1581          */
1582         if (priv.qid == 0) {
1583                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1584                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1585         } else {
1586                 /*
1587                  * current interpretation of the fabrics spec
1588                  * is at minimum you make hrqsize sqsize+1, or a
1589                  * 1's based representation of sqsize.
1590                  */
1591                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1592                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1593         }
1594
1595         ret = rdma_connect(queue->cm_id, &param);
1596         if (ret) {
1597                 dev_err(ctrl->ctrl.device,
1598                         "rdma_connect failed (%d).\n", ret);
1599                 goto out_destroy_queue_ib;
1600         }
1601
1602         return 0;
1603
1604 out_destroy_queue_ib:
1605         nvme_rdma_destroy_queue_ib(queue);
1606         return ret;
1607 }
1608
1609 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1610                 struct rdma_cm_event *ev)
1611 {
1612         struct nvme_rdma_queue *queue = cm_id->context;
1613         int cm_error = 0;
1614
1615         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1616                 rdma_event_msg(ev->event), ev->event,
1617                 ev->status, cm_id);
1618
1619         switch (ev->event) {
1620         case RDMA_CM_EVENT_ADDR_RESOLVED:
1621                 cm_error = nvme_rdma_addr_resolved(queue);
1622                 break;
1623         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1624                 cm_error = nvme_rdma_route_resolved(queue);
1625                 break;
1626         case RDMA_CM_EVENT_ESTABLISHED:
1627                 queue->cm_error = nvme_rdma_conn_established(queue);
1628                 /* complete cm_done regardless of success/failure */
1629                 complete(&queue->cm_done);
1630                 return 0;
1631         case RDMA_CM_EVENT_REJECTED:
1632                 nvme_rdma_destroy_queue_ib(queue);
1633                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1634                 break;
1635         case RDMA_CM_EVENT_ROUTE_ERROR:
1636         case RDMA_CM_EVENT_CONNECT_ERROR:
1637         case RDMA_CM_EVENT_UNREACHABLE:
1638                 nvme_rdma_destroy_queue_ib(queue);
1639                 /* fall through */
1640         case RDMA_CM_EVENT_ADDR_ERROR:
1641                 dev_dbg(queue->ctrl->ctrl.device,
1642                         "CM error event %d\n", ev->event);
1643                 cm_error = -ECONNRESET;
1644                 break;
1645         case RDMA_CM_EVENT_DISCONNECTED:
1646         case RDMA_CM_EVENT_ADDR_CHANGE:
1647         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1648                 dev_dbg(queue->ctrl->ctrl.device,
1649                         "disconnect received - connection closed\n");
1650                 nvme_rdma_error_recovery(queue->ctrl);
1651                 break;
1652         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1653                 /* device removal is handled via the ib_client API */
1654                 break;
1655         default:
1656                 dev_err(queue->ctrl->ctrl.device,
1657                         "Unexpected RDMA CM event (%d)\n", ev->event);
1658                 nvme_rdma_error_recovery(queue->ctrl);
1659                 break;
1660         }
1661
1662         if (cm_error) {
1663                 queue->cm_error = cm_error;
1664                 complete(&queue->cm_done);
1665         }
1666
1667         return 0;
1668 }
1669
1670 static enum blk_eh_timer_return
1671 nvme_rdma_timeout(struct request *rq, bool reserved)
1672 {
1673         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1674
1675         dev_warn(req->queue->ctrl->ctrl.device,
1676                  "I/O %d QID %d timeout, reset controller\n",
1677                  rq->tag, nvme_rdma_queue_idx(req->queue));
1678
1679         /* queue error recovery */
1680         nvme_rdma_error_recovery(req->queue->ctrl);
1681
1682         /* fail with DNR on cmd timeout */
1683         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1684
1685         return BLK_EH_DONE;
1686 }
1687
1688 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1689                 const struct blk_mq_queue_data *bd)
1690 {
1691         struct nvme_ns *ns = hctx->queue->queuedata;
1692         struct nvme_rdma_queue *queue = hctx->driver_data;
1693         struct request *rq = bd->rq;
1694         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1695         struct nvme_rdma_qe *sqe = &req->sqe;
1696         struct nvme_command *c = sqe->data;
1697         struct ib_device *dev;
1698         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1699         blk_status_t ret;
1700         int err;
1701
1702         WARN_ON_ONCE(rq->tag < 0);
1703
1704         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1705                 return nvmf_fail_nonready_command(rq);
1706
1707         dev = queue->device->dev;
1708         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1709                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1710
1711         ret = nvme_setup_cmd(ns, rq, c);
1712         if (ret)
1713                 return ret;
1714
1715         blk_mq_start_request(rq);
1716
1717         err = nvme_rdma_map_data(queue, rq, c);
1718         if (unlikely(err < 0)) {
1719                 dev_err(queue->ctrl->ctrl.device,
1720                              "Failed to map data (%d)\n", err);
1721                 nvme_cleanup_cmd(rq);
1722                 goto err;
1723         }
1724
1725         sqe->cqe.done = nvme_rdma_send_done;
1726
1727         ib_dma_sync_single_for_device(dev, sqe->dma,
1728                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1729
1730         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1731                         req->mr ? &req->reg_wr.wr : NULL);
1732         if (unlikely(err)) {
1733                 nvme_rdma_unmap_data(queue, rq);
1734                 goto err;
1735         }
1736
1737         return BLK_STS_OK;
1738 err:
1739         if (err == -ENOMEM || err == -EAGAIN)
1740                 return BLK_STS_RESOURCE;
1741         return BLK_STS_IOERR;
1742 }
1743
1744 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1745 {
1746         struct nvme_rdma_queue *queue = hctx->driver_data;
1747         struct ib_cq *cq = queue->ib_cq;
1748         struct ib_wc wc;
1749         int found = 0;
1750
1751         while (ib_poll_cq(cq, 1, &wc) > 0) {
1752                 struct ib_cqe *cqe = wc.wr_cqe;
1753
1754                 if (cqe) {
1755                         if (cqe->done == nvme_rdma_recv_done)
1756                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1757                         else
1758                                 cqe->done(cq, &wc);
1759                 }
1760         }
1761
1762         return found;
1763 }
1764
1765 static void nvme_rdma_complete_rq(struct request *rq)
1766 {
1767         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1768
1769         nvme_rdma_unmap_data(req->queue, rq);
1770         nvme_complete_rq(rq);
1771 }
1772
1773 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1774 {
1775         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1776
1777         return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1778 }
1779
1780 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1781         .queue_rq       = nvme_rdma_queue_rq,
1782         .complete       = nvme_rdma_complete_rq,
1783         .init_request   = nvme_rdma_init_request,
1784         .exit_request   = nvme_rdma_exit_request,
1785         .init_hctx      = nvme_rdma_init_hctx,
1786         .poll           = nvme_rdma_poll,
1787         .timeout        = nvme_rdma_timeout,
1788         .map_queues     = nvme_rdma_map_queues,
1789 };
1790
1791 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1792         .queue_rq       = nvme_rdma_queue_rq,
1793         .complete       = nvme_rdma_complete_rq,
1794         .init_request   = nvme_rdma_init_request,
1795         .exit_request   = nvme_rdma_exit_request,
1796         .init_hctx      = nvme_rdma_init_admin_hctx,
1797         .timeout        = nvme_rdma_timeout,
1798 };
1799
1800 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1801 {
1802         nvme_rdma_teardown_io_queues(ctrl, shutdown);
1803         if (shutdown)
1804                 nvme_shutdown_ctrl(&ctrl->ctrl);
1805         else
1806                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1807         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1808 }
1809
1810 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1811 {
1812         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1813 }
1814
1815 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1816 {
1817         struct nvme_rdma_ctrl *ctrl =
1818                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1819
1820         nvme_stop_ctrl(&ctrl->ctrl);
1821         nvme_rdma_shutdown_ctrl(ctrl, false);
1822
1823         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1824                 /* state change failure should never happen */
1825                 WARN_ON_ONCE(1);
1826                 return;
1827         }
1828
1829         if (nvme_rdma_setup_ctrl(ctrl, false))
1830                 goto out_fail;
1831
1832         return;
1833
1834 out_fail:
1835         ++ctrl->ctrl.nr_reconnects;
1836         nvme_rdma_reconnect_or_remove(ctrl);
1837 }
1838
1839 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1840         .name                   = "rdma",
1841         .module                 = THIS_MODULE,
1842         .flags                  = NVME_F_FABRICS,
1843         .reg_read32             = nvmf_reg_read32,
1844         .reg_read64             = nvmf_reg_read64,
1845         .reg_write32            = nvmf_reg_write32,
1846         .free_ctrl              = nvme_rdma_free_ctrl,
1847         .submit_async_event     = nvme_rdma_submit_async_event,
1848         .delete_ctrl            = nvme_rdma_delete_ctrl,
1849         .get_address            = nvmf_get_address,
1850         .stop_ctrl              = nvme_rdma_stop_ctrl,
1851 };
1852
1853 static inline bool
1854 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1855         struct nvmf_ctrl_options *opts)
1856 {
1857         char *stdport = __stringify(NVME_RDMA_IP_PORT);
1858
1859
1860         if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1861             strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1862                 return false;
1863
1864         if (opts->mask & NVMF_OPT_TRSVCID &&
1865             ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1866                 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1867                         return false;
1868         } else if (opts->mask & NVMF_OPT_TRSVCID) {
1869                 if (strcmp(opts->trsvcid, stdport))
1870                         return false;
1871         } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1872                 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1873                         return false;
1874         }
1875         /* else, it's a match as both have stdport. Fall to next checks */
1876
1877         /*
1878          * checking the local address is rough. In most cases, one
1879          * is not specified and the host port is selected by the stack.
1880          *
1881          * Assume no match if:
1882          *  local address is specified and address is not the same
1883          *  local address is not specified but remote is, or vice versa
1884          *    (admin using specific host_traddr when it matters).
1885          */
1886         if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1887             ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1888                 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1889                         return false;
1890         } else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1891                    ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1892                 return false;
1893         /*
1894          * if neither controller had an host port specified, assume it's
1895          * a match as everything else matched.
1896          */
1897
1898         return true;
1899 }
1900
1901 /*
1902  * Fails a connection request if it matches an existing controller
1903  * (association) with the same tuple:
1904  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1905  *
1906  * if local address is not specified in the request, it will match an
1907  * existing controller with all the other parameters the same and no
1908  * local port address specified as well.
1909  *
1910  * The ports don't need to be compared as they are intrinsically
1911  * already matched by the port pointers supplied.
1912  */
1913 static bool
1914 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1915 {
1916         struct nvme_rdma_ctrl *ctrl;
1917         bool found = false;
1918
1919         mutex_lock(&nvme_rdma_ctrl_mutex);
1920         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1921                 found = __nvme_rdma_options_match(ctrl, opts);
1922                 if (found)
1923                         break;
1924         }
1925         mutex_unlock(&nvme_rdma_ctrl_mutex);
1926
1927         return found;
1928 }
1929
1930 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1931                 struct nvmf_ctrl_options *opts)
1932 {
1933         struct nvme_rdma_ctrl *ctrl;
1934         int ret;
1935         bool changed;
1936         char *port;
1937
1938         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1939         if (!ctrl)
1940                 return ERR_PTR(-ENOMEM);
1941         ctrl->ctrl.opts = opts;
1942         INIT_LIST_HEAD(&ctrl->list);
1943
1944         if (opts->mask & NVMF_OPT_TRSVCID)
1945                 port = opts->trsvcid;
1946         else
1947                 port = __stringify(NVME_RDMA_IP_PORT);
1948
1949         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1950                         opts->traddr, port, &ctrl->addr);
1951         if (ret) {
1952                 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1953                 goto out_free_ctrl;
1954         }
1955
1956         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1957                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1958                         opts->host_traddr, NULL, &ctrl->src_addr);
1959                 if (ret) {
1960                         pr_err("malformed src address passed: %s\n",
1961                                opts->host_traddr);
1962                         goto out_free_ctrl;
1963                 }
1964         }
1965
1966         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1967                 ret = -EALREADY;
1968                 goto out_free_ctrl;
1969         }
1970
1971         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1972                         nvme_rdma_reconnect_ctrl_work);
1973         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1974         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1975
1976         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1977         ctrl->ctrl.sqsize = opts->queue_size - 1;
1978         ctrl->ctrl.kato = opts->kato;
1979
1980         ret = -ENOMEM;
1981         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1982                                 GFP_KERNEL);
1983         if (!ctrl->queues)
1984                 goto out_free_ctrl;
1985
1986         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1987                                 0 /* no quirks, we're perfect! */);
1988         if (ret)
1989                 goto out_kfree_queues;
1990
1991         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1992         WARN_ON_ONCE(!changed);
1993
1994         ret = nvme_rdma_setup_ctrl(ctrl, true);
1995         if (ret)
1996                 goto out_uninit_ctrl;
1997
1998         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1999                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2000
2001         nvme_get_ctrl(&ctrl->ctrl);
2002
2003         mutex_lock(&nvme_rdma_ctrl_mutex);
2004         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2005         mutex_unlock(&nvme_rdma_ctrl_mutex);
2006
2007         return &ctrl->ctrl;
2008
2009 out_uninit_ctrl:
2010         nvme_uninit_ctrl(&ctrl->ctrl);
2011         nvme_put_ctrl(&ctrl->ctrl);
2012         if (ret > 0)
2013                 ret = -EIO;
2014         return ERR_PTR(ret);
2015 out_kfree_queues:
2016         kfree(ctrl->queues);
2017 out_free_ctrl:
2018         kfree(ctrl);
2019         return ERR_PTR(ret);
2020 }
2021
2022 static struct nvmf_transport_ops nvme_rdma_transport = {
2023         .name           = "rdma",
2024         .module         = THIS_MODULE,
2025         .required_opts  = NVMF_OPT_TRADDR,
2026         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2027                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2028         .create_ctrl    = nvme_rdma_create_ctrl,
2029 };
2030
2031 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2032 {
2033         struct nvme_rdma_ctrl *ctrl;
2034         struct nvme_rdma_device *ndev;
2035         bool found = false;
2036
2037         mutex_lock(&device_list_mutex);
2038         list_for_each_entry(ndev, &device_list, entry) {
2039                 if (ndev->dev == ib_device) {
2040                         found = true;
2041                         break;
2042                 }
2043         }
2044         mutex_unlock(&device_list_mutex);
2045
2046         if (!found)
2047                 return;
2048
2049         /* Delete all controllers using this device */
2050         mutex_lock(&nvme_rdma_ctrl_mutex);
2051         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2052                 if (ctrl->device->dev != ib_device)
2053                         continue;
2054                 nvme_delete_ctrl(&ctrl->ctrl);
2055         }
2056         mutex_unlock(&nvme_rdma_ctrl_mutex);
2057
2058         flush_workqueue(nvme_delete_wq);
2059 }
2060
2061 static struct ib_client nvme_rdma_ib_client = {
2062         .name   = "nvme_rdma",
2063         .remove = nvme_rdma_remove_one
2064 };
2065
2066 static int __init nvme_rdma_init_module(void)
2067 {
2068         int ret;
2069
2070         ret = ib_register_client(&nvme_rdma_ib_client);
2071         if (ret)
2072                 return ret;
2073
2074         ret = nvmf_register_transport(&nvme_rdma_transport);
2075         if (ret)
2076                 goto err_unreg_client;
2077
2078         return 0;
2079
2080 err_unreg_client:
2081         ib_unregister_client(&nvme_rdma_ib_client);
2082         return ret;
2083 }
2084
2085 static void __exit nvme_rdma_cleanup_module(void)
2086 {
2087         nvmf_unregister_transport(&nvme_rdma_transport);
2088         ib_unregister_client(&nvme_rdma_ib_client);
2089 }
2090
2091 module_init(nvme_rdma_init_module);
2092 module_exit(nvme_rdma_cleanup_module);
2093
2094 MODULE_LICENSE("GPL v2");