drivers/nvme/host/rdma.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics RDMA host code.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/slab.h>
  10 #include <rdma/mr_pool.h>
  11 #include <linux/err.h>
  12 #include <linux/string.h>
  13 #include <linux/atomic.h>
  14 #include <linux/blk-mq.h>
  15 #include <linux/blk-mq-rdma.h>
  16 #include <linux/types.h>
  17 #include <linux/list.h>
  18 #include <linux/mutex.h>
  19 #include <linux/scatterlist.h>
  20 #include <linux/nvme.h>
  21 #include <asm/unaligned.h>
  22
  23 #include <rdma/ib_verbs.h>
  24 #include <rdma/rdma_cm.h>
  25 #include <linux/nvme-rdma.h>
  26
  27 #include "nvme.h"
  28 #include "fabrics.h"
  29
  30
  31 #define NVME_RDMA_CONNECT_TIMEOUT_MS    3000            /* 3 second */
  32
  33 #define NVME_RDMA_MAX_SEGMENTS          256
  34
  35 #define NVME_RDMA_MAX_INLINE_SEGMENTS   4
  36
  37 struct nvme_rdma_device {
  38         struct ib_device        *dev;
  39         struct ib_pd            *pd;
  40         struct kref             ref;
  41         struct list_head        entry;
  42         unsigned int            num_inline_segments;
  43 };
  44
  45 struct nvme_rdma_qe {
  46         struct ib_cqe           cqe;
  47         void                    *data;
  48         u64                     dma;
  49 };
  50
  51 struct nvme_rdma_queue;
  52 struct nvme_rdma_request {
  53         struct nvme_request     req;
  54         struct ib_mr            *mr;
  55         struct nvme_rdma_qe     sqe;
  56         union nvme_result       result;
  57         __le16                  status;
  58         refcount_t              ref;
  59         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  60         u32                     num_sge;
  61         int                     nents;
  62         struct ib_reg_wr        reg_wr;
  63         struct ib_cqe           reg_cqe;
  64         struct nvme_rdma_queue  *queue;
  65         struct sg_table         sg_table;
  66         struct scatterlist      first_sgl[];
  67 };
  68
  69 enum nvme_rdma_queue_flags {
  70         NVME_RDMA_Q_ALLOCATED           = 0,
  71         NVME_RDMA_Q_LIVE                = 1,
  72         NVME_RDMA_Q_TR_READY            = 2,
  73 };
  74
  75 struct nvme_rdma_queue {
  76         struct nvme_rdma_qe     *rsp_ring;
  77         int                     queue_size;
  78         size_t                  cmnd_capsule_len;
  79         struct nvme_rdma_ctrl   *ctrl;
  80         struct nvme_rdma_device *device;
  81         struct ib_cq            *ib_cq;
  82         struct ib_qp            *qp;
  83
  84         unsigned long           flags;
  85         struct rdma_cm_id       *cm_id;
  86         int                     cm_error;
  87         struct completion       cm_done;
  88 };
  89
  90 struct nvme_rdma_ctrl {
  91         /* read only in the hot path */
  92         struct nvme_rdma_queue  *queues;
  93
  94         /* other member variables */
  95         struct blk_mq_tag_set   tag_set;
  96         struct work_struct      err_work;
  97
  98         struct nvme_rdma_qe     async_event_sqe;
  99
 100         struct delayed_work     reconnect_work;
 101
 102         struct list_head        list;
 103
 104         struct blk_mq_tag_set   admin_tag_set;
 105         struct nvme_rdma_device *device;
 106
 107         u32                     max_fr_pages;
 108
 109         struct sockaddr_storage addr;
 110         struct sockaddr_storage src_addr;
 111
 112         struct nvme_ctrl        ctrl;
 113         bool                    use_inline_data;
 114         u32                     io_queues[HCTX_MAX_TYPES];
 115 };
 116
 117 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 118 {
 119         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 120 }
 121
 122 static LIST_HEAD(device_list);
 123 static DEFINE_MUTEX(device_list_mutex);
 124
 125 static LIST_HEAD(nvme_rdma_ctrl_list);
 126 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 127
 128 /*
 129  * Disabling this option makes small I/O goes faster, but is fundamentally
 130  * unsafe.  With it turned off we will have to register a global rkey that
 131  * allows read and write access to all physical memory.
 132  */
 133 static bool register_always = true;
 134 module_param(register_always, bool, 0444);
 135 MODULE_PARM_DESC(register_always,
 136          "Use memory registration even for contiguous memory regions");
 137
 138 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 139                 struct rdma_cm_event *event);
 140 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 141
 142 static const struct blk_mq_ops nvme_rdma_mq_ops;
 143 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 144
 145 /* XXX: really should move to a generic header sooner or later.. */
 146 static inline void put_unaligned_le24(u32 val, u8 *p)
 147 {
 148         *p++ = val;
 149         *p++ = val >> 8;
 150         *p++ = val >> 16;
 151 }
 152
 153 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 154 {
 155         return queue - queue->ctrl->queues;
 156 }
 157
 158 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
 159 {
 160         return nvme_rdma_queue_idx(queue) >
 161                 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
 162                 queue->ctrl->io_queues[HCTX_TYPE_READ];
 163 }
 164
 165 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 166 {
 167         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 168 }
 169
 170 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 171                 size_t capsule_size, enum dma_data_direction dir)
 172 {
 173         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 174         kfree(qe->data);
 175 }
 176
 177 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 178                 size_t capsule_size, enum dma_data_direction dir)
 179 {
 180         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 181         if (!qe->data)
 182                 return -ENOMEM;
 183
 184         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 185         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 186                 kfree(qe->data);
 187                 qe->data = NULL;
 188                 return -ENOMEM;
 189         }
 190
 191         return 0;
 192 }
 193
 194 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 195                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 196                 size_t capsule_size, enum dma_data_direction dir)
 197 {
 198         int i;
 199
 200         for (i = 0; i < ib_queue_size; i++)
 201                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 202         kfree(ring);
 203 }
 204
 205 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 206                 size_t ib_queue_size, size_t capsule_size,
 207                 enum dma_data_direction dir)
 208 {
 209         struct nvme_rdma_qe *ring;
 210         int i;
 211
 212         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 213         if (!ring)
 214                 return NULL;
 215
 216         for (i = 0; i < ib_queue_size; i++) {
 217                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 218                         goto out_free_ring;
 219         }
 220
 221         return ring;
 222
 223 out_free_ring:
 224         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 225         return NULL;
 226 }
 227
 228 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 229 {
 230         pr_debug("QP event %s (%d)\n",
 231                  ib_event_msg(event->event), event->event);
 232
 233 }
 234
 235 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 236 {
 237         int ret;
 238
 239         ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
 240                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 241         if (ret < 0)
 242                 return ret;
 243         if (ret == 0)
 244                 return -ETIMEDOUT;
 245         WARN_ON_ONCE(queue->cm_error > 0);
 246         return queue->cm_error;
 247 }
 248
 249 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 250 {
 251         struct nvme_rdma_device *dev = queue->device;
 252         struct ib_qp_init_attr init_attr;
 253         int ret;
 254
 255         memset(&init_attr, 0, sizeof(init_attr));
 256         init_attr.event_handler = nvme_rdma_qp_event;
 257         /* +1 for drain */
 258         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 259         /* +1 for drain */
 260         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 261         init_attr.cap.max_recv_sge = 1;
 262         init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
 263         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 264         init_attr.qp_type = IB_QPT_RC;
 265         init_attr.send_cq = queue->ib_cq;
 266         init_attr.recv_cq = queue->ib_cq;
 267
 268         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 269
 270         queue->qp = queue->cm_id->qp;
 271         return ret;
 272 }
 273
 274 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 275                 struct request *rq, unsigned int hctx_idx)
 276 {
 277         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 278         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 279         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 280         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 281         struct nvme_rdma_device *dev = queue->device;
 282
 283         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 284                         DMA_TO_DEVICE);
 285 }
 286
 287 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 288                 struct request *rq, unsigned int hctx_idx,
 289                 unsigned int numa_node)
 290 {
 291         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 292         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 293         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 294         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 295         struct nvme_rdma_device *dev = queue->device;
 296         struct ib_device *ibdev = dev->dev;
 297         int ret;
 298
 299         nvme_req(rq)->ctrl = &ctrl->ctrl;
 300         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 301                         DMA_TO_DEVICE);
 302         if (ret)
 303                 return ret;
 304
 305         req->queue = queue;
 306
 307         return 0;
 308 }
 309
 310 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 311                 unsigned int hctx_idx)
 312 {
 313         struct nvme_rdma_ctrl *ctrl = data;
 314         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 315
 316         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 317
 318         hctx->driver_data = queue;
 319         return 0;
 320 }
 321
 322 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 323                 unsigned int hctx_idx)
 324 {
 325         struct nvme_rdma_ctrl *ctrl = data;
 326         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 327
 328         BUG_ON(hctx_idx != 0);
 329
 330         hctx->driver_data = queue;
 331         return 0;
 332 }
 333
 334 static void nvme_rdma_free_dev(struct kref *ref)
 335 {
 336         struct nvme_rdma_device *ndev =
 337                 container_of(ref, struct nvme_rdma_device, ref);
 338
 339         mutex_lock(&device_list_mutex);
 340         list_del(&ndev->entry);
 341         mutex_unlock(&device_list_mutex);
 342
 343         ib_dealloc_pd(ndev->pd);
 344         kfree(ndev);
 345 }
 346
 347 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 348 {
 349         kref_put(&dev->ref, nvme_rdma_free_dev);
 350 }
 351
 352 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 353 {
 354         return kref_get_unless_zero(&dev->ref);
 355 }
 356
 357 static struct nvme_rdma_device *
 358 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 359 {
 360         struct nvme_rdma_device *ndev;
 361
 362         mutex_lock(&device_list_mutex);
 363         list_for_each_entry(ndev, &device_list, entry) {
 364                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 365                     nvme_rdma_dev_get(ndev))
 366                         goto out_unlock;
 367         }
 368
 369         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 370         if (!ndev)
 371                 goto out_err;
 372
 373         ndev->dev = cm_id->device;
 374         kref_init(&ndev->ref);
 375
 376         ndev->pd = ib_alloc_pd(ndev->dev,
 377                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 378         if (IS_ERR(ndev->pd))
 379                 goto out_free_dev;
 380
 381         if (!(ndev->dev->attrs.device_cap_flags &
 382               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 383                 dev_err(&ndev->dev->dev,
 384                         "Memory registrations not supported.\n");
 385                 goto out_free_pd;
 386         }
 387
 388         ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
 389                                         ndev->dev->attrs.max_send_sge - 1);
 390         list_add(&ndev->entry, &device_list);
 391 out_unlock:
 392         mutex_unlock(&device_list_mutex);
 393         return ndev;
 394
 395 out_free_pd:
 396         ib_dealloc_pd(ndev->pd);
 397 out_free_dev:
 398         kfree(ndev);
 399 out_err:
 400         mutex_unlock(&device_list_mutex);
 401         return NULL;
 402 }
 403
 404 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 405 {
 406         struct nvme_rdma_device *dev;
 407         struct ib_device *ibdev;
 408
 409         if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
 410                 return;
 411
 412         dev = queue->device;
 413         ibdev = dev->dev;
 414
 415         ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
 416
 417         /*
 418          * The cm_id object might have been destroyed during RDMA connection
 419          * establishment error flow to avoid getting other cma events, thus
 420          * the destruction of the QP shouldn't use rdma_cm API.
 421          */
 422         ib_destroy_qp(queue->qp);
 423         ib_free_cq(queue->ib_cq);
 424
 425         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 426                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 427
 428         nvme_rdma_dev_put(dev);
 429 }
 430
 431 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
 432 {
 433         return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
 434                      ibdev->attrs.max_fast_reg_page_list_len);
 435 }
 436
 437 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 438 {
 439         struct ib_device *ibdev;
 440         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 441         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 442         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 443         enum ib_poll_context poll_ctx;
 444         int ret;
 445
 446         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 447         if (!queue->device) {
 448                 dev_err(queue->cm_id->device->dev.parent,
 449                         "no client data found!\n");
 450                 return -ECONNREFUSED;
 451         }
 452         ibdev = queue->device->dev;
 453
 454         /*
 455          * Spread I/O queues completion vectors according their queue index.
 456          * Admin queues can always go on completion vector 0.
 457          */
 458         comp_vector = idx == 0 ? idx : idx - 1;
 459
 460         /* Polling queues need direct cq polling context */
 461         if (nvme_rdma_poll_queue(queue))
 462                 poll_ctx = IB_POLL_DIRECT;
 463         else
 464                 poll_ctx = IB_POLL_SOFTIRQ;
 465
 466         /* +1 for ib_stop_cq */
 467         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 468                                 cq_factor * queue->queue_size + 1,
 469                                 comp_vector, poll_ctx);
 470         if (IS_ERR(queue->ib_cq)) {
 471                 ret = PTR_ERR(queue->ib_cq);
 472                 goto out_put_dev;
 473         }
 474
 475         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 476         if (ret)
 477                 goto out_destroy_ib_cq;
 478
 479         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 480                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 481         if (!queue->rsp_ring) {
 482                 ret = -ENOMEM;
 483                 goto out_destroy_qp;
 484         }
 485
 486         ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
 487                               queue->queue_size,
 488                               IB_MR_TYPE_MEM_REG,
 489                               nvme_rdma_get_max_fr_pages(ibdev));
 490         if (ret) {
 491                 dev_err(queue->ctrl->ctrl.device,
 492                         "failed to initialize MR pool sized %d for QID %d\n",
 493                         queue->queue_size, idx);
 494                 goto out_destroy_ring;
 495         }
 496
 497         set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
 498
 499         return 0;
 500
 501 out_destroy_ring:
 502         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 503                             sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 504 out_destroy_qp:
 505         rdma_destroy_qp(queue->cm_id);
 506 out_destroy_ib_cq:
 507         ib_free_cq(queue->ib_cq);
 508 out_put_dev:
 509         nvme_rdma_dev_put(queue->device);
 510         return ret;
 511 }
 512
 513 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 514                 int idx, size_t queue_size)
 515 {
 516         struct nvme_rdma_queue *queue;
 517         struct sockaddr *src_addr = NULL;
 518         int ret;
 519
 520         queue = &ctrl->queues[idx];
 521         queue->ctrl = ctrl;
 522         init_completion(&queue->cm_done);
 523
 524         if (idx > 0)
 525                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 526         else
 527                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 528
 529         queue->queue_size = queue_size;
 530
 531         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 532                         RDMA_PS_TCP, IB_QPT_RC);
 533         if (IS_ERR(queue->cm_id)) {
 534                 dev_info(ctrl->ctrl.device,
 535                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 536                 return PTR_ERR(queue->cm_id);
 537         }
 538
 539         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 540                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 541
 542         queue->cm_error = -ETIMEDOUT;
 543         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 544                         (struct sockaddr *)&ctrl->addr,
 545                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 546         if (ret) {
 547                 dev_info(ctrl->ctrl.device,
 548                         "rdma_resolve_addr failed (%d).\n", ret);
 549                 goto out_destroy_cm_id;
 550         }
 551
 552         ret = nvme_rdma_wait_for_cm(queue);
 553         if (ret) {
 554                 dev_info(ctrl->ctrl.device,
 555                         "rdma connection establishment failed (%d)\n", ret);
 556                 goto out_destroy_cm_id;
 557         }
 558
 559         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 560
 561         return 0;
 562
 563 out_destroy_cm_id:
 564         rdma_destroy_id(queue->cm_id);
 565         nvme_rdma_destroy_queue_ib(queue);
 566         return ret;
 567 }
 568
 569 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 570 {
 571         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 572                 return;
 573
 574         rdma_disconnect(queue->cm_id);
 575         ib_drain_qp(queue->qp);
 576 }
 577
 578 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 579 {
 580         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 581                 return;
 582
 583         nvme_rdma_destroy_queue_ib(queue);
 584         rdma_destroy_id(queue->cm_id);
 585 }
 586
 587 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 588 {
 589         int i;
 590
 591         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 592                 nvme_rdma_free_queue(&ctrl->queues[i]);
 593 }
 594
 595 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 596 {
 597         int i;
 598
 599         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 600                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 601 }
 602
 603 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 604 {
 605         struct nvme_rdma_queue *queue = &ctrl->queues[idx];
 606         bool poll = nvme_rdma_poll_queue(queue);
 607         int ret;
 608
 609         if (idx)
 610                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
 611         else
 612                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 613
 614         if (!ret)
 615                 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
 616         else
 617                 dev_info(ctrl->ctrl.device,
 618                         "failed to connect queue: %d ret=%d\n", idx, ret);
 619         return ret;
 620 }
 621
 622 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 623 {
 624         int i, ret = 0;
 625
 626         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 627                 ret = nvme_rdma_start_queue(ctrl, i);
 628                 if (ret)
 629                         goto out_stop_queues;
 630         }
 631
 632         return 0;
 633
 634 out_stop_queues:
 635         for (i--; i >= 1; i--)
 636                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 637         return ret;
 638 }
 639
 640 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 641 {
 642         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 643         struct ib_device *ibdev = ctrl->device->dev;
 644         unsigned int nr_io_queues;
 645         int i, ret;
 646
 647         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 648
 649         /*
 650          * we map queues according to the device irq vectors for
 651          * optimal locality so we don't need more queues than
 652          * completion vectors.
 653          */
 654         nr_io_queues = min_t(unsigned int, nr_io_queues,
 655                                 ibdev->num_comp_vectors);
 656
 657         if (opts->nr_write_queues) {
 658                 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
 659                                 min(opts->nr_write_queues, nr_io_queues);
 660                 nr_io_queues += ctrl->io_queues[HCTX_TYPE_DEFAULT];
 661         } else {
 662                 ctrl->io_queues[HCTX_TYPE_DEFAULT] = nr_io_queues;
 663         }
 664
 665         ctrl->io_queues[HCTX_TYPE_READ] = nr_io_queues;
 666
 667         if (opts->nr_poll_queues) {
 668                 ctrl->io_queues[HCTX_TYPE_POLL] =
 669                         min(opts->nr_poll_queues, num_online_cpus());
 670                 nr_io_queues += ctrl->io_queues[HCTX_TYPE_POLL];
 671         }
 672
 673         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 674         if (ret)
 675                 return ret;
 676
 677         ctrl->ctrl.queue_count = nr_io_queues + 1;
 678         if (ctrl->ctrl.queue_count < 2)
 679                 return 0;
 680
 681         dev_info(ctrl->ctrl.device,
 682                 "creating %d I/O queues.\n", nr_io_queues);
 683
 684         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 685                 ret = nvme_rdma_alloc_queue(ctrl, i,
 686                                 ctrl->ctrl.sqsize + 1);
 687                 if (ret)
 688                         goto out_free_queues;
 689         }
 690
 691         return 0;
 692
 693 out_free_queues:
 694         for (i--; i >= 1; i--)
 695                 nvme_rdma_free_queue(&ctrl->queues[i]);
 696
 697         return ret;
 698 }
 699
 700 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 701                 bool admin)
 702 {
 703         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 704         struct blk_mq_tag_set *set;
 705         int ret;
 706
 707         if (admin) {
 708                 set = &ctrl->admin_tag_set;
 709                 memset(set, 0, sizeof(*set));
 710                 set->ops = &nvme_rdma_admin_mq_ops;
 711                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 712                 set->reserved_tags = 2; /* connect + keep-alive */
 713                 set->numa_node = nctrl->numa_node;
 714                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 715                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 716                 set->driver_data = ctrl;
 717                 set->nr_hw_queues = 1;
 718                 set->timeout = ADMIN_TIMEOUT;
 719                 set->flags = BLK_MQ_F_NO_SCHED;
 720         } else {
 721                 set = &ctrl->tag_set;
 722                 memset(set, 0, sizeof(*set));
 723                 set->ops = &nvme_rdma_mq_ops;
 724                 set->queue_depth = nctrl->sqsize + 1;
 725                 set->reserved_tags = 1; /* fabric connect */
 726                 set->numa_node = nctrl->numa_node;
 727                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 728                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 729                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 730                 set->driver_data = ctrl;
 731                 set->nr_hw_queues = nctrl->queue_count - 1;
 732                 set->timeout = NVME_IO_TIMEOUT;
 733                 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
 734         }
 735
 736         ret = blk_mq_alloc_tag_set(set);
 737         if (ret)
 738                 return ERR_PTR(ret);
 739
 740         return set;
 741 }
 742
 743 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 744                 bool remove)
 745 {
 746         if (remove) {
 747                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 748                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 749         }
 750         if (ctrl->async_event_sqe.data) {
 751                 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 752                                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 753                 ctrl->async_event_sqe.data = NULL;
 754         }
 755         nvme_rdma_free_queue(&ctrl->queues[0]);
 756 }
 757
 758 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 759                 bool new)
 760 {
 761         int error;
 762
 763         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 764         if (error)
 765                 return error;
 766
 767         ctrl->device = ctrl->queues[0].device;
 768         ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
 769
 770         ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
 771
 772         error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 773                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 774         if (error)
 775                 goto out_free_queue;
 776
 777         if (new) {
 778                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 779                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 780                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 781                         goto out_free_async_qe;
 782                 }
 783
 784                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 785                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 786                         error = PTR_ERR(ctrl->ctrl.admin_q);
 787                         goto out_free_tagset;
 788                 }
 789         }
 790
 791         error = nvme_rdma_start_queue(ctrl, 0);
 792         if (error)
 793                 goto out_cleanup_queue;
 794
 795         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 796                         &ctrl->ctrl.cap);
 797         if (error) {
 798                 dev_err(ctrl->ctrl.device,
 799                         "prop_get NVME_REG_CAP failed\n");
 800                 goto out_stop_queue;
 801         }
 802
 803         ctrl->ctrl.sqsize =
 804                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 805
 806         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 807         if (error)
 808                 goto out_stop_queue;
 809
 810         ctrl->ctrl.max_hw_sectors =
 811                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 812
 813         error = nvme_init_identify(&ctrl->ctrl);
 814         if (error)
 815                 goto out_stop_queue;
 816
 817         return 0;
 818
 819 out_stop_queue:
 820         nvme_rdma_stop_queue(&ctrl->queues[0]);
 821 out_cleanup_queue:
 822         if (new)
 823                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 824 out_free_tagset:
 825         if (new)
 826                 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
 827 out_free_async_qe:
 828         nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 829                 sizeof(struct nvme_command), DMA_TO_DEVICE);
 830         ctrl->async_event_sqe.data = NULL;
 831 out_free_queue:
 832         nvme_rdma_free_queue(&ctrl->queues[0]);
 833         return error;
 834 }
 835
 836 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 837                 bool remove)
 838 {
 839         if (remove) {
 840                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 841                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 842         }
 843         nvme_rdma_free_io_queues(ctrl);
 844 }
 845
 846 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 847 {
 848         int ret;
 849
 850         ret = nvme_rdma_alloc_io_queues(ctrl);
 851         if (ret)
 852                 return ret;
 853
 854         if (new) {
 855                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 856                 if (IS_ERR(ctrl->ctrl.tagset)) {
 857                         ret = PTR_ERR(ctrl->ctrl.tagset);
 858                         goto out_free_io_queues;
 859                 }
 860
 861                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 862                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 863                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 864                         goto out_free_tag_set;
 865                 }
 866         } else {
 867                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 868                         ctrl->ctrl.queue_count - 1);
 869         }
 870
 871         ret = nvme_rdma_start_io_queues(ctrl);
 872         if (ret)
 873                 goto out_cleanup_connect_q;
 874
 875         return 0;
 876
 877 out_cleanup_connect_q:
 878         if (new)
 879                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 880 out_free_tag_set:
 881         if (new)
 882                 blk_mq_free_tag_set(ctrl->ctrl.tagset);
 883 out_free_io_queues:
 884         nvme_rdma_free_io_queues(ctrl);
 885         return ret;
 886 }
 887
 888 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
 889                 bool remove)
 890 {
 891         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 892         nvme_rdma_stop_queue(&ctrl->queues[0]);
 893         if (ctrl->ctrl.admin_tagset)
 894                 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
 895                         nvme_cancel_request, &ctrl->ctrl);
 896         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 897         nvme_rdma_destroy_admin_queue(ctrl, remove);
 898 }
 899
 900 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
 901                 bool remove)
 902 {
 903         if (ctrl->ctrl.queue_count > 1) {
 904                 nvme_stop_queues(&ctrl->ctrl);
 905                 nvme_rdma_stop_io_queues(ctrl);
 906                 if (ctrl->ctrl.tagset)
 907                         blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
 908                                 nvme_cancel_request, &ctrl->ctrl);
 909                 if (remove)
 910                         nvme_start_queues(&ctrl->ctrl);
 911                 nvme_rdma_destroy_io_queues(ctrl, remove);
 912         }
 913 }
 914
 915 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 916 {
 917         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 918
 919         if (list_empty(&ctrl->list))
 920                 goto free_ctrl;
 921
 922         mutex_lock(&nvme_rdma_ctrl_mutex);
 923         list_del(&ctrl->list);
 924         mutex_unlock(&nvme_rdma_ctrl_mutex);
 925
 926         nvmf_free_options(nctrl->opts);
 927 free_ctrl:
 928         kfree(ctrl->queues);
 929         kfree(ctrl);
 930 }
 931
 932 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 933 {
 934         /* If we are resetting/deleting then do nothing */
 935         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
 936                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 937                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 938                 return;
 939         }
 940
 941         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 942                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 943                         ctrl->ctrl.opts->reconnect_delay);
 944                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 945                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 946         } else {
 947                 nvme_delete_ctrl(&ctrl->ctrl);
 948         }
 949 }
 950
 951 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
 952 {
 953         int ret = -EINVAL;
 954         bool changed;
 955
 956         ret = nvme_rdma_configure_admin_queue(ctrl, new);
 957         if (ret)
 958                 return ret;
 959
 960         if (ctrl->ctrl.icdoff) {
 961                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
 962                 goto destroy_admin;
 963         }
 964
 965         if (!(ctrl->ctrl.sgls & (1 << 2))) {
 966                 dev_err(ctrl->ctrl.device,
 967                         "Mandatory keyed sgls are not supported!\n");
 968                 goto destroy_admin;
 969         }
 970
 971         if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
 972                 dev_warn(ctrl->ctrl.device,
 973                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
 974                         ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
 975         }
 976
 977         if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
 978                 dev_warn(ctrl->ctrl.device,
 979                         "sqsize %u > ctrl maxcmd %u, clamping down\n",
 980                         ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
 981                 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
 982         }
 983
 984         if (ctrl->ctrl.sgls & (1 << 20))
 985                 ctrl->use_inline_data = true;
 986
 987         if (ctrl->ctrl.queue_count > 1) {
 988                 ret = nvme_rdma_configure_io_queues(ctrl, new);
 989                 if (ret)
 990                         goto destroy_admin;
 991         }
 992
 993         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 994         if (!changed) {
 995                 /* state change failure is ok if we're in DELETING state */
 996                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
 997                 ret = -EINVAL;
 998                 goto destroy_io;
 999         }
1000
1001         nvme_start_ctrl(&ctrl->ctrl);
1002         return 0;
1003
1004 destroy_io:
1005         if (ctrl->ctrl.queue_count > 1)
1006                 nvme_rdma_destroy_io_queues(ctrl, new);
1007 destroy_admin:
1008         nvme_rdma_stop_queue(&ctrl->queues[0]);
1009         nvme_rdma_destroy_admin_queue(ctrl, new);
1010         return ret;
1011 }
1012
1013 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1014 {
1015         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1016                         struct nvme_rdma_ctrl, reconnect_work);
1017
1018         ++ctrl->ctrl.nr_reconnects;
1019
1020         if (nvme_rdma_setup_ctrl(ctrl, false))
1021                 goto requeue;
1022
1023         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1024                         ctrl->ctrl.nr_reconnects);
1025
1026         ctrl->ctrl.nr_reconnects = 0;
1027
1028         return;
1029
1030 requeue:
1031         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1032                         ctrl->ctrl.nr_reconnects);
1033         nvme_rdma_reconnect_or_remove(ctrl);
1034 }
1035
1036 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1037 {
1038         struct nvme_rdma_ctrl *ctrl = container_of(work,
1039                         struct nvme_rdma_ctrl, err_work);
1040
1041         nvme_stop_keep_alive(&ctrl->ctrl);
1042         nvme_rdma_teardown_io_queues(ctrl, false);
1043         nvme_start_queues(&ctrl->ctrl);
1044         nvme_rdma_teardown_admin_queue(ctrl, false);
1045
1046         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1047                 /* state change failure is ok if we're in DELETING state */
1048                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1049                 return;
1050         }
1051
1052         nvme_rdma_reconnect_or_remove(ctrl);
1053 }
1054
1055 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1056 {
1057         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1058                 return;
1059
1060         queue_work(nvme_wq, &ctrl->err_work);
1061 }
1062
1063 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1064                 const char *op)
1065 {
1066         struct nvme_rdma_queue *queue = cq->cq_context;
1067         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1068
1069         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1070                 dev_info(ctrl->ctrl.device,
1071                              "%s for CQE 0x%p failed with status %s (%d)\n",
1072                              op, wc->wr_cqe,
1073                              ib_wc_status_msg(wc->status), wc->status);
1074         nvme_rdma_error_recovery(ctrl);
1075 }
1076
1077 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1078 {
1079         if (unlikely(wc->status != IB_WC_SUCCESS))
1080                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1081 }
1082
1083 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1084 {
1085         struct nvme_rdma_request *req =
1086                 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1087         struct request *rq = blk_mq_rq_from_pdu(req);
1088
1089         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1090                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1091                 return;
1092         }
1093
1094         if (refcount_dec_and_test(&req->ref))
1095                 nvme_end_request(rq, req->status, req->result);
1096
1097 }
1098
1099 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1100                 struct nvme_rdma_request *req)
1101 {
1102         struct ib_send_wr wr = {
1103                 .opcode             = IB_WR_LOCAL_INV,
1104                 .next               = NULL,
1105                 .num_sge            = 0,
1106                 .send_flags         = IB_SEND_SIGNALED,
1107                 .ex.invalidate_rkey = req->mr->rkey,
1108         };
1109
1110         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1111         wr.wr_cqe = &req->reg_cqe;
1112
1113         return ib_post_send(queue->qp, &wr, NULL);
1114 }
1115
1116 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1117                 struct request *rq)
1118 {
1119         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1120         struct nvme_rdma_device *dev = queue->device;
1121         struct ib_device *ibdev = dev->dev;
1122
1123         if (!blk_rq_nr_phys_segments(rq))
1124                 return;
1125
1126         if (req->mr) {
1127                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1128                 req->mr = NULL;
1129         }
1130
1131         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1132                         req->nents, rq_data_dir(rq) ==
1133                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1134
1135         nvme_cleanup_cmd(rq);
1136         sg_free_table_chained(&req->sg_table, true);
1137 }
1138
1139 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1140 {
1141         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1142
1143         sg->addr = 0;
1144         put_unaligned_le24(0, sg->length);
1145         put_unaligned_le32(0, sg->key);
1146         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1147         return 0;
1148 }
1149
1150 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1151                 struct nvme_rdma_request *req, struct nvme_command *c,
1152                 int count)
1153 {
1154         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1155         struct scatterlist *sgl = req->sg_table.sgl;
1156         struct ib_sge *sge = &req->sge[1];
1157         u32 len = 0;
1158         int i;
1159
1160         for (i = 0; i < count; i++, sgl++, sge++) {
1161                 sge->addr = sg_dma_address(sgl);
1162                 sge->length = sg_dma_len(sgl);
1163                 sge->lkey = queue->device->pd->local_dma_lkey;
1164                 len += sge->length;
1165         }
1166
1167         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1168         sg->length = cpu_to_le32(len);
1169         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1170
1171         req->num_sge += count;
1172         return 0;
1173 }
1174
1175 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1176                 struct nvme_rdma_request *req, struct nvme_command *c)
1177 {
1178         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1179
1180         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1181         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1182         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1183         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1184         return 0;
1185 }
1186
1187 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1188                 struct nvme_rdma_request *req, struct nvme_command *c,
1189                 int count)
1190 {
1191         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1192         int nr;
1193
1194         req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1195         if (WARN_ON_ONCE(!req->mr))
1196                 return -EAGAIN;
1197
1198         /*
1199          * Align the MR to a 4K page size to match the ctrl page size and
1200          * the block virtual boundary.
1201          */
1202         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1203         if (unlikely(nr < count)) {
1204                 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1205                 req->mr = NULL;
1206                 if (nr < 0)
1207                         return nr;
1208                 return -EINVAL;
1209         }
1210
1211         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1212
1213         req->reg_cqe.done = nvme_rdma_memreg_done;
1214         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1215         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1216         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1217         req->reg_wr.wr.num_sge = 0;
1218         req->reg_wr.mr = req->mr;
1219         req->reg_wr.key = req->mr->rkey;
1220         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1221                              IB_ACCESS_REMOTE_READ |
1222                              IB_ACCESS_REMOTE_WRITE;
1223
1224         sg->addr = cpu_to_le64(req->mr->iova);
1225         put_unaligned_le24(req->mr->length, sg->length);
1226         put_unaligned_le32(req->mr->rkey, sg->key);
1227         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1228                         NVME_SGL_FMT_INVALIDATE;
1229
1230         return 0;
1231 }
1232
1233 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1234                 struct request *rq, struct nvme_command *c)
1235 {
1236         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1237         struct nvme_rdma_device *dev = queue->device;
1238         struct ib_device *ibdev = dev->dev;
1239         int count, ret;
1240
1241         req->num_sge = 1;
1242         refcount_set(&req->ref, 2); /* send and recv completions */
1243
1244         c->common.flags |= NVME_CMD_SGL_METABUF;
1245
1246         if (!blk_rq_nr_phys_segments(rq))
1247                 return nvme_rdma_set_sg_null(c);
1248
1249         req->sg_table.sgl = req->first_sgl;
1250         ret = sg_alloc_table_chained(&req->sg_table,
1251                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1252         if (ret)
1253                 return -ENOMEM;
1254
1255         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1256
1257         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1258                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1259         if (unlikely(count <= 0)) {
1260                 ret = -EIO;
1261                 goto out_free_table;
1262         }
1263
1264         if (count <= dev->num_inline_segments) {
1265                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1266                     queue->ctrl->use_inline_data &&
1267                     blk_rq_payload_bytes(rq) <=
1268                                 nvme_rdma_inline_data_size(queue)) {
1269                         ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1270                         goto out;
1271                 }
1272
1273                 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1274                         ret = nvme_rdma_map_sg_single(queue, req, c);
1275                         goto out;
1276                 }
1277         }
1278
1279         ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1280 out:
1281         if (unlikely(ret))
1282                 goto out_unmap_sg;
1283
1284         return 0;
1285
1286 out_unmap_sg:
1287         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1288                         req->nents, rq_data_dir(rq) ==
1289                         WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1290 out_free_table:
1291         sg_free_table_chained(&req->sg_table, true);
1292         return ret;
1293 }
1294
1295 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1296 {
1297         struct nvme_rdma_qe *qe =
1298                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1299         struct nvme_rdma_request *req =
1300                 container_of(qe, struct nvme_rdma_request, sqe);
1301         struct request *rq = blk_mq_rq_from_pdu(req);
1302
1303         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1304                 nvme_rdma_wr_error(cq, wc, "SEND");
1305                 return;
1306         }
1307
1308         if (refcount_dec_and_test(&req->ref))
1309                 nvme_end_request(rq, req->status, req->result);
1310 }
1311
1312 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1313                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1314                 struct ib_send_wr *first)
1315 {
1316         struct ib_send_wr wr;
1317         int ret;
1318
1319         sge->addr   = qe->dma;
1320         sge->length = sizeof(struct nvme_command),
1321         sge->lkey   = queue->device->pd->local_dma_lkey;
1322
1323         wr.next       = NULL;
1324         wr.wr_cqe     = &qe->cqe;
1325         wr.sg_list    = sge;
1326         wr.num_sge    = num_sge;
1327         wr.opcode     = IB_WR_SEND;
1328         wr.send_flags = IB_SEND_SIGNALED;
1329
1330         if (first)
1331                 first->next = &wr;
1332         else
1333                 first = &wr;
1334
1335         ret = ib_post_send(queue->qp, first, NULL);
1336         if (unlikely(ret)) {
1337                 dev_err(queue->ctrl->ctrl.device,
1338                              "%s failed with error code %d\n", __func__, ret);
1339         }
1340         return ret;
1341 }
1342
1343 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1344                 struct nvme_rdma_qe *qe)
1345 {
1346         struct ib_recv_wr wr;
1347         struct ib_sge list;
1348         int ret;
1349
1350         list.addr   = qe->dma;
1351         list.length = sizeof(struct nvme_completion);
1352         list.lkey   = queue->device->pd->local_dma_lkey;
1353
1354         qe->cqe.done = nvme_rdma_recv_done;
1355
1356         wr.next     = NULL;
1357         wr.wr_cqe   = &qe->cqe;
1358         wr.sg_list  = &list;
1359         wr.num_sge  = 1;
1360
1361         ret = ib_post_recv(queue->qp, &wr, NULL);
1362         if (unlikely(ret)) {
1363                 dev_err(queue->ctrl->ctrl.device,
1364                         "%s failed with error code %d\n", __func__, ret);
1365         }
1366         return ret;
1367 }
1368
1369 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1370 {
1371         u32 queue_idx = nvme_rdma_queue_idx(queue);
1372
1373         if (queue_idx == 0)
1374                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1375         return queue->ctrl->tag_set.tags[queue_idx - 1];
1376 }
1377
1378 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1379 {
1380         if (unlikely(wc->status != IB_WC_SUCCESS))
1381                 nvme_rdma_wr_error(cq, wc, "ASYNC");
1382 }
1383
1384 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1385 {
1386         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1387         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1388         struct ib_device *dev = queue->device->dev;
1389         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1390         struct nvme_command *cmd = sqe->data;
1391         struct ib_sge sge;
1392         int ret;
1393
1394         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1395
1396         memset(cmd, 0, sizeof(*cmd));
1397         cmd->common.opcode = nvme_admin_async_event;
1398         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1399         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1400         nvme_rdma_set_sg_null(cmd);
1401
1402         sqe->cqe.done = nvme_rdma_async_done;
1403
1404         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1405                         DMA_TO_DEVICE);
1406
1407         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1408         WARN_ON_ONCE(ret);
1409 }
1410
1411 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1412                 struct nvme_completion *cqe, struct ib_wc *wc)
1413 {
1414         struct request *rq;
1415         struct nvme_rdma_request *req;
1416
1417         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1418         if (!rq) {
1419                 dev_err(queue->ctrl->ctrl.device,
1420                         "tag 0x%x on QP %#x not found\n",
1421                         cqe->command_id, queue->qp->qp_num);
1422                 nvme_rdma_error_recovery(queue->ctrl);
1423                 return;
1424         }
1425         req = blk_mq_rq_to_pdu(rq);
1426
1427         req->status = cqe->status;
1428         req->result = cqe->result;
1429
1430         if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1431                 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1432                         dev_err(queue->ctrl->ctrl.device,
1433                                 "Bogus remote invalidation for rkey %#x\n",
1434                                 req->mr->rkey);
1435                         nvme_rdma_error_recovery(queue->ctrl);
1436                 }
1437         } else if (req->mr) {
1438                 int ret;
1439
1440                 ret = nvme_rdma_inv_rkey(queue, req);
1441                 if (unlikely(ret < 0)) {
1442                         dev_err(queue->ctrl->ctrl.device,
1443                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1444                                 req->mr->rkey, ret);
1445                         nvme_rdma_error_recovery(queue->ctrl);
1446                 }
1447                 /* the local invalidation completion will end the request */
1448                 return;
1449         }
1450
1451         if (refcount_dec_and_test(&req->ref))
1452                 nvme_end_request(rq, req->status, req->result);
1453 }
1454
1455 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1456 {
1457         struct nvme_rdma_qe *qe =
1458                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1459         struct nvme_rdma_queue *queue = cq->cq_context;
1460         struct ib_device *ibdev = queue->device->dev;
1461         struct nvme_completion *cqe = qe->data;
1462         const size_t len = sizeof(struct nvme_completion);
1463
1464         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1465                 nvme_rdma_wr_error(cq, wc, "RECV");
1466                 return;
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                 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1482         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1483
1484         nvme_rdma_post_recv(queue, qe);
1485 }
1486
1487 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1488 {
1489         int ret, i;
1490
1491         for (i = 0; i < queue->queue_size; i++) {
1492                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1493                 if (ret)
1494                         goto out_destroy_queue_ib;
1495         }
1496
1497         return 0;
1498
1499 out_destroy_queue_ib:
1500         nvme_rdma_destroy_queue_ib(queue);
1501         return ret;
1502 }
1503
1504 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1505                 struct rdma_cm_event *ev)
1506 {
1507         struct rdma_cm_id *cm_id = queue->cm_id;
1508         int status = ev->status;
1509         const char *rej_msg;
1510         const struct nvme_rdma_cm_rej *rej_data;
1511         u8 rej_data_len;
1512
1513         rej_msg = rdma_reject_msg(cm_id, status);
1514         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1515
1516         if (rej_data && rej_data_len >= sizeof(u16)) {
1517                 u16 sts = le16_to_cpu(rej_data->sts);
1518
1519                 dev_err(queue->ctrl->ctrl.device,
1520                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1521                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1522         } else {
1523                 dev_err(queue->ctrl->ctrl.device,
1524                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1525         }
1526
1527         return -ECONNRESET;
1528 }
1529
1530 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1531 {
1532         int ret;
1533
1534         ret = nvme_rdma_create_queue_ib(queue);
1535         if (ret)
1536                 return ret;
1537
1538         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1539         if (ret) {
1540                 dev_err(queue->ctrl->ctrl.device,
1541                         "rdma_resolve_route failed (%d).\n",
1542                         queue->cm_error);
1543                 goto out_destroy_queue;
1544         }
1545
1546         return 0;
1547
1548 out_destroy_queue:
1549         nvme_rdma_destroy_queue_ib(queue);
1550         return ret;
1551 }
1552
1553 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1554 {
1555         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1556         struct rdma_conn_param param = { };
1557         struct nvme_rdma_cm_req priv = { };
1558         int ret;
1559
1560         param.qp_num = queue->qp->qp_num;
1561         param.flow_control = 1;
1562
1563         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1564         /* maximum retry count */
1565         param.retry_count = 7;
1566         param.rnr_retry_count = 7;
1567         param.private_data = &priv;
1568         param.private_data_len = sizeof(priv);
1569
1570         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1571         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1572         /*
1573          * set the admin queue depth to the minimum size
1574          * specified by the Fabrics standard.
1575          */
1576         if (priv.qid == 0) {
1577                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1578                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1579         } else {
1580                 /*
1581                  * current interpretation of the fabrics spec
1582                  * is at minimum you make hrqsize sqsize+1, or a
1583                  * 1's based representation of sqsize.
1584                  */
1585                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1586                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1587         }
1588
1589         ret = rdma_connect(queue->cm_id, &param);
1590         if (ret) {
1591                 dev_err(ctrl->ctrl.device,
1592                         "rdma_connect failed (%d).\n", ret);
1593                 goto out_destroy_queue_ib;
1594         }
1595
1596         return 0;
1597
1598 out_destroy_queue_ib:
1599         nvme_rdma_destroy_queue_ib(queue);
1600         return ret;
1601 }
1602
1603 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1604                 struct rdma_cm_event *ev)
1605 {
1606         struct nvme_rdma_queue *queue = cm_id->context;
1607         int cm_error = 0;
1608
1609         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1610                 rdma_event_msg(ev->event), ev->event,
1611                 ev->status, cm_id);
1612
1613         switch (ev->event) {
1614         case RDMA_CM_EVENT_ADDR_RESOLVED:
1615                 cm_error = nvme_rdma_addr_resolved(queue);
1616                 break;
1617         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1618                 cm_error = nvme_rdma_route_resolved(queue);
1619                 break;
1620         case RDMA_CM_EVENT_ESTABLISHED:
1621                 queue->cm_error = nvme_rdma_conn_established(queue);
1622                 /* complete cm_done regardless of success/failure */
1623                 complete(&queue->cm_done);
1624                 return 0;
1625         case RDMA_CM_EVENT_REJECTED:
1626                 nvme_rdma_destroy_queue_ib(queue);
1627                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1628                 break;
1629         case RDMA_CM_EVENT_ROUTE_ERROR:
1630         case RDMA_CM_EVENT_CONNECT_ERROR:
1631         case RDMA_CM_EVENT_UNREACHABLE:
1632                 nvme_rdma_destroy_queue_ib(queue);
1633                 /* fall through */
1634         case RDMA_CM_EVENT_ADDR_ERROR:
1635                 dev_dbg(queue->ctrl->ctrl.device,
1636                         "CM error event %d\n", ev->event);
1637                 cm_error = -ECONNRESET;
1638                 break;
1639         case RDMA_CM_EVENT_DISCONNECTED:
1640         case RDMA_CM_EVENT_ADDR_CHANGE:
1641         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1642                 dev_dbg(queue->ctrl->ctrl.device,
1643                         "disconnect received - connection closed\n");
1644                 nvme_rdma_error_recovery(queue->ctrl);
1645                 break;
1646         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1647                 /* device removal is handled via the ib_client API */
1648                 break;
1649         default:
1650                 dev_err(queue->ctrl->ctrl.device,
1651                         "Unexpected RDMA CM event (%d)\n", ev->event);
1652                 nvme_rdma_error_recovery(queue->ctrl);
1653                 break;
1654         }
1655
1656         if (cm_error) {
1657                 queue->cm_error = cm_error;
1658                 complete(&queue->cm_done);
1659         }
1660
1661         return 0;
1662 }
1663
1664 static enum blk_eh_timer_return
1665 nvme_rdma_timeout(struct request *rq, bool reserved)
1666 {
1667         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1668         struct nvme_rdma_queue *queue = req->queue;
1669         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1670
1671         dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1672                  rq->tag, nvme_rdma_queue_idx(queue));
1673
1674         if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1675                 /*
1676                  * Teardown immediately if controller times out while starting
1677                  * or we are already started error recovery. all outstanding
1678                  * requests are completed on shutdown, so we return BLK_EH_DONE.
1679                  */
1680                 flush_work(&ctrl->err_work);
1681                 nvme_rdma_teardown_io_queues(ctrl, false);
1682                 nvme_rdma_teardown_admin_queue(ctrl, false);
1683                 return BLK_EH_DONE;
1684         }
1685
1686         dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1687         nvme_rdma_error_recovery(ctrl);
1688
1689         return BLK_EH_RESET_TIMER;
1690 }
1691
1692 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1693                 const struct blk_mq_queue_data *bd)
1694 {
1695         struct nvme_ns *ns = hctx->queue->queuedata;
1696         struct nvme_rdma_queue *queue = hctx->driver_data;
1697         struct request *rq = bd->rq;
1698         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1699         struct nvme_rdma_qe *sqe = &req->sqe;
1700         struct nvme_command *c = sqe->data;
1701         struct ib_device *dev;
1702         bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1703         blk_status_t ret;
1704         int err;
1705
1706         WARN_ON_ONCE(rq->tag < 0);
1707
1708         if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1709                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1710
1711         dev = queue->device->dev;
1712         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1713                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1714
1715         ret = nvme_setup_cmd(ns, rq, c);
1716         if (ret)
1717                 return ret;
1718
1719         blk_mq_start_request(rq);
1720
1721         err = nvme_rdma_map_data(queue, rq, c);
1722         if (unlikely(err < 0)) {
1723                 dev_err(queue->ctrl->ctrl.device,
1724                              "Failed to map data (%d)\n", err);
1725                 nvme_cleanup_cmd(rq);
1726                 goto err;
1727         }
1728
1729         sqe->cqe.done = nvme_rdma_send_done;
1730
1731         ib_dma_sync_single_for_device(dev, sqe->dma,
1732                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1733
1734         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1735                         req->mr ? &req->reg_wr.wr : NULL);
1736         if (unlikely(err)) {
1737                 nvme_rdma_unmap_data(queue, rq);
1738                 goto err;
1739         }
1740
1741         return BLK_STS_OK;
1742 err:
1743         if (err == -ENOMEM || err == -EAGAIN)
1744                 return BLK_STS_RESOURCE;
1745         return BLK_STS_IOERR;
1746 }
1747
1748 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
1749 {
1750         struct nvme_rdma_queue *queue = hctx->driver_data;
1751
1752         return ib_process_cq_direct(queue->ib_cq, -1);
1753 }
1754
1755 static void nvme_rdma_complete_rq(struct request *rq)
1756 {
1757         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1758
1759         nvme_rdma_unmap_data(req->queue, rq);
1760         nvme_complete_rq(rq);
1761 }
1762
1763 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1764 {
1765         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1766
1767         set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1768         set->map[HCTX_TYPE_DEFAULT].nr_queues =
1769                         ctrl->io_queues[HCTX_TYPE_DEFAULT];
1770         set->map[HCTX_TYPE_READ].nr_queues = ctrl->io_queues[HCTX_TYPE_READ];
1771         if (ctrl->ctrl.opts->nr_write_queues) {
1772                 /* separate read/write queues */
1773                 set->map[HCTX_TYPE_READ].queue_offset =
1774                                 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1775         } else {
1776                 /* mixed read/write queues */
1777                 set->map[HCTX_TYPE_READ].queue_offset = 0;
1778         }
1779         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1780                         ctrl->device->dev, 0);
1781         blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1782                         ctrl->device->dev, 0);
1783
1784         if (ctrl->ctrl.opts->nr_poll_queues) {
1785                 set->map[HCTX_TYPE_POLL].nr_queues =
1786                                 ctrl->io_queues[HCTX_TYPE_POLL];
1787                 set->map[HCTX_TYPE_POLL].queue_offset =
1788                                 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1789                 if (ctrl->ctrl.opts->nr_write_queues)
1790                         set->map[HCTX_TYPE_POLL].queue_offset +=
1791                                 ctrl->io_queues[HCTX_TYPE_READ];
1792                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1793         }
1794         return 0;
1795 }
1796
1797 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1798         .queue_rq       = nvme_rdma_queue_rq,
1799         .complete       = nvme_rdma_complete_rq,
1800         .init_request   = nvme_rdma_init_request,
1801         .exit_request   = nvme_rdma_exit_request,
1802         .init_hctx      = nvme_rdma_init_hctx,
1803         .timeout        = nvme_rdma_timeout,
1804         .map_queues     = nvme_rdma_map_queues,
1805         .poll           = nvme_rdma_poll,
1806 };
1807
1808 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1809         .queue_rq       = nvme_rdma_queue_rq,
1810         .complete       = nvme_rdma_complete_rq,
1811         .init_request   = nvme_rdma_init_request,
1812         .exit_request   = nvme_rdma_exit_request,
1813         .init_hctx      = nvme_rdma_init_admin_hctx,
1814         .timeout        = nvme_rdma_timeout,
1815 };
1816
1817 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1818 {
1819         cancel_work_sync(&ctrl->err_work);
1820         cancel_delayed_work_sync(&ctrl->reconnect_work);
1821
1822         nvme_rdma_teardown_io_queues(ctrl, shutdown);
1823         if (shutdown)
1824                 nvme_shutdown_ctrl(&ctrl->ctrl);
1825         else
1826                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1827         nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1828 }
1829
1830 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1831 {
1832         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1833 }
1834
1835 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1836 {
1837         struct nvme_rdma_ctrl *ctrl =
1838                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1839
1840         nvme_stop_ctrl(&ctrl->ctrl);
1841         nvme_rdma_shutdown_ctrl(ctrl, false);
1842
1843         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1844                 /* state change failure should never happen */
1845                 WARN_ON_ONCE(1);
1846                 return;
1847         }
1848
1849         if (nvme_rdma_setup_ctrl(ctrl, false))
1850                 goto out_fail;
1851
1852         return;
1853
1854 out_fail:
1855         ++ctrl->ctrl.nr_reconnects;
1856         nvme_rdma_reconnect_or_remove(ctrl);
1857 }
1858
1859 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1860         .name                   = "rdma",
1861         .module                 = THIS_MODULE,
1862         .flags                  = NVME_F_FABRICS,
1863         .reg_read32             = nvmf_reg_read32,
1864         .reg_read64             = nvmf_reg_read64,
1865         .reg_write32            = nvmf_reg_write32,
1866         .free_ctrl              = nvme_rdma_free_ctrl,
1867         .submit_async_event     = nvme_rdma_submit_async_event,
1868         .delete_ctrl            = nvme_rdma_delete_ctrl,
1869         .get_address            = nvmf_get_address,
1870 };
1871
1872 /*
1873  * Fails a connection request if it matches an existing controller
1874  * (association) with the same tuple:
1875  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1876  *
1877  * if local address is not specified in the request, it will match an
1878  * existing controller with all the other parameters the same and no
1879  * local port address specified as well.
1880  *
1881  * The ports don't need to be compared as they are intrinsically
1882  * already matched by the port pointers supplied.
1883  */
1884 static bool
1885 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1886 {
1887         struct nvme_rdma_ctrl *ctrl;
1888         bool found = false;
1889
1890         mutex_lock(&nvme_rdma_ctrl_mutex);
1891         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1892                 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1893                 if (found)
1894                         break;
1895         }
1896         mutex_unlock(&nvme_rdma_ctrl_mutex);
1897
1898         return found;
1899 }
1900
1901 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1902                 struct nvmf_ctrl_options *opts)
1903 {
1904         struct nvme_rdma_ctrl *ctrl;
1905         int ret;
1906         bool changed;
1907
1908         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1909         if (!ctrl)
1910                 return ERR_PTR(-ENOMEM);
1911         ctrl->ctrl.opts = opts;
1912         INIT_LIST_HEAD(&ctrl->list);
1913
1914         if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1915                 opts->trsvcid =
1916                         kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1917                 if (!opts->trsvcid) {
1918                         ret = -ENOMEM;
1919                         goto out_free_ctrl;
1920                 }
1921                 opts->mask |= NVMF_OPT_TRSVCID;
1922         }
1923
1924         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1925                         opts->traddr, opts->trsvcid, &ctrl->addr);
1926         if (ret) {
1927                 pr_err("malformed address passed: %s:%s\n",
1928                         opts->traddr, opts->trsvcid);
1929                 goto out_free_ctrl;
1930         }
1931
1932         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1933                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1934                         opts->host_traddr, NULL, &ctrl->src_addr);
1935                 if (ret) {
1936                         pr_err("malformed src address passed: %s\n",
1937                                opts->host_traddr);
1938                         goto out_free_ctrl;
1939                 }
1940         }
1941
1942         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1943                 ret = -EALREADY;
1944                 goto out_free_ctrl;
1945         }
1946
1947         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1948                         nvme_rdma_reconnect_ctrl_work);
1949         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1950         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1951
1952         ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
1953                                 opts->nr_poll_queues + 1;
1954         ctrl->ctrl.sqsize = opts->queue_size - 1;
1955         ctrl->ctrl.kato = opts->kato;
1956
1957         ret = -ENOMEM;
1958         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1959                                 GFP_KERNEL);
1960         if (!ctrl->queues)
1961                 goto out_free_ctrl;
1962
1963         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1964                                 0 /* no quirks, we're perfect! */);
1965         if (ret)
1966                 goto out_kfree_queues;
1967
1968         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1969         WARN_ON_ONCE(!changed);
1970
1971         ret = nvme_rdma_setup_ctrl(ctrl, true);
1972         if (ret)
1973                 goto out_uninit_ctrl;
1974
1975         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1976                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1977
1978         nvme_get_ctrl(&ctrl->ctrl);
1979
1980         mutex_lock(&nvme_rdma_ctrl_mutex);
1981         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1982         mutex_unlock(&nvme_rdma_ctrl_mutex);
1983
1984         return &ctrl->ctrl;
1985
1986 out_uninit_ctrl:
1987         nvme_uninit_ctrl(&ctrl->ctrl);
1988         nvme_put_ctrl(&ctrl->ctrl);
1989         if (ret > 0)
1990                 ret = -EIO;
1991         return ERR_PTR(ret);
1992 out_kfree_queues:
1993         kfree(ctrl->queues);
1994 out_free_ctrl:
1995         kfree(ctrl);
1996         return ERR_PTR(ret);
1997 }
1998
1999 static struct nvmf_transport_ops nvme_rdma_transport = {
2000         .name           = "rdma",
2001         .module         = THIS_MODULE,
2002         .required_opts  = NVMF_OPT_TRADDR,
2003         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2004                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2005                           NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES,
2006         .create_ctrl    = nvme_rdma_create_ctrl,
2007 };
2008
2009 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2010 {
2011         struct nvme_rdma_ctrl *ctrl;
2012         struct nvme_rdma_device *ndev;
2013         bool found = false;
2014
2015         mutex_lock(&device_list_mutex);
2016         list_for_each_entry(ndev, &device_list, entry) {
2017                 if (ndev->dev == ib_device) {
2018                         found = true;
2019                         break;
2020                 }
2021         }
2022         mutex_unlock(&device_list_mutex);
2023
2024         if (!found)
2025                 return;
2026
2027         /* Delete all controllers using this device */
2028         mutex_lock(&nvme_rdma_ctrl_mutex);
2029         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2030                 if (ctrl->device->dev != ib_device)
2031                         continue;
2032                 nvme_delete_ctrl(&ctrl->ctrl);
2033         }
2034         mutex_unlock(&nvme_rdma_ctrl_mutex);
2035
2036         flush_workqueue(nvme_delete_wq);
2037 }
2038
2039 static struct ib_client nvme_rdma_ib_client = {
2040         .name   = "nvme_rdma",
2041         .remove = nvme_rdma_remove_one
2042 };
2043
2044 static int __init nvme_rdma_init_module(void)
2045 {
2046         int ret;
2047
2048         ret = ib_register_client(&nvme_rdma_ib_client);
2049         if (ret)
2050                 return ret;
2051
2052         ret = nvmf_register_transport(&nvme_rdma_transport);
2053         if (ret)
2054                 goto err_unreg_client;
2055
2056         return 0;
2057
2058 err_unreg_client:
2059         ib_unregister_client(&nvme_rdma_ib_client);
2060         return ret;
2061 }
2062
2063 static void __exit nvme_rdma_cleanup_module(void)
2064 {
2065         nvmf_unregister_transport(&nvme_rdma_transport);
2066         ib_unregister_client(&nvme_rdma_ib_client);
2067 }
2068
2069 module_init(nvme_rdma_init_module);
2070 module_exit(nvme_rdma_cleanup_module);
2071
2072 MODULE_LICENSE("GPL v2");