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 <linux/err.h>
  19 #include <linux/string.h>
  20 #include <linux/atomic.h>
  21 #include <linux/blk-mq.h>
  22 #include <linux/types.h>
  23 #include <linux/list.h>
  24 #include <linux/mutex.h>
  25 #include <linux/scatterlist.h>
  26 #include <linux/nvme.h>
  27 #include <asm/unaligned.h>
  28
  29 #include <rdma/ib_verbs.h>
  30 #include <rdma/rdma_cm.h>
  31 #include <linux/nvme-rdma.h>
  32
  33 #include "nvme.h"
  34 #include "fabrics.h"
  35
  36
  37 #define NVME_RDMA_CONNECT_TIMEOUT_MS    3000            /* 3 second */
  38
  39 #define NVME_RDMA_MAX_SEGMENT_SIZE      0xffffff        /* 24-bit SGL field */
  40
  41 #define NVME_RDMA_MAX_SEGMENTS          256
  42
  43 #define NVME_RDMA_MAX_INLINE_SEGMENTS   1
  44
  45 /*
  46  * We handle AEN commands ourselves and don't even let the
  47  * block layer know about them.
  48  */
  49 #define NVME_RDMA_NR_AEN_COMMANDS      1
  50 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
  51         (NVME_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
  52
  53 struct nvme_rdma_device {
  54         struct ib_device       *dev;
  55         struct ib_pd           *pd;
  56         struct kref             ref;
  57         struct list_head        entry;
  58 };
  59
  60 struct nvme_rdma_qe {
  61         struct ib_cqe           cqe;
  62         void                    *data;
  63         u64                     dma;
  64 };
  65
  66 struct nvme_rdma_queue;
  67 struct nvme_rdma_request {
  68         struct nvme_request     req;
  69         struct ib_mr            *mr;
  70         struct nvme_rdma_qe     sqe;
  71         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  72         u32                     num_sge;
  73         int                     nents;
  74         bool                    inline_data;
  75         struct ib_reg_wr        reg_wr;
  76         struct ib_cqe           reg_cqe;
  77         struct nvme_rdma_queue  *queue;
  78         struct sg_table         sg_table;
  79         struct scatterlist      first_sgl[];
  80 };
  81
  82 enum nvme_rdma_queue_flags {
  83         NVME_RDMA_Q_LIVE                = 0,
  84         NVME_RDMA_Q_DELETING            = 1,
  85 };
  86
  87 struct nvme_rdma_queue {
  88         struct nvme_rdma_qe     *rsp_ring;
  89         atomic_t                sig_count;
  90         int                     queue_size;
  91         size_t                  cmnd_capsule_len;
  92         struct nvme_rdma_ctrl   *ctrl;
  93         struct nvme_rdma_device *device;
  94         struct ib_cq            *ib_cq;
  95         struct ib_qp            *qp;
  96
  97         unsigned long           flags;
  98         struct rdma_cm_id       *cm_id;
  99         int                     cm_error;
 100         struct completion       cm_done;
 101 };
 102
 103 struct nvme_rdma_ctrl {
 104         /* read only in the hot path */
 105         struct nvme_rdma_queue  *queues;
 106
 107         /* other member variables */
 108         struct blk_mq_tag_set   tag_set;
 109         struct work_struct      delete_work;
 110         struct work_struct      err_work;
 111
 112         struct nvme_rdma_qe     async_event_sqe;
 113
 114         struct delayed_work     reconnect_work;
 115
 116         struct list_head        list;
 117
 118         struct blk_mq_tag_set   admin_tag_set;
 119         struct nvme_rdma_device *device;
 120
 121         u32                     max_fr_pages;
 122
 123         struct sockaddr_storage addr;
 124         struct sockaddr_storage src_addr;
 125
 126         struct nvme_ctrl        ctrl;
 127 };
 128
 129 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 130 {
 131         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 132 }
 133
 134 static LIST_HEAD(device_list);
 135 static DEFINE_MUTEX(device_list_mutex);
 136
 137 static LIST_HEAD(nvme_rdma_ctrl_list);
 138 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 139
 140 /*
 141  * Disabling this option makes small I/O goes faster, but is fundamentally
 142  * unsafe.  With it turned off we will have to register a global rkey that
 143  * allows read and write access to all physical memory.
 144  */
 145 static bool register_always = true;
 146 module_param(register_always, bool, 0444);
 147 MODULE_PARM_DESC(register_always,
 148          "Use memory registration even for contiguous memory regions");
 149
 150 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 151                 struct rdma_cm_event *event);
 152 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 153
 154 /* XXX: really should move to a generic header sooner or later.. */
 155 static inline void put_unaligned_le24(u32 val, u8 *p)
 156 {
 157         *p++ = val;
 158         *p++ = val >> 8;
 159         *p++ = val >> 16;
 160 }
 161
 162 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 163 {
 164         return queue - queue->ctrl->queues;
 165 }
 166
 167 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 168 {
 169         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 170 }
 171
 172 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 173                 size_t capsule_size, enum dma_data_direction dir)
 174 {
 175         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 176         kfree(qe->data);
 177 }
 178
 179 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 180                 size_t capsule_size, enum dma_data_direction dir)
 181 {
 182         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 183         if (!qe->data)
 184                 return -ENOMEM;
 185
 186         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 187         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 188                 kfree(qe->data);
 189                 return -ENOMEM;
 190         }
 191
 192         return 0;
 193 }
 194
 195 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 196                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 197                 size_t capsule_size, enum dma_data_direction dir)
 198 {
 199         int i;
 200
 201         for (i = 0; i < ib_queue_size; i++)
 202                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 203         kfree(ring);
 204 }
 205
 206 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 207                 size_t ib_queue_size, size_t capsule_size,
 208                 enum dma_data_direction dir)
 209 {
 210         struct nvme_rdma_qe *ring;
 211         int i;
 212
 213         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 214         if (!ring)
 215                 return NULL;
 216
 217         for (i = 0; i < ib_queue_size; i++) {
 218                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 219                         goto out_free_ring;
 220         }
 221
 222         return ring;
 223
 224 out_free_ring:
 225         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 226         return NULL;
 227 }
 228
 229 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 230 {
 231         pr_debug("QP event %s (%d)\n",
 232                  ib_event_msg(event->event), event->event);
 233
 234 }
 235
 236 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 237 {
 238         wait_for_completion_interruptible_timeout(&queue->cm_done,
 239                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 240         return queue->cm_error;
 241 }
 242
 243 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 244 {
 245         struct nvme_rdma_device *dev = queue->device;
 246         struct ib_qp_init_attr init_attr;
 247         int ret;
 248
 249         memset(&init_attr, 0, sizeof(init_attr));
 250         init_attr.event_handler = nvme_rdma_qp_event;
 251         /* +1 for drain */
 252         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 253         /* +1 for drain */
 254         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 255         init_attr.cap.max_recv_sge = 1;
 256         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 257         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 258         init_attr.qp_type = IB_QPT_RC;
 259         init_attr.send_cq = queue->ib_cq;
 260         init_attr.recv_cq = queue->ib_cq;
 261
 262         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 263
 264         queue->qp = queue->cm_id->qp;
 265         return ret;
 266 }
 267
 268 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 269 {
 270         struct nvme_rdma_ctrl *ctrl = data;
 271         struct nvme_rdma_device *dev = ctrl->device;
 272         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 273         int ret = 0;
 274
 275         ib_dereg_mr(req->mr);
 276
 277         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 278                         ctrl->max_fr_pages);
 279         if (IS_ERR(req->mr)) {
 280                 ret = PTR_ERR(req->mr);
 281                 req->mr = NULL;
 282                 goto out;
 283         }
 284
 285         req->mr->need_inval = false;
 286
 287 out:
 288         return ret;
 289 }
 290
 291 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 292                 struct request *rq, unsigned int hctx_idx)
 293 {
 294         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 295         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 296         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 297         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 298         struct nvme_rdma_device *dev = queue->device;
 299
 300         if (req->mr)
 301                 ib_dereg_mr(req->mr);
 302
 303         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 304                         DMA_TO_DEVICE);
 305 }
 306
 307 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 308                 struct request *rq, unsigned int hctx_idx,
 309                 unsigned int numa_node)
 310 {
 311         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 312         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 313         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 314         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 315         struct nvme_rdma_device *dev = queue->device;
 316         struct ib_device *ibdev = dev->dev;
 317         int ret;
 318
 319         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 320                         DMA_TO_DEVICE);
 321         if (ret)
 322                 return ret;
 323
 324         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 325                         ctrl->max_fr_pages);
 326         if (IS_ERR(req->mr)) {
 327                 ret = PTR_ERR(req->mr);
 328                 goto out_free_qe;
 329         }
 330
 331         req->queue = queue;
 332
 333         return 0;
 334
 335 out_free_qe:
 336         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 337                         DMA_TO_DEVICE);
 338         return -ENOMEM;
 339 }
 340
 341 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 342                 unsigned int hctx_idx)
 343 {
 344         struct nvme_rdma_ctrl *ctrl = data;
 345         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 346
 347         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 348
 349         hctx->driver_data = queue;
 350         return 0;
 351 }
 352
 353 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 354                 unsigned int hctx_idx)
 355 {
 356         struct nvme_rdma_ctrl *ctrl = data;
 357         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 358
 359         BUG_ON(hctx_idx != 0);
 360
 361         hctx->driver_data = queue;
 362         return 0;
 363 }
 364
 365 static void nvme_rdma_free_dev(struct kref *ref)
 366 {
 367         struct nvme_rdma_device *ndev =
 368                 container_of(ref, struct nvme_rdma_device, ref);
 369
 370         mutex_lock(&device_list_mutex);
 371         list_del(&ndev->entry);
 372         mutex_unlock(&device_list_mutex);
 373
 374         ib_dealloc_pd(ndev->pd);
 375         kfree(ndev);
 376 }
 377
 378 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 379 {
 380         kref_put(&dev->ref, nvme_rdma_free_dev);
 381 }
 382
 383 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 384 {
 385         return kref_get_unless_zero(&dev->ref);
 386 }
 387
 388 static struct nvme_rdma_device *
 389 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 390 {
 391         struct nvme_rdma_device *ndev;
 392
 393         mutex_lock(&device_list_mutex);
 394         list_for_each_entry(ndev, &device_list, entry) {
 395                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 396                     nvme_rdma_dev_get(ndev))
 397                         goto out_unlock;
 398         }
 399
 400         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 401         if (!ndev)
 402                 goto out_err;
 403
 404         ndev->dev = cm_id->device;
 405         kref_init(&ndev->ref);
 406
 407         ndev->pd = ib_alloc_pd(ndev->dev,
 408                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 409         if (IS_ERR(ndev->pd))
 410                 goto out_free_dev;
 411
 412         if (!(ndev->dev->attrs.device_cap_flags &
 413               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 414                 dev_err(&ndev->dev->dev,
 415                         "Memory registrations not supported.\n");
 416                 goto out_free_pd;
 417         }
 418
 419         list_add(&ndev->entry, &device_list);
 420 out_unlock:
 421         mutex_unlock(&device_list_mutex);
 422         return ndev;
 423
 424 out_free_pd:
 425         ib_dealloc_pd(ndev->pd);
 426 out_free_dev:
 427         kfree(ndev);
 428 out_err:
 429         mutex_unlock(&device_list_mutex);
 430         return NULL;
 431 }
 432
 433 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 434 {
 435         struct nvme_rdma_device *dev;
 436         struct ib_device *ibdev;
 437
 438         dev = queue->device;
 439         ibdev = dev->dev;
 440         rdma_destroy_qp(queue->cm_id);
 441         ib_free_cq(queue->ib_cq);
 442
 443         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 444                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 445
 446         nvme_rdma_dev_put(dev);
 447 }
 448
 449 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 450 {
 451         struct ib_device *ibdev;
 452         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 453         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 454         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 455         int ret;
 456
 457         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 458         if (!queue->device) {
 459                 dev_err(queue->cm_id->device->dev.parent,
 460                         "no client data found!\n");
 461                 return -ECONNREFUSED;
 462         }
 463         ibdev = queue->device->dev;
 464
 465         /*
 466          * The admin queue is barely used once the controller is live, so don't
 467          * bother to spread it out.
 468          */
 469         if (idx == 0)
 470                 comp_vector = 0;
 471         else
 472                 comp_vector = idx % ibdev->num_comp_vectors;
 473
 474
 475         /* +1 for ib_stop_cq */
 476         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 477                                 cq_factor * queue->queue_size + 1,
 478                                 comp_vector, IB_POLL_SOFTIRQ);
 479         if (IS_ERR(queue->ib_cq)) {
 480                 ret = PTR_ERR(queue->ib_cq);
 481                 goto out_put_dev;
 482         }
 483
 484         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 485         if (ret)
 486                 goto out_destroy_ib_cq;
 487
 488         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 489                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 490         if (!queue->rsp_ring) {
 491                 ret = -ENOMEM;
 492                 goto out_destroy_qp;
 493         }
 494
 495         return 0;
 496
 497 out_destroy_qp:
 498         ib_destroy_qp(queue->qp);
 499 out_destroy_ib_cq:
 500         ib_free_cq(queue->ib_cq);
 501 out_put_dev:
 502         nvme_rdma_dev_put(queue->device);
 503         return ret;
 504 }
 505
 506 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
 507                 int idx, size_t queue_size)
 508 {
 509         struct nvme_rdma_queue *queue;
 510         struct sockaddr *src_addr = NULL;
 511         int ret;
 512
 513         queue = &ctrl->queues[idx];
 514         queue->ctrl = ctrl;
 515         init_completion(&queue->cm_done);
 516
 517         if (idx > 0)
 518                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 519         else
 520                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 521
 522         queue->queue_size = queue_size;
 523         atomic_set(&queue->sig_count, 0);
 524
 525         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 526                         RDMA_PS_TCP, IB_QPT_RC);
 527         if (IS_ERR(queue->cm_id)) {
 528                 dev_info(ctrl->ctrl.device,
 529                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 530                 return PTR_ERR(queue->cm_id);
 531         }
 532
 533         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 534                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 535
 536         queue->cm_error = -ETIMEDOUT;
 537         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 538                         (struct sockaddr *)&ctrl->addr,
 539                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 540         if (ret) {
 541                 dev_info(ctrl->ctrl.device,
 542                         "rdma_resolve_addr failed (%d).\n", ret);
 543                 goto out_destroy_cm_id;
 544         }
 545
 546         ret = nvme_rdma_wait_for_cm(queue);
 547         if (ret) {
 548                 dev_info(ctrl->ctrl.device,
 549                         "rdma_resolve_addr wait failed (%d).\n", ret);
 550                 goto out_destroy_cm_id;
 551         }
 552
 553         clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
 554
 555         return 0;
 556
 557 out_destroy_cm_id:
 558         rdma_destroy_id(queue->cm_id);
 559         return ret;
 560 }
 561
 562 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 563 {
 564         rdma_disconnect(queue->cm_id);
 565         ib_drain_qp(queue->qp);
 566 }
 567
 568 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 569 {
 570         nvme_rdma_destroy_queue_ib(queue);
 571         rdma_destroy_id(queue->cm_id);
 572 }
 573
 574 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
 575 {
 576         if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
 577                 return;
 578         nvme_rdma_stop_queue(queue);
 579         nvme_rdma_free_queue(queue);
 580 }
 581
 582 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 583 {
 584         int i;
 585
 586         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 587                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 588 }
 589
 590 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
 591 {
 592         int i, ret = 0;
 593
 594         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 595                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 596                 if (ret) {
 597                         dev_info(ctrl->ctrl.device,
 598                                 "failed to connect i/o queue: %d\n", ret);
 599                         goto out_free_queues;
 600                 }
 601                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 602         }
 603
 604         return 0;
 605
 606 out_free_queues:
 607         nvme_rdma_free_io_queues(ctrl);
 608         return ret;
 609 }
 610
 611 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
 612 {
 613         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 614         unsigned int nr_io_queues;
 615         int i, ret;
 616
 617         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 618         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 619         if (ret)
 620                 return ret;
 621
 622         ctrl->ctrl.queue_count = nr_io_queues + 1;
 623         if (ctrl->ctrl.queue_count < 2)
 624                 return 0;
 625
 626         dev_info(ctrl->ctrl.device,
 627                 "creating %d I/O queues.\n", nr_io_queues);
 628
 629         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 630                 ret = nvme_rdma_init_queue(ctrl, i,
 631                                            ctrl->ctrl.opts->queue_size);
 632                 if (ret) {
 633                         dev_info(ctrl->ctrl.device,
 634                                 "failed to initialize i/o queue: %d\n", ret);
 635                         goto out_free_queues;
 636                 }
 637         }
 638
 639         return 0;
 640
 641 out_free_queues:
 642         for (i--; i >= 1; i--)
 643                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 644
 645         return ret;
 646 }
 647
 648 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
 649 {
 650         nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
 651                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 652         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 653         blk_cleanup_queue(ctrl->ctrl.admin_q);
 654         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 655         nvme_rdma_dev_put(ctrl->device);
 656 }
 657
 658 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 659 {
 660         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 661
 662         if (list_empty(&ctrl->list))
 663                 goto free_ctrl;
 664
 665         mutex_lock(&nvme_rdma_ctrl_mutex);
 666         list_del(&ctrl->list);
 667         mutex_unlock(&nvme_rdma_ctrl_mutex);
 668
 669         kfree(ctrl->queues);
 670         nvmf_free_options(nctrl->opts);
 671 free_ctrl:
 672         kfree(ctrl);
 673 }
 674
 675 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 676 {
 677         /* If we are resetting/deleting then do nothing */
 678         if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
 679                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 680                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 681                 return;
 682         }
 683
 684         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 685                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 686                         ctrl->ctrl.opts->reconnect_delay);
 687                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 688                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 689         } else {
 690                 dev_info(ctrl->ctrl.device, "Removing controller...\n");
 691                 queue_work(nvme_wq, &ctrl->delete_work);
 692         }
 693 }
 694
 695 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 696 {
 697         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 698                         struct nvme_rdma_ctrl, reconnect_work);
 699         bool changed;
 700         int ret;
 701
 702         ++ctrl->ctrl.nr_reconnects;
 703
 704         if (ctrl->ctrl.queue_count > 1) {
 705                 nvme_rdma_free_io_queues(ctrl);
 706
 707                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
 708                 if (ret)
 709                         goto requeue;
 710         }
 711
 712         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 713
 714         ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
 715         if (ret)
 716                 goto requeue;
 717
 718         ret = nvme_rdma_init_queue(ctrl, 0, NVME_AQ_DEPTH);
 719         if (ret)
 720                 goto requeue;
 721
 722         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 723         if (ret)
 724                 goto requeue;
 725
 726         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
 727
 728         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 729         if (ret)
 730                 goto requeue;
 731
 732         if (ctrl->ctrl.queue_count > 1) {
 733                 ret = nvme_rdma_init_io_queues(ctrl);
 734                 if (ret)
 735                         goto requeue;
 736
 737                 ret = nvme_rdma_connect_io_queues(ctrl);
 738                 if (ret)
 739                         goto requeue;
 740
 741                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 742                                 ctrl->ctrl.queue_count - 1);
 743         }
 744
 745         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 746         WARN_ON_ONCE(!changed);
 747         ctrl->ctrl.nr_reconnects = 0;
 748
 749         nvme_start_ctrl(&ctrl->ctrl);
 750
 751         dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
 752
 753         return;
 754
 755 requeue:
 756         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
 757                         ctrl->ctrl.nr_reconnects);
 758         nvme_rdma_reconnect_or_remove(ctrl);
 759 }
 760
 761 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 762 {
 763         struct nvme_rdma_ctrl *ctrl = container_of(work,
 764                         struct nvme_rdma_ctrl, err_work);
 765         int i;
 766
 767         nvme_stop_ctrl(&ctrl->ctrl);
 768
 769         for (i = 0; i < ctrl->ctrl.queue_count; i++)
 770                 clear_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 771
 772         if (ctrl->ctrl.queue_count > 1)
 773                 nvme_stop_queues(&ctrl->ctrl);
 774         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 775
 776         /* We must take care of fastfail/requeue all our inflight requests */
 777         if (ctrl->ctrl.queue_count > 1)
 778                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 779                                         nvme_cancel_request, &ctrl->ctrl);
 780         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 781                                 nvme_cancel_request, &ctrl->ctrl);
 782
 783         /*
 784          * queues are not a live anymore, so restart the queues to fail fast
 785          * new IO
 786          */
 787         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 788         nvme_start_queues(&ctrl->ctrl);
 789
 790         nvme_rdma_reconnect_or_remove(ctrl);
 791 }
 792
 793 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 794 {
 795         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 796                 return;
 797
 798         queue_work(nvme_wq, &ctrl->err_work);
 799 }
 800
 801 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 802                 const char *op)
 803 {
 804         struct nvme_rdma_queue *queue = cq->cq_context;
 805         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 806
 807         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 808                 dev_info(ctrl->ctrl.device,
 809                              "%s for CQE 0x%p failed with status %s (%d)\n",
 810                              op, wc->wr_cqe,
 811                              ib_wc_status_msg(wc->status), wc->status);
 812         nvme_rdma_error_recovery(ctrl);
 813 }
 814
 815 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
 816 {
 817         if (unlikely(wc->status != IB_WC_SUCCESS))
 818                 nvme_rdma_wr_error(cq, wc, "MEMREG");
 819 }
 820
 821 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 822 {
 823         if (unlikely(wc->status != IB_WC_SUCCESS))
 824                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
 825 }
 826
 827 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
 828                 struct nvme_rdma_request *req)
 829 {
 830         struct ib_send_wr *bad_wr;
 831         struct ib_send_wr wr = {
 832                 .opcode             = IB_WR_LOCAL_INV,
 833                 .next               = NULL,
 834                 .num_sge            = 0,
 835                 .send_flags         = 0,
 836                 .ex.invalidate_rkey = req->mr->rkey,
 837         };
 838
 839         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
 840         wr.wr_cqe = &req->reg_cqe;
 841
 842         return ib_post_send(queue->qp, &wr, &bad_wr);
 843 }
 844
 845 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 846                 struct request *rq)
 847 {
 848         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 849         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 850         struct nvme_rdma_device *dev = queue->device;
 851         struct ib_device *ibdev = dev->dev;
 852         int res;
 853
 854         if (!blk_rq_bytes(rq))
 855                 return;
 856
 857         if (req->mr->need_inval) {
 858                 res = nvme_rdma_inv_rkey(queue, req);
 859                 if (res < 0) {
 860                         dev_err(ctrl->ctrl.device,
 861                                 "Queueing INV WR for rkey %#x failed (%d)\n",
 862                                 req->mr->rkey, res);
 863                         nvme_rdma_error_recovery(queue->ctrl);
 864                 }
 865         }
 866
 867         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 868                         req->nents, rq_data_dir(rq) ==
 869                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 870
 871         nvme_cleanup_cmd(rq);
 872         sg_free_table_chained(&req->sg_table, true);
 873 }
 874
 875 static int nvme_rdma_set_sg_null(struct nvme_command *c)
 876 {
 877         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 878
 879         sg->addr = 0;
 880         put_unaligned_le24(0, sg->length);
 881         put_unaligned_le32(0, sg->key);
 882         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 883         return 0;
 884 }
 885
 886 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
 887                 struct nvme_rdma_request *req, struct nvme_command *c)
 888 {
 889         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
 890
 891         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
 892         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
 893         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
 894
 895         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
 896         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
 897         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
 898
 899         req->inline_data = true;
 900         req->num_sge++;
 901         return 0;
 902 }
 903
 904 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
 905                 struct nvme_rdma_request *req, struct nvme_command *c)
 906 {
 907         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 908
 909         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
 910         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
 911         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
 912         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 913         return 0;
 914 }
 915
 916 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 917                 struct nvme_rdma_request *req, struct nvme_command *c,
 918                 int count)
 919 {
 920         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 921         int nr;
 922
 923         /*
 924          * Align the MR to a 4K page size to match the ctrl page size and
 925          * the block virtual boundary.
 926          */
 927         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
 928         if (nr < count) {
 929                 if (nr < 0)
 930                         return nr;
 931                 return -EINVAL;
 932         }
 933
 934         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
 935
 936         req->reg_cqe.done = nvme_rdma_memreg_done;
 937         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
 938         req->reg_wr.wr.opcode = IB_WR_REG_MR;
 939         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
 940         req->reg_wr.wr.num_sge = 0;
 941         req->reg_wr.mr = req->mr;
 942         req->reg_wr.key = req->mr->rkey;
 943         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 944                              IB_ACCESS_REMOTE_READ |
 945                              IB_ACCESS_REMOTE_WRITE;
 946
 947         req->mr->need_inval = true;
 948
 949         sg->addr = cpu_to_le64(req->mr->iova);
 950         put_unaligned_le24(req->mr->length, sg->length);
 951         put_unaligned_le32(req->mr->rkey, sg->key);
 952         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
 953                         NVME_SGL_FMT_INVALIDATE;
 954
 955         return 0;
 956 }
 957
 958 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 959                 struct request *rq, struct nvme_command *c)
 960 {
 961         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 962         struct nvme_rdma_device *dev = queue->device;
 963         struct ib_device *ibdev = dev->dev;
 964         int count, ret;
 965
 966         req->num_sge = 1;
 967         req->inline_data = false;
 968         req->mr->need_inval = false;
 969
 970         c->common.flags |= NVME_CMD_SGL_METABUF;
 971
 972         if (!blk_rq_bytes(rq))
 973                 return nvme_rdma_set_sg_null(c);
 974
 975         req->sg_table.sgl = req->first_sgl;
 976         ret = sg_alloc_table_chained(&req->sg_table,
 977                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
 978         if (ret)
 979                 return -ENOMEM;
 980
 981         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
 982
 983         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
 984                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 985         if (unlikely(count <= 0)) {
 986                 sg_free_table_chained(&req->sg_table, true);
 987                 return -EIO;
 988         }
 989
 990         if (count == 1) {
 991                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
 992                     blk_rq_payload_bytes(rq) <=
 993                                 nvme_rdma_inline_data_size(queue))
 994                         return nvme_rdma_map_sg_inline(queue, req, c);
 995
 996                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
 997                         return nvme_rdma_map_sg_single(queue, req, c);
 998         }
 999
1000         return nvme_rdma_map_sg_fr(queue, req, c, count);
1001 }
1002
1003 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1004 {
1005         if (unlikely(wc->status != IB_WC_SUCCESS))
1006                 nvme_rdma_wr_error(cq, wc, "SEND");
1007 }
1008
1009 /*
1010  * We want to signal completion at least every queue depth/2.  This returns the
1011  * largest power of two that is not above half of (queue size + 1) to optimize
1012  * (avoid divisions).
1013  */
1014 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1015 {
1016         int limit = 1 << ilog2((queue->queue_size + 1) / 2);
1017
1018         return (atomic_inc_return(&queue->sig_count) & (limit - 1)) == 0;
1019 }
1020
1021 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1022                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1023                 struct ib_send_wr *first, bool flush)
1024 {
1025         struct ib_send_wr wr, *bad_wr;
1026         int ret;
1027
1028         sge->addr   = qe->dma;
1029         sge->length = sizeof(struct nvme_command),
1030         sge->lkey   = queue->device->pd->local_dma_lkey;
1031
1032         qe->cqe.done = nvme_rdma_send_done;
1033
1034         wr.next       = NULL;
1035         wr.wr_cqe     = &qe->cqe;
1036         wr.sg_list    = sge;
1037         wr.num_sge    = num_sge;
1038         wr.opcode     = IB_WR_SEND;
1039         wr.send_flags = 0;
1040
1041         /*
1042          * Unsignalled send completions are another giant desaster in the
1043          * IB Verbs spec:  If we don't regularly post signalled sends
1044          * the send queue will fill up and only a QP reset will rescue us.
1045          * Would have been way to obvious to handle this in hardware or
1046          * at least the RDMA stack..
1047          *
1048          * Always signal the flushes. The magic request used for the flush
1049          * sequencer is not allocated in our driver's tagset and it's
1050          * triggered to be freed by blk_cleanup_queue(). So we need to
1051          * always mark it as signaled to ensure that the "wr_cqe", which is
1052          * embedded in request's payload, is not freed when __ib_process_cq()
1053          * calls wr_cqe->done().
1054          */
1055         if (nvme_rdma_queue_sig_limit(queue) || flush)
1056                 wr.send_flags |= IB_SEND_SIGNALED;
1057
1058         if (first)
1059                 first->next = &wr;
1060         else
1061                 first = &wr;
1062
1063         ret = ib_post_send(queue->qp, first, &bad_wr);
1064         if (ret) {
1065                 dev_err(queue->ctrl->ctrl.device,
1066                              "%s failed with error code %d\n", __func__, ret);
1067         }
1068         return ret;
1069 }
1070
1071 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1072                 struct nvme_rdma_qe *qe)
1073 {
1074         struct ib_recv_wr wr, *bad_wr;
1075         struct ib_sge list;
1076         int ret;
1077
1078         list.addr   = qe->dma;
1079         list.length = sizeof(struct nvme_completion);
1080         list.lkey   = queue->device->pd->local_dma_lkey;
1081
1082         qe->cqe.done = nvme_rdma_recv_done;
1083
1084         wr.next     = NULL;
1085         wr.wr_cqe   = &qe->cqe;
1086         wr.sg_list  = &list;
1087         wr.num_sge  = 1;
1088
1089         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1090         if (ret) {
1091                 dev_err(queue->ctrl->ctrl.device,
1092                         "%s failed with error code %d\n", __func__, ret);
1093         }
1094         return ret;
1095 }
1096
1097 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1098 {
1099         u32 queue_idx = nvme_rdma_queue_idx(queue);
1100
1101         if (queue_idx == 0)
1102                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1103         return queue->ctrl->tag_set.tags[queue_idx - 1];
1104 }
1105
1106 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1107 {
1108         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1109         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1110         struct ib_device *dev = queue->device->dev;
1111         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1112         struct nvme_command *cmd = sqe->data;
1113         struct ib_sge sge;
1114         int ret;
1115
1116         if (WARN_ON_ONCE(aer_idx != 0))
1117                 return;
1118
1119         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1120
1121         memset(cmd, 0, sizeof(*cmd));
1122         cmd->common.opcode = nvme_admin_async_event;
1123         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1124         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1125         nvme_rdma_set_sg_null(cmd);
1126
1127         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1128                         DMA_TO_DEVICE);
1129
1130         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1131         WARN_ON_ONCE(ret);
1132 }
1133
1134 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1135                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1136 {
1137         struct request *rq;
1138         struct nvme_rdma_request *req;
1139         int ret = 0;
1140
1141         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1142         if (!rq) {
1143                 dev_err(queue->ctrl->ctrl.device,
1144                         "tag 0x%x on QP %#x not found\n",
1145                         cqe->command_id, queue->qp->qp_num);
1146                 nvme_rdma_error_recovery(queue->ctrl);
1147                 return ret;
1148         }
1149         req = blk_mq_rq_to_pdu(rq);
1150
1151         if (rq->tag == tag)
1152                 ret = 1;
1153
1154         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1155             wc->ex.invalidate_rkey == req->mr->rkey)
1156                 req->mr->need_inval = false;
1157
1158         nvme_end_request(rq, cqe->status, cqe->result);
1159         return ret;
1160 }
1161
1162 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1163 {
1164         struct nvme_rdma_qe *qe =
1165                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1166         struct nvme_rdma_queue *queue = cq->cq_context;
1167         struct ib_device *ibdev = queue->device->dev;
1168         struct nvme_completion *cqe = qe->data;
1169         const size_t len = sizeof(struct nvme_completion);
1170         int ret = 0;
1171
1172         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1173                 nvme_rdma_wr_error(cq, wc, "RECV");
1174                 return 0;
1175         }
1176
1177         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1178         /*
1179          * AEN requests are special as they don't time out and can
1180          * survive any kind of queue freeze and often don't respond to
1181          * aborts.  We don't even bother to allocate a struct request
1182          * for them but rather special case them here.
1183          */
1184         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1185                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1186                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1187                                 &cqe->result);
1188         else
1189                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1190         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1191
1192         nvme_rdma_post_recv(queue, qe);
1193         return ret;
1194 }
1195
1196 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1197 {
1198         __nvme_rdma_recv_done(cq, wc, -1);
1199 }
1200
1201 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1202 {
1203         int ret, i;
1204
1205         for (i = 0; i < queue->queue_size; i++) {
1206                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1207                 if (ret)
1208                         goto out_destroy_queue_ib;
1209         }
1210
1211         return 0;
1212
1213 out_destroy_queue_ib:
1214         nvme_rdma_destroy_queue_ib(queue);
1215         return ret;
1216 }
1217
1218 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1219                 struct rdma_cm_event *ev)
1220 {
1221         struct rdma_cm_id *cm_id = queue->cm_id;
1222         int status = ev->status;
1223         const char *rej_msg;
1224         const struct nvme_rdma_cm_rej *rej_data;
1225         u8 rej_data_len;
1226
1227         rej_msg = rdma_reject_msg(cm_id, status);
1228         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1229
1230         if (rej_data && rej_data_len >= sizeof(u16)) {
1231                 u16 sts = le16_to_cpu(rej_data->sts);
1232
1233                 dev_err(queue->ctrl->ctrl.device,
1234                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1235                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1236         } else {
1237                 dev_err(queue->ctrl->ctrl.device,
1238                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1239         }
1240
1241         return -ECONNRESET;
1242 }
1243
1244 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1245 {
1246         int ret;
1247
1248         ret = nvme_rdma_create_queue_ib(queue);
1249         if (ret)
1250                 return ret;
1251
1252         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1253         if (ret) {
1254                 dev_err(queue->ctrl->ctrl.device,
1255                         "rdma_resolve_route failed (%d).\n",
1256                         queue->cm_error);
1257                 goto out_destroy_queue;
1258         }
1259
1260         return 0;
1261
1262 out_destroy_queue:
1263         nvme_rdma_destroy_queue_ib(queue);
1264         return ret;
1265 }
1266
1267 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1268 {
1269         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1270         struct rdma_conn_param param = { };
1271         struct nvme_rdma_cm_req priv = { };
1272         int ret;
1273
1274         param.qp_num = queue->qp->qp_num;
1275         param.flow_control = 1;
1276
1277         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1278         /* maximum retry count */
1279         param.retry_count = 7;
1280         param.rnr_retry_count = 7;
1281         param.private_data = &priv;
1282         param.private_data_len = sizeof(priv);
1283
1284         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1285         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1286         /*
1287          * set the admin queue depth to the minimum size
1288          * specified by the Fabrics standard.
1289          */
1290         if (priv.qid == 0) {
1291                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1292                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1293         } else {
1294                 /*
1295                  * current interpretation of the fabrics spec
1296                  * is at minimum you make hrqsize sqsize+1, or a
1297                  * 1's based representation of sqsize.
1298                  */
1299                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1300                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1301         }
1302
1303         ret = rdma_connect(queue->cm_id, &param);
1304         if (ret) {
1305                 dev_err(ctrl->ctrl.device,
1306                         "rdma_connect failed (%d).\n", ret);
1307                 goto out_destroy_queue_ib;
1308         }
1309
1310         return 0;
1311
1312 out_destroy_queue_ib:
1313         nvme_rdma_destroy_queue_ib(queue);
1314         return ret;
1315 }
1316
1317 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1318                 struct rdma_cm_event *ev)
1319 {
1320         struct nvme_rdma_queue *queue = cm_id->context;
1321         int cm_error = 0;
1322
1323         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1324                 rdma_event_msg(ev->event), ev->event,
1325                 ev->status, cm_id);
1326
1327         switch (ev->event) {
1328         case RDMA_CM_EVENT_ADDR_RESOLVED:
1329                 cm_error = nvme_rdma_addr_resolved(queue);
1330                 break;
1331         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1332                 cm_error = nvme_rdma_route_resolved(queue);
1333                 break;
1334         case RDMA_CM_EVENT_ESTABLISHED:
1335                 queue->cm_error = nvme_rdma_conn_established(queue);
1336                 /* complete cm_done regardless of success/failure */
1337                 complete(&queue->cm_done);
1338                 return 0;
1339         case RDMA_CM_EVENT_REJECTED:
1340                 nvme_rdma_destroy_queue_ib(queue);
1341                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1342                 break;
1343         case RDMA_CM_EVENT_ROUTE_ERROR:
1344         case RDMA_CM_EVENT_CONNECT_ERROR:
1345         case RDMA_CM_EVENT_UNREACHABLE:
1346                 nvme_rdma_destroy_queue_ib(queue);
1347         case RDMA_CM_EVENT_ADDR_ERROR:
1348                 dev_dbg(queue->ctrl->ctrl.device,
1349                         "CM error event %d\n", ev->event);
1350                 cm_error = -ECONNRESET;
1351                 break;
1352         case RDMA_CM_EVENT_DISCONNECTED:
1353         case RDMA_CM_EVENT_ADDR_CHANGE:
1354         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1355                 dev_dbg(queue->ctrl->ctrl.device,
1356                         "disconnect received - connection closed\n");
1357                 nvme_rdma_error_recovery(queue->ctrl);
1358                 break;
1359         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1360                 /* device removal is handled via the ib_client API */
1361                 break;
1362         default:
1363                 dev_err(queue->ctrl->ctrl.device,
1364                         "Unexpected RDMA CM event (%d)\n", ev->event);
1365                 nvme_rdma_error_recovery(queue->ctrl);
1366                 break;
1367         }
1368
1369         if (cm_error) {
1370                 queue->cm_error = cm_error;
1371                 complete(&queue->cm_done);
1372         }
1373
1374         return 0;
1375 }
1376
1377 static enum blk_eh_timer_return
1378 nvme_rdma_timeout(struct request *rq, bool reserved)
1379 {
1380         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1381
1382         /* queue error recovery */
1383         nvme_rdma_error_recovery(req->queue->ctrl);
1384
1385         /* fail with DNR on cmd timeout */
1386         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1387
1388         return BLK_EH_HANDLED;
1389 }
1390
1391 /*
1392  * We cannot accept any other command until the Connect command has completed.
1393  */
1394 static inline blk_status_t
1395 nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1396 {
1397         if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1398                 struct nvme_command *cmd = nvme_req(rq)->cmd;
1399
1400                 if (!blk_rq_is_passthrough(rq) ||
1401                     cmd->common.opcode != nvme_fabrics_command ||
1402                     cmd->fabrics.fctype != nvme_fabrics_type_connect) {
1403                         /*
1404                          * reconnecting state means transport disruption, which
1405                          * can take a long time and even might fail permanently,
1406                          * so we can't let incoming I/O be requeued forever.
1407                          * fail it fast to allow upper layers a chance to
1408                          * failover.
1409                          */
1410                         if (queue->ctrl->ctrl.state == NVME_CTRL_RECONNECTING)
1411                                 return BLK_STS_IOERR;
1412                         return BLK_STS_RESOURCE; /* try again later */
1413                 }
1414         }
1415
1416         return 0;
1417 }
1418
1419 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1420                 const struct blk_mq_queue_data *bd)
1421 {
1422         struct nvme_ns *ns = hctx->queue->queuedata;
1423         struct nvme_rdma_queue *queue = hctx->driver_data;
1424         struct request *rq = bd->rq;
1425         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1426         struct nvme_rdma_qe *sqe = &req->sqe;
1427         struct nvme_command *c = sqe->data;
1428         bool flush = false;
1429         struct ib_device *dev;
1430         blk_status_t ret;
1431         int err;
1432
1433         WARN_ON_ONCE(rq->tag < 0);
1434
1435         ret = nvme_rdma_queue_is_ready(queue, rq);
1436         if (unlikely(ret))
1437                 return ret;
1438
1439         dev = queue->device->dev;
1440         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1441                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1442
1443         ret = nvme_setup_cmd(ns, rq, c);
1444         if (ret)
1445                 return ret;
1446
1447         blk_mq_start_request(rq);
1448
1449         err = nvme_rdma_map_data(queue, rq, c);
1450         if (err < 0) {
1451                 dev_err(queue->ctrl->ctrl.device,
1452                              "Failed to map data (%d)\n", err);
1453                 nvme_cleanup_cmd(rq);
1454                 goto err;
1455         }
1456
1457         ib_dma_sync_single_for_device(dev, sqe->dma,
1458                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1459
1460         if (req_op(rq) == REQ_OP_FLUSH)
1461                 flush = true;
1462         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1463                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1464         if (err) {
1465                 nvme_rdma_unmap_data(queue, rq);
1466                 goto err;
1467         }
1468
1469         return BLK_STS_OK;
1470 err:
1471         if (err == -ENOMEM || err == -EAGAIN)
1472                 return BLK_STS_RESOURCE;
1473         return BLK_STS_IOERR;
1474 }
1475
1476 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1477 {
1478         struct nvme_rdma_queue *queue = hctx->driver_data;
1479         struct ib_cq *cq = queue->ib_cq;
1480         struct ib_wc wc;
1481         int found = 0;
1482
1483         while (ib_poll_cq(cq, 1, &wc) > 0) {
1484                 struct ib_cqe *cqe = wc.wr_cqe;
1485
1486                 if (cqe) {
1487                         if (cqe->done == nvme_rdma_recv_done)
1488                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1489                         else
1490                                 cqe->done(cq, &wc);
1491                 }
1492         }
1493
1494         return found;
1495 }
1496
1497 static void nvme_rdma_complete_rq(struct request *rq)
1498 {
1499         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1500
1501         nvme_rdma_unmap_data(req->queue, rq);
1502         nvme_complete_rq(rq);
1503 }
1504
1505 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1506         .queue_rq       = nvme_rdma_queue_rq,
1507         .complete       = nvme_rdma_complete_rq,
1508         .init_request   = nvme_rdma_init_request,
1509         .exit_request   = nvme_rdma_exit_request,
1510         .reinit_request = nvme_rdma_reinit_request,
1511         .init_hctx      = nvme_rdma_init_hctx,
1512         .poll           = nvme_rdma_poll,
1513         .timeout        = nvme_rdma_timeout,
1514 };
1515
1516 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1517         .queue_rq       = nvme_rdma_queue_rq,
1518         .complete       = nvme_rdma_complete_rq,
1519         .init_request   = nvme_rdma_init_request,
1520         .exit_request   = nvme_rdma_exit_request,
1521         .reinit_request = nvme_rdma_reinit_request,
1522         .init_hctx      = nvme_rdma_init_admin_hctx,
1523         .timeout        = nvme_rdma_timeout,
1524 };
1525
1526 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1527 {
1528         int error;
1529
1530         error = nvme_rdma_init_queue(ctrl, 0, NVME_AQ_DEPTH);
1531         if (error)
1532                 return error;
1533
1534         ctrl->device = ctrl->queues[0].device;
1535
1536         /*
1537          * We need a reference on the device as long as the tag_set is alive,
1538          * as the MRs in the request structures need a valid ib_device.
1539          */
1540         error = -EINVAL;
1541         if (!nvme_rdma_dev_get(ctrl->device))
1542                 goto out_free_queue;
1543
1544         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1545                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1546
1547         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1548         ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1549         ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1550         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1551         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1552         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1553                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1554         ctrl->admin_tag_set.driver_data = ctrl;
1555         ctrl->admin_tag_set.nr_hw_queues = 1;
1556         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1557
1558         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1559         if (error)
1560                 goto out_put_dev;
1561
1562         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1563         if (IS_ERR(ctrl->ctrl.admin_q)) {
1564                 error = PTR_ERR(ctrl->ctrl.admin_q);
1565                 goto out_free_tagset;
1566         }
1567
1568         error = nvmf_connect_admin_queue(&ctrl->ctrl);
1569         if (error)
1570                 goto out_cleanup_queue;
1571
1572         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
1573
1574         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP,
1575                         &ctrl->ctrl.cap);
1576         if (error) {
1577                 dev_err(ctrl->ctrl.device,
1578                         "prop_get NVME_REG_CAP failed\n");
1579                 goto out_cleanup_queue;
1580         }
1581
1582         ctrl->ctrl.sqsize =
1583                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
1584
1585         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1586         if (error)
1587                 goto out_cleanup_queue;
1588
1589         ctrl->ctrl.max_hw_sectors =
1590                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
1591
1592         error = nvme_init_identify(&ctrl->ctrl);
1593         if (error)
1594                 goto out_cleanup_queue;
1595
1596         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1597                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
1598                         DMA_TO_DEVICE);
1599         if (error)
1600                 goto out_cleanup_queue;
1601
1602         return 0;
1603
1604 out_cleanup_queue:
1605         blk_cleanup_queue(ctrl->ctrl.admin_q);
1606 out_free_tagset:
1607         /* disconnect and drain the queue before freeing the tagset */
1608         nvme_rdma_stop_queue(&ctrl->queues[0]);
1609         blk_mq_free_tag_set(&ctrl->admin_tag_set);
1610 out_put_dev:
1611         nvme_rdma_dev_put(ctrl->device);
1612 out_free_queue:
1613         nvme_rdma_free_queue(&ctrl->queues[0]);
1614         return error;
1615 }
1616
1617 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1618 {
1619         cancel_work_sync(&ctrl->err_work);
1620         cancel_delayed_work_sync(&ctrl->reconnect_work);
1621
1622         if (ctrl->ctrl.queue_count > 1) {
1623                 nvme_stop_queues(&ctrl->ctrl);
1624                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1625                                         nvme_cancel_request, &ctrl->ctrl);
1626                 nvme_rdma_free_io_queues(ctrl);
1627         }
1628
1629         if (test_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags))
1630                 nvme_shutdown_ctrl(&ctrl->ctrl);
1631
1632         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1633         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1634                                 nvme_cancel_request, &ctrl->ctrl);
1635         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1636         nvme_rdma_destroy_admin_queue(ctrl);
1637 }
1638
1639 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1640 {
1641         nvme_stop_ctrl(&ctrl->ctrl);
1642         nvme_remove_namespaces(&ctrl->ctrl);
1643         if (shutdown)
1644                 nvme_rdma_shutdown_ctrl(ctrl);
1645
1646         nvme_uninit_ctrl(&ctrl->ctrl);
1647         if (ctrl->ctrl.tagset) {
1648                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1649                 blk_mq_free_tag_set(&ctrl->tag_set);
1650                 nvme_rdma_dev_put(ctrl->device);
1651         }
1652
1653         nvme_put_ctrl(&ctrl->ctrl);
1654 }
1655
1656 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1657 {
1658         struct nvme_rdma_ctrl *ctrl = container_of(work,
1659                                 struct nvme_rdma_ctrl, delete_work);
1660
1661         __nvme_rdma_remove_ctrl(ctrl, true);
1662 }
1663
1664 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1665 {
1666         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1667                 return -EBUSY;
1668
1669         if (!queue_work(nvme_wq, &ctrl->delete_work))
1670                 return -EBUSY;
1671
1672         return 0;
1673 }
1674
1675 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1676 {
1677         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1678         int ret = 0;
1679
1680         /*
1681          * Keep a reference until all work is flushed since
1682          * __nvme_rdma_del_ctrl can free the ctrl mem
1683          */
1684         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1685                 return -EBUSY;
1686         ret = __nvme_rdma_del_ctrl(ctrl);
1687         if (!ret)
1688                 flush_work(&ctrl->delete_work);
1689         nvme_put_ctrl(&ctrl->ctrl);
1690         return ret;
1691 }
1692
1693 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1694 {
1695         struct nvme_rdma_ctrl *ctrl = container_of(work,
1696                                 struct nvme_rdma_ctrl, delete_work);
1697
1698         __nvme_rdma_remove_ctrl(ctrl, false);
1699 }
1700
1701 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1702 {
1703         struct nvme_rdma_ctrl *ctrl =
1704                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1705         int ret;
1706         bool changed;
1707
1708         nvme_stop_ctrl(&ctrl->ctrl);
1709         nvme_rdma_shutdown_ctrl(ctrl);
1710
1711         ret = nvme_rdma_configure_admin_queue(ctrl);
1712         if (ret) {
1713                 /* ctrl is already shutdown, just remove the ctrl */
1714                 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1715                 goto del_dead_ctrl;
1716         }
1717
1718         if (ctrl->ctrl.queue_count > 1) {
1719                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1720                 if (ret)
1721                         goto del_dead_ctrl;
1722
1723                 ret = nvme_rdma_init_io_queues(ctrl);
1724                 if (ret)
1725                         goto del_dead_ctrl;
1726
1727                 ret = nvme_rdma_connect_io_queues(ctrl);
1728                 if (ret)
1729                         goto del_dead_ctrl;
1730
1731                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
1732                                 ctrl->ctrl.queue_count - 1);
1733         }
1734
1735         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1736         WARN_ON_ONCE(!changed);
1737
1738         nvme_start_ctrl(&ctrl->ctrl);
1739
1740         return;
1741
1742 del_dead_ctrl:
1743         /* Deleting this dead controller... */
1744         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1745         WARN_ON(!queue_work(nvme_wq, &ctrl->delete_work));
1746 }
1747
1748 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1749         .name                   = "rdma",
1750         .module                 = THIS_MODULE,
1751         .flags                  = NVME_F_FABRICS,
1752         .reg_read32             = nvmf_reg_read32,
1753         .reg_read64             = nvmf_reg_read64,
1754         .reg_write32            = nvmf_reg_write32,
1755         .free_ctrl              = nvme_rdma_free_ctrl,
1756         .submit_async_event     = nvme_rdma_submit_async_event,
1757         .delete_ctrl            = nvme_rdma_del_ctrl,
1758         .get_address            = nvmf_get_address,
1759 };
1760
1761 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1762 {
1763         int ret;
1764
1765         ret = nvme_rdma_init_io_queues(ctrl);
1766         if (ret)
1767                 return ret;
1768
1769         /*
1770          * We need a reference on the device as long as the tag_set is alive,
1771          * as the MRs in the request structures need a valid ib_device.
1772          */
1773         ret = -EINVAL;
1774         if (!nvme_rdma_dev_get(ctrl->device))
1775                 goto out_free_io_queues;
1776
1777         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1778         ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1779         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1780         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1781         ctrl->tag_set.numa_node = NUMA_NO_NODE;
1782         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1783         ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1784                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1785         ctrl->tag_set.driver_data = ctrl;
1786         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
1787         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1788
1789         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1790         if (ret)
1791                 goto out_put_dev;
1792         ctrl->ctrl.tagset = &ctrl->tag_set;
1793
1794         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1795         if (IS_ERR(ctrl->ctrl.connect_q)) {
1796                 ret = PTR_ERR(ctrl->ctrl.connect_q);
1797                 goto out_free_tag_set;
1798         }
1799
1800         ret = nvme_rdma_connect_io_queues(ctrl);
1801         if (ret)
1802                 goto out_cleanup_connect_q;
1803
1804         return 0;
1805
1806 out_cleanup_connect_q:
1807         blk_cleanup_queue(ctrl->ctrl.connect_q);
1808 out_free_tag_set:
1809         blk_mq_free_tag_set(&ctrl->tag_set);
1810 out_put_dev:
1811         nvme_rdma_dev_put(ctrl->device);
1812 out_free_io_queues:
1813         nvme_rdma_free_io_queues(ctrl);
1814         return ret;
1815 }
1816
1817 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1818                 struct nvmf_ctrl_options *opts)
1819 {
1820         struct nvme_rdma_ctrl *ctrl;
1821         int ret;
1822         bool changed;
1823         char *port;
1824
1825         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1826         if (!ctrl)
1827                 return ERR_PTR(-ENOMEM);
1828         ctrl->ctrl.opts = opts;
1829         INIT_LIST_HEAD(&ctrl->list);
1830
1831         if (opts->mask & NVMF_OPT_TRSVCID)
1832                 port = opts->trsvcid;
1833         else
1834                 port = __stringify(NVME_RDMA_IP_PORT);
1835
1836         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1837                         opts->traddr, port, &ctrl->addr);
1838         if (ret) {
1839                 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1840                 goto out_free_ctrl;
1841         }
1842
1843         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1844                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1845                         opts->host_traddr, NULL, &ctrl->src_addr);
1846                 if (ret) {
1847                         pr_err("malformed src address passed: %s\n",
1848                                opts->host_traddr);
1849                         goto out_free_ctrl;
1850                 }
1851         }
1852
1853         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1854                                 0 /* no quirks, we're perfect! */);
1855         if (ret)
1856                 goto out_free_ctrl;
1857
1858         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1859                         nvme_rdma_reconnect_ctrl_work);
1860         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1861         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1862         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1863
1864         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1865         ctrl->ctrl.sqsize = opts->queue_size - 1;
1866         ctrl->ctrl.kato = opts->kato;
1867
1868         ret = -ENOMEM;
1869         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1870                                 GFP_KERNEL);
1871         if (!ctrl->queues)
1872                 goto out_uninit_ctrl;
1873
1874         ret = nvme_rdma_configure_admin_queue(ctrl);
1875         if (ret)
1876                 goto out_kfree_queues;
1877
1878         /* sanity check icdoff */
1879         if (ctrl->ctrl.icdoff) {
1880                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1881                 ret = -EINVAL;
1882                 goto out_remove_admin_queue;
1883         }
1884
1885         /* sanity check keyed sgls */
1886         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1887                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1888                 ret = -EINVAL;
1889                 goto out_remove_admin_queue;
1890         }
1891
1892         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1893                 /* warn if maxcmd is lower than queue_size */
1894                 dev_warn(ctrl->ctrl.device,
1895                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1896                         opts->queue_size, ctrl->ctrl.maxcmd);
1897                 opts->queue_size = ctrl->ctrl.maxcmd;
1898         }
1899
1900         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1901                 /* warn if sqsize is lower than queue_size */
1902                 dev_warn(ctrl->ctrl.device,
1903                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1904                         opts->queue_size, ctrl->ctrl.sqsize + 1);
1905                 opts->queue_size = ctrl->ctrl.sqsize + 1;
1906         }
1907
1908         if (opts->nr_io_queues) {
1909                 ret = nvme_rdma_create_io_queues(ctrl);
1910                 if (ret)
1911                         goto out_remove_admin_queue;
1912         }
1913
1914         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1915         WARN_ON_ONCE(!changed);
1916
1917         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1918                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1919
1920         kref_get(&ctrl->ctrl.kref);
1921
1922         mutex_lock(&nvme_rdma_ctrl_mutex);
1923         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1924         mutex_unlock(&nvme_rdma_ctrl_mutex);
1925
1926         nvme_start_ctrl(&ctrl->ctrl);
1927
1928         return &ctrl->ctrl;
1929
1930 out_remove_admin_queue:
1931         nvme_rdma_destroy_admin_queue(ctrl);
1932 out_kfree_queues:
1933         kfree(ctrl->queues);
1934 out_uninit_ctrl:
1935         nvme_uninit_ctrl(&ctrl->ctrl);
1936         nvme_put_ctrl(&ctrl->ctrl);
1937         if (ret > 0)
1938                 ret = -EIO;
1939         return ERR_PTR(ret);
1940 out_free_ctrl:
1941         kfree(ctrl);
1942         return ERR_PTR(ret);
1943 }
1944
1945 static struct nvmf_transport_ops nvme_rdma_transport = {
1946         .name           = "rdma",
1947         .required_opts  = NVMF_OPT_TRADDR,
1948         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1949                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1950         .create_ctrl    = nvme_rdma_create_ctrl,
1951 };
1952
1953 static void nvme_rdma_add_one(struct ib_device *ib_device)
1954 {
1955 }
1956
1957 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1958 {
1959         struct nvme_rdma_ctrl *ctrl;
1960
1961         /* Delete all controllers using this device */
1962         mutex_lock(&nvme_rdma_ctrl_mutex);
1963         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1964                 if (ctrl->device->dev != ib_device)
1965                         continue;
1966                 dev_info(ctrl->ctrl.device,
1967                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
1968                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1969                 __nvme_rdma_del_ctrl(ctrl);
1970         }
1971         mutex_unlock(&nvme_rdma_ctrl_mutex);
1972
1973         flush_workqueue(nvme_wq);
1974 }
1975
1976 static struct ib_client nvme_rdma_ib_client = {
1977         .name   = "nvme_rdma",
1978         .add = nvme_rdma_add_one,
1979         .remove = nvme_rdma_remove_one
1980 };
1981
1982 static int __init nvme_rdma_init_module(void)
1983 {
1984         int ret;
1985
1986         ret = ib_register_client(&nvme_rdma_ib_client);
1987         if (ret)
1988                 return ret;
1989
1990         ret = nvmf_register_transport(&nvme_rdma_transport);
1991         if (ret)
1992                 goto err_unreg_client;
1993
1994         return 0;
1995
1996 err_unreg_client:
1997         ib_unregister_client(&nvme_rdma_ib_client);
1998         return ret;
1999 }
2000
2001 static void __exit nvme_rdma_cleanup_module(void)
2002 {
2003         nvmf_unregister_transport(&nvme_rdma_transport);
2004         ib_unregister_client(&nvme_rdma_ib_client);
2005 }
2006
2007 module_init(nvme_rdma_init_module);
2008 module_exit(nvme_rdma_cleanup_module);
2009
2010 MODULE_LICENSE("GPL v2");