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