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 <rdma/ib_cm.h>
  32 #include <linux/nvme-rdma.h>
  33
  34 #include "nvme.h"
  35 #include "fabrics.h"
  36
  37
  38 #define NVME_RDMA_CONNECT_TIMEOUT_MS    1000            /* 1 second */
  39
  40 #define NVME_RDMA_MAX_SEGMENT_SIZE      0xffffff        /* 24-bit SGL field */
  41
  42 #define NVME_RDMA_MAX_SEGMENTS          256
  43
  44 #define NVME_RDMA_MAX_INLINE_SEGMENTS   1
  45
  46 /*
  47  * We handle AEN commands ourselves and don't even let the
  48  * block layer know about them.
  49  */
  50 #define NVME_RDMA_NR_AEN_COMMANDS      1
  51 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
  52         (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
  53
  54 struct nvme_rdma_device {
  55         struct ib_device       *dev;
  56         struct ib_pd           *pd;
  57         struct ib_mr           *mr;
  58         struct kref             ref;
  59         struct list_head        entry;
  60 };
  61
  62 struct nvme_rdma_qe {
  63         struct ib_cqe           cqe;
  64         void                    *data;
  65         u64                     dma;
  66 };
  67
  68 struct nvme_rdma_queue;
  69 struct nvme_rdma_request {
  70         struct ib_mr            *mr;
  71         struct nvme_rdma_qe     sqe;
  72         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  73         u32                     num_sge;
  74         int                     nents;
  75         bool                    inline_data;
  76         struct ib_reg_wr        reg_wr;
  77         struct ib_cqe           reg_cqe;
  78         struct nvme_rdma_queue  *queue;
  79         struct sg_table         sg_table;
  80         struct scatterlist      first_sgl[];
  81 };
  82
  83 enum nvme_rdma_queue_flags {
  84         NVME_RDMA_Q_CONNECTED = (1 << 0),
  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 and written in the hot path */
 105         spinlock_t              lock;
 106
 107         /* read only in the hot path */
 108         struct nvme_rdma_queue  *queues;
 109         u32                     queue_count;
 110
 111         /* other member variables */
 112         struct blk_mq_tag_set   tag_set;
 113         struct work_struct      delete_work;
 114         struct work_struct      reset_work;
 115         struct work_struct      err_work;
 116
 117         struct nvme_rdma_qe     async_event_sqe;
 118
 119         int                     reconnect_delay;
 120         struct delayed_work     reconnect_work;
 121
 122         struct list_head        list;
 123
 124         struct blk_mq_tag_set   admin_tag_set;
 125         struct nvme_rdma_device *device;
 126
 127         u64                     cap;
 128         u32                     max_fr_pages;
 129
 130         union {
 131                 struct sockaddr addr;
 132                 struct sockaddr_in addr_in;
 133         };
 134
 135         struct nvme_ctrl        ctrl;
 136 };
 137
 138 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 139 {
 140         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 141 }
 142
 143 static LIST_HEAD(device_list);
 144 static DEFINE_MUTEX(device_list_mutex);
 145
 146 static LIST_HEAD(nvme_rdma_ctrl_list);
 147 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 148
 149 static struct workqueue_struct *nvme_rdma_wq;
 150
 151 /*
 152  * Disabling this option makes small I/O goes faster, but is fundamentally
 153  * unsafe.  With it turned off we will have to register a global rkey that
 154  * allows read and write access to all physical memory.
 155  */
 156 static bool register_always = true;
 157 module_param(register_always, bool, 0444);
 158 MODULE_PARM_DESC(register_always,
 159          "Use memory registration even for contiguous memory regions");
 160
 161 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 162                 struct rdma_cm_event *event);
 163 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 164
 165 /* XXX: really should move to a generic header sooner or later.. */
 166 static inline void put_unaligned_le24(u32 val, u8 *p)
 167 {
 168         *p++ = val;
 169         *p++ = val >> 8;
 170         *p++ = val >> 16;
 171 }
 172
 173 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 174 {
 175         return queue - queue->ctrl->queues;
 176 }
 177
 178 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 179 {
 180         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 181 }
 182
 183 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 184                 size_t capsule_size, enum dma_data_direction dir)
 185 {
 186         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 187         kfree(qe->data);
 188 }
 189
 190 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 191                 size_t capsule_size, enum dma_data_direction dir)
 192 {
 193         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 194         if (!qe->data)
 195                 return -ENOMEM;
 196
 197         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 198         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 199                 kfree(qe->data);
 200                 return -ENOMEM;
 201         }
 202
 203         return 0;
 204 }
 205
 206 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 207                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 208                 size_t capsule_size, enum dma_data_direction dir)
 209 {
 210         int i;
 211
 212         for (i = 0; i < ib_queue_size; i++)
 213                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 214         kfree(ring);
 215 }
 216
 217 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 218                 size_t ib_queue_size, size_t capsule_size,
 219                 enum dma_data_direction dir)
 220 {
 221         struct nvme_rdma_qe *ring;
 222         int i;
 223
 224         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 225         if (!ring)
 226                 return NULL;
 227
 228         for (i = 0; i < ib_queue_size; i++) {
 229                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 230                         goto out_free_ring;
 231         }
 232
 233         return ring;
 234
 235 out_free_ring:
 236         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 237         return NULL;
 238 }
 239
 240 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 241 {
 242         pr_debug("QP event %d\n", event->event);
 243 }
 244
 245 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 246 {
 247         wait_for_completion_interruptible_timeout(&queue->cm_done,
 248                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 249         return queue->cm_error;
 250 }
 251
 252 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 253 {
 254         struct nvme_rdma_device *dev = queue->device;
 255         struct ib_qp_init_attr init_attr;
 256         int ret;
 257
 258         memset(&init_attr, 0, sizeof(init_attr));
 259         init_attr.event_handler = nvme_rdma_qp_event;
 260         /* +1 for drain */
 261         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 262         /* +1 for drain */
 263         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 264         init_attr.cap.max_recv_sge = 1;
 265         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 266         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 267         init_attr.qp_type = IB_QPT_RC;
 268         init_attr.send_cq = queue->ib_cq;
 269         init_attr.recv_cq = queue->ib_cq;
 270
 271         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 272
 273         queue->qp = queue->cm_id->qp;
 274         return ret;
 275 }
 276
 277 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 278 {
 279         struct nvme_rdma_ctrl *ctrl = data;
 280         struct nvme_rdma_device *dev = ctrl->device;
 281         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 282         int ret = 0;
 283
 284         if (!req->mr->need_inval)
 285                 goto out;
 286
 287         ib_dereg_mr(req->mr);
 288
 289         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 290                         ctrl->max_fr_pages);
 291         if (IS_ERR(req->mr)) {
 292                 ret = PTR_ERR(req->mr);
 293                 req->mr = NULL;
 294         }
 295
 296         req->mr->need_inval = false;
 297
 298 out:
 299         return ret;
 300 }
 301
 302 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
 303                 struct request *rq, unsigned int queue_idx)
 304 {
 305         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 306         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 307         struct nvme_rdma_device *dev = queue->device;
 308
 309         if (req->mr)
 310                 ib_dereg_mr(req->mr);
 311
 312         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 313                         DMA_TO_DEVICE);
 314 }
 315
 316 static void nvme_rdma_exit_request(void *data, struct request *rq,
 317                                 unsigned int hctx_idx, unsigned int rq_idx)
 318 {
 319         return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
 320 }
 321
 322 static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
 323                                 unsigned int hctx_idx, unsigned int rq_idx)
 324 {
 325         return __nvme_rdma_exit_request(data, rq, 0);
 326 }
 327
 328 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
 329                 struct request *rq, unsigned int queue_idx)
 330 {
 331         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 332         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 333         struct nvme_rdma_device *dev = queue->device;
 334         struct ib_device *ibdev = dev->dev;
 335         int ret;
 336
 337         BUG_ON(queue_idx >= ctrl->queue_count);
 338
 339         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 340                         DMA_TO_DEVICE);
 341         if (ret)
 342                 return ret;
 343
 344         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 345                         ctrl->max_fr_pages);
 346         if (IS_ERR(req->mr)) {
 347                 ret = PTR_ERR(req->mr);
 348                 goto out_free_qe;
 349         }
 350
 351         req->queue = queue;
 352
 353         return 0;
 354
 355 out_free_qe:
 356         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 357                         DMA_TO_DEVICE);
 358         return -ENOMEM;
 359 }
 360
 361 static int nvme_rdma_init_request(void *data, struct request *rq,
 362                                 unsigned int hctx_idx, unsigned int rq_idx,
 363                                 unsigned int numa_node)
 364 {
 365         return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
 366 }
 367
 368 static int nvme_rdma_init_admin_request(void *data, struct request *rq,
 369                                 unsigned int hctx_idx, unsigned int rq_idx,
 370                                 unsigned int numa_node)
 371 {
 372         return __nvme_rdma_init_request(data, rq, 0);
 373 }
 374
 375 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 376                 unsigned int hctx_idx)
 377 {
 378         struct nvme_rdma_ctrl *ctrl = data;
 379         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 380
 381         BUG_ON(hctx_idx >= ctrl->queue_count);
 382
 383         hctx->driver_data = queue;
 384         return 0;
 385 }
 386
 387 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 388                 unsigned int hctx_idx)
 389 {
 390         struct nvme_rdma_ctrl *ctrl = data;
 391         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 392
 393         BUG_ON(hctx_idx != 0);
 394
 395         hctx->driver_data = queue;
 396         return 0;
 397 }
 398
 399 static void nvme_rdma_free_dev(struct kref *ref)
 400 {
 401         struct nvme_rdma_device *ndev =
 402                 container_of(ref, struct nvme_rdma_device, ref);
 403
 404         mutex_lock(&device_list_mutex);
 405         list_del(&ndev->entry);
 406         mutex_unlock(&device_list_mutex);
 407
 408         if (!register_always)
 409                 ib_dereg_mr(ndev->mr);
 410         ib_dealloc_pd(ndev->pd);
 411
 412         kfree(ndev);
 413 }
 414
 415 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 416 {
 417         kref_put(&dev->ref, nvme_rdma_free_dev);
 418 }
 419
 420 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 421 {
 422         return kref_get_unless_zero(&dev->ref);
 423 }
 424
 425 static struct nvme_rdma_device *
 426 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 427 {
 428         struct nvme_rdma_device *ndev;
 429
 430         mutex_lock(&device_list_mutex);
 431         list_for_each_entry(ndev, &device_list, entry) {
 432                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 433                     nvme_rdma_dev_get(ndev))
 434                         goto out_unlock;
 435         }
 436
 437         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 438         if (!ndev)
 439                 goto out_err;
 440
 441         ndev->dev = cm_id->device;
 442         kref_init(&ndev->ref);
 443
 444         ndev->pd = ib_alloc_pd(ndev->dev);
 445         if (IS_ERR(ndev->pd))
 446                 goto out_free_dev;
 447
 448         if (!register_always) {
 449                 ndev->mr = ib_get_dma_mr(ndev->pd,
 450                                             IB_ACCESS_LOCAL_WRITE |
 451                                             IB_ACCESS_REMOTE_READ |
 452                                             IB_ACCESS_REMOTE_WRITE);
 453                 if (IS_ERR(ndev->mr))
 454                         goto out_free_pd;
 455         }
 456
 457         if (!(ndev->dev->attrs.device_cap_flags &
 458               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 459                 dev_err(&ndev->dev->dev,
 460                         "Memory registrations not supported.\n");
 461                 goto out_free_mr;
 462         }
 463
 464         list_add(&ndev->entry, &device_list);
 465 out_unlock:
 466         mutex_unlock(&device_list_mutex);
 467         return ndev;
 468
 469 out_free_mr:
 470         if (!register_always)
 471                 ib_dereg_mr(ndev->mr);
 472 out_free_pd:
 473         ib_dealloc_pd(ndev->pd);
 474 out_free_dev:
 475         kfree(ndev);
 476 out_err:
 477         mutex_unlock(&device_list_mutex);
 478         return NULL;
 479 }
 480
 481 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 482 {
 483         struct nvme_rdma_device *dev = queue->device;
 484         struct ib_device *ibdev = dev->dev;
 485
 486         rdma_destroy_qp(queue->cm_id);
 487         ib_free_cq(queue->ib_cq);
 488
 489         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 490                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 491
 492         nvme_rdma_dev_put(dev);
 493 }
 494
 495 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
 496                 struct nvme_rdma_device *dev)
 497 {
 498         struct ib_device *ibdev = dev->dev;
 499         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 500         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 501         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 502
 503         int ret;
 504
 505         queue->device = dev;
 506
 507         /*
 508          * The admin queue is barely used once the controller is live, so don't
 509          * bother to spread it out.
 510          */
 511         if (idx == 0)
 512                 comp_vector = 0;
 513         else
 514                 comp_vector = idx % ibdev->num_comp_vectors;
 515
 516
 517         /* +1 for ib_stop_cq */
 518         queue->ib_cq = ib_alloc_cq(dev->dev, queue,
 519                                 cq_factor * queue->queue_size + 1, comp_vector,
 520                                 IB_POLL_SOFTIRQ);
 521         if (IS_ERR(queue->ib_cq)) {
 522                 ret = PTR_ERR(queue->ib_cq);
 523                 goto out;
 524         }
 525
 526         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 527         if (ret)
 528                 goto out_destroy_ib_cq;
 529
 530         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 531                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 532         if (!queue->rsp_ring) {
 533                 ret = -ENOMEM;
 534                 goto out_destroy_qp;
 535         }
 536
 537         return 0;
 538
 539 out_destroy_qp:
 540         ib_destroy_qp(queue->qp);
 541 out_destroy_ib_cq:
 542         ib_free_cq(queue->ib_cq);
 543 out:
 544         return ret;
 545 }
 546
 547 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
 548                 int idx, size_t queue_size)
 549 {
 550         struct nvme_rdma_queue *queue;
 551         int ret;
 552
 553         queue = &ctrl->queues[idx];
 554         queue->ctrl = ctrl;
 555         init_completion(&queue->cm_done);
 556
 557         if (idx > 0)
 558                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 559         else
 560                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 561
 562         queue->queue_size = queue_size;
 563
 564         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 565                         RDMA_PS_TCP, IB_QPT_RC);
 566         if (IS_ERR(queue->cm_id)) {
 567                 dev_info(ctrl->ctrl.device,
 568                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 569                 return PTR_ERR(queue->cm_id);
 570         }
 571
 572         queue->cm_error = -ETIMEDOUT;
 573         ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
 574                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 575         if (ret) {
 576                 dev_info(ctrl->ctrl.device,
 577                         "rdma_resolve_addr failed (%d).\n", ret);
 578                 goto out_destroy_cm_id;
 579         }
 580
 581         ret = nvme_rdma_wait_for_cm(queue);
 582         if (ret) {
 583                 dev_info(ctrl->ctrl.device,
 584                         "rdma_resolve_addr wait failed (%d).\n", ret);
 585                 goto out_destroy_cm_id;
 586         }
 587
 588         set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
 589
 590         return 0;
 591
 592 out_destroy_cm_id:
 593         rdma_destroy_id(queue->cm_id);
 594         return ret;
 595 }
 596
 597 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 598 {
 599         rdma_disconnect(queue->cm_id);
 600         ib_drain_qp(queue->qp);
 601 }
 602
 603 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 604 {
 605         nvme_rdma_destroy_queue_ib(queue);
 606         rdma_destroy_id(queue->cm_id);
 607 }
 608
 609 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
 610 {
 611         if (!test_and_clear_bit(NVME_RDMA_Q_CONNECTED, &queue->flags))
 612                 return;
 613         nvme_rdma_stop_queue(queue);
 614         nvme_rdma_free_queue(queue);
 615 }
 616
 617 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 618 {
 619         int i;
 620
 621         for (i = 1; i < ctrl->queue_count; i++)
 622                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 623 }
 624
 625 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
 626 {
 627         int i, ret = 0;
 628
 629         for (i = 1; i < ctrl->queue_count; i++) {
 630                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 631                 if (ret)
 632                         break;
 633         }
 634
 635         return ret;
 636 }
 637
 638 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
 639 {
 640         int i, ret;
 641
 642         for (i = 1; i < ctrl->queue_count; i++) {
 643                 ret = nvme_rdma_init_queue(ctrl, i,
 644                                            ctrl->ctrl.opts->queue_size);
 645                 if (ret) {
 646                         dev_info(ctrl->ctrl.device,
 647                                 "failed to initialize i/o queue: %d\n", ret);
 648                         goto out_free_queues;
 649                 }
 650         }
 651
 652         return 0;
 653
 654 out_free_queues:
 655         for (; i >= 1; i--)
 656                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 657
 658         return ret;
 659 }
 660
 661 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
 662 {
 663         nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
 664                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 665         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 666         blk_cleanup_queue(ctrl->ctrl.admin_q);
 667         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 668         nvme_rdma_dev_put(ctrl->device);
 669 }
 670
 671 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 672 {
 673         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 674
 675         if (list_empty(&ctrl->list))
 676                 goto free_ctrl;
 677
 678         mutex_lock(&nvme_rdma_ctrl_mutex);
 679         list_del(&ctrl->list);
 680         mutex_unlock(&nvme_rdma_ctrl_mutex);
 681
 682         kfree(ctrl->queues);
 683         nvmf_free_options(nctrl->opts);
 684 free_ctrl:
 685         kfree(ctrl);
 686 }
 687
 688 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 689 {
 690         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 691                         struct nvme_rdma_ctrl, reconnect_work);
 692         bool changed;
 693         int ret;
 694
 695         if (ctrl->queue_count > 1) {
 696                 nvme_rdma_free_io_queues(ctrl);
 697
 698                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
 699                 if (ret)
 700                         goto requeue;
 701         }
 702
 703         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 704
 705         ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
 706         if (ret)
 707                 goto requeue;
 708
 709         ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
 710         if (ret)
 711                 goto requeue;
 712
 713         blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
 714
 715         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 716         if (ret)
 717                 goto stop_admin_q;
 718
 719         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 720         if (ret)
 721                 goto stop_admin_q;
 722
 723         nvme_start_keep_alive(&ctrl->ctrl);
 724
 725         if (ctrl->queue_count > 1) {
 726                 ret = nvme_rdma_init_io_queues(ctrl);
 727                 if (ret)
 728                         goto stop_admin_q;
 729
 730                 ret = nvme_rdma_connect_io_queues(ctrl);
 731                 if (ret)
 732                         goto stop_admin_q;
 733         }
 734
 735         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 736         WARN_ON_ONCE(!changed);
 737
 738         if (ctrl->queue_count > 1) {
 739                 nvme_start_queues(&ctrl->ctrl);
 740                 nvme_queue_scan(&ctrl->ctrl);
 741                 nvme_queue_async_events(&ctrl->ctrl);
 742         }
 743
 744         dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
 745
 746         return;
 747
 748 stop_admin_q:
 749         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 750 requeue:
 751         /* Make sure we are not resetting/deleting */
 752         if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
 753                 dev_info(ctrl->ctrl.device,
 754                         "Failed reconnect attempt, requeueing...\n");
 755                 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 756                                         ctrl->reconnect_delay * HZ);
 757         }
 758 }
 759
 760 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 761 {
 762         struct nvme_rdma_ctrl *ctrl = container_of(work,
 763                         struct nvme_rdma_ctrl, err_work);
 764
 765         nvme_stop_keep_alive(&ctrl->ctrl);
 766         if (ctrl->queue_count > 1)
 767                 nvme_stop_queues(&ctrl->ctrl);
 768         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 769
 770         /* We must take care of fastfail/requeue all our inflight requests */
 771         if (ctrl->queue_count > 1)
 772                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 773                                         nvme_cancel_request, &ctrl->ctrl);
 774         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 775                                 nvme_cancel_request, &ctrl->ctrl);
 776
 777         dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
 778                 ctrl->reconnect_delay);
 779
 780         queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 781                                 ctrl->reconnect_delay * HZ);
 782 }
 783
 784 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 785 {
 786         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 787                 return;
 788
 789         queue_work(nvme_rdma_wq, &ctrl->err_work);
 790 }
 791
 792 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 793                 const char *op)
 794 {
 795         struct nvme_rdma_queue *queue = cq->cq_context;
 796         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 797
 798         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 799                 dev_info(ctrl->ctrl.device,
 800                              "%s for CQE 0x%p failed with status %s (%d)\n",
 801                              op, wc->wr_cqe,
 802                              ib_wc_status_msg(wc->status), wc->status);
 803         nvme_rdma_error_recovery(ctrl);
 804 }
 805
 806 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
 807 {
 808         if (unlikely(wc->status != IB_WC_SUCCESS))
 809                 nvme_rdma_wr_error(cq, wc, "MEMREG");
 810 }
 811
 812 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 813 {
 814         if (unlikely(wc->status != IB_WC_SUCCESS))
 815                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
 816 }
 817
 818 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
 819                 struct nvme_rdma_request *req)
 820 {
 821         struct ib_send_wr *bad_wr;
 822         struct ib_send_wr wr = {
 823                 .opcode             = IB_WR_LOCAL_INV,
 824                 .next               = NULL,
 825                 .num_sge            = 0,
 826                 .send_flags         = 0,
 827                 .ex.invalidate_rkey = req->mr->rkey,
 828         };
 829
 830         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
 831         wr.wr_cqe = &req->reg_cqe;
 832
 833         return ib_post_send(queue->qp, &wr, &bad_wr);
 834 }
 835
 836 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 837                 struct request *rq)
 838 {
 839         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 840         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 841         struct nvme_rdma_device *dev = queue->device;
 842         struct ib_device *ibdev = dev->dev;
 843         int res;
 844
 845         if (!blk_rq_bytes(rq))
 846                 return;
 847
 848         if (req->mr->need_inval) {
 849                 res = nvme_rdma_inv_rkey(queue, req);
 850                 if (res < 0) {
 851                         dev_err(ctrl->ctrl.device,
 852                                 "Queueing INV WR for rkey %#x failed (%d)\n",
 853                                 req->mr->rkey, res);
 854                         nvme_rdma_error_recovery(queue->ctrl);
 855                 }
 856         }
 857
 858         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 859                         req->nents, rq_data_dir(rq) ==
 860                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 861
 862         nvme_cleanup_cmd(rq);
 863         sg_free_table_chained(&req->sg_table, true);
 864 }
 865
 866 static int nvme_rdma_set_sg_null(struct nvme_command *c)
 867 {
 868         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 869
 870         sg->addr = 0;
 871         put_unaligned_le24(0, sg->length);
 872         put_unaligned_le32(0, sg->key);
 873         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 874         return 0;
 875 }
 876
 877 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
 878                 struct nvme_rdma_request *req, struct nvme_command *c)
 879 {
 880         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
 881
 882         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
 883         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
 884         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
 885
 886         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
 887         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
 888         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
 889
 890         req->inline_data = true;
 891         req->num_sge++;
 892         return 0;
 893 }
 894
 895 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
 896                 struct nvme_rdma_request *req, struct nvme_command *c)
 897 {
 898         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 899
 900         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
 901         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
 902         put_unaligned_le32(queue->device->mr->rkey, sg->key);
 903         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 904         return 0;
 905 }
 906
 907 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 908                 struct nvme_rdma_request *req, struct nvme_command *c,
 909                 int count)
 910 {
 911         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 912         int nr;
 913
 914         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
 915         if (nr < count) {
 916                 if (nr < 0)
 917                         return nr;
 918                 return -EINVAL;
 919         }
 920
 921         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
 922
 923         req->reg_cqe.done = nvme_rdma_memreg_done;
 924         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
 925         req->reg_wr.wr.opcode = IB_WR_REG_MR;
 926         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
 927         req->reg_wr.wr.num_sge = 0;
 928         req->reg_wr.mr = req->mr;
 929         req->reg_wr.key = req->mr->rkey;
 930         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 931                              IB_ACCESS_REMOTE_READ |
 932                              IB_ACCESS_REMOTE_WRITE;
 933
 934         req->mr->need_inval = true;
 935
 936         sg->addr = cpu_to_le64(req->mr->iova);
 937         put_unaligned_le24(req->mr->length, sg->length);
 938         put_unaligned_le32(req->mr->rkey, sg->key);
 939         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
 940                         NVME_SGL_FMT_INVALIDATE;
 941
 942         return 0;
 943 }
 944
 945 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 946                 struct request *rq, unsigned int map_len,
 947                 struct nvme_command *c)
 948 {
 949         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 950         struct nvme_rdma_device *dev = queue->device;
 951         struct ib_device *ibdev = dev->dev;
 952         int nents, count;
 953         int ret;
 954
 955         req->num_sge = 1;
 956         req->inline_data = false;
 957         req->mr->need_inval = false;
 958
 959         c->common.flags |= NVME_CMD_SGL_METABUF;
 960
 961         if (!blk_rq_bytes(rq))
 962                 return nvme_rdma_set_sg_null(c);
 963
 964         req->sg_table.sgl = req->first_sgl;
 965         ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
 966                                 req->sg_table.sgl);
 967         if (ret)
 968                 return -ENOMEM;
 969
 970         nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
 971         BUG_ON(nents > rq->nr_phys_segments);
 972         req->nents = nents;
 973
 974         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
 975                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 976         if (unlikely(count <= 0)) {
 977                 sg_free_table_chained(&req->sg_table, true);
 978                 return -EIO;
 979         }
 980
 981         if (count == 1) {
 982                 if (rq_data_dir(rq) == WRITE &&
 983                     map_len <= nvme_rdma_inline_data_size(queue) &&
 984                     nvme_rdma_queue_idx(queue))
 985                         return nvme_rdma_map_sg_inline(queue, req, c);
 986
 987                 if (!register_always)
 988                         return nvme_rdma_map_sg_single(queue, req, c);
 989         }
 990
 991         return nvme_rdma_map_sg_fr(queue, req, c, count);
 992 }
 993
 994 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 995 {
 996         if (unlikely(wc->status != IB_WC_SUCCESS))
 997                 nvme_rdma_wr_error(cq, wc, "SEND");
 998 }
 999
1000 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1001                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1002                 struct ib_send_wr *first, bool flush)
1003 {
1004         struct ib_send_wr wr, *bad_wr;
1005         int ret;
1006
1007         sge->addr   = qe->dma;
1008         sge->length = sizeof(struct nvme_command),
1009         sge->lkey   = queue->device->pd->local_dma_lkey;
1010
1011         qe->cqe.done = nvme_rdma_send_done;
1012
1013         wr.next       = NULL;
1014         wr.wr_cqe     = &qe->cqe;
1015         wr.sg_list    = sge;
1016         wr.num_sge    = num_sge;
1017         wr.opcode     = IB_WR_SEND;
1018         wr.send_flags = 0;
1019
1020         /*
1021          * Unsignalled send completions are another giant desaster in the
1022          * IB Verbs spec:  If we don't regularly post signalled sends
1023          * the send queue will fill up and only a QP reset will rescue us.
1024          * Would have been way to obvious to handle this in hardware or
1025          * at least the RDMA stack..
1026          *
1027          * This messy and racy code sniplet is copy and pasted from the iSER
1028          * initiator, and the magic '32' comes from there as well.
1029          *
1030          * Always signal the flushes. The magic request used for the flush
1031          * sequencer is not allocated in our driver's tagset and it's
1032          * triggered to be freed by blk_cleanup_queue(). So we need to
1033          * always mark it as signaled to ensure that the "wr_cqe", which is
1034          * embeded in request's payload, is not freed when __ib_process_cq()
1035          * calls wr_cqe->done().
1036          */
1037         if ((++queue->sig_count % 32) == 0 || flush)
1038                 wr.send_flags |= IB_SEND_SIGNALED;
1039
1040         if (first)
1041                 first->next = &wr;
1042         else
1043                 first = &wr;
1044
1045         ret = ib_post_send(queue->qp, first, &bad_wr);
1046         if (ret) {
1047                 dev_err(queue->ctrl->ctrl.device,
1048                              "%s failed with error code %d\n", __func__, ret);
1049         }
1050         return ret;
1051 }
1052
1053 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1054                 struct nvme_rdma_qe *qe)
1055 {
1056         struct ib_recv_wr wr, *bad_wr;
1057         struct ib_sge list;
1058         int ret;
1059
1060         list.addr   = qe->dma;
1061         list.length = sizeof(struct nvme_completion);
1062         list.lkey   = queue->device->pd->local_dma_lkey;
1063
1064         qe->cqe.done = nvme_rdma_recv_done;
1065
1066         wr.next     = NULL;
1067         wr.wr_cqe   = &qe->cqe;
1068         wr.sg_list  = &list;
1069         wr.num_sge  = 1;
1070
1071         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1072         if (ret) {
1073                 dev_err(queue->ctrl->ctrl.device,
1074                         "%s failed with error code %d\n", __func__, ret);
1075         }
1076         return ret;
1077 }
1078
1079 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1080 {
1081         u32 queue_idx = nvme_rdma_queue_idx(queue);
1082
1083         if (queue_idx == 0)
1084                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1085         return queue->ctrl->tag_set.tags[queue_idx - 1];
1086 }
1087
1088 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1089 {
1090         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1091         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1092         struct ib_device *dev = queue->device->dev;
1093         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1094         struct nvme_command *cmd = sqe->data;
1095         struct ib_sge sge;
1096         int ret;
1097
1098         if (WARN_ON_ONCE(aer_idx != 0))
1099                 return;
1100
1101         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1102
1103         memset(cmd, 0, sizeof(*cmd));
1104         cmd->common.opcode = nvme_admin_async_event;
1105         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1106         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1107         nvme_rdma_set_sg_null(cmd);
1108
1109         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1110                         DMA_TO_DEVICE);
1111
1112         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1113         WARN_ON_ONCE(ret);
1114 }
1115
1116 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1117                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1118 {
1119         u16 status = le16_to_cpu(cqe->status);
1120         struct request *rq;
1121         struct nvme_rdma_request *req;
1122         int ret = 0;
1123
1124         status >>= 1;
1125
1126         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1127         if (!rq) {
1128                 dev_err(queue->ctrl->ctrl.device,
1129                         "tag 0x%x on QP %#x not found\n",
1130                         cqe->command_id, queue->qp->qp_num);
1131                 nvme_rdma_error_recovery(queue->ctrl);
1132                 return ret;
1133         }
1134         req = blk_mq_rq_to_pdu(rq);
1135
1136         if (rq->cmd_type == REQ_TYPE_DRV_PRIV && rq->special)
1137                 memcpy(rq->special, cqe, sizeof(*cqe));
1138
1139         if (rq->tag == tag)
1140                 ret = 1;
1141
1142         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1143             wc->ex.invalidate_rkey == req->mr->rkey)
1144                 req->mr->need_inval = false;
1145
1146         blk_mq_complete_request(rq, status);
1147
1148         return ret;
1149 }
1150
1151 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1152 {
1153         struct nvme_rdma_qe *qe =
1154                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1155         struct nvme_rdma_queue *queue = cq->cq_context;
1156         struct ib_device *ibdev = queue->device->dev;
1157         struct nvme_completion *cqe = qe->data;
1158         const size_t len = sizeof(struct nvme_completion);
1159         int ret = 0;
1160
1161         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1162                 nvme_rdma_wr_error(cq, wc, "RECV");
1163                 return 0;
1164         }
1165
1166         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1167         /*
1168          * AEN requests are special as they don't time out and can
1169          * survive any kind of queue freeze and often don't respond to
1170          * aborts.  We don't even bother to allocate a struct request
1171          * for them but rather special case them here.
1172          */
1173         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1174                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1175                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe);
1176         else
1177                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1178         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1179
1180         nvme_rdma_post_recv(queue, qe);
1181         return ret;
1182 }
1183
1184 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1185 {
1186         __nvme_rdma_recv_done(cq, wc, -1);
1187 }
1188
1189 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1190 {
1191         int ret, i;
1192
1193         for (i = 0; i < queue->queue_size; i++) {
1194                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1195                 if (ret)
1196                         goto out_destroy_queue_ib;
1197         }
1198
1199         return 0;
1200
1201 out_destroy_queue_ib:
1202         nvme_rdma_destroy_queue_ib(queue);
1203         return ret;
1204 }
1205
1206 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1207                 struct rdma_cm_event *ev)
1208 {
1209         if (ev->param.conn.private_data_len) {
1210                 struct nvme_rdma_cm_rej *rej =
1211                         (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1212
1213                 dev_err(queue->ctrl->ctrl.device,
1214                         "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1215                 /* XXX: Think of something clever to do here... */
1216         } else {
1217                 dev_err(queue->ctrl->ctrl.device,
1218                         "Connect rejected, no private data.\n");
1219         }
1220
1221         return -ECONNRESET;
1222 }
1223
1224 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1225 {
1226         struct nvme_rdma_device *dev;
1227         int ret;
1228
1229         dev = nvme_rdma_find_get_device(queue->cm_id);
1230         if (!dev) {
1231                 dev_err(queue->cm_id->device->dma_device,
1232                         "no client data found!\n");
1233                 return -ECONNREFUSED;
1234         }
1235
1236         ret = nvme_rdma_create_queue_ib(queue, dev);
1237         if (ret) {
1238                 nvme_rdma_dev_put(dev);
1239                 goto out;
1240         }
1241
1242         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1243         if (ret) {
1244                 dev_err(queue->ctrl->ctrl.device,
1245                         "rdma_resolve_route failed (%d).\n",
1246                         queue->cm_error);
1247                 goto out_destroy_queue;
1248         }
1249
1250         return 0;
1251
1252 out_destroy_queue:
1253         nvme_rdma_destroy_queue_ib(queue);
1254 out:
1255         return ret;
1256 }
1257
1258 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1259 {
1260         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1261         struct rdma_conn_param param = { };
1262         struct nvme_rdma_cm_req priv = { };
1263         int ret;
1264
1265         param.qp_num = queue->qp->qp_num;
1266         param.flow_control = 1;
1267
1268         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1269         /* maximum retry count */
1270         param.retry_count = 7;
1271         param.rnr_retry_count = 7;
1272         param.private_data = &priv;
1273         param.private_data_len = sizeof(priv);
1274
1275         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1276         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1277         /*
1278          * set the admin queue depth to the minimum size
1279          * specified by the Fabrics standard.
1280          */
1281         if (priv.qid == 0) {
1282                 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1283                 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1284         } else {
1285                 /*
1286                  * current interpretation of the fabrics spec
1287                  * is at minimum you make hrqsize sqsize+1, or a
1288                  * 1's based representation of sqsize.
1289                  */
1290                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1291                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1292         }
1293
1294         ret = rdma_connect(queue->cm_id, &param);
1295         if (ret) {
1296                 dev_err(ctrl->ctrl.device,
1297                         "rdma_connect failed (%d).\n", ret);
1298                 goto out_destroy_queue_ib;
1299         }
1300
1301         return 0;
1302
1303 out_destroy_queue_ib:
1304         nvme_rdma_destroy_queue_ib(queue);
1305         return ret;
1306 }
1307
1308 /**
1309  * nvme_rdma_device_unplug() - Handle RDMA device unplug
1310  * @queue:      Queue that owns the cm_id that caught the event
1311  *
1312  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1313  * to unplug so we should take care of destroying our RDMA resources.
1314  * This event will be generated for each allocated cm_id.
1315  *
1316  * In our case, the RDMA resources are managed per controller and not
1317  * only per queue. So the way we handle this is we trigger an implicit
1318  * controller deletion upon the first DEVICE_REMOVAL event we see, and
1319  * hold the event inflight until the controller deletion is completed.
1320  *
1321  * One exception that we need to handle is the destruction of the cm_id
1322  * that caught the event. Since we hold the callout until the controller
1323  * deletion is completed, we'll deadlock if the controller deletion will
1324  * call rdma_destroy_id on this queue's cm_id. Thus, we claim ownership
1325  * of destroying this queue before-hand, destroy the queue resources,
1326  * then queue the controller deletion which won't destroy this queue and
1327  * we destroy the cm_id implicitely by returning a non-zero rc to the callout.
1328  */
1329 static int nvme_rdma_device_unplug(struct nvme_rdma_queue *queue)
1330 {
1331         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1332         int ret = 0;
1333
1334         /* Own the controller deletion */
1335         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1336                 return 0;
1337
1338         dev_warn(ctrl->ctrl.device,
1339                 "Got rdma device removal event, deleting ctrl\n");
1340
1341         /* Get rid of reconnect work if its running */
1342         cancel_delayed_work_sync(&ctrl->reconnect_work);
1343
1344         /* Disable the queue so ctrl delete won't free it */
1345         if (test_and_clear_bit(NVME_RDMA_Q_CONNECTED, &queue->flags)) {
1346                 /* Free this queue ourselves */
1347                 nvme_rdma_stop_queue(queue);
1348                 nvme_rdma_destroy_queue_ib(queue);
1349
1350                 /* Return non-zero so the cm_id will destroy implicitly */
1351                 ret = 1;
1352         }
1353
1354         /*
1355          * Queue controller deletion. Keep a reference until all
1356          * work is flushed since delete_work will free the ctrl mem
1357          */
1358         kref_get(&ctrl->ctrl.kref);
1359         queue_work(nvme_rdma_wq, &ctrl->delete_work);
1360         flush_work(&ctrl->delete_work);
1361         nvme_put_ctrl(&ctrl->ctrl);
1362
1363         return ret;
1364 }
1365
1366 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1367                 struct rdma_cm_event *ev)
1368 {
1369         struct nvme_rdma_queue *queue = cm_id->context;
1370         int cm_error = 0;
1371
1372         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1373                 rdma_event_msg(ev->event), ev->event,
1374                 ev->status, cm_id);
1375
1376         switch (ev->event) {
1377         case RDMA_CM_EVENT_ADDR_RESOLVED:
1378                 cm_error = nvme_rdma_addr_resolved(queue);
1379                 break;
1380         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1381                 cm_error = nvme_rdma_route_resolved(queue);
1382                 break;
1383         case RDMA_CM_EVENT_ESTABLISHED:
1384                 queue->cm_error = nvme_rdma_conn_established(queue);
1385                 /* complete cm_done regardless of success/failure */
1386                 complete(&queue->cm_done);
1387                 return 0;
1388         case RDMA_CM_EVENT_REJECTED:
1389                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1390                 break;
1391         case RDMA_CM_EVENT_ADDR_ERROR:
1392         case RDMA_CM_EVENT_ROUTE_ERROR:
1393         case RDMA_CM_EVENT_CONNECT_ERROR:
1394         case RDMA_CM_EVENT_UNREACHABLE:
1395                 dev_dbg(queue->ctrl->ctrl.device,
1396                         "CM error event %d\n", ev->event);
1397                 cm_error = -ECONNRESET;
1398                 break;
1399         case RDMA_CM_EVENT_DISCONNECTED:
1400         case RDMA_CM_EVENT_ADDR_CHANGE:
1401         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1402                 dev_dbg(queue->ctrl->ctrl.device,
1403                         "disconnect received - connection closed\n");
1404                 nvme_rdma_error_recovery(queue->ctrl);
1405                 break;
1406         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1407                 /* return 1 means impliciy CM ID destroy */
1408                 return nvme_rdma_device_unplug(queue);
1409         default:
1410                 dev_err(queue->ctrl->ctrl.device,
1411                         "Unexpected RDMA CM event (%d)\n", ev->event);
1412                 nvme_rdma_error_recovery(queue->ctrl);
1413                 break;
1414         }
1415
1416         if (cm_error) {
1417                 queue->cm_error = cm_error;
1418                 complete(&queue->cm_done);
1419         }
1420
1421         return 0;
1422 }
1423
1424 static enum blk_eh_timer_return
1425 nvme_rdma_timeout(struct request *rq, bool reserved)
1426 {
1427         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1428
1429         /* queue error recovery */
1430         nvme_rdma_error_recovery(req->queue->ctrl);
1431
1432         /* fail with DNR on cmd timeout */
1433         rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1434
1435         return BLK_EH_HANDLED;
1436 }
1437
1438 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1439                 const struct blk_mq_queue_data *bd)
1440 {
1441         struct nvme_ns *ns = hctx->queue->queuedata;
1442         struct nvme_rdma_queue *queue = hctx->driver_data;
1443         struct request *rq = bd->rq;
1444         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1445         struct nvme_rdma_qe *sqe = &req->sqe;
1446         struct nvme_command *c = sqe->data;
1447         bool flush = false;
1448         struct ib_device *dev;
1449         unsigned int map_len;
1450         int ret;
1451
1452         WARN_ON_ONCE(rq->tag < 0);
1453
1454         dev = queue->device->dev;
1455         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1456                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1457
1458         ret = nvme_setup_cmd(ns, rq, c);
1459         if (ret)
1460                 return ret;
1461
1462         c->common.command_id = rq->tag;
1463         blk_mq_start_request(rq);
1464
1465         map_len = nvme_map_len(rq);
1466         ret = nvme_rdma_map_data(queue, rq, map_len, c);
1467         if (ret < 0) {
1468                 dev_err(queue->ctrl->ctrl.device,
1469                              "Failed to map data (%d)\n", ret);
1470                 nvme_cleanup_cmd(rq);
1471                 goto err;
1472         }
1473
1474         ib_dma_sync_single_for_device(dev, sqe->dma,
1475                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1476
1477         if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1478                 flush = true;
1479         ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1480                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1481         if (ret) {
1482                 nvme_rdma_unmap_data(queue, rq);
1483                 goto err;
1484         }
1485
1486         return BLK_MQ_RQ_QUEUE_OK;
1487 err:
1488         return (ret == -ENOMEM || ret == -EAGAIN) ?
1489                 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1490 }
1491
1492 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1493 {
1494         struct nvme_rdma_queue *queue = hctx->driver_data;
1495         struct ib_cq *cq = queue->ib_cq;
1496         struct ib_wc wc;
1497         int found = 0;
1498
1499         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1500         while (ib_poll_cq(cq, 1, &wc) > 0) {
1501                 struct ib_cqe *cqe = wc.wr_cqe;
1502
1503                 if (cqe) {
1504                         if (cqe->done == nvme_rdma_recv_done)
1505                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1506                         else
1507                                 cqe->done(cq, &wc);
1508                 }
1509         }
1510
1511         return found;
1512 }
1513
1514 static void nvme_rdma_complete_rq(struct request *rq)
1515 {
1516         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1517         struct nvme_rdma_queue *queue = req->queue;
1518         int error = 0;
1519
1520         nvme_rdma_unmap_data(queue, rq);
1521
1522         if (unlikely(rq->errors)) {
1523                 if (nvme_req_needs_retry(rq, rq->errors)) {
1524                         nvme_requeue_req(rq);
1525                         return;
1526                 }
1527
1528                 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1529                         error = rq->errors;
1530                 else
1531                         error = nvme_error_status(rq->errors);
1532         }
1533
1534         blk_mq_end_request(rq, error);
1535 }
1536
1537 static struct blk_mq_ops nvme_rdma_mq_ops = {
1538         .queue_rq       = nvme_rdma_queue_rq,
1539         .complete       = nvme_rdma_complete_rq,
1540         .map_queue      = blk_mq_map_queue,
1541         .init_request   = nvme_rdma_init_request,
1542         .exit_request   = nvme_rdma_exit_request,
1543         .reinit_request = nvme_rdma_reinit_request,
1544         .init_hctx      = nvme_rdma_init_hctx,
1545         .poll           = nvme_rdma_poll,
1546         .timeout        = nvme_rdma_timeout,
1547 };
1548
1549 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1550         .queue_rq       = nvme_rdma_queue_rq,
1551         .complete       = nvme_rdma_complete_rq,
1552         .map_queue      = blk_mq_map_queue,
1553         .init_request   = nvme_rdma_init_admin_request,
1554         .exit_request   = nvme_rdma_exit_admin_request,
1555         .reinit_request = nvme_rdma_reinit_request,
1556         .init_hctx      = nvme_rdma_init_admin_hctx,
1557         .timeout        = nvme_rdma_timeout,
1558 };
1559
1560 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1561 {
1562         int error;
1563
1564         error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1565         if (error)
1566                 return error;
1567
1568         ctrl->device = ctrl->queues[0].device;
1569
1570         /*
1571          * We need a reference on the device as long as the tag_set is alive,
1572          * as the MRs in the request structures need a valid ib_device.
1573          */
1574         error = -EINVAL;
1575         if (!nvme_rdma_dev_get(ctrl->device))
1576                 goto out_free_queue;
1577
1578         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1579                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1580
1581         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1582         ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1583         ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1584         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1585         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1586         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1587                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1588         ctrl->admin_tag_set.driver_data = ctrl;
1589         ctrl->admin_tag_set.nr_hw_queues = 1;
1590         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1591
1592         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1593         if (error)
1594                 goto out_put_dev;
1595
1596         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1597         if (IS_ERR(ctrl->ctrl.admin_q)) {
1598                 error = PTR_ERR(ctrl->ctrl.admin_q);
1599                 goto out_free_tagset;
1600         }
1601
1602         error = nvmf_connect_admin_queue(&ctrl->ctrl);
1603         if (error)
1604                 goto out_cleanup_queue;
1605
1606         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1607         if (error) {
1608                 dev_err(ctrl->ctrl.device,
1609                         "prop_get NVME_REG_CAP failed\n");
1610                 goto out_cleanup_queue;
1611         }
1612
1613         ctrl->ctrl.sqsize =
1614                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1615
1616         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1617         if (error)
1618                 goto out_cleanup_queue;
1619
1620         ctrl->ctrl.max_hw_sectors =
1621                 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1622
1623         error = nvme_init_identify(&ctrl->ctrl);
1624         if (error)
1625                 goto out_cleanup_queue;
1626
1627         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1628                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
1629                         DMA_TO_DEVICE);
1630         if (error)
1631                 goto out_cleanup_queue;
1632
1633         nvme_start_keep_alive(&ctrl->ctrl);
1634
1635         return 0;
1636
1637 out_cleanup_queue:
1638         blk_cleanup_queue(ctrl->ctrl.admin_q);
1639 out_free_tagset:
1640         /* disconnect and drain the queue before freeing the tagset */
1641         nvme_rdma_stop_queue(&ctrl->queues[0]);
1642         blk_mq_free_tag_set(&ctrl->admin_tag_set);
1643 out_put_dev:
1644         nvme_rdma_dev_put(ctrl->device);
1645 out_free_queue:
1646         nvme_rdma_free_queue(&ctrl->queues[0]);
1647         return error;
1648 }
1649
1650 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1651 {
1652         nvme_stop_keep_alive(&ctrl->ctrl);
1653         cancel_work_sync(&ctrl->err_work);
1654         cancel_delayed_work_sync(&ctrl->reconnect_work);
1655
1656         if (ctrl->queue_count > 1) {
1657                 nvme_stop_queues(&ctrl->ctrl);
1658                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1659                                         nvme_cancel_request, &ctrl->ctrl);
1660                 nvme_rdma_free_io_queues(ctrl);
1661         }
1662
1663         if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1664                 nvme_shutdown_ctrl(&ctrl->ctrl);
1665
1666         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1667         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1668                                 nvme_cancel_request, &ctrl->ctrl);
1669         nvme_rdma_destroy_admin_queue(ctrl);
1670 }
1671
1672 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1673 {
1674         nvme_uninit_ctrl(&ctrl->ctrl);
1675         if (shutdown)
1676                 nvme_rdma_shutdown_ctrl(ctrl);
1677
1678         if (ctrl->ctrl.tagset) {
1679                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1680                 blk_mq_free_tag_set(&ctrl->tag_set);
1681                 nvme_rdma_dev_put(ctrl->device);
1682         }
1683
1684         nvme_put_ctrl(&ctrl->ctrl);
1685 }
1686
1687 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1688 {
1689         struct nvme_rdma_ctrl *ctrl = container_of(work,
1690                                 struct nvme_rdma_ctrl, delete_work);
1691
1692         __nvme_rdma_remove_ctrl(ctrl, true);
1693 }
1694
1695 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1696 {
1697         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1698                 return -EBUSY;
1699
1700         if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1701                 return -EBUSY;
1702
1703         return 0;
1704 }
1705
1706 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1707 {
1708         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1709         int ret = 0;
1710
1711         /*
1712          * Keep a reference until all work is flushed since
1713          * __nvme_rdma_del_ctrl can free the ctrl mem
1714          */
1715         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1716                 return -EBUSY;
1717         ret = __nvme_rdma_del_ctrl(ctrl);
1718         if (!ret)
1719                 flush_work(&ctrl->delete_work);
1720         nvme_put_ctrl(&ctrl->ctrl);
1721         return ret;
1722 }
1723
1724 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1725 {
1726         struct nvme_rdma_ctrl *ctrl = container_of(work,
1727                                 struct nvme_rdma_ctrl, delete_work);
1728
1729         __nvme_rdma_remove_ctrl(ctrl, false);
1730 }
1731
1732 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1733 {
1734         struct nvme_rdma_ctrl *ctrl = container_of(work,
1735                                         struct nvme_rdma_ctrl, reset_work);
1736         int ret;
1737         bool changed;
1738
1739         nvme_rdma_shutdown_ctrl(ctrl);
1740
1741         ret = nvme_rdma_configure_admin_queue(ctrl);
1742         if (ret) {
1743                 /* ctrl is already shutdown, just remove the ctrl */
1744                 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1745                 goto del_dead_ctrl;
1746         }
1747
1748         if (ctrl->queue_count > 1) {
1749                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1750                 if (ret)
1751                         goto del_dead_ctrl;
1752
1753                 ret = nvme_rdma_init_io_queues(ctrl);
1754                 if (ret)
1755                         goto del_dead_ctrl;
1756
1757                 ret = nvme_rdma_connect_io_queues(ctrl);
1758                 if (ret)
1759                         goto del_dead_ctrl;
1760         }
1761
1762         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1763         WARN_ON_ONCE(!changed);
1764
1765         if (ctrl->queue_count > 1) {
1766                 nvme_start_queues(&ctrl->ctrl);
1767                 nvme_queue_scan(&ctrl->ctrl);
1768                 nvme_queue_async_events(&ctrl->ctrl);
1769         }
1770
1771         return;
1772
1773 del_dead_ctrl:
1774         /* Deleting this dead controller... */
1775         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1776         WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1777 }
1778
1779 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1780 {
1781         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1782
1783         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1784                 return -EBUSY;
1785
1786         if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1787                 return -EBUSY;
1788
1789         flush_work(&ctrl->reset_work);
1790
1791         return 0;
1792 }
1793
1794 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1795         .name                   = "rdma",
1796         .module                 = THIS_MODULE,
1797         .is_fabrics             = true,
1798         .reg_read32             = nvmf_reg_read32,
1799         .reg_read64             = nvmf_reg_read64,
1800         .reg_write32            = nvmf_reg_write32,
1801         .reset_ctrl             = nvme_rdma_reset_ctrl,
1802         .free_ctrl              = nvme_rdma_free_ctrl,
1803         .submit_async_event     = nvme_rdma_submit_async_event,
1804         .delete_ctrl            = nvme_rdma_del_ctrl,
1805         .get_subsysnqn          = nvmf_get_subsysnqn,
1806         .get_address            = nvmf_get_address,
1807 };
1808
1809 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1810 {
1811         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1812         int ret;
1813
1814         ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1815         if (ret)
1816                 return ret;
1817
1818         ctrl->queue_count = opts->nr_io_queues + 1;
1819         if (ctrl->queue_count < 2)
1820                 return 0;
1821
1822         dev_info(ctrl->ctrl.device,
1823                 "creating %d I/O queues.\n", opts->nr_io_queues);
1824
1825         ret = nvme_rdma_init_io_queues(ctrl);
1826         if (ret)
1827                 return ret;
1828
1829         /*
1830          * We need a reference on the device as long as the tag_set is alive,
1831          * as the MRs in the request structures need a valid ib_device.
1832          */
1833         ret = -EINVAL;
1834         if (!nvme_rdma_dev_get(ctrl->device))
1835                 goto out_free_io_queues;
1836
1837         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1838         ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1839         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1840         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1841         ctrl->tag_set.numa_node = NUMA_NO_NODE;
1842         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1843         ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1844                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1845         ctrl->tag_set.driver_data = ctrl;
1846         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1847         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1848
1849         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1850         if (ret)
1851                 goto out_put_dev;
1852         ctrl->ctrl.tagset = &ctrl->tag_set;
1853
1854         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1855         if (IS_ERR(ctrl->ctrl.connect_q)) {
1856                 ret = PTR_ERR(ctrl->ctrl.connect_q);
1857                 goto out_free_tag_set;
1858         }
1859
1860         ret = nvme_rdma_connect_io_queues(ctrl);
1861         if (ret)
1862                 goto out_cleanup_connect_q;
1863
1864         return 0;
1865
1866 out_cleanup_connect_q:
1867         blk_cleanup_queue(ctrl->ctrl.connect_q);
1868 out_free_tag_set:
1869         blk_mq_free_tag_set(&ctrl->tag_set);
1870 out_put_dev:
1871         nvme_rdma_dev_put(ctrl->device);
1872 out_free_io_queues:
1873         nvme_rdma_free_io_queues(ctrl);
1874         return ret;
1875 }
1876
1877 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1878 {
1879         u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1880         size_t buflen = strlen(p);
1881
1882         /* XXX: handle IPv6 addresses */
1883
1884         if (buflen > INET_ADDRSTRLEN)
1885                 return -EINVAL;
1886         if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1887                 return -EINVAL;
1888         in_addr->sin_family = AF_INET;
1889         return 0;
1890 }
1891
1892 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1893                 struct nvmf_ctrl_options *opts)
1894 {
1895         struct nvme_rdma_ctrl *ctrl;
1896         int ret;
1897         bool changed;
1898
1899         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1900         if (!ctrl)
1901                 return ERR_PTR(-ENOMEM);
1902         ctrl->ctrl.opts = opts;
1903         INIT_LIST_HEAD(&ctrl->list);
1904
1905         ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1906         if (ret) {
1907                 pr_err("malformed IP address passed: %s\n", opts->traddr);
1908                 goto out_free_ctrl;
1909         }
1910
1911         if (opts->mask & NVMF_OPT_TRSVCID) {
1912                 u16 port;
1913
1914                 ret = kstrtou16(opts->trsvcid, 0, &port);
1915                 if (ret)
1916                         goto out_free_ctrl;
1917
1918                 ctrl->addr_in.sin_port = cpu_to_be16(port);
1919         } else {
1920                 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1921         }
1922
1923         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1924                                 0 /* no quirks, we're perfect! */);
1925         if (ret)
1926                 goto out_free_ctrl;
1927
1928         ctrl->reconnect_delay = opts->reconnect_delay;
1929         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1930                         nvme_rdma_reconnect_ctrl_work);
1931         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1932         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1933         INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1934         spin_lock_init(&ctrl->lock);
1935
1936         ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1937         ctrl->ctrl.sqsize = opts->queue_size - 1;
1938         ctrl->ctrl.kato = opts->kato;
1939
1940         ret = -ENOMEM;
1941         ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1942                                 GFP_KERNEL);
1943         if (!ctrl->queues)
1944                 goto out_uninit_ctrl;
1945
1946         ret = nvme_rdma_configure_admin_queue(ctrl);
1947         if (ret)
1948                 goto out_kfree_queues;
1949
1950         /* sanity check icdoff */
1951         if (ctrl->ctrl.icdoff) {
1952                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1953                 goto out_remove_admin_queue;
1954         }
1955
1956         /* sanity check keyed sgls */
1957         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1958                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1959                 goto out_remove_admin_queue;
1960         }
1961
1962         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1963                 /* warn if maxcmd is lower than queue_size */
1964                 dev_warn(ctrl->ctrl.device,
1965                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1966                         opts->queue_size, ctrl->ctrl.maxcmd);
1967                 opts->queue_size = ctrl->ctrl.maxcmd;
1968         }
1969
1970         if (opts->nr_io_queues) {
1971                 ret = nvme_rdma_create_io_queues(ctrl);
1972                 if (ret)
1973                         goto out_remove_admin_queue;
1974         }
1975
1976         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1977         WARN_ON_ONCE(!changed);
1978
1979         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1980                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1981
1982         kref_get(&ctrl->ctrl.kref);
1983
1984         mutex_lock(&nvme_rdma_ctrl_mutex);
1985         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1986         mutex_unlock(&nvme_rdma_ctrl_mutex);
1987
1988         if (opts->nr_io_queues) {
1989                 nvme_queue_scan(&ctrl->ctrl);
1990                 nvme_queue_async_events(&ctrl->ctrl);
1991         }
1992
1993         return &ctrl->ctrl;
1994
1995 out_remove_admin_queue:
1996         nvme_stop_keep_alive(&ctrl->ctrl);
1997         nvme_rdma_destroy_admin_queue(ctrl);
1998 out_kfree_queues:
1999         kfree(ctrl->queues);
2000 out_uninit_ctrl:
2001         nvme_uninit_ctrl(&ctrl->ctrl);
2002         nvme_put_ctrl(&ctrl->ctrl);
2003         if (ret > 0)
2004                 ret = -EIO;
2005         return ERR_PTR(ret);
2006 out_free_ctrl:
2007         kfree(ctrl);
2008         return ERR_PTR(ret);
2009 }
2010
2011 static struct nvmf_transport_ops nvme_rdma_transport = {
2012         .name           = "rdma",
2013         .required_opts  = NVMF_OPT_TRADDR,
2014         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
2015         .create_ctrl    = nvme_rdma_create_ctrl,
2016 };
2017
2018 static int __init nvme_rdma_init_module(void)
2019 {
2020         nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2021         if (!nvme_rdma_wq)
2022                 return -ENOMEM;
2023
2024         nvmf_register_transport(&nvme_rdma_transport);
2025         return 0;
2026 }
2027
2028 static void __exit nvme_rdma_cleanup_module(void)
2029 {
2030         struct nvme_rdma_ctrl *ctrl;
2031
2032         nvmf_unregister_transport(&nvme_rdma_transport);
2033
2034         mutex_lock(&nvme_rdma_ctrl_mutex);
2035         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2036                 __nvme_rdma_del_ctrl(ctrl);
2037         mutex_unlock(&nvme_rdma_ctrl_mutex);
2038
2039         destroy_workqueue(nvme_rdma_wq);
2040 }
2041
2042 module_init(nvme_rdma_init_module);
2043 module_exit(nvme_rdma_cleanup_module);
2044
2045 MODULE_LICENSE("GPL v2");