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         NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1),
  86         NVME_RDMA_Q_DELETING = (1 << 2),
  87 };
  88
  89 struct nvme_rdma_queue {
  90         struct nvme_rdma_qe     *rsp_ring;
  91         u8                      sig_count;
  92         int                     queue_size;
  93         size_t                  cmnd_capsule_len;
  94         struct nvme_rdma_ctrl   *ctrl;
  95         struct nvme_rdma_device *device;
  96         struct ib_cq            *ib_cq;
  97         struct ib_qp            *qp;
  98
  99         unsigned long           flags;
 100         struct rdma_cm_id       *cm_id;
 101         int                     cm_error;
 102         struct completion       cm_done;
 103 };
 104
 105 struct nvme_rdma_ctrl {
 106         /* read and written in the hot path */
 107         spinlock_t              lock;
 108
 109         /* read only in the hot path */
 110         struct nvme_rdma_queue  *queues;
 111         u32                     queue_count;
 112
 113         /* other member variables */
 114         struct blk_mq_tag_set   tag_set;
 115         struct work_struct      delete_work;
 116         struct work_struct      reset_work;
 117         struct work_struct      err_work;
 118
 119         struct nvme_rdma_qe     async_event_sqe;
 120
 121         int                     reconnect_delay;
 122         struct delayed_work     reconnect_work;
 123
 124         struct list_head        list;
 125
 126         struct blk_mq_tag_set   admin_tag_set;
 127         struct nvme_rdma_device *device;
 128
 129         u64                     cap;
 130         u32                     max_fr_pages;
 131
 132         union {
 133                 struct sockaddr addr;
 134                 struct sockaddr_in addr_in;
 135         };
 136
 137         struct nvme_ctrl        ctrl;
 138 };
 139
 140 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 141 {
 142         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 143 }
 144
 145 static LIST_HEAD(device_list);
 146 static DEFINE_MUTEX(device_list_mutex);
 147
 148 static LIST_HEAD(nvme_rdma_ctrl_list);
 149 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 150
 151 static struct workqueue_struct *nvme_rdma_wq;
 152
 153 /*
 154  * Disabling this option makes small I/O goes faster, but is fundamentally
 155  * unsafe.  With it turned off we will have to register a global rkey that
 156  * allows read and write access to all physical memory.
 157  */
 158 static bool register_always = true;
 159 module_param(register_always, bool, 0444);
 160 MODULE_PARM_DESC(register_always,
 161          "Use memory registration even for contiguous memory regions");
 162
 163 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 164                 struct rdma_cm_event *event);
 165 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 166
 167 /* XXX: really should move to a generic header sooner or later.. */
 168 static inline void put_unaligned_le24(u32 val, u8 *p)
 169 {
 170         *p++ = val;
 171         *p++ = val >> 8;
 172         *p++ = val >> 16;
 173 }
 174
 175 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 176 {
 177         return queue - queue->ctrl->queues;
 178 }
 179
 180 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 181 {
 182         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 183 }
 184
 185 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 186                 size_t capsule_size, enum dma_data_direction dir)
 187 {
 188         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 189         kfree(qe->data);
 190 }
 191
 192 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 193                 size_t capsule_size, enum dma_data_direction dir)
 194 {
 195         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 196         if (!qe->data)
 197                 return -ENOMEM;
 198
 199         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 200         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 201                 kfree(qe->data);
 202                 return -ENOMEM;
 203         }
 204
 205         return 0;
 206 }
 207
 208 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 209                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 210                 size_t capsule_size, enum dma_data_direction dir)
 211 {
 212         int i;
 213
 214         for (i = 0; i < ib_queue_size; i++)
 215                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 216         kfree(ring);
 217 }
 218
 219 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 220                 size_t ib_queue_size, size_t capsule_size,
 221                 enum dma_data_direction dir)
 222 {
 223         struct nvme_rdma_qe *ring;
 224         int i;
 225
 226         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 227         if (!ring)
 228                 return NULL;
 229
 230         for (i = 0; i < ib_queue_size; i++) {
 231                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 232                         goto out_free_ring;
 233         }
 234
 235         return ring;
 236
 237 out_free_ring:
 238         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 239         return NULL;
 240 }
 241
 242 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 243 {
 244         pr_debug("QP event %d\n", event->event);
 245 }
 246
 247 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 248 {
 249         wait_for_completion_interruptible_timeout(&queue->cm_done,
 250                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 251         return queue->cm_error;
 252 }
 253
 254 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 255 {
 256         struct nvme_rdma_device *dev = queue->device;
 257         struct ib_qp_init_attr init_attr;
 258         int ret;
 259
 260         memset(&init_attr, 0, sizeof(init_attr));
 261         init_attr.event_handler = nvme_rdma_qp_event;
 262         /* +1 for drain */
 263         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 264         /* +1 for drain */
 265         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 266         init_attr.cap.max_recv_sge = 1;
 267         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 268         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 269         init_attr.qp_type = IB_QPT_RC;
 270         init_attr.send_cq = queue->ib_cq;
 271         init_attr.recv_cq = queue->ib_cq;
 272
 273         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 274
 275         queue->qp = queue->cm_id->qp;
 276         return ret;
 277 }
 278
 279 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 280 {
 281         struct nvme_rdma_ctrl *ctrl = data;
 282         struct nvme_rdma_device *dev = ctrl->device;
 283         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 284         int ret = 0;
 285
 286         if (!req->mr->need_inval)
 287                 goto out;
 288
 289         ib_dereg_mr(req->mr);
 290
 291         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 292                         ctrl->max_fr_pages);
 293         if (IS_ERR(req->mr)) {
 294                 ret = PTR_ERR(req->mr);
 295                 req->mr = NULL;
 296                 goto out;
 297         }
 298
 299         req->mr->need_inval = false;
 300
 301 out:
 302         return ret;
 303 }
 304
 305 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
 306                 struct request *rq, unsigned int queue_idx)
 307 {
 308         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 309         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 310         struct nvme_rdma_device *dev = queue->device;
 311
 312         if (req->mr)
 313                 ib_dereg_mr(req->mr);
 314
 315         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 316                         DMA_TO_DEVICE);
 317 }
 318
 319 static void nvme_rdma_exit_request(void *data, struct request *rq,
 320                                 unsigned int hctx_idx, unsigned int rq_idx)
 321 {
 322         return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
 323 }
 324
 325 static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
 326                                 unsigned int hctx_idx, unsigned int rq_idx)
 327 {
 328         return __nvme_rdma_exit_request(data, rq, 0);
 329 }
 330
 331 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
 332                 struct request *rq, unsigned int queue_idx)
 333 {
 334         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 335         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 336         struct nvme_rdma_device *dev = queue->device;
 337         struct ib_device *ibdev = dev->dev;
 338         int ret;
 339
 340         BUG_ON(queue_idx >= ctrl->queue_count);
 341
 342         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 343                         DMA_TO_DEVICE);
 344         if (ret)
 345                 return ret;
 346
 347         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 348                         ctrl->max_fr_pages);
 349         if (IS_ERR(req->mr)) {
 350                 ret = PTR_ERR(req->mr);
 351                 goto out_free_qe;
 352         }
 353
 354         req->queue = queue;
 355
 356         return 0;
 357
 358 out_free_qe:
 359         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 360                         DMA_TO_DEVICE);
 361         return -ENOMEM;
 362 }
 363
 364 static int nvme_rdma_init_request(void *data, struct request *rq,
 365                                 unsigned int hctx_idx, unsigned int rq_idx,
 366                                 unsigned int numa_node)
 367 {
 368         return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
 369 }
 370
 371 static int nvme_rdma_init_admin_request(void *data, struct request *rq,
 372                                 unsigned int hctx_idx, unsigned int rq_idx,
 373                                 unsigned int numa_node)
 374 {
 375         return __nvme_rdma_init_request(data, rq, 0);
 376 }
 377
 378 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 379                 unsigned int hctx_idx)
 380 {
 381         struct nvme_rdma_ctrl *ctrl = data;
 382         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 383
 384         BUG_ON(hctx_idx >= ctrl->queue_count);
 385
 386         hctx->driver_data = queue;
 387         return 0;
 388 }
 389
 390 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 391                 unsigned int hctx_idx)
 392 {
 393         struct nvme_rdma_ctrl *ctrl = data;
 394         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 395
 396         BUG_ON(hctx_idx != 0);
 397
 398         hctx->driver_data = queue;
 399         return 0;
 400 }
 401
 402 static void nvme_rdma_free_dev(struct kref *ref)
 403 {
 404         struct nvme_rdma_device *ndev =
 405                 container_of(ref, struct nvme_rdma_device, ref);
 406
 407         mutex_lock(&device_list_mutex);
 408         list_del(&ndev->entry);
 409         mutex_unlock(&device_list_mutex);
 410
 411         if (!register_always)
 412                 ib_dereg_mr(ndev->mr);
 413         ib_dealloc_pd(ndev->pd);
 414
 415         kfree(ndev);
 416 }
 417
 418 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 419 {
 420         kref_put(&dev->ref, nvme_rdma_free_dev);
 421 }
 422
 423 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 424 {
 425         return kref_get_unless_zero(&dev->ref);
 426 }
 427
 428 static struct nvme_rdma_device *
 429 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 430 {
 431         struct nvme_rdma_device *ndev;
 432
 433         mutex_lock(&device_list_mutex);
 434         list_for_each_entry(ndev, &device_list, entry) {
 435                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 436                     nvme_rdma_dev_get(ndev))
 437                         goto out_unlock;
 438         }
 439
 440         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 441         if (!ndev)
 442                 goto out_err;
 443
 444         ndev->dev = cm_id->device;
 445         kref_init(&ndev->ref);
 446
 447         ndev->pd = ib_alloc_pd(ndev->dev);
 448         if (IS_ERR(ndev->pd))
 449                 goto out_free_dev;
 450
 451         if (!register_always) {
 452                 ndev->mr = ib_get_dma_mr(ndev->pd,
 453                                             IB_ACCESS_LOCAL_WRITE |
 454                                             IB_ACCESS_REMOTE_READ |
 455                                             IB_ACCESS_REMOTE_WRITE);
 456                 if (IS_ERR(ndev->mr))
 457                         goto out_free_pd;
 458         }
 459
 460         if (!(ndev->dev->attrs.device_cap_flags &
 461               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 462                 dev_err(&ndev->dev->dev,
 463                         "Memory registrations not supported.\n");
 464                 goto out_free_mr;
 465         }
 466
 467         list_add(&ndev->entry, &device_list);
 468 out_unlock:
 469         mutex_unlock(&device_list_mutex);
 470         return ndev;
 471
 472 out_free_mr:
 473         if (!register_always)
 474                 ib_dereg_mr(ndev->mr);
 475 out_free_pd:
 476         ib_dealloc_pd(ndev->pd);
 477 out_free_dev:
 478         kfree(ndev);
 479 out_err:
 480         mutex_unlock(&device_list_mutex);
 481         return NULL;
 482 }
 483
 484 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 485 {
 486         struct nvme_rdma_device *dev;
 487         struct ib_device *ibdev;
 488
 489         if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags))
 490                 return;
 491
 492         dev = queue->device;
 493         ibdev = dev->dev;
 494         rdma_destroy_qp(queue->cm_id);
 495         ib_free_cq(queue->ib_cq);
 496
 497         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 498                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 499
 500         nvme_rdma_dev_put(dev);
 501 }
 502
 503 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
 504                 struct nvme_rdma_device *dev)
 505 {
 506         struct ib_device *ibdev = dev->dev;
 507         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 508         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 509         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 510
 511         int ret;
 512
 513         queue->device = dev;
 514
 515         /*
 516          * The admin queue is barely used once the controller is live, so don't
 517          * bother to spread it out.
 518          */
 519         if (idx == 0)
 520                 comp_vector = 0;
 521         else
 522                 comp_vector = idx % ibdev->num_comp_vectors;
 523
 524
 525         /* +1 for ib_stop_cq */
 526         queue->ib_cq = ib_alloc_cq(dev->dev, queue,
 527                                 cq_factor * queue->queue_size + 1, comp_vector,
 528                                 IB_POLL_SOFTIRQ);
 529         if (IS_ERR(queue->ib_cq)) {
 530                 ret = PTR_ERR(queue->ib_cq);
 531                 goto out;
 532         }
 533
 534         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 535         if (ret)
 536                 goto out_destroy_ib_cq;
 537
 538         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 539                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 540         if (!queue->rsp_ring) {
 541                 ret = -ENOMEM;
 542                 goto out_destroy_qp;
 543         }
 544         set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags);
 545
 546         return 0;
 547
 548 out_destroy_qp:
 549         ib_destroy_qp(queue->qp);
 550 out_destroy_ib_cq:
 551         ib_free_cq(queue->ib_cq);
 552 out:
 553         return ret;
 554 }
 555
 556 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
 557                 int idx, size_t queue_size)
 558 {
 559         struct nvme_rdma_queue *queue;
 560         int ret;
 561
 562         queue = &ctrl->queues[idx];
 563         queue->ctrl = ctrl;
 564         init_completion(&queue->cm_done);
 565
 566         if (idx > 0)
 567                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 568         else
 569                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 570
 571         queue->queue_size = queue_size;
 572
 573         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 574                         RDMA_PS_TCP, IB_QPT_RC);
 575         if (IS_ERR(queue->cm_id)) {
 576                 dev_info(ctrl->ctrl.device,
 577                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 578                 return PTR_ERR(queue->cm_id);
 579         }
 580
 581         queue->cm_error = -ETIMEDOUT;
 582         ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
 583                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 584         if (ret) {
 585                 dev_info(ctrl->ctrl.device,
 586                         "rdma_resolve_addr failed (%d).\n", ret);
 587                 goto out_destroy_cm_id;
 588         }
 589
 590         ret = nvme_rdma_wait_for_cm(queue);
 591         if (ret) {
 592                 dev_info(ctrl->ctrl.device,
 593                         "rdma_resolve_addr wait failed (%d).\n", ret);
 594                 goto out_destroy_cm_id;
 595         }
 596
 597         clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
 598         set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
 599
 600         return 0;
 601
 602 out_destroy_cm_id:
 603         nvme_rdma_destroy_queue_ib(queue);
 604         rdma_destroy_id(queue->cm_id);
 605         return ret;
 606 }
 607
 608 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 609 {
 610         rdma_disconnect(queue->cm_id);
 611         ib_drain_qp(queue->qp);
 612 }
 613
 614 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 615 {
 616         nvme_rdma_destroy_queue_ib(queue);
 617         rdma_destroy_id(queue->cm_id);
 618 }
 619
 620 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
 621 {
 622         if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
 623                 return;
 624         nvme_rdma_stop_queue(queue);
 625         nvme_rdma_free_queue(queue);
 626 }
 627
 628 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 629 {
 630         int i;
 631
 632         for (i = 1; i < ctrl->queue_count; i++)
 633                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 634 }
 635
 636 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
 637 {
 638         int i, ret = 0;
 639
 640         for (i = 1; i < ctrl->queue_count; i++) {
 641                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 642                 if (ret)
 643                         break;
 644         }
 645
 646         return ret;
 647 }
 648
 649 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
 650 {
 651         int i, ret;
 652
 653         for (i = 1; i < ctrl->queue_count; i++) {
 654                 ret = nvme_rdma_init_queue(ctrl, i,
 655                                            ctrl->ctrl.opts->queue_size);
 656                 if (ret) {
 657                         dev_info(ctrl->ctrl.device,
 658                                 "failed to initialize i/o queue: %d\n", ret);
 659                         goto out_free_queues;
 660                 }
 661         }
 662
 663         return 0;
 664
 665 out_free_queues:
 666         for (i--; i >= 1; i--)
 667                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 668
 669         return ret;
 670 }
 671
 672 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
 673 {
 674         nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
 675                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 676         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 677         blk_cleanup_queue(ctrl->ctrl.admin_q);
 678         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 679         nvme_rdma_dev_put(ctrl->device);
 680 }
 681
 682 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 683 {
 684         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 685
 686         if (list_empty(&ctrl->list))
 687                 goto free_ctrl;
 688
 689         mutex_lock(&nvme_rdma_ctrl_mutex);
 690         list_del(&ctrl->list);
 691         mutex_unlock(&nvme_rdma_ctrl_mutex);
 692
 693         kfree(ctrl->queues);
 694         nvmf_free_options(nctrl->opts);
 695 free_ctrl:
 696         kfree(ctrl);
 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         if (ctrl->queue_count > 1) {
 707                 nvme_rdma_free_io_queues(ctrl);
 708
 709                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
 710                 if (ret)
 711                         goto requeue;
 712         }
 713
 714         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 715
 716         ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
 717         if (ret)
 718                 goto requeue;
 719
 720         ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
 721         if (ret)
 722                 goto requeue;
 723
 724         blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
 725
 726         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 727         if (ret)
 728                 goto stop_admin_q;
 729
 730         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 731         if (ret)
 732                 goto stop_admin_q;
 733
 734         nvme_start_keep_alive(&ctrl->ctrl);
 735
 736         if (ctrl->queue_count > 1) {
 737                 ret = nvme_rdma_init_io_queues(ctrl);
 738                 if (ret)
 739                         goto stop_admin_q;
 740
 741                 ret = nvme_rdma_connect_io_queues(ctrl);
 742                 if (ret)
 743                         goto stop_admin_q;
 744         }
 745
 746         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 747         WARN_ON_ONCE(!changed);
 748
 749         if (ctrl->queue_count > 1) {
 750                 nvme_start_queues(&ctrl->ctrl);
 751                 nvme_queue_scan(&ctrl->ctrl);
 752                 nvme_queue_async_events(&ctrl->ctrl);
 753         }
 754
 755         dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
 756
 757         return;
 758
 759 stop_admin_q:
 760         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 761 requeue:
 762         /* Make sure we are not resetting/deleting */
 763         if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
 764                 dev_info(ctrl->ctrl.device,
 765                         "Failed reconnect attempt, requeueing...\n");
 766                 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 767                                         ctrl->reconnect_delay * HZ);
 768         }
 769 }
 770
 771 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 772 {
 773         struct nvme_rdma_ctrl *ctrl = container_of(work,
 774                         struct nvme_rdma_ctrl, err_work);
 775         int i;
 776
 777         nvme_stop_keep_alive(&ctrl->ctrl);
 778
 779         for (i = 0; i < ctrl->queue_count; i++)
 780                 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
 781
 782         if (ctrl->queue_count > 1)
 783                 nvme_stop_queues(&ctrl->ctrl);
 784         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 785
 786         /* We must take care of fastfail/requeue all our inflight requests */
 787         if (ctrl->queue_count > 1)
 788                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 789                                         nvme_cancel_request, &ctrl->ctrl);
 790         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 791                                 nvme_cancel_request, &ctrl->ctrl);
 792
 793         dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
 794                 ctrl->reconnect_delay);
 795
 796         queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 797                                 ctrl->reconnect_delay * HZ);
 798 }
 799
 800 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 801 {
 802         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 803                 return;
 804
 805         queue_work(nvme_rdma_wq, &ctrl->err_work);
 806 }
 807
 808 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 809                 const char *op)
 810 {
 811         struct nvme_rdma_queue *queue = cq->cq_context;
 812         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 813
 814         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 815                 dev_info(ctrl->ctrl.device,
 816                              "%s for CQE 0x%p failed with status %s (%d)\n",
 817                              op, wc->wr_cqe,
 818                              ib_wc_status_msg(wc->status), wc->status);
 819         nvme_rdma_error_recovery(ctrl);
 820 }
 821
 822 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
 823 {
 824         if (unlikely(wc->status != IB_WC_SUCCESS))
 825                 nvme_rdma_wr_error(cq, wc, "MEMREG");
 826 }
 827
 828 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 829 {
 830         if (unlikely(wc->status != IB_WC_SUCCESS))
 831                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
 832 }
 833
 834 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
 835                 struct nvme_rdma_request *req)
 836 {
 837         struct ib_send_wr *bad_wr;
 838         struct ib_send_wr wr = {
 839                 .opcode             = IB_WR_LOCAL_INV,
 840                 .next               = NULL,
 841                 .num_sge            = 0,
 842                 .send_flags         = 0,
 843                 .ex.invalidate_rkey = req->mr->rkey,
 844         };
 845
 846         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
 847         wr.wr_cqe = &req->reg_cqe;
 848
 849         return ib_post_send(queue->qp, &wr, &bad_wr);
 850 }
 851
 852 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 853                 struct request *rq)
 854 {
 855         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 856         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 857         struct nvme_rdma_device *dev = queue->device;
 858         struct ib_device *ibdev = dev->dev;
 859         int res;
 860
 861         if (!blk_rq_bytes(rq))
 862                 return;
 863
 864         if (req->mr->need_inval) {
 865                 res = nvme_rdma_inv_rkey(queue, req);
 866                 if (res < 0) {
 867                         dev_err(ctrl->ctrl.device,
 868                                 "Queueing INV WR for rkey %#x failed (%d)\n",
 869                                 req->mr->rkey, res);
 870                         nvme_rdma_error_recovery(queue->ctrl);
 871                 }
 872         }
 873
 874         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 875                         req->nents, rq_data_dir(rq) ==
 876                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 877
 878         nvme_cleanup_cmd(rq);
 879         sg_free_table_chained(&req->sg_table, true);
 880 }
 881
 882 static int nvme_rdma_set_sg_null(struct nvme_command *c)
 883 {
 884         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 885
 886         sg->addr = 0;
 887         put_unaligned_le24(0, sg->length);
 888         put_unaligned_le32(0, sg->key);
 889         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 890         return 0;
 891 }
 892
 893 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
 894                 struct nvme_rdma_request *req, struct nvme_command *c)
 895 {
 896         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
 897
 898         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
 899         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
 900         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
 901
 902         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
 903         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
 904         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
 905
 906         req->inline_data = true;
 907         req->num_sge++;
 908         return 0;
 909 }
 910
 911 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
 912                 struct nvme_rdma_request *req, struct nvme_command *c)
 913 {
 914         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 915
 916         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
 917         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
 918         put_unaligned_le32(queue->device->mr->rkey, sg->key);
 919         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 920         return 0;
 921 }
 922
 923 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 924                 struct nvme_rdma_request *req, struct nvme_command *c,
 925                 int count)
 926 {
 927         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 928         int nr;
 929
 930         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
 931         if (nr < count) {
 932                 if (nr < 0)
 933                         return nr;
 934                 return -EINVAL;
 935         }
 936
 937         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
 938
 939         req->reg_cqe.done = nvme_rdma_memreg_done;
 940         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
 941         req->reg_wr.wr.opcode = IB_WR_REG_MR;
 942         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
 943         req->reg_wr.wr.num_sge = 0;
 944         req->reg_wr.mr = req->mr;
 945         req->reg_wr.key = req->mr->rkey;
 946         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 947                              IB_ACCESS_REMOTE_READ |
 948                              IB_ACCESS_REMOTE_WRITE;
 949
 950         req->mr->need_inval = true;
 951
 952         sg->addr = cpu_to_le64(req->mr->iova);
 953         put_unaligned_le24(req->mr->length, sg->length);
 954         put_unaligned_le32(req->mr->rkey, sg->key);
 955         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
 956                         NVME_SGL_FMT_INVALIDATE;
 957
 958         return 0;
 959 }
 960
 961 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 962                 struct request *rq, unsigned int map_len,
 963                 struct nvme_command *c)
 964 {
 965         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 966         struct nvme_rdma_device *dev = queue->device;
 967         struct ib_device *ibdev = dev->dev;
 968         int nents, count;
 969         int ret;
 970
 971         req->num_sge = 1;
 972         req->inline_data = false;
 973         req->mr->need_inval = false;
 974
 975         c->common.flags |= NVME_CMD_SGL_METABUF;
 976
 977         if (!blk_rq_bytes(rq))
 978                 return nvme_rdma_set_sg_null(c);
 979
 980         req->sg_table.sgl = req->first_sgl;
 981         ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
 982                                 req->sg_table.sgl);
 983         if (ret)
 984                 return -ENOMEM;
 985
 986         nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
 987         BUG_ON(nents > rq->nr_phys_segments);
 988         req->nents = nents;
 989
 990         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
 991                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 992         if (unlikely(count <= 0)) {
 993                 sg_free_table_chained(&req->sg_table, true);
 994                 return -EIO;
 995         }
 996
 997         if (count == 1) {
 998                 if (rq_data_dir(rq) == WRITE &&
 999                     map_len <= nvme_rdma_inline_data_size(queue) &&
1000                     nvme_rdma_queue_idx(queue))
1001                         return nvme_rdma_map_sg_inline(queue, req, c);
1002
1003                 if (!register_always)
1004                         return nvme_rdma_map_sg_single(queue, req, c);
1005         }
1006
1007         return nvme_rdma_map_sg_fr(queue, req, c, count);
1008 }
1009
1010 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1011 {
1012         if (unlikely(wc->status != IB_WC_SUCCESS))
1013                 nvme_rdma_wr_error(cq, wc, "SEND");
1014 }
1015
1016 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1017                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1018                 struct ib_send_wr *first, bool flush)
1019 {
1020         struct ib_send_wr wr, *bad_wr;
1021         int ret;
1022
1023         sge->addr   = qe->dma;
1024         sge->length = sizeof(struct nvme_command),
1025         sge->lkey   = queue->device->pd->local_dma_lkey;
1026
1027         qe->cqe.done = nvme_rdma_send_done;
1028
1029         wr.next       = NULL;
1030         wr.wr_cqe     = &qe->cqe;
1031         wr.sg_list    = sge;
1032         wr.num_sge    = num_sge;
1033         wr.opcode     = IB_WR_SEND;
1034         wr.send_flags = 0;
1035
1036         /*
1037          * Unsignalled send completions are another giant desaster in the
1038          * IB Verbs spec:  If we don't regularly post signalled sends
1039          * the send queue will fill up and only a QP reset will rescue us.
1040          * Would have been way to obvious to handle this in hardware or
1041          * at least the RDMA stack..
1042          *
1043          * This messy and racy code sniplet is copy and pasted from the iSER
1044          * initiator, and the magic '32' comes from there as well.
1045          *
1046          * Always signal the flushes. The magic request used for the flush
1047          * sequencer is not allocated in our driver's tagset and it's
1048          * triggered to be freed by blk_cleanup_queue(). So we need to
1049          * always mark it as signaled to ensure that the "wr_cqe", which is
1050          * embeded in request's payload, is not freed when __ib_process_cq()
1051          * calls wr_cqe->done().
1052          */
1053         if ((++queue->sig_count % 32) == 0 || flush)
1054                 wr.send_flags |= IB_SEND_SIGNALED;
1055
1056         if (first)
1057                 first->next = &wr;
1058         else
1059                 first = &wr;
1060
1061         ret = ib_post_send(queue->qp, first, &bad_wr);
1062         if (ret) {
1063                 dev_err(queue->ctrl->ctrl.device,
1064                              "%s failed with error code %d\n", __func__, ret);
1065         }
1066         return ret;
1067 }
1068
1069 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1070                 struct nvme_rdma_qe *qe)
1071 {
1072         struct ib_recv_wr wr, *bad_wr;
1073         struct ib_sge list;
1074         int ret;
1075
1076         list.addr   = qe->dma;
1077         list.length = sizeof(struct nvme_completion);
1078         list.lkey   = queue->device->pd->local_dma_lkey;
1079
1080         qe->cqe.done = nvme_rdma_recv_done;
1081
1082         wr.next     = NULL;
1083         wr.wr_cqe   = &qe->cqe;
1084         wr.sg_list  = &list;
1085         wr.num_sge  = 1;
1086
1087         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1088         if (ret) {
1089                 dev_err(queue->ctrl->ctrl.device,
1090                         "%s failed with error code %d\n", __func__, ret);
1091         }
1092         return ret;
1093 }
1094
1095 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1096 {
1097         u32 queue_idx = nvme_rdma_queue_idx(queue);
1098
1099         if (queue_idx == 0)
1100                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1101         return queue->ctrl->tag_set.tags[queue_idx - 1];
1102 }
1103
1104 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1105 {
1106         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1107         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1108         struct ib_device *dev = queue->device->dev;
1109         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1110         struct nvme_command *cmd = sqe->data;
1111         struct ib_sge sge;
1112         int ret;
1113
1114         if (WARN_ON_ONCE(aer_idx != 0))
1115                 return;
1116
1117         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1118
1119         memset(cmd, 0, sizeof(*cmd));
1120         cmd->common.opcode = nvme_admin_async_event;
1121         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1122         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1123         nvme_rdma_set_sg_null(cmd);
1124
1125         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1126                         DMA_TO_DEVICE);
1127
1128         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1129         WARN_ON_ONCE(ret);
1130 }
1131
1132 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1133                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1134 {
1135         u16 status = le16_to_cpu(cqe->status);
1136         struct request *rq;
1137         struct nvme_rdma_request *req;
1138         int ret = 0;
1139
1140         status >>= 1;
1141
1142         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1143         if (!rq) {
1144                 dev_err(queue->ctrl->ctrl.device,
1145                         "tag 0x%x on QP %#x not found\n",
1146                         cqe->command_id, queue->qp->qp_num);
1147                 nvme_rdma_error_recovery(queue->ctrl);
1148                 return ret;
1149         }
1150         req = blk_mq_rq_to_pdu(rq);
1151
1152         if (rq->cmd_type == REQ_TYPE_DRV_PRIV && rq->special)
1153                 memcpy(rq->special, cqe, sizeof(*cqe));
1154
1155         if (rq->tag == tag)
1156                 ret = 1;
1157
1158         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1159             wc->ex.invalidate_rkey == req->mr->rkey)
1160                 req->mr->need_inval = false;
1161
1162         blk_mq_complete_request(rq, status);
1163
1164         return ret;
1165 }
1166
1167 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1168 {
1169         struct nvme_rdma_qe *qe =
1170                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1171         struct nvme_rdma_queue *queue = cq->cq_context;
1172         struct ib_device *ibdev = queue->device->dev;
1173         struct nvme_completion *cqe = qe->data;
1174         const size_t len = sizeof(struct nvme_completion);
1175         int ret = 0;
1176
1177         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1178                 nvme_rdma_wr_error(cq, wc, "RECV");
1179                 return 0;
1180         }
1181
1182         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1183         /*
1184          * AEN requests are special as they don't time out and can
1185          * survive any kind of queue freeze and often don't respond to
1186          * aborts.  We don't even bother to allocate a struct request
1187          * for them but rather special case them here.
1188          */
1189         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1190                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1191                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe);
1192         else
1193                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1194         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1195
1196         nvme_rdma_post_recv(queue, qe);
1197         return ret;
1198 }
1199
1200 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1201 {
1202         __nvme_rdma_recv_done(cq, wc, -1);
1203 }
1204
1205 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1206 {
1207         int ret, i;
1208
1209         for (i = 0; i < queue->queue_size; i++) {
1210                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1211                 if (ret)
1212                         goto out_destroy_queue_ib;
1213         }
1214
1215         return 0;
1216
1217 out_destroy_queue_ib:
1218         nvme_rdma_destroy_queue_ib(queue);
1219         return ret;
1220 }
1221
1222 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1223                 struct rdma_cm_event *ev)
1224 {
1225         if (ev->param.conn.private_data_len) {
1226                 struct nvme_rdma_cm_rej *rej =
1227                         (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1228
1229                 dev_err(queue->ctrl->ctrl.device,
1230                         "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1231                 /* XXX: Think of something clever to do here... */
1232         } else {
1233                 dev_err(queue->ctrl->ctrl.device,
1234                         "Connect rejected, no private data.\n");
1235         }
1236
1237         return -ECONNRESET;
1238 }
1239
1240 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1241 {
1242         struct nvme_rdma_device *dev;
1243         int ret;
1244
1245         dev = nvme_rdma_find_get_device(queue->cm_id);
1246         if (!dev) {
1247                 dev_err(queue->cm_id->device->dma_device,
1248                         "no client data found!\n");
1249                 return -ECONNREFUSED;
1250         }
1251
1252         ret = nvme_rdma_create_queue_ib(queue, dev);
1253         if (ret) {
1254                 nvme_rdma_dev_put(dev);
1255                 goto out;
1256         }
1257
1258         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1259         if (ret) {
1260                 dev_err(queue->ctrl->ctrl.device,
1261                         "rdma_resolve_route failed (%d).\n",
1262                         queue->cm_error);
1263                 goto out_destroy_queue;
1264         }
1265
1266         return 0;
1267
1268 out_destroy_queue:
1269         nvme_rdma_destroy_queue_ib(queue);
1270 out:
1271         return ret;
1272 }
1273
1274 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1275 {
1276         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1277         struct rdma_conn_param param = { };
1278         struct nvme_rdma_cm_req priv = { };
1279         int ret;
1280
1281         param.qp_num = queue->qp->qp_num;
1282         param.flow_control = 1;
1283
1284         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1285         /* maximum retry count */
1286         param.retry_count = 7;
1287         param.rnr_retry_count = 7;
1288         param.private_data = &priv;
1289         param.private_data_len = sizeof(priv);
1290
1291         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1292         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1293         /*
1294          * set the admin queue depth to the minimum size
1295          * specified by the Fabrics standard.
1296          */
1297         if (priv.qid == 0) {
1298                 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1299                 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1300         } else {
1301                 /*
1302                  * current interpretation of the fabrics spec
1303                  * is at minimum you make hrqsize sqsize+1, or a
1304                  * 1's based representation of sqsize.
1305                  */
1306                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1307                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1308         }
1309
1310         ret = rdma_connect(queue->cm_id, &param);
1311         if (ret) {
1312                 dev_err(ctrl->ctrl.device,
1313                         "rdma_connect failed (%d).\n", ret);
1314                 goto out_destroy_queue_ib;
1315         }
1316
1317         return 0;
1318
1319 out_destroy_queue_ib:
1320         nvme_rdma_destroy_queue_ib(queue);
1321         return ret;
1322 }
1323
1324 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1325                 struct rdma_cm_event *ev)
1326 {
1327         struct nvme_rdma_queue *queue = cm_id->context;
1328         int cm_error = 0;
1329
1330         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1331                 rdma_event_msg(ev->event), ev->event,
1332                 ev->status, cm_id);
1333
1334         switch (ev->event) {
1335         case RDMA_CM_EVENT_ADDR_RESOLVED:
1336                 cm_error = nvme_rdma_addr_resolved(queue);
1337                 break;
1338         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1339                 cm_error = nvme_rdma_route_resolved(queue);
1340                 break;
1341         case RDMA_CM_EVENT_ESTABLISHED:
1342                 queue->cm_error = nvme_rdma_conn_established(queue);
1343                 /* complete cm_done regardless of success/failure */
1344                 complete(&queue->cm_done);
1345                 return 0;
1346         case RDMA_CM_EVENT_REJECTED:
1347                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1348                 break;
1349         case RDMA_CM_EVENT_ADDR_ERROR:
1350         case RDMA_CM_EVENT_ROUTE_ERROR:
1351         case RDMA_CM_EVENT_CONNECT_ERROR:
1352         case RDMA_CM_EVENT_UNREACHABLE:
1353                 dev_dbg(queue->ctrl->ctrl.device,
1354                         "CM error event %d\n", ev->event);
1355                 cm_error = -ECONNRESET;
1356                 break;
1357         case RDMA_CM_EVENT_DISCONNECTED:
1358         case RDMA_CM_EVENT_ADDR_CHANGE:
1359         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1360                 dev_dbg(queue->ctrl->ctrl.device,
1361                         "disconnect received - connection closed\n");
1362                 nvme_rdma_error_recovery(queue->ctrl);
1363                 break;
1364         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1365                 /* device removal is handled via the ib_client API */
1366                 break;
1367         default:
1368                 dev_err(queue->ctrl->ctrl.device,
1369                         "Unexpected RDMA CM event (%d)\n", ev->event);
1370                 nvme_rdma_error_recovery(queue->ctrl);
1371                 break;
1372         }
1373
1374         if (cm_error) {
1375                 queue->cm_error = cm_error;
1376                 complete(&queue->cm_done);
1377         }
1378
1379         return 0;
1380 }
1381
1382 static enum blk_eh_timer_return
1383 nvme_rdma_timeout(struct request *rq, bool reserved)
1384 {
1385         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1386
1387         /* queue error recovery */
1388         nvme_rdma_error_recovery(req->queue->ctrl);
1389
1390         /* fail with DNR on cmd timeout */
1391         rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1392
1393         return BLK_EH_HANDLED;
1394 }
1395
1396 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1397                 const struct blk_mq_queue_data *bd)
1398 {
1399         struct nvme_ns *ns = hctx->queue->queuedata;
1400         struct nvme_rdma_queue *queue = hctx->driver_data;
1401         struct request *rq = bd->rq;
1402         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1403         struct nvme_rdma_qe *sqe = &req->sqe;
1404         struct nvme_command *c = sqe->data;
1405         bool flush = false;
1406         struct ib_device *dev;
1407         unsigned int map_len;
1408         int ret;
1409
1410         WARN_ON_ONCE(rq->tag < 0);
1411
1412         dev = queue->device->dev;
1413         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1414                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1415
1416         ret = nvme_setup_cmd(ns, rq, c);
1417         if (ret)
1418                 return ret;
1419
1420         c->common.command_id = rq->tag;
1421         blk_mq_start_request(rq);
1422
1423         map_len = nvme_map_len(rq);
1424         ret = nvme_rdma_map_data(queue, rq, map_len, c);
1425         if (ret < 0) {
1426                 dev_err(queue->ctrl->ctrl.device,
1427                              "Failed to map data (%d)\n", ret);
1428                 nvme_cleanup_cmd(rq);
1429                 goto err;
1430         }
1431
1432         ib_dma_sync_single_for_device(dev, sqe->dma,
1433                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1434
1435         if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1436                 flush = true;
1437         ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1438                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1439         if (ret) {
1440                 nvme_rdma_unmap_data(queue, rq);
1441                 goto err;
1442         }
1443
1444         return BLK_MQ_RQ_QUEUE_OK;
1445 err:
1446         return (ret == -ENOMEM || ret == -EAGAIN) ?
1447                 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1448 }
1449
1450 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1451 {
1452         struct nvme_rdma_queue *queue = hctx->driver_data;
1453         struct ib_cq *cq = queue->ib_cq;
1454         struct ib_wc wc;
1455         int found = 0;
1456
1457         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1458         while (ib_poll_cq(cq, 1, &wc) > 0) {
1459                 struct ib_cqe *cqe = wc.wr_cqe;
1460
1461                 if (cqe) {
1462                         if (cqe->done == nvme_rdma_recv_done)
1463                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1464                         else
1465                                 cqe->done(cq, &wc);
1466                 }
1467         }
1468
1469         return found;
1470 }
1471
1472 static void nvme_rdma_complete_rq(struct request *rq)
1473 {
1474         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1475         struct nvme_rdma_queue *queue = req->queue;
1476         int error = 0;
1477
1478         nvme_rdma_unmap_data(queue, rq);
1479
1480         if (unlikely(rq->errors)) {
1481                 if (nvme_req_needs_retry(rq, rq->errors)) {
1482                         nvme_requeue_req(rq);
1483                         return;
1484                 }
1485
1486                 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1487                         error = rq->errors;
1488                 else
1489                         error = nvme_error_status(rq->errors);
1490         }
1491
1492         blk_mq_end_request(rq, error);
1493 }
1494
1495 static struct blk_mq_ops nvme_rdma_mq_ops = {
1496         .queue_rq       = nvme_rdma_queue_rq,
1497         .complete       = nvme_rdma_complete_rq,
1498         .map_queue      = blk_mq_map_queue,
1499         .init_request   = nvme_rdma_init_request,
1500         .exit_request   = nvme_rdma_exit_request,
1501         .reinit_request = nvme_rdma_reinit_request,
1502         .init_hctx      = nvme_rdma_init_hctx,
1503         .poll           = nvme_rdma_poll,
1504         .timeout        = nvme_rdma_timeout,
1505 };
1506
1507 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1508         .queue_rq       = nvme_rdma_queue_rq,
1509         .complete       = nvme_rdma_complete_rq,
1510         .map_queue      = blk_mq_map_queue,
1511         .init_request   = nvme_rdma_init_admin_request,
1512         .exit_request   = nvme_rdma_exit_admin_request,
1513         .reinit_request = nvme_rdma_reinit_request,
1514         .init_hctx      = nvme_rdma_init_admin_hctx,
1515         .timeout        = nvme_rdma_timeout,
1516 };
1517
1518 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1519 {
1520         int error;
1521
1522         error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1523         if (error)
1524                 return error;
1525
1526         ctrl->device = ctrl->queues[0].device;
1527
1528         /*
1529          * We need a reference on the device as long as the tag_set is alive,
1530          * as the MRs in the request structures need a valid ib_device.
1531          */
1532         error = -EINVAL;
1533         if (!nvme_rdma_dev_get(ctrl->device))
1534                 goto out_free_queue;
1535
1536         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1537                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1538
1539         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1540         ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1541         ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1542         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1543         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1544         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1545                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1546         ctrl->admin_tag_set.driver_data = ctrl;
1547         ctrl->admin_tag_set.nr_hw_queues = 1;
1548         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1549
1550         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1551         if (error)
1552                 goto out_put_dev;
1553
1554         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1555         if (IS_ERR(ctrl->ctrl.admin_q)) {
1556                 error = PTR_ERR(ctrl->ctrl.admin_q);
1557                 goto out_free_tagset;
1558         }
1559
1560         error = nvmf_connect_admin_queue(&ctrl->ctrl);
1561         if (error)
1562                 goto out_cleanup_queue;
1563
1564         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1565         if (error) {
1566                 dev_err(ctrl->ctrl.device,
1567                         "prop_get NVME_REG_CAP failed\n");
1568                 goto out_cleanup_queue;
1569         }
1570
1571         ctrl->ctrl.sqsize =
1572                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1573
1574         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1575         if (error)
1576                 goto out_cleanup_queue;
1577
1578         ctrl->ctrl.max_hw_sectors =
1579                 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1580
1581         error = nvme_init_identify(&ctrl->ctrl);
1582         if (error)
1583                 goto out_cleanup_queue;
1584
1585         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1586                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
1587                         DMA_TO_DEVICE);
1588         if (error)
1589                 goto out_cleanup_queue;
1590
1591         nvme_start_keep_alive(&ctrl->ctrl);
1592
1593         return 0;
1594
1595 out_cleanup_queue:
1596         blk_cleanup_queue(ctrl->ctrl.admin_q);
1597 out_free_tagset:
1598         /* disconnect and drain the queue before freeing the tagset */
1599         nvme_rdma_stop_queue(&ctrl->queues[0]);
1600         blk_mq_free_tag_set(&ctrl->admin_tag_set);
1601 out_put_dev:
1602         nvme_rdma_dev_put(ctrl->device);
1603 out_free_queue:
1604         nvme_rdma_free_queue(&ctrl->queues[0]);
1605         return error;
1606 }
1607
1608 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1609 {
1610         nvme_stop_keep_alive(&ctrl->ctrl);
1611         cancel_work_sync(&ctrl->err_work);
1612         cancel_delayed_work_sync(&ctrl->reconnect_work);
1613
1614         if (ctrl->queue_count > 1) {
1615                 nvme_stop_queues(&ctrl->ctrl);
1616                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1617                                         nvme_cancel_request, &ctrl->ctrl);
1618                 nvme_rdma_free_io_queues(ctrl);
1619         }
1620
1621         if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1622                 nvme_shutdown_ctrl(&ctrl->ctrl);
1623
1624         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1625         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1626                                 nvme_cancel_request, &ctrl->ctrl);
1627         nvme_rdma_destroy_admin_queue(ctrl);
1628 }
1629
1630 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1631 {
1632         nvme_uninit_ctrl(&ctrl->ctrl);
1633         if (shutdown)
1634                 nvme_rdma_shutdown_ctrl(ctrl);
1635
1636         if (ctrl->ctrl.tagset) {
1637                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1638                 blk_mq_free_tag_set(&ctrl->tag_set);
1639                 nvme_rdma_dev_put(ctrl->device);
1640         }
1641
1642         nvme_put_ctrl(&ctrl->ctrl);
1643 }
1644
1645 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1646 {
1647         struct nvme_rdma_ctrl *ctrl = container_of(work,
1648                                 struct nvme_rdma_ctrl, delete_work);
1649
1650         __nvme_rdma_remove_ctrl(ctrl, true);
1651 }
1652
1653 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1654 {
1655         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1656                 return -EBUSY;
1657
1658         if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1659                 return -EBUSY;
1660
1661         return 0;
1662 }
1663
1664 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1665 {
1666         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1667         int ret = 0;
1668
1669         /*
1670          * Keep a reference until all work is flushed since
1671          * __nvme_rdma_del_ctrl can free the ctrl mem
1672          */
1673         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1674                 return -EBUSY;
1675         ret = __nvme_rdma_del_ctrl(ctrl);
1676         if (!ret)
1677                 flush_work(&ctrl->delete_work);
1678         nvme_put_ctrl(&ctrl->ctrl);
1679         return ret;
1680 }
1681
1682 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1683 {
1684         struct nvme_rdma_ctrl *ctrl = container_of(work,
1685                                 struct nvme_rdma_ctrl, delete_work);
1686
1687         __nvme_rdma_remove_ctrl(ctrl, false);
1688 }
1689
1690 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1691 {
1692         struct nvme_rdma_ctrl *ctrl = container_of(work,
1693                                         struct nvme_rdma_ctrl, reset_work);
1694         int ret;
1695         bool changed;
1696
1697         nvme_rdma_shutdown_ctrl(ctrl);
1698
1699         ret = nvme_rdma_configure_admin_queue(ctrl);
1700         if (ret) {
1701                 /* ctrl is already shutdown, just remove the ctrl */
1702                 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1703                 goto del_dead_ctrl;
1704         }
1705
1706         if (ctrl->queue_count > 1) {
1707                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1708                 if (ret)
1709                         goto del_dead_ctrl;
1710
1711                 ret = nvme_rdma_init_io_queues(ctrl);
1712                 if (ret)
1713                         goto del_dead_ctrl;
1714
1715                 ret = nvme_rdma_connect_io_queues(ctrl);
1716                 if (ret)
1717                         goto del_dead_ctrl;
1718         }
1719
1720         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1721         WARN_ON_ONCE(!changed);
1722
1723         if (ctrl->queue_count > 1) {
1724                 nvme_start_queues(&ctrl->ctrl);
1725                 nvme_queue_scan(&ctrl->ctrl);
1726                 nvme_queue_async_events(&ctrl->ctrl);
1727         }
1728
1729         return;
1730
1731 del_dead_ctrl:
1732         /* Deleting this dead controller... */
1733         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1734         WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1735 }
1736
1737 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1738 {
1739         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1740
1741         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1742                 return -EBUSY;
1743
1744         if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1745                 return -EBUSY;
1746
1747         flush_work(&ctrl->reset_work);
1748
1749         return 0;
1750 }
1751
1752 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1753         .name                   = "rdma",
1754         .module                 = THIS_MODULE,
1755         .is_fabrics             = true,
1756         .reg_read32             = nvmf_reg_read32,
1757         .reg_read64             = nvmf_reg_read64,
1758         .reg_write32            = nvmf_reg_write32,
1759         .reset_ctrl             = nvme_rdma_reset_ctrl,
1760         .free_ctrl              = nvme_rdma_free_ctrl,
1761         .submit_async_event     = nvme_rdma_submit_async_event,
1762         .delete_ctrl            = nvme_rdma_del_ctrl,
1763         .get_subsysnqn          = nvmf_get_subsysnqn,
1764         .get_address            = nvmf_get_address,
1765 };
1766
1767 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1768 {
1769         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1770         int ret;
1771
1772         ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1773         if (ret)
1774                 return ret;
1775
1776         ctrl->queue_count = opts->nr_io_queues + 1;
1777         if (ctrl->queue_count < 2)
1778                 return 0;
1779
1780         dev_info(ctrl->ctrl.device,
1781                 "creating %d I/O queues.\n", opts->nr_io_queues);
1782
1783         ret = nvme_rdma_init_io_queues(ctrl);
1784         if (ret)
1785                 return ret;
1786
1787         /*
1788          * We need a reference on the device as long as the tag_set is alive,
1789          * as the MRs in the request structures need a valid ib_device.
1790          */
1791         ret = -EINVAL;
1792         if (!nvme_rdma_dev_get(ctrl->device))
1793                 goto out_free_io_queues;
1794
1795         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1796         ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1797         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1798         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1799         ctrl->tag_set.numa_node = NUMA_NO_NODE;
1800         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1801         ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1802                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1803         ctrl->tag_set.driver_data = ctrl;
1804         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1805         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1806
1807         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1808         if (ret)
1809                 goto out_put_dev;
1810         ctrl->ctrl.tagset = &ctrl->tag_set;
1811
1812         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1813         if (IS_ERR(ctrl->ctrl.connect_q)) {
1814                 ret = PTR_ERR(ctrl->ctrl.connect_q);
1815                 goto out_free_tag_set;
1816         }
1817
1818         ret = nvme_rdma_connect_io_queues(ctrl);
1819         if (ret)
1820                 goto out_cleanup_connect_q;
1821
1822         return 0;
1823
1824 out_cleanup_connect_q:
1825         blk_cleanup_queue(ctrl->ctrl.connect_q);
1826 out_free_tag_set:
1827         blk_mq_free_tag_set(&ctrl->tag_set);
1828 out_put_dev:
1829         nvme_rdma_dev_put(ctrl->device);
1830 out_free_io_queues:
1831         nvme_rdma_free_io_queues(ctrl);
1832         return ret;
1833 }
1834
1835 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1836 {
1837         u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1838         size_t buflen = strlen(p);
1839
1840         /* XXX: handle IPv6 addresses */
1841
1842         if (buflen > INET_ADDRSTRLEN)
1843                 return -EINVAL;
1844         if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1845                 return -EINVAL;
1846         in_addr->sin_family = AF_INET;
1847         return 0;
1848 }
1849
1850 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1851                 struct nvmf_ctrl_options *opts)
1852 {
1853         struct nvme_rdma_ctrl *ctrl;
1854         int ret;
1855         bool changed;
1856
1857         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1858         if (!ctrl)
1859                 return ERR_PTR(-ENOMEM);
1860         ctrl->ctrl.opts = opts;
1861         INIT_LIST_HEAD(&ctrl->list);
1862
1863         ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1864         if (ret) {
1865                 pr_err("malformed IP address passed: %s\n", opts->traddr);
1866                 goto out_free_ctrl;
1867         }
1868
1869         if (opts->mask & NVMF_OPT_TRSVCID) {
1870                 u16 port;
1871
1872                 ret = kstrtou16(opts->trsvcid, 0, &port);
1873                 if (ret)
1874                         goto out_free_ctrl;
1875
1876                 ctrl->addr_in.sin_port = cpu_to_be16(port);
1877         } else {
1878                 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1879         }
1880
1881         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1882                                 0 /* no quirks, we're perfect! */);
1883         if (ret)
1884                 goto out_free_ctrl;
1885
1886         ctrl->reconnect_delay = opts->reconnect_delay;
1887         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1888                         nvme_rdma_reconnect_ctrl_work);
1889         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1890         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1891         INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1892         spin_lock_init(&ctrl->lock);
1893
1894         ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1895         ctrl->ctrl.sqsize = opts->queue_size - 1;
1896         ctrl->ctrl.kato = opts->kato;
1897
1898         ret = -ENOMEM;
1899         ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1900                                 GFP_KERNEL);
1901         if (!ctrl->queues)
1902                 goto out_uninit_ctrl;
1903
1904         ret = nvme_rdma_configure_admin_queue(ctrl);
1905         if (ret)
1906                 goto out_kfree_queues;
1907
1908         /* sanity check icdoff */
1909         if (ctrl->ctrl.icdoff) {
1910                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1911                 goto out_remove_admin_queue;
1912         }
1913
1914         /* sanity check keyed sgls */
1915         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1916                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1917                 goto out_remove_admin_queue;
1918         }
1919
1920         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1921                 /* warn if maxcmd is lower than queue_size */
1922                 dev_warn(ctrl->ctrl.device,
1923                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1924                         opts->queue_size, ctrl->ctrl.maxcmd);
1925                 opts->queue_size = ctrl->ctrl.maxcmd;
1926         }
1927
1928         if (opts->nr_io_queues) {
1929                 ret = nvme_rdma_create_io_queues(ctrl);
1930                 if (ret)
1931                         goto out_remove_admin_queue;
1932         }
1933
1934         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1935         WARN_ON_ONCE(!changed);
1936
1937         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1938                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1939
1940         kref_get(&ctrl->ctrl.kref);
1941
1942         mutex_lock(&nvme_rdma_ctrl_mutex);
1943         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1944         mutex_unlock(&nvme_rdma_ctrl_mutex);
1945
1946         if (opts->nr_io_queues) {
1947                 nvme_queue_scan(&ctrl->ctrl);
1948                 nvme_queue_async_events(&ctrl->ctrl);
1949         }
1950
1951         return &ctrl->ctrl;
1952
1953 out_remove_admin_queue:
1954         nvme_stop_keep_alive(&ctrl->ctrl);
1955         nvme_rdma_destroy_admin_queue(ctrl);
1956 out_kfree_queues:
1957         kfree(ctrl->queues);
1958 out_uninit_ctrl:
1959         nvme_uninit_ctrl(&ctrl->ctrl);
1960         nvme_put_ctrl(&ctrl->ctrl);
1961         if (ret > 0)
1962                 ret = -EIO;
1963         return ERR_PTR(ret);
1964 out_free_ctrl:
1965         kfree(ctrl);
1966         return ERR_PTR(ret);
1967 }
1968
1969 static struct nvmf_transport_ops nvme_rdma_transport = {
1970         .name           = "rdma",
1971         .required_opts  = NVMF_OPT_TRADDR,
1972         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
1973         .create_ctrl    = nvme_rdma_create_ctrl,
1974 };
1975
1976 static void nvme_rdma_add_one(struct ib_device *ib_device)
1977 {
1978 }
1979
1980 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1981 {
1982         struct nvme_rdma_ctrl *ctrl;
1983
1984         /* Delete all controllers using this device */
1985         mutex_lock(&nvme_rdma_ctrl_mutex);
1986         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1987                 if (ctrl->device->dev != ib_device)
1988                         continue;
1989                 dev_info(ctrl->ctrl.device,
1990                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
1991                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1992                 __nvme_rdma_del_ctrl(ctrl);
1993         }
1994         mutex_unlock(&nvme_rdma_ctrl_mutex);
1995
1996         flush_workqueue(nvme_rdma_wq);
1997 }
1998
1999 static struct ib_client nvme_rdma_ib_client = {
2000         .name   = "nvme_rdma",
2001         .add = nvme_rdma_add_one,
2002         .remove = nvme_rdma_remove_one
2003 };
2004
2005 static int __init nvme_rdma_init_module(void)
2006 {
2007         int ret;
2008
2009         nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2010         if (!nvme_rdma_wq)
2011                 return -ENOMEM;
2012
2013         ret = ib_register_client(&nvme_rdma_ib_client);
2014         if (ret) {
2015                 destroy_workqueue(nvme_rdma_wq);
2016                 return ret;
2017         }
2018
2019         nvmf_register_transport(&nvme_rdma_transport);
2020         return 0;
2021 }
2022
2023 static void __exit nvme_rdma_cleanup_module(void)
2024 {
2025         nvmf_unregister_transport(&nvme_rdma_transport);
2026         ib_unregister_client(&nvme_rdma_ib_client);
2027         destroy_workqueue(nvme_rdma_wq);
2028 }
2029
2030 module_init(nvme_rdma_init_module);
2031 module_exit(nvme_rdma_cleanup_module);
2032
2033 MODULE_LICENSE("GPL v2");