drivers/nvme/host/rdma.c

   1 /*
   2  * NVMe over Fabrics RDMA host code.
   3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15 #include <linux/module.h>
  16 #include <linux/init.h>
  17 #include <linux/slab.h>
  18 #include <linux/err.h>
  19 #include <linux/string.h>
  20 #include <linux/atomic.h>
  21 #include <linux/blk-mq.h>
  22 #include <linux/types.h>
  23 #include <linux/list.h>
  24 #include <linux/mutex.h>
  25 #include <linux/scatterlist.h>
  26 #include <linux/nvme.h>
  27 #include <asm/unaligned.h>
  28
  29 #include <rdma/ib_verbs.h>
  30 #include <rdma/rdma_cm.h>
  31 #include <linux/nvme-rdma.h>
  32
  33 #include "nvme.h"
  34 #include "fabrics.h"
  35
  36
  37 #define NVME_RDMA_CONNECT_TIMEOUT_MS    1000            /* 1 second */
  38
  39 #define NVME_RDMA_MAX_SEGMENT_SIZE      0xffffff        /* 24-bit SGL field */
  40
  41 #define NVME_RDMA_MAX_SEGMENTS          256
  42
  43 #define NVME_RDMA_MAX_INLINE_SEGMENTS   1
  44
  45 static const char *const nvme_rdma_cm_status_strs[] = {
  46         [NVME_RDMA_CM_INVALID_LEN]      = "invalid length",
  47         [NVME_RDMA_CM_INVALID_RECFMT]   = "invalid record format",
  48         [NVME_RDMA_CM_INVALID_QID]      = "invalid queue ID",
  49         [NVME_RDMA_CM_INVALID_HSQSIZE]  = "invalid host SQ size",
  50         [NVME_RDMA_CM_INVALID_HRQSIZE]  = "invalid host RQ size",
  51         [NVME_RDMA_CM_NO_RSC]           = "resource not found",
  52         [NVME_RDMA_CM_INVALID_IRD]      = "invalid IRD",
  53         [NVME_RDMA_CM_INVALID_ORD]      = "Invalid ORD",
  54 };
  55
  56 static const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status)
  57 {
  58         size_t index = status;
  59
  60         if (index < ARRAY_SIZE(nvme_rdma_cm_status_strs) &&
  61             nvme_rdma_cm_status_strs[index])
  62                 return nvme_rdma_cm_status_strs[index];
  63         else
  64                 return "unrecognized reason";
  65 };
  66
  67 /*
  68  * We handle AEN commands ourselves and don't even let the
  69  * block layer know about them.
  70  */
  71 #define NVME_RDMA_NR_AEN_COMMANDS      1
  72 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
  73         (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
  74
  75 struct nvme_rdma_device {
  76         struct ib_device       *dev;
  77         struct ib_pd           *pd;
  78         struct kref             ref;
  79         struct list_head        entry;
  80 };
  81
  82 struct nvme_rdma_qe {
  83         struct ib_cqe           cqe;
  84         void                    *data;
  85         u64                     dma;
  86 };
  87
  88 struct nvme_rdma_queue;
  89 struct nvme_rdma_request {
  90         struct nvme_request     req;
  91         struct ib_mr            *mr;
  92         struct nvme_rdma_qe     sqe;
  93         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  94         u32                     num_sge;
  95         int                     nents;
  96         bool                    inline_data;
  97         struct ib_reg_wr        reg_wr;
  98         struct ib_cqe           reg_cqe;
  99         struct nvme_rdma_queue  *queue;
 100         struct sg_table         sg_table;
 101         struct scatterlist      first_sgl[];
 102 };
 103
 104 enum nvme_rdma_queue_flags {
 105         NVME_RDMA_Q_CONNECTED = (1 << 0),
 106         NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1),
 107         NVME_RDMA_Q_DELETING = (1 << 2),
 108         NVME_RDMA_Q_LIVE = (1 << 3),
 109 };
 110
 111 struct nvme_rdma_queue {
 112         struct nvme_rdma_qe     *rsp_ring;
 113         u8                      sig_count;
 114         int                     queue_size;
 115         size_t                  cmnd_capsule_len;
 116         struct nvme_rdma_ctrl   *ctrl;
 117         struct nvme_rdma_device *device;
 118         struct ib_cq            *ib_cq;
 119         struct ib_qp            *qp;
 120
 121         unsigned long           flags;
 122         struct rdma_cm_id       *cm_id;
 123         int                     cm_error;
 124         struct completion       cm_done;
 125 };
 126
 127 struct nvme_rdma_ctrl {
 128         /* read and written in the hot path */
 129         spinlock_t              lock;
 130
 131         /* read only in the hot path */
 132         struct nvme_rdma_queue  *queues;
 133         u32                     queue_count;
 134
 135         /* other member variables */
 136         struct blk_mq_tag_set   tag_set;
 137         struct work_struct      delete_work;
 138         struct work_struct      reset_work;
 139         struct work_struct      err_work;
 140
 141         struct nvme_rdma_qe     async_event_sqe;
 142
 143         int                     reconnect_delay;
 144         struct delayed_work     reconnect_work;
 145
 146         struct list_head        list;
 147
 148         struct blk_mq_tag_set   admin_tag_set;
 149         struct nvme_rdma_device *device;
 150
 151         u64                     cap;
 152         u32                     max_fr_pages;
 153
 154         union {
 155                 struct sockaddr addr;
 156                 struct sockaddr_in addr_in;
 157         };
 158
 159         struct nvme_ctrl        ctrl;
 160 };
 161
 162 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 163 {
 164         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 165 }
 166
 167 static LIST_HEAD(device_list);
 168 static DEFINE_MUTEX(device_list_mutex);
 169
 170 static LIST_HEAD(nvme_rdma_ctrl_list);
 171 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 172
 173 static struct workqueue_struct *nvme_rdma_wq;
 174
 175 /*
 176  * Disabling this option makes small I/O goes faster, but is fundamentally
 177  * unsafe.  With it turned off we will have to register a global rkey that
 178  * allows read and write access to all physical memory.
 179  */
 180 static bool register_always = true;
 181 module_param(register_always, bool, 0444);
 182 MODULE_PARM_DESC(register_always,
 183          "Use memory registration even for contiguous memory regions");
 184
 185 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 186                 struct rdma_cm_event *event);
 187 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 188
 189 /* XXX: really should move to a generic header sooner or later.. */
 190 static inline void put_unaligned_le24(u32 val, u8 *p)
 191 {
 192         *p++ = val;
 193         *p++ = val >> 8;
 194         *p++ = val >> 16;
 195 }
 196
 197 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 198 {
 199         return queue - queue->ctrl->queues;
 200 }
 201
 202 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 203 {
 204         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 205 }
 206
 207 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 208                 size_t capsule_size, enum dma_data_direction dir)
 209 {
 210         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 211         kfree(qe->data);
 212 }
 213
 214 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 215                 size_t capsule_size, enum dma_data_direction dir)
 216 {
 217         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 218         if (!qe->data)
 219                 return -ENOMEM;
 220
 221         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 222         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 223                 kfree(qe->data);
 224                 return -ENOMEM;
 225         }
 226
 227         return 0;
 228 }
 229
 230 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 231                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 232                 size_t capsule_size, enum dma_data_direction dir)
 233 {
 234         int i;
 235
 236         for (i = 0; i < ib_queue_size; i++)
 237                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 238         kfree(ring);
 239 }
 240
 241 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 242                 size_t ib_queue_size, size_t capsule_size,
 243                 enum dma_data_direction dir)
 244 {
 245         struct nvme_rdma_qe *ring;
 246         int i;
 247
 248         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 249         if (!ring)
 250                 return NULL;
 251
 252         for (i = 0; i < ib_queue_size; i++) {
 253                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 254                         goto out_free_ring;
 255         }
 256
 257         return ring;
 258
 259 out_free_ring:
 260         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 261         return NULL;
 262 }
 263
 264 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 265 {
 266         pr_debug("QP event %s (%d)\n",
 267                  ib_event_msg(event->event), event->event);
 268
 269 }
 270
 271 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 272 {
 273         wait_for_completion_interruptible_timeout(&queue->cm_done,
 274                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 275         return queue->cm_error;
 276 }
 277
 278 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 279 {
 280         struct nvme_rdma_device *dev = queue->device;
 281         struct ib_qp_init_attr init_attr;
 282         int ret;
 283
 284         memset(&init_attr, 0, sizeof(init_attr));
 285         init_attr.event_handler = nvme_rdma_qp_event;
 286         /* +1 for drain */
 287         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 288         /* +1 for drain */
 289         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 290         init_attr.cap.max_recv_sge = 1;
 291         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 292         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 293         init_attr.qp_type = IB_QPT_RC;
 294         init_attr.send_cq = queue->ib_cq;
 295         init_attr.recv_cq = queue->ib_cq;
 296
 297         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 298
 299         queue->qp = queue->cm_id->qp;
 300         return ret;
 301 }
 302
 303 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 304 {
 305         struct nvme_rdma_ctrl *ctrl = data;
 306         struct nvme_rdma_device *dev = ctrl->device;
 307         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 308         int ret = 0;
 309
 310         if (!req->mr->need_inval)
 311                 goto out;
 312
 313         ib_dereg_mr(req->mr);
 314
 315         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 316                         ctrl->max_fr_pages);
 317         if (IS_ERR(req->mr)) {
 318                 ret = PTR_ERR(req->mr);
 319                 req->mr = NULL;
 320                 goto out;
 321         }
 322
 323         req->mr->need_inval = false;
 324
 325 out:
 326         return ret;
 327 }
 328
 329 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
 330                 struct request *rq, unsigned int queue_idx)
 331 {
 332         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 333         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 334         struct nvme_rdma_device *dev = queue->device;
 335
 336         if (req->mr)
 337                 ib_dereg_mr(req->mr);
 338
 339         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 340                         DMA_TO_DEVICE);
 341 }
 342
 343 static void nvme_rdma_exit_request(void *data, struct request *rq,
 344                                 unsigned int hctx_idx, unsigned int rq_idx)
 345 {
 346         return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
 347 }
 348
 349 static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
 350                                 unsigned int hctx_idx, unsigned int rq_idx)
 351 {
 352         return __nvme_rdma_exit_request(data, rq, 0);
 353 }
 354
 355 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
 356                 struct request *rq, unsigned int queue_idx)
 357 {
 358         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 359         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 360         struct nvme_rdma_device *dev = queue->device;
 361         struct ib_device *ibdev = dev->dev;
 362         int ret;
 363
 364         BUG_ON(queue_idx >= ctrl->queue_count);
 365
 366         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 367                         DMA_TO_DEVICE);
 368         if (ret)
 369                 return ret;
 370
 371         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 372                         ctrl->max_fr_pages);
 373         if (IS_ERR(req->mr)) {
 374                 ret = PTR_ERR(req->mr);
 375                 goto out_free_qe;
 376         }
 377
 378         req->queue = queue;
 379
 380         return 0;
 381
 382 out_free_qe:
 383         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 384                         DMA_TO_DEVICE);
 385         return -ENOMEM;
 386 }
 387
 388 static int nvme_rdma_init_request(void *data, struct request *rq,
 389                                 unsigned int hctx_idx, unsigned int rq_idx,
 390                                 unsigned int numa_node)
 391 {
 392         return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
 393 }
 394
 395 static int nvme_rdma_init_admin_request(void *data, struct request *rq,
 396                                 unsigned int hctx_idx, unsigned int rq_idx,
 397                                 unsigned int numa_node)
 398 {
 399         return __nvme_rdma_init_request(data, rq, 0);
 400 }
 401
 402 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 403                 unsigned int hctx_idx)
 404 {
 405         struct nvme_rdma_ctrl *ctrl = data;
 406         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 407
 408         BUG_ON(hctx_idx >= ctrl->queue_count);
 409
 410         hctx->driver_data = queue;
 411         return 0;
 412 }
 413
 414 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 415                 unsigned int hctx_idx)
 416 {
 417         struct nvme_rdma_ctrl *ctrl = data;
 418         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 419
 420         BUG_ON(hctx_idx != 0);
 421
 422         hctx->driver_data = queue;
 423         return 0;
 424 }
 425
 426 static void nvme_rdma_free_dev(struct kref *ref)
 427 {
 428         struct nvme_rdma_device *ndev =
 429                 container_of(ref, struct nvme_rdma_device, ref);
 430
 431         mutex_lock(&device_list_mutex);
 432         list_del(&ndev->entry);
 433         mutex_unlock(&device_list_mutex);
 434
 435         ib_dealloc_pd(ndev->pd);
 436         kfree(ndev);
 437 }
 438
 439 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 440 {
 441         kref_put(&dev->ref, nvme_rdma_free_dev);
 442 }
 443
 444 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 445 {
 446         return kref_get_unless_zero(&dev->ref);
 447 }
 448
 449 static struct nvme_rdma_device *
 450 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 451 {
 452         struct nvme_rdma_device *ndev;
 453
 454         mutex_lock(&device_list_mutex);
 455         list_for_each_entry(ndev, &device_list, entry) {
 456                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 457                     nvme_rdma_dev_get(ndev))
 458                         goto out_unlock;
 459         }
 460
 461         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 462         if (!ndev)
 463                 goto out_err;
 464
 465         ndev->dev = cm_id->device;
 466         kref_init(&ndev->ref);
 467
 468         ndev->pd = ib_alloc_pd(ndev->dev,
 469                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 470         if (IS_ERR(ndev->pd))
 471                 goto out_free_dev;
 472
 473         if (!(ndev->dev->attrs.device_cap_flags &
 474               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 475                 dev_err(&ndev->dev->dev,
 476                         "Memory registrations not supported.\n");
 477                 goto out_free_pd;
 478         }
 479
 480         list_add(&ndev->entry, &device_list);
 481 out_unlock:
 482         mutex_unlock(&device_list_mutex);
 483         return ndev;
 484
 485 out_free_pd:
 486         ib_dealloc_pd(ndev->pd);
 487 out_free_dev:
 488         kfree(ndev);
 489 out_err:
 490         mutex_unlock(&device_list_mutex);
 491         return NULL;
 492 }
 493
 494 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 495 {
 496         struct nvme_rdma_device *dev;
 497         struct ib_device *ibdev;
 498
 499         if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags))
 500                 return;
 501
 502         dev = queue->device;
 503         ibdev = dev->dev;
 504         rdma_destroy_qp(queue->cm_id);
 505         ib_free_cq(queue->ib_cq);
 506
 507         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 508                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 509
 510         nvme_rdma_dev_put(dev);
 511 }
 512
 513 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
 514                 struct nvme_rdma_device *dev)
 515 {
 516         struct ib_device *ibdev = dev->dev;
 517         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 518         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 519         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 520
 521         int ret;
 522
 523         queue->device = dev;
 524
 525         /*
 526          * The admin queue is barely used once the controller is live, so don't
 527          * bother to spread it out.
 528          */
 529         if (idx == 0)
 530                 comp_vector = 0;
 531         else
 532                 comp_vector = idx % ibdev->num_comp_vectors;
 533
 534
 535         /* +1 for ib_stop_cq */
 536         queue->ib_cq = ib_alloc_cq(dev->dev, queue,
 537                                 cq_factor * queue->queue_size + 1, comp_vector,
 538                                 IB_POLL_SOFTIRQ);
 539         if (IS_ERR(queue->ib_cq)) {
 540                 ret = PTR_ERR(queue->ib_cq);
 541                 goto out;
 542         }
 543
 544         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 545         if (ret)
 546                 goto out_destroy_ib_cq;
 547
 548         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 549                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 550         if (!queue->rsp_ring) {
 551                 ret = -ENOMEM;
 552                 goto out_destroy_qp;
 553         }
 554         set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags);
 555
 556         return 0;
 557
 558 out_destroy_qp:
 559         ib_destroy_qp(queue->qp);
 560 out_destroy_ib_cq:
 561         ib_free_cq(queue->ib_cq);
 562 out:
 563         return ret;
 564 }
 565
 566 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
 567                 int idx, size_t queue_size)
 568 {
 569         struct nvme_rdma_queue *queue;
 570         int ret;
 571
 572         queue = &ctrl->queues[idx];
 573         queue->ctrl = ctrl;
 574         init_completion(&queue->cm_done);
 575
 576         if (idx > 0)
 577                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 578         else
 579                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 580
 581         queue->queue_size = queue_size;
 582
 583         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 584                         RDMA_PS_TCP, IB_QPT_RC);
 585         if (IS_ERR(queue->cm_id)) {
 586                 dev_info(ctrl->ctrl.device,
 587                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 588                 return PTR_ERR(queue->cm_id);
 589         }
 590
 591         queue->cm_error = -ETIMEDOUT;
 592         ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
 593                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 594         if (ret) {
 595                 dev_info(ctrl->ctrl.device,
 596                         "rdma_resolve_addr failed (%d).\n", ret);
 597                 goto out_destroy_cm_id;
 598         }
 599
 600         ret = nvme_rdma_wait_for_cm(queue);
 601         if (ret) {
 602                 dev_info(ctrl->ctrl.device,
 603                         "rdma_resolve_addr wait failed (%d).\n", ret);
 604                 goto out_destroy_cm_id;
 605         }
 606
 607         clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
 608         set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
 609
 610         return 0;
 611
 612 out_destroy_cm_id:
 613         nvme_rdma_destroy_queue_ib(queue);
 614         rdma_destroy_id(queue->cm_id);
 615         return ret;
 616 }
 617
 618 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 619 {
 620         rdma_disconnect(queue->cm_id);
 621         ib_drain_qp(queue->qp);
 622 }
 623
 624 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 625 {
 626         nvme_rdma_destroy_queue_ib(queue);
 627         rdma_destroy_id(queue->cm_id);
 628 }
 629
 630 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
 631 {
 632         if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
 633                 return;
 634         nvme_rdma_stop_queue(queue);
 635         nvme_rdma_free_queue(queue);
 636 }
 637
 638 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 639 {
 640         int i;
 641
 642         for (i = 1; i < ctrl->queue_count; i++)
 643                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 644 }
 645
 646 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
 647 {
 648         int i, ret = 0;
 649
 650         for (i = 1; i < ctrl->queue_count; i++) {
 651                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 652                 if (ret) {
 653                         dev_info(ctrl->ctrl.device,
 654                                 "failed to connect i/o queue: %d\n", ret);
 655                         goto out_free_queues;
 656                 }
 657                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 658         }
 659
 660         return 0;
 661
 662 out_free_queues:
 663         nvme_rdma_free_io_queues(ctrl);
 664         return ret;
 665 }
 666
 667 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
 668 {
 669         int i, ret;
 670
 671         for (i = 1; i < ctrl->queue_count; i++) {
 672                 ret = nvme_rdma_init_queue(ctrl, i,
 673                                            ctrl->ctrl.opts->queue_size);
 674                 if (ret) {
 675                         dev_info(ctrl->ctrl.device,
 676                                 "failed to initialize i/o queue: %d\n", ret);
 677                         goto out_free_queues;
 678                 }
 679         }
 680
 681         return 0;
 682
 683 out_free_queues:
 684         for (i--; i >= 1; i--)
 685                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 686
 687         return ret;
 688 }
 689
 690 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
 691 {
 692         nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
 693                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 694         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 695         blk_cleanup_queue(ctrl->ctrl.admin_q);
 696         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 697         nvme_rdma_dev_put(ctrl->device);
 698 }
 699
 700 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 701 {
 702         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 703
 704         if (list_empty(&ctrl->list))
 705                 goto free_ctrl;
 706
 707         mutex_lock(&nvme_rdma_ctrl_mutex);
 708         list_del(&ctrl->list);
 709         mutex_unlock(&nvme_rdma_ctrl_mutex);
 710
 711         kfree(ctrl->queues);
 712         nvmf_free_options(nctrl->opts);
 713 free_ctrl:
 714         kfree(ctrl);
 715 }
 716
 717 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 718 {
 719         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 720                         struct nvme_rdma_ctrl, reconnect_work);
 721         bool changed;
 722         int ret;
 723
 724         if (ctrl->queue_count > 1) {
 725                 nvme_rdma_free_io_queues(ctrl);
 726
 727                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
 728                 if (ret)
 729                         goto requeue;
 730         }
 731
 732         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 733
 734         ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
 735         if (ret)
 736                 goto requeue;
 737
 738         ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
 739         if (ret)
 740                 goto requeue;
 741
 742         blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
 743
 744         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 745         if (ret)
 746                 goto stop_admin_q;
 747
 748         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
 749
 750         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 751         if (ret)
 752                 goto stop_admin_q;
 753
 754         nvme_start_keep_alive(&ctrl->ctrl);
 755
 756         if (ctrl->queue_count > 1) {
 757                 ret = nvme_rdma_init_io_queues(ctrl);
 758                 if (ret)
 759                         goto stop_admin_q;
 760
 761                 ret = nvme_rdma_connect_io_queues(ctrl);
 762                 if (ret)
 763                         goto stop_admin_q;
 764         }
 765
 766         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 767         WARN_ON_ONCE(!changed);
 768
 769         if (ctrl->queue_count > 1) {
 770                 nvme_start_queues(&ctrl->ctrl);
 771                 nvme_queue_scan(&ctrl->ctrl);
 772                 nvme_queue_async_events(&ctrl->ctrl);
 773         }
 774
 775         dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
 776
 777         return;
 778
 779 stop_admin_q:
 780         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 781 requeue:
 782         /* Make sure we are not resetting/deleting */
 783         if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
 784                 dev_info(ctrl->ctrl.device,
 785                         "Failed reconnect attempt, requeueing...\n");
 786                 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 787                                         ctrl->reconnect_delay * HZ);
 788         }
 789 }
 790
 791 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 792 {
 793         struct nvme_rdma_ctrl *ctrl = container_of(work,
 794                         struct nvme_rdma_ctrl, err_work);
 795         int i;
 796
 797         nvme_stop_keep_alive(&ctrl->ctrl);
 798
 799         for (i = 0; i < ctrl->queue_count; i++) {
 800                 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
 801                 clear_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 802         }
 803
 804         if (ctrl->queue_count > 1)
 805                 nvme_stop_queues(&ctrl->ctrl);
 806         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 807
 808         /* We must take care of fastfail/requeue all our inflight requests */
 809         if (ctrl->queue_count > 1)
 810                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 811                                         nvme_cancel_request, &ctrl->ctrl);
 812         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 813                                 nvme_cancel_request, &ctrl->ctrl);
 814
 815         dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
 816                 ctrl->reconnect_delay);
 817
 818         queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 819                                 ctrl->reconnect_delay * HZ);
 820 }
 821
 822 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 823 {
 824         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 825                 return;
 826
 827         queue_work(nvme_rdma_wq, &ctrl->err_work);
 828 }
 829
 830 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 831                 const char *op)
 832 {
 833         struct nvme_rdma_queue *queue = cq->cq_context;
 834         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 835
 836         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 837                 dev_info(ctrl->ctrl.device,
 838                              "%s for CQE 0x%p failed with status %s (%d)\n",
 839                              op, wc->wr_cqe,
 840                              ib_wc_status_msg(wc->status), wc->status);
 841         nvme_rdma_error_recovery(ctrl);
 842 }
 843
 844 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
 845 {
 846         if (unlikely(wc->status != IB_WC_SUCCESS))
 847                 nvme_rdma_wr_error(cq, wc, "MEMREG");
 848 }
 849
 850 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 851 {
 852         if (unlikely(wc->status != IB_WC_SUCCESS))
 853                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
 854 }
 855
 856 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
 857                 struct nvme_rdma_request *req)
 858 {
 859         struct ib_send_wr *bad_wr;
 860         struct ib_send_wr wr = {
 861                 .opcode             = IB_WR_LOCAL_INV,
 862                 .next               = NULL,
 863                 .num_sge            = 0,
 864                 .send_flags         = 0,
 865                 .ex.invalidate_rkey = req->mr->rkey,
 866         };
 867
 868         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
 869         wr.wr_cqe = &req->reg_cqe;
 870
 871         return ib_post_send(queue->qp, &wr, &bad_wr);
 872 }
 873
 874 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 875                 struct request *rq)
 876 {
 877         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 878         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 879         struct nvme_rdma_device *dev = queue->device;
 880         struct ib_device *ibdev = dev->dev;
 881         int res;
 882
 883         if (!blk_rq_bytes(rq))
 884                 return;
 885
 886         if (req->mr->need_inval) {
 887                 res = nvme_rdma_inv_rkey(queue, req);
 888                 if (res < 0) {
 889                         dev_err(ctrl->ctrl.device,
 890                                 "Queueing INV WR for rkey %#x failed (%d)\n",
 891                                 req->mr->rkey, res);
 892                         nvme_rdma_error_recovery(queue->ctrl);
 893                 }
 894         }
 895
 896         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 897                         req->nents, rq_data_dir(rq) ==
 898                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 899
 900         nvme_cleanup_cmd(rq);
 901         sg_free_table_chained(&req->sg_table, true);
 902 }
 903
 904 static int nvme_rdma_set_sg_null(struct nvme_command *c)
 905 {
 906         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 907
 908         sg->addr = 0;
 909         put_unaligned_le24(0, sg->length);
 910         put_unaligned_le32(0, sg->key);
 911         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 912         return 0;
 913 }
 914
 915 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
 916                 struct nvme_rdma_request *req, struct nvme_command *c)
 917 {
 918         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
 919
 920         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
 921         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
 922         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
 923
 924         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
 925         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
 926         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
 927
 928         req->inline_data = true;
 929         req->num_sge++;
 930         return 0;
 931 }
 932
 933 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
 934                 struct nvme_rdma_request *req, struct nvme_command *c)
 935 {
 936         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 937
 938         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
 939         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
 940         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
 941         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 942         return 0;
 943 }
 944
 945 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 946                 struct nvme_rdma_request *req, struct nvme_command *c,
 947                 int count)
 948 {
 949         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 950         int nr;
 951
 952         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
 953         if (nr < count) {
 954                 if (nr < 0)
 955                         return nr;
 956                 return -EINVAL;
 957         }
 958
 959         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
 960
 961         req->reg_cqe.done = nvme_rdma_memreg_done;
 962         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
 963         req->reg_wr.wr.opcode = IB_WR_REG_MR;
 964         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
 965         req->reg_wr.wr.num_sge = 0;
 966         req->reg_wr.mr = req->mr;
 967         req->reg_wr.key = req->mr->rkey;
 968         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 969                              IB_ACCESS_REMOTE_READ |
 970                              IB_ACCESS_REMOTE_WRITE;
 971
 972         req->mr->need_inval = true;
 973
 974         sg->addr = cpu_to_le64(req->mr->iova);
 975         put_unaligned_le24(req->mr->length, sg->length);
 976         put_unaligned_le32(req->mr->rkey, sg->key);
 977         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
 978                         NVME_SGL_FMT_INVALIDATE;
 979
 980         return 0;
 981 }
 982
 983 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 984                 struct request *rq, struct nvme_command *c)
 985 {
 986         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 987         struct nvme_rdma_device *dev = queue->device;
 988         struct ib_device *ibdev = dev->dev;
 989         int count, ret;
 990
 991         req->num_sge = 1;
 992         req->inline_data = false;
 993         req->mr->need_inval = false;
 994
 995         c->common.flags |= NVME_CMD_SGL_METABUF;
 996
 997         if (!blk_rq_bytes(rq))
 998                 return nvme_rdma_set_sg_null(c);
 999
1000         req->sg_table.sgl = req->first_sgl;
1001         ret = sg_alloc_table_chained(&req->sg_table,
1002                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1003         if (ret)
1004                 return -ENOMEM;
1005
1006         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1007
1008         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1009                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1010         if (unlikely(count <= 0)) {
1011                 sg_free_table_chained(&req->sg_table, true);
1012                 return -EIO;
1013         }
1014
1015         if (count == 1) {
1016                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1017                     blk_rq_payload_bytes(rq) <=
1018                                 nvme_rdma_inline_data_size(queue))
1019                         return nvme_rdma_map_sg_inline(queue, req, c);
1020
1021                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1022                         return nvme_rdma_map_sg_single(queue, req, c);
1023         }
1024
1025         return nvme_rdma_map_sg_fr(queue, req, c, count);
1026 }
1027
1028 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1029 {
1030         if (unlikely(wc->status != IB_WC_SUCCESS))
1031                 nvme_rdma_wr_error(cq, wc, "SEND");
1032 }
1033
1034 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1035                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1036                 struct ib_send_wr *first, bool flush)
1037 {
1038         struct ib_send_wr wr, *bad_wr;
1039         int ret;
1040
1041         sge->addr   = qe->dma;
1042         sge->length = sizeof(struct nvme_command),
1043         sge->lkey   = queue->device->pd->local_dma_lkey;
1044
1045         qe->cqe.done = nvme_rdma_send_done;
1046
1047         wr.next       = NULL;
1048         wr.wr_cqe     = &qe->cqe;
1049         wr.sg_list    = sge;
1050         wr.num_sge    = num_sge;
1051         wr.opcode     = IB_WR_SEND;
1052         wr.send_flags = 0;
1053
1054         /*
1055          * Unsignalled send completions are another giant desaster in the
1056          * IB Verbs spec:  If we don't regularly post signalled sends
1057          * the send queue will fill up and only a QP reset will rescue us.
1058          * Would have been way to obvious to handle this in hardware or
1059          * at least the RDMA stack..
1060          *
1061          * This messy and racy code sniplet is copy and pasted from the iSER
1062          * initiator, and the magic '32' comes from there as well.
1063          *
1064          * Always signal the flushes. The magic request used for the flush
1065          * sequencer is not allocated in our driver's tagset and it's
1066          * triggered to be freed by blk_cleanup_queue(). So we need to
1067          * always mark it as signaled to ensure that the "wr_cqe", which is
1068          * embeded in request's payload, is not freed when __ib_process_cq()
1069          * calls wr_cqe->done().
1070          */
1071         if ((++queue->sig_count % 32) == 0 || flush)
1072                 wr.send_flags |= IB_SEND_SIGNALED;
1073
1074         if (first)
1075                 first->next = &wr;
1076         else
1077                 first = &wr;
1078
1079         ret = ib_post_send(queue->qp, first, &bad_wr);
1080         if (ret) {
1081                 dev_err(queue->ctrl->ctrl.device,
1082                              "%s failed with error code %d\n", __func__, ret);
1083         }
1084         return ret;
1085 }
1086
1087 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1088                 struct nvme_rdma_qe *qe)
1089 {
1090         struct ib_recv_wr wr, *bad_wr;
1091         struct ib_sge list;
1092         int ret;
1093
1094         list.addr   = qe->dma;
1095         list.length = sizeof(struct nvme_completion);
1096         list.lkey   = queue->device->pd->local_dma_lkey;
1097
1098         qe->cqe.done = nvme_rdma_recv_done;
1099
1100         wr.next     = NULL;
1101         wr.wr_cqe   = &qe->cqe;
1102         wr.sg_list  = &list;
1103         wr.num_sge  = 1;
1104
1105         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1106         if (ret) {
1107                 dev_err(queue->ctrl->ctrl.device,
1108                         "%s failed with error code %d\n", __func__, ret);
1109         }
1110         return ret;
1111 }
1112
1113 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1114 {
1115         u32 queue_idx = nvme_rdma_queue_idx(queue);
1116
1117         if (queue_idx == 0)
1118                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1119         return queue->ctrl->tag_set.tags[queue_idx - 1];
1120 }
1121
1122 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1123 {
1124         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1125         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1126         struct ib_device *dev = queue->device->dev;
1127         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1128         struct nvme_command *cmd = sqe->data;
1129         struct ib_sge sge;
1130         int ret;
1131
1132         if (WARN_ON_ONCE(aer_idx != 0))
1133                 return;
1134
1135         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1136
1137         memset(cmd, 0, sizeof(*cmd));
1138         cmd->common.opcode = nvme_admin_async_event;
1139         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1140         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1141         nvme_rdma_set_sg_null(cmd);
1142
1143         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1144                         DMA_TO_DEVICE);
1145
1146         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1147         WARN_ON_ONCE(ret);
1148 }
1149
1150 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1151                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1152 {
1153         struct request *rq;
1154         struct nvme_rdma_request *req;
1155         int ret = 0;
1156
1157         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1158         if (!rq) {
1159                 dev_err(queue->ctrl->ctrl.device,
1160                         "tag 0x%x on QP %#x not found\n",
1161                         cqe->command_id, queue->qp->qp_num);
1162                 nvme_rdma_error_recovery(queue->ctrl);
1163                 return ret;
1164         }
1165         req = blk_mq_rq_to_pdu(rq);
1166
1167         if (rq->tag == tag)
1168                 ret = 1;
1169
1170         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1171             wc->ex.invalidate_rkey == req->mr->rkey)
1172                 req->mr->need_inval = false;
1173
1174         req->req.result = cqe->result;
1175         blk_mq_complete_request(rq, le16_to_cpu(cqe->status) >> 1);
1176         return ret;
1177 }
1178
1179 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1180 {
1181         struct nvme_rdma_qe *qe =
1182                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1183         struct nvme_rdma_queue *queue = cq->cq_context;
1184         struct ib_device *ibdev = queue->device->dev;
1185         struct nvme_completion *cqe = qe->data;
1186         const size_t len = sizeof(struct nvme_completion);
1187         int ret = 0;
1188
1189         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1190                 nvme_rdma_wr_error(cq, wc, "RECV");
1191                 return 0;
1192         }
1193
1194         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1195         /*
1196          * AEN requests are special as they don't time out and can
1197          * survive any kind of queue freeze and often don't respond to
1198          * aborts.  We don't even bother to allocate a struct request
1199          * for them but rather special case them here.
1200          */
1201         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1202                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1203                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1204                                 &cqe->result);
1205         else
1206                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1207         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1208
1209         nvme_rdma_post_recv(queue, qe);
1210         return ret;
1211 }
1212
1213 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1214 {
1215         __nvme_rdma_recv_done(cq, wc, -1);
1216 }
1217
1218 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1219 {
1220         int ret, i;
1221
1222         for (i = 0; i < queue->queue_size; i++) {
1223                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1224                 if (ret)
1225                         goto out_destroy_queue_ib;
1226         }
1227
1228         return 0;
1229
1230 out_destroy_queue_ib:
1231         nvme_rdma_destroy_queue_ib(queue);
1232         return ret;
1233 }
1234
1235 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1236                 struct rdma_cm_event *ev)
1237 {
1238         struct rdma_cm_id *cm_id = queue->cm_id;
1239         int status = ev->status;
1240         const char *rej_msg;
1241         const struct nvme_rdma_cm_rej *rej_data;
1242         u8 rej_data_len;
1243
1244         rej_msg = rdma_reject_msg(cm_id, status);
1245         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1246
1247         if (rej_data && rej_data_len >= sizeof(u16)) {
1248                 u16 sts = le16_to_cpu(rej_data->sts);
1249
1250                 dev_err(queue->ctrl->ctrl.device,
1251                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1252                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1253         } else {
1254                 dev_err(queue->ctrl->ctrl.device,
1255                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1256         }
1257
1258         return -ECONNRESET;
1259 }
1260
1261 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1262 {
1263         struct nvme_rdma_device *dev;
1264         int ret;
1265
1266         dev = nvme_rdma_find_get_device(queue->cm_id);
1267         if (!dev) {
1268                 dev_err(queue->cm_id->device->dma_device,
1269                         "no client data found!\n");
1270                 return -ECONNREFUSED;
1271         }
1272
1273         ret = nvme_rdma_create_queue_ib(queue, dev);
1274         if (ret) {
1275                 nvme_rdma_dev_put(dev);
1276                 goto out;
1277         }
1278
1279         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1280         if (ret) {
1281                 dev_err(queue->ctrl->ctrl.device,
1282                         "rdma_resolve_route failed (%d).\n",
1283                         queue->cm_error);
1284                 goto out_destroy_queue;
1285         }
1286
1287         return 0;
1288
1289 out_destroy_queue:
1290         nvme_rdma_destroy_queue_ib(queue);
1291 out:
1292         return ret;
1293 }
1294
1295 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1296 {
1297         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1298         struct rdma_conn_param param = { };
1299         struct nvme_rdma_cm_req priv = { };
1300         int ret;
1301
1302         param.qp_num = queue->qp->qp_num;
1303         param.flow_control = 1;
1304
1305         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1306         /* maximum retry count */
1307         param.retry_count = 7;
1308         param.rnr_retry_count = 7;
1309         param.private_data = &priv;
1310         param.private_data_len = sizeof(priv);
1311
1312         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1313         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1314         /*
1315          * set the admin queue depth to the minimum size
1316          * specified by the Fabrics standard.
1317          */
1318         if (priv.qid == 0) {
1319                 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1320                 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1321         } else {
1322                 /*
1323                  * current interpretation of the fabrics spec
1324                  * is at minimum you make hrqsize sqsize+1, or a
1325                  * 1's based representation of sqsize.
1326                  */
1327                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1328                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1329         }
1330
1331         ret = rdma_connect(queue->cm_id, &param);
1332         if (ret) {
1333                 dev_err(ctrl->ctrl.device,
1334                         "rdma_connect failed (%d).\n", ret);
1335                 goto out_destroy_queue_ib;
1336         }
1337
1338         return 0;
1339
1340 out_destroy_queue_ib:
1341         nvme_rdma_destroy_queue_ib(queue);
1342         return ret;
1343 }
1344
1345 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1346                 struct rdma_cm_event *ev)
1347 {
1348         struct nvme_rdma_queue *queue = cm_id->context;
1349         int cm_error = 0;
1350
1351         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1352                 rdma_event_msg(ev->event), ev->event,
1353                 ev->status, cm_id);
1354
1355         switch (ev->event) {
1356         case RDMA_CM_EVENT_ADDR_RESOLVED:
1357                 cm_error = nvme_rdma_addr_resolved(queue);
1358                 break;
1359         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1360                 cm_error = nvme_rdma_route_resolved(queue);
1361                 break;
1362         case RDMA_CM_EVENT_ESTABLISHED:
1363                 queue->cm_error = nvme_rdma_conn_established(queue);
1364                 /* complete cm_done regardless of success/failure */
1365                 complete(&queue->cm_done);
1366                 return 0;
1367         case RDMA_CM_EVENT_REJECTED:
1368                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1369                 break;
1370         case RDMA_CM_EVENT_ADDR_ERROR:
1371         case RDMA_CM_EVENT_ROUTE_ERROR:
1372         case RDMA_CM_EVENT_CONNECT_ERROR:
1373         case RDMA_CM_EVENT_UNREACHABLE:
1374                 dev_dbg(queue->ctrl->ctrl.device,
1375                         "CM error event %d\n", ev->event);
1376                 cm_error = -ECONNRESET;
1377                 break;
1378         case RDMA_CM_EVENT_DISCONNECTED:
1379         case RDMA_CM_EVENT_ADDR_CHANGE:
1380         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1381                 dev_dbg(queue->ctrl->ctrl.device,
1382                         "disconnect received - connection closed\n");
1383                 nvme_rdma_error_recovery(queue->ctrl);
1384                 break;
1385         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1386                 /* device removal is handled via the ib_client API */
1387                 break;
1388         default:
1389                 dev_err(queue->ctrl->ctrl.device,
1390                         "Unexpected RDMA CM event (%d)\n", ev->event);
1391                 nvme_rdma_error_recovery(queue->ctrl);
1392                 break;
1393         }
1394
1395         if (cm_error) {
1396                 queue->cm_error = cm_error;
1397                 complete(&queue->cm_done);
1398         }
1399
1400         return 0;
1401 }
1402
1403 static enum blk_eh_timer_return
1404 nvme_rdma_timeout(struct request *rq, bool reserved)
1405 {
1406         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1407
1408         /* queue error recovery */
1409         nvme_rdma_error_recovery(req->queue->ctrl);
1410
1411         /* fail with DNR on cmd timeout */
1412         rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1413
1414         return BLK_EH_HANDLED;
1415 }
1416
1417 /*
1418  * We cannot accept any other command until the Connect command has completed.
1419  */
1420 static inline bool nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue,
1421                 struct request *rq)
1422 {
1423         if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1424                 struct nvme_command *cmd = nvme_req(rq)->cmd;
1425
1426                 if (rq->cmd_type != REQ_TYPE_DRV_PRIV ||
1427                     cmd->common.opcode != nvme_fabrics_command ||
1428                     cmd->fabrics.fctype != nvme_fabrics_type_connect)
1429                         return false;
1430         }
1431
1432         return true;
1433 }
1434
1435 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1436                 const struct blk_mq_queue_data *bd)
1437 {
1438         struct nvme_ns *ns = hctx->queue->queuedata;
1439         struct nvme_rdma_queue *queue = hctx->driver_data;
1440         struct request *rq = bd->rq;
1441         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1442         struct nvme_rdma_qe *sqe = &req->sqe;
1443         struct nvme_command *c = sqe->data;
1444         bool flush = false;
1445         struct ib_device *dev;
1446         int ret;
1447
1448         WARN_ON_ONCE(rq->tag < 0);
1449
1450         if (!nvme_rdma_queue_is_ready(queue, rq))
1451                 return BLK_MQ_RQ_QUEUE_BUSY;
1452
1453         dev = queue->device->dev;
1454         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1455                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1456
1457         ret = nvme_setup_cmd(ns, rq, c);
1458         if (ret != BLK_MQ_RQ_QUEUE_OK)
1459                 return ret;
1460
1461         blk_mq_start_request(rq);
1462
1463         ret = nvme_rdma_map_data(queue, rq, c);
1464         if (ret < 0) {
1465                 dev_err(queue->ctrl->ctrl.device,
1466                              "Failed to map data (%d)\n", ret);
1467                 nvme_cleanup_cmd(rq);
1468                 goto err;
1469         }
1470
1471         ib_dma_sync_single_for_device(dev, sqe->dma,
1472                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1473
1474         if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1475                 flush = true;
1476         ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1477                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1478         if (ret) {
1479                 nvme_rdma_unmap_data(queue, rq);
1480                 goto err;
1481         }
1482
1483         return BLK_MQ_RQ_QUEUE_OK;
1484 err:
1485         return (ret == -ENOMEM || ret == -EAGAIN) ?
1486                 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1487 }
1488
1489 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1490 {
1491         struct nvme_rdma_queue *queue = hctx->driver_data;
1492         struct ib_cq *cq = queue->ib_cq;
1493         struct ib_wc wc;
1494         int found = 0;
1495
1496         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1497         while (ib_poll_cq(cq, 1, &wc) > 0) {
1498                 struct ib_cqe *cqe = wc.wr_cqe;
1499
1500                 if (cqe) {
1501                         if (cqe->done == nvme_rdma_recv_done)
1502                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1503                         else
1504                                 cqe->done(cq, &wc);
1505                 }
1506         }
1507
1508         return found;
1509 }
1510
1511 static void nvme_rdma_complete_rq(struct request *rq)
1512 {
1513         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1514         struct nvme_rdma_queue *queue = req->queue;
1515         int error = 0;
1516
1517         nvme_rdma_unmap_data(queue, rq);
1518
1519         if (unlikely(rq->errors)) {
1520                 if (nvme_req_needs_retry(rq, rq->errors)) {
1521                         nvme_requeue_req(rq);
1522                         return;
1523                 }
1524
1525                 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1526                         error = rq->errors;
1527                 else
1528                         error = nvme_error_status(rq->errors);
1529         }
1530
1531         blk_mq_end_request(rq, error);
1532 }
1533
1534 static struct blk_mq_ops nvme_rdma_mq_ops = {
1535         .queue_rq       = nvme_rdma_queue_rq,
1536         .complete       = nvme_rdma_complete_rq,
1537         .init_request   = nvme_rdma_init_request,
1538         .exit_request   = nvme_rdma_exit_request,
1539         .reinit_request = nvme_rdma_reinit_request,
1540         .init_hctx      = nvme_rdma_init_hctx,
1541         .poll           = nvme_rdma_poll,
1542         .timeout        = nvme_rdma_timeout,
1543 };
1544
1545 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1546         .queue_rq       = nvme_rdma_queue_rq,
1547         .complete       = nvme_rdma_complete_rq,
1548         .init_request   = nvme_rdma_init_admin_request,
1549         .exit_request   = nvme_rdma_exit_admin_request,
1550         .reinit_request = nvme_rdma_reinit_request,
1551         .init_hctx      = nvme_rdma_init_admin_hctx,
1552         .timeout        = nvme_rdma_timeout,
1553 };
1554
1555 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1556 {
1557         int error;
1558
1559         error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1560         if (error)
1561                 return error;
1562
1563         ctrl->device = ctrl->queues[0].device;
1564
1565         /*
1566          * We need a reference on the device as long as the tag_set is alive,
1567          * as the MRs in the request structures need a valid ib_device.
1568          */
1569         error = -EINVAL;
1570         if (!nvme_rdma_dev_get(ctrl->device))
1571                 goto out_free_queue;
1572
1573         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1574                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1575
1576         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1577         ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1578         ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1579         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1580         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1581         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1582                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1583         ctrl->admin_tag_set.driver_data = ctrl;
1584         ctrl->admin_tag_set.nr_hw_queues = 1;
1585         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1586
1587         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1588         if (error)
1589                 goto out_put_dev;
1590
1591         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1592         if (IS_ERR(ctrl->ctrl.admin_q)) {
1593                 error = PTR_ERR(ctrl->ctrl.admin_q);
1594                 goto out_free_tagset;
1595         }
1596
1597         error = nvmf_connect_admin_queue(&ctrl->ctrl);
1598         if (error)
1599                 goto out_cleanup_queue;
1600
1601         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
1602
1603         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1604         if (error) {
1605                 dev_err(ctrl->ctrl.device,
1606                         "prop_get NVME_REG_CAP failed\n");
1607                 goto out_cleanup_queue;
1608         }
1609
1610         ctrl->ctrl.sqsize =
1611                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1612
1613         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1614         if (error)
1615                 goto out_cleanup_queue;
1616
1617         ctrl->ctrl.max_hw_sectors =
1618                 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1619
1620         error = nvme_init_identify(&ctrl->ctrl);
1621         if (error)
1622                 goto out_cleanup_queue;
1623
1624         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1625                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
1626                         DMA_TO_DEVICE);
1627         if (error)
1628                 goto out_cleanup_queue;
1629
1630         nvme_start_keep_alive(&ctrl->ctrl);
1631
1632         return 0;
1633
1634 out_cleanup_queue:
1635         blk_cleanup_queue(ctrl->ctrl.admin_q);
1636 out_free_tagset:
1637         /* disconnect and drain the queue before freeing the tagset */
1638         nvme_rdma_stop_queue(&ctrl->queues[0]);
1639         blk_mq_free_tag_set(&ctrl->admin_tag_set);
1640 out_put_dev:
1641         nvme_rdma_dev_put(ctrl->device);
1642 out_free_queue:
1643         nvme_rdma_free_queue(&ctrl->queues[0]);
1644         return error;
1645 }
1646
1647 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1648 {
1649         nvme_stop_keep_alive(&ctrl->ctrl);
1650         cancel_work_sync(&ctrl->err_work);
1651         cancel_delayed_work_sync(&ctrl->reconnect_work);
1652
1653         if (ctrl->queue_count > 1) {
1654                 nvme_stop_queues(&ctrl->ctrl);
1655                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1656                                         nvme_cancel_request, &ctrl->ctrl);
1657                 nvme_rdma_free_io_queues(ctrl);
1658         }
1659
1660         if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1661                 nvme_shutdown_ctrl(&ctrl->ctrl);
1662
1663         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1664         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1665                                 nvme_cancel_request, &ctrl->ctrl);
1666         nvme_rdma_destroy_admin_queue(ctrl);
1667 }
1668
1669 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1670 {
1671         nvme_uninit_ctrl(&ctrl->ctrl);
1672         if (shutdown)
1673                 nvme_rdma_shutdown_ctrl(ctrl);
1674
1675         if (ctrl->ctrl.tagset) {
1676                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1677                 blk_mq_free_tag_set(&ctrl->tag_set);
1678                 nvme_rdma_dev_put(ctrl->device);
1679         }
1680
1681         nvme_put_ctrl(&ctrl->ctrl);
1682 }
1683
1684 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1685 {
1686         struct nvme_rdma_ctrl *ctrl = container_of(work,
1687                                 struct nvme_rdma_ctrl, delete_work);
1688
1689         __nvme_rdma_remove_ctrl(ctrl, true);
1690 }
1691
1692 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1693 {
1694         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1695                 return -EBUSY;
1696
1697         if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1698                 return -EBUSY;
1699
1700         return 0;
1701 }
1702
1703 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1704 {
1705         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1706         int ret = 0;
1707
1708         /*
1709          * Keep a reference until all work is flushed since
1710          * __nvme_rdma_del_ctrl can free the ctrl mem
1711          */
1712         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1713                 return -EBUSY;
1714         ret = __nvme_rdma_del_ctrl(ctrl);
1715         if (!ret)
1716                 flush_work(&ctrl->delete_work);
1717         nvme_put_ctrl(&ctrl->ctrl);
1718         return ret;
1719 }
1720
1721 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1722 {
1723         struct nvme_rdma_ctrl *ctrl = container_of(work,
1724                                 struct nvme_rdma_ctrl, delete_work);
1725
1726         __nvme_rdma_remove_ctrl(ctrl, false);
1727 }
1728
1729 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1730 {
1731         struct nvme_rdma_ctrl *ctrl = container_of(work,
1732                                         struct nvme_rdma_ctrl, reset_work);
1733         int ret;
1734         bool changed;
1735
1736         nvme_rdma_shutdown_ctrl(ctrl);
1737
1738         ret = nvme_rdma_configure_admin_queue(ctrl);
1739         if (ret) {
1740                 /* ctrl is already shutdown, just remove the ctrl */
1741                 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1742                 goto del_dead_ctrl;
1743         }
1744
1745         if (ctrl->queue_count > 1) {
1746                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1747                 if (ret)
1748                         goto del_dead_ctrl;
1749
1750                 ret = nvme_rdma_init_io_queues(ctrl);
1751                 if (ret)
1752                         goto del_dead_ctrl;
1753
1754                 ret = nvme_rdma_connect_io_queues(ctrl);
1755                 if (ret)
1756                         goto del_dead_ctrl;
1757         }
1758
1759         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1760         WARN_ON_ONCE(!changed);
1761
1762         if (ctrl->queue_count > 1) {
1763                 nvme_start_queues(&ctrl->ctrl);
1764                 nvme_queue_scan(&ctrl->ctrl);
1765                 nvme_queue_async_events(&ctrl->ctrl);
1766         }
1767
1768         return;
1769
1770 del_dead_ctrl:
1771         /* Deleting this dead controller... */
1772         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1773         WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1774 }
1775
1776 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1777 {
1778         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1779
1780         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1781                 return -EBUSY;
1782
1783         if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1784                 return -EBUSY;
1785
1786         flush_work(&ctrl->reset_work);
1787
1788         return 0;
1789 }
1790
1791 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1792         .name                   = "rdma",
1793         .module                 = THIS_MODULE,
1794         .is_fabrics             = true,
1795         .reg_read32             = nvmf_reg_read32,
1796         .reg_read64             = nvmf_reg_read64,
1797         .reg_write32            = nvmf_reg_write32,
1798         .reset_ctrl             = nvme_rdma_reset_ctrl,
1799         .free_ctrl              = nvme_rdma_free_ctrl,
1800         .submit_async_event     = nvme_rdma_submit_async_event,
1801         .delete_ctrl            = nvme_rdma_del_ctrl,
1802         .get_subsysnqn          = nvmf_get_subsysnqn,
1803         .get_address            = nvmf_get_address,
1804 };
1805
1806 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1807 {
1808         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1809         int ret;
1810
1811         ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1812         if (ret)
1813                 return ret;
1814
1815         ctrl->queue_count = opts->nr_io_queues + 1;
1816         if (ctrl->queue_count < 2)
1817                 return 0;
1818
1819         dev_info(ctrl->ctrl.device,
1820                 "creating %d I/O queues.\n", opts->nr_io_queues);
1821
1822         ret = nvme_rdma_init_io_queues(ctrl);
1823         if (ret)
1824                 return ret;
1825
1826         /*
1827          * We need a reference on the device as long as the tag_set is alive,
1828          * as the MRs in the request structures need a valid ib_device.
1829          */
1830         ret = -EINVAL;
1831         if (!nvme_rdma_dev_get(ctrl->device))
1832                 goto out_free_io_queues;
1833
1834         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1835         ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1836         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1837         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1838         ctrl->tag_set.numa_node = NUMA_NO_NODE;
1839         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1840         ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1841                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1842         ctrl->tag_set.driver_data = ctrl;
1843         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1844         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1845
1846         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1847         if (ret)
1848                 goto out_put_dev;
1849         ctrl->ctrl.tagset = &ctrl->tag_set;
1850
1851         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1852         if (IS_ERR(ctrl->ctrl.connect_q)) {
1853                 ret = PTR_ERR(ctrl->ctrl.connect_q);
1854                 goto out_free_tag_set;
1855         }
1856
1857         ret = nvme_rdma_connect_io_queues(ctrl);
1858         if (ret)
1859                 goto out_cleanup_connect_q;
1860
1861         return 0;
1862
1863 out_cleanup_connect_q:
1864         blk_cleanup_queue(ctrl->ctrl.connect_q);
1865 out_free_tag_set:
1866         blk_mq_free_tag_set(&ctrl->tag_set);
1867 out_put_dev:
1868         nvme_rdma_dev_put(ctrl->device);
1869 out_free_io_queues:
1870         nvme_rdma_free_io_queues(ctrl);
1871         return ret;
1872 }
1873
1874 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1875 {
1876         u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1877         size_t buflen = strlen(p);
1878
1879         /* XXX: handle IPv6 addresses */
1880
1881         if (buflen > INET_ADDRSTRLEN)
1882                 return -EINVAL;
1883         if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1884                 return -EINVAL;
1885         in_addr->sin_family = AF_INET;
1886         return 0;
1887 }
1888
1889 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1890                 struct nvmf_ctrl_options *opts)
1891 {
1892         struct nvme_rdma_ctrl *ctrl;
1893         int ret;
1894         bool changed;
1895
1896         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1897         if (!ctrl)
1898                 return ERR_PTR(-ENOMEM);
1899         ctrl->ctrl.opts = opts;
1900         INIT_LIST_HEAD(&ctrl->list);
1901
1902         ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1903         if (ret) {
1904                 pr_err("malformed IP address passed: %s\n", opts->traddr);
1905                 goto out_free_ctrl;
1906         }
1907
1908         if (opts->mask & NVMF_OPT_TRSVCID) {
1909                 u16 port;
1910
1911                 ret = kstrtou16(opts->trsvcid, 0, &port);
1912                 if (ret)
1913                         goto out_free_ctrl;
1914
1915                 ctrl->addr_in.sin_port = cpu_to_be16(port);
1916         } else {
1917                 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1918         }
1919
1920         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1921                                 0 /* no quirks, we're perfect! */);
1922         if (ret)
1923                 goto out_free_ctrl;
1924
1925         ctrl->reconnect_delay = opts->reconnect_delay;
1926         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1927                         nvme_rdma_reconnect_ctrl_work);
1928         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1929         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1930         INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1931         spin_lock_init(&ctrl->lock);
1932
1933         ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1934         ctrl->ctrl.sqsize = opts->queue_size - 1;
1935         ctrl->ctrl.kato = opts->kato;
1936
1937         ret = -ENOMEM;
1938         ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1939                                 GFP_KERNEL);
1940         if (!ctrl->queues)
1941                 goto out_uninit_ctrl;
1942
1943         ret = nvme_rdma_configure_admin_queue(ctrl);
1944         if (ret)
1945                 goto out_kfree_queues;
1946
1947         /* sanity check icdoff */
1948         if (ctrl->ctrl.icdoff) {
1949                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1950                 goto out_remove_admin_queue;
1951         }
1952
1953         /* sanity check keyed sgls */
1954         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1955                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1956                 goto out_remove_admin_queue;
1957         }
1958
1959         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1960                 /* warn if maxcmd is lower than queue_size */
1961                 dev_warn(ctrl->ctrl.device,
1962                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1963                         opts->queue_size, ctrl->ctrl.maxcmd);
1964                 opts->queue_size = ctrl->ctrl.maxcmd;
1965         }
1966
1967         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1968                 /* warn if sqsize is lower than queue_size */
1969                 dev_warn(ctrl->ctrl.device,
1970                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1971                         opts->queue_size, ctrl->ctrl.sqsize + 1);
1972                 opts->queue_size = ctrl->ctrl.sqsize + 1;
1973         }
1974
1975         if (opts->nr_io_queues) {
1976                 ret = nvme_rdma_create_io_queues(ctrl);
1977                 if (ret)
1978                         goto out_remove_admin_queue;
1979         }
1980
1981         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1982         WARN_ON_ONCE(!changed);
1983
1984         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1985                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1986
1987         kref_get(&ctrl->ctrl.kref);
1988
1989         mutex_lock(&nvme_rdma_ctrl_mutex);
1990         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1991         mutex_unlock(&nvme_rdma_ctrl_mutex);
1992
1993         if (opts->nr_io_queues) {
1994                 nvme_queue_scan(&ctrl->ctrl);
1995                 nvme_queue_async_events(&ctrl->ctrl);
1996         }
1997
1998         return &ctrl->ctrl;
1999
2000 out_remove_admin_queue:
2001         nvme_stop_keep_alive(&ctrl->ctrl);
2002         nvme_rdma_destroy_admin_queue(ctrl);
2003 out_kfree_queues:
2004         kfree(ctrl->queues);
2005 out_uninit_ctrl:
2006         nvme_uninit_ctrl(&ctrl->ctrl);
2007         nvme_put_ctrl(&ctrl->ctrl);
2008         if (ret > 0)
2009                 ret = -EIO;
2010         return ERR_PTR(ret);
2011 out_free_ctrl:
2012         kfree(ctrl);
2013         return ERR_PTR(ret);
2014 }
2015
2016 static struct nvmf_transport_ops nvme_rdma_transport = {
2017         .name           = "rdma",
2018         .required_opts  = NVMF_OPT_TRADDR,
2019         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
2020         .create_ctrl    = nvme_rdma_create_ctrl,
2021 };
2022
2023 static void nvme_rdma_add_one(struct ib_device *ib_device)
2024 {
2025 }
2026
2027 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2028 {
2029         struct nvme_rdma_ctrl *ctrl;
2030
2031         /* Delete all controllers using this device */
2032         mutex_lock(&nvme_rdma_ctrl_mutex);
2033         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2034                 if (ctrl->device->dev != ib_device)
2035                         continue;
2036                 dev_info(ctrl->ctrl.device,
2037                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
2038                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2039                 __nvme_rdma_del_ctrl(ctrl);
2040         }
2041         mutex_unlock(&nvme_rdma_ctrl_mutex);
2042
2043         flush_workqueue(nvme_rdma_wq);
2044 }
2045
2046 static struct ib_client nvme_rdma_ib_client = {
2047         .name   = "nvme_rdma",
2048         .add = nvme_rdma_add_one,
2049         .remove = nvme_rdma_remove_one
2050 };
2051
2052 static int __init nvme_rdma_init_module(void)
2053 {
2054         int ret;
2055
2056         nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2057         if (!nvme_rdma_wq)
2058                 return -ENOMEM;
2059
2060         ret = ib_register_client(&nvme_rdma_ib_client);
2061         if (ret) {
2062                 destroy_workqueue(nvme_rdma_wq);
2063                 return ret;
2064         }
2065
2066         nvmf_register_transport(&nvme_rdma_transport);
2067         return 0;
2068 }
2069
2070 static void __exit nvme_rdma_cleanup_module(void)
2071 {
2072         nvmf_unregister_transport(&nvme_rdma_transport);
2073         ib_unregister_client(&nvme_rdma_ib_client);
2074         destroy_workqueue(nvme_rdma_wq);
2075 }
2076
2077 module_init(nvme_rdma_init_module);
2078 module_exit(nvme_rdma_cleanup_module);
2079
2080 MODULE_LICENSE("GPL v2");