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/blk-mq-rdma.h>
  23 #include <linux/types.h>
  24 #include <linux/list.h>
  25 #include <linux/mutex.h>
  26 #include <linux/scatterlist.h>
  27 #include <linux/nvme.h>
  28 #include <asm/unaligned.h>
  29
  30 #include <rdma/ib_verbs.h>
  31 #include <rdma/rdma_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    3000            /* 3 second */
  39
  40 #define NVME_RDMA_MAX_SEGMENTS          256
  41
  42 #define NVME_RDMA_MAX_INLINE_SEGMENTS   1
  43
  44 /*
  45  * We handle AEN commands ourselves and don't even let the
  46  * block layer know about them.
  47  */
  48 #define NVME_RDMA_NR_AEN_COMMANDS      1
  49 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
  50         (NVME_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
  51
  52 struct nvme_rdma_device {
  53         struct ib_device       *dev;
  54         struct ib_pd           *pd;
  55         struct kref             ref;
  56         struct list_head        entry;
  57 };
  58
  59 struct nvme_rdma_qe {
  60         struct ib_cqe           cqe;
  61         void                    *data;
  62         u64                     dma;
  63 };
  64
  65 struct nvme_rdma_queue;
  66 struct nvme_rdma_request {
  67         struct nvme_request     req;
  68         struct ib_mr            *mr;
  69         struct nvme_rdma_qe     sqe;
  70         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  71         u32                     num_sge;
  72         int                     nents;
  73         bool                    inline_data;
  74         struct ib_reg_wr        reg_wr;
  75         struct ib_cqe           reg_cqe;
  76         struct nvme_rdma_queue  *queue;
  77         struct sg_table         sg_table;
  78         struct scatterlist      first_sgl[];
  79 };
  80
  81 enum nvme_rdma_queue_flags {
  82         NVME_RDMA_Q_ALLOCATED           = 0,
  83         NVME_RDMA_Q_LIVE                = 1,
  84 };
  85
  86 struct nvme_rdma_queue {
  87         struct nvme_rdma_qe     *rsp_ring;
  88         atomic_t                sig_count;
  89         int                     queue_size;
  90         size_t                  cmnd_capsule_len;
  91         struct nvme_rdma_ctrl   *ctrl;
  92         struct nvme_rdma_device *device;
  93         struct ib_cq            *ib_cq;
  94         struct ib_qp            *qp;
  95
  96         unsigned long           flags;
  97         struct rdma_cm_id       *cm_id;
  98         int                     cm_error;
  99         struct completion       cm_done;
 100 };
 101
 102 struct nvme_rdma_ctrl {
 103         /* read only in the hot path */
 104         struct nvme_rdma_queue  *queues;
 105
 106         /* other member variables */
 107         struct blk_mq_tag_set   tag_set;
 108         struct work_struct      delete_work;
 109         struct work_struct      err_work;
 110
 111         struct nvme_rdma_qe     async_event_sqe;
 112
 113         struct delayed_work     reconnect_work;
 114
 115         struct list_head        list;
 116
 117         struct blk_mq_tag_set   admin_tag_set;
 118         struct nvme_rdma_device *device;
 119
 120         u32                     max_fr_pages;
 121
 122         struct sockaddr_storage addr;
 123         struct sockaddr_storage src_addr;
 124
 125         struct nvme_ctrl        ctrl;
 126 };
 127
 128 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 129 {
 130         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 131 }
 132
 133 static LIST_HEAD(device_list);
 134 static DEFINE_MUTEX(device_list_mutex);
 135
 136 static LIST_HEAD(nvme_rdma_ctrl_list);
 137 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 138
 139 /*
 140  * Disabling this option makes small I/O goes faster, but is fundamentally
 141  * unsafe.  With it turned off we will have to register a global rkey that
 142  * allows read and write access to all physical memory.
 143  */
 144 static bool register_always = true;
 145 module_param(register_always, bool, 0444);
 146 MODULE_PARM_DESC(register_always,
 147          "Use memory registration even for contiguous memory regions");
 148
 149 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 150                 struct rdma_cm_event *event);
 151 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 152
 153 static const struct blk_mq_ops nvme_rdma_mq_ops;
 154 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 155
 156 /* XXX: really should move to a generic header sooner or later.. */
 157 static inline void put_unaligned_le24(u32 val, u8 *p)
 158 {
 159         *p++ = val;
 160         *p++ = val >> 8;
 161         *p++ = val >> 16;
 162 }
 163
 164 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 165 {
 166         return queue - queue->ctrl->queues;
 167 }
 168
 169 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 170 {
 171         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 172 }
 173
 174 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 175                 size_t capsule_size, enum dma_data_direction dir)
 176 {
 177         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 178         kfree(qe->data);
 179 }
 180
 181 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 182                 size_t capsule_size, enum dma_data_direction dir)
 183 {
 184         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 185         if (!qe->data)
 186                 return -ENOMEM;
 187
 188         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 189         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 190                 kfree(qe->data);
 191                 return -ENOMEM;
 192         }
 193
 194         return 0;
 195 }
 196
 197 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 198                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 199                 size_t capsule_size, enum dma_data_direction dir)
 200 {
 201         int i;
 202
 203         for (i = 0; i < ib_queue_size; i++)
 204                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 205         kfree(ring);
 206 }
 207
 208 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 209                 size_t ib_queue_size, size_t capsule_size,
 210                 enum dma_data_direction dir)
 211 {
 212         struct nvme_rdma_qe *ring;
 213         int i;
 214
 215         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 216         if (!ring)
 217                 return NULL;
 218
 219         for (i = 0; i < ib_queue_size; i++) {
 220                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 221                         goto out_free_ring;
 222         }
 223
 224         return ring;
 225
 226 out_free_ring:
 227         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 228         return NULL;
 229 }
 230
 231 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 232 {
 233         pr_debug("QP event %s (%d)\n",
 234                  ib_event_msg(event->event), event->event);
 235
 236 }
 237
 238 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 239 {
 240         wait_for_completion_interruptible_timeout(&queue->cm_done,
 241                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 242         return queue->cm_error;
 243 }
 244
 245 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 246 {
 247         struct nvme_rdma_device *dev = queue->device;
 248         struct ib_qp_init_attr init_attr;
 249         int ret;
 250
 251         memset(&init_attr, 0, sizeof(init_attr));
 252         init_attr.event_handler = nvme_rdma_qp_event;
 253         /* +1 for drain */
 254         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 255         /* +1 for drain */
 256         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 257         init_attr.cap.max_recv_sge = 1;
 258         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 259         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 260         init_attr.qp_type = IB_QPT_RC;
 261         init_attr.send_cq = queue->ib_cq;
 262         init_attr.recv_cq = queue->ib_cq;
 263
 264         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 265
 266         queue->qp = queue->cm_id->qp;
 267         return ret;
 268 }
 269
 270 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 271 {
 272         struct nvme_rdma_ctrl *ctrl = data;
 273         struct nvme_rdma_device *dev = ctrl->device;
 274         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 275         int ret = 0;
 276
 277         if (WARN_ON_ONCE(!req->mr))
 278                 return 0;
 279
 280         ib_dereg_mr(req->mr);
 281
 282         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 283                         ctrl->max_fr_pages);
 284         if (IS_ERR(req->mr)) {
 285                 ret = PTR_ERR(req->mr);
 286                 req->mr = NULL;
 287                 goto out;
 288         }
 289
 290         req->mr->need_inval = false;
 291
 292 out:
 293         return ret;
 294 }
 295
 296 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 297                 struct request *rq, unsigned int hctx_idx)
 298 {
 299         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 300         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 301         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 302         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 303         struct nvme_rdma_device *dev = queue->device;
 304
 305         if (req->mr)
 306                 ib_dereg_mr(req->mr);
 307
 308         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 309                         DMA_TO_DEVICE);
 310 }
 311
 312 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 313                 struct request *rq, unsigned int hctx_idx,
 314                 unsigned int numa_node)
 315 {
 316         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 317         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 318         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 319         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 320         struct nvme_rdma_device *dev = queue->device;
 321         struct ib_device *ibdev = dev->dev;
 322         int ret;
 323
 324         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 325                         DMA_TO_DEVICE);
 326         if (ret)
 327                 return ret;
 328
 329         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 330                         ctrl->max_fr_pages);
 331         if (IS_ERR(req->mr)) {
 332                 ret = PTR_ERR(req->mr);
 333                 goto out_free_qe;
 334         }
 335
 336         req->queue = queue;
 337
 338         return 0;
 339
 340 out_free_qe:
 341         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 342                         DMA_TO_DEVICE);
 343         return -ENOMEM;
 344 }
 345
 346 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 347                 unsigned int hctx_idx)
 348 {
 349         struct nvme_rdma_ctrl *ctrl = data;
 350         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 351
 352         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 353
 354         hctx->driver_data = queue;
 355         return 0;
 356 }
 357
 358 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 359                 unsigned int hctx_idx)
 360 {
 361         struct nvme_rdma_ctrl *ctrl = data;
 362         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 363
 364         BUG_ON(hctx_idx != 0);
 365
 366         hctx->driver_data = queue;
 367         return 0;
 368 }
 369
 370 static void nvme_rdma_free_dev(struct kref *ref)
 371 {
 372         struct nvme_rdma_device *ndev =
 373                 container_of(ref, struct nvme_rdma_device, ref);
 374
 375         mutex_lock(&device_list_mutex);
 376         list_del(&ndev->entry);
 377         mutex_unlock(&device_list_mutex);
 378
 379         ib_dealloc_pd(ndev->pd);
 380         kfree(ndev);
 381 }
 382
 383 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 384 {
 385         kref_put(&dev->ref, nvme_rdma_free_dev);
 386 }
 387
 388 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 389 {
 390         return kref_get_unless_zero(&dev->ref);
 391 }
 392
 393 static struct nvme_rdma_device *
 394 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 395 {
 396         struct nvme_rdma_device *ndev;
 397
 398         mutex_lock(&device_list_mutex);
 399         list_for_each_entry(ndev, &device_list, entry) {
 400                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 401                     nvme_rdma_dev_get(ndev))
 402                         goto out_unlock;
 403         }
 404
 405         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 406         if (!ndev)
 407                 goto out_err;
 408
 409         ndev->dev = cm_id->device;
 410         kref_init(&ndev->ref);
 411
 412         ndev->pd = ib_alloc_pd(ndev->dev,
 413                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 414         if (IS_ERR(ndev->pd))
 415                 goto out_free_dev;
 416
 417         if (!(ndev->dev->attrs.device_cap_flags &
 418               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 419                 dev_err(&ndev->dev->dev,
 420                         "Memory registrations not supported.\n");
 421                 goto out_free_pd;
 422         }
 423
 424         list_add(&ndev->entry, &device_list);
 425 out_unlock:
 426         mutex_unlock(&device_list_mutex);
 427         return ndev;
 428
 429 out_free_pd:
 430         ib_dealloc_pd(ndev->pd);
 431 out_free_dev:
 432         kfree(ndev);
 433 out_err:
 434         mutex_unlock(&device_list_mutex);
 435         return NULL;
 436 }
 437
 438 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 439 {
 440         struct nvme_rdma_device *dev = queue->device;
 441         struct ib_device *ibdev = dev->dev;
 442
 443         rdma_destroy_qp(queue->cm_id);
 444         ib_free_cq(queue->ib_cq);
 445
 446         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 447                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 448
 449         nvme_rdma_dev_put(dev);
 450 }
 451
 452 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 453 {
 454         struct ib_device *ibdev;
 455         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 456         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 457         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 458         int ret;
 459
 460         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 461         if (!queue->device) {
 462                 dev_err(queue->cm_id->device->dev.parent,
 463                         "no client data found!\n");
 464                 return -ECONNREFUSED;
 465         }
 466         ibdev = queue->device->dev;
 467
 468         /*
 469          * Spread I/O queues completion vectors according their queue index.
 470          * Admin queues can always go on completion vector 0.
 471          */
 472         comp_vector = idx == 0 ? idx : idx - 1;
 473
 474         /* +1 for ib_stop_cq */
 475         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 476                                 cq_factor * queue->queue_size + 1,
 477                                 comp_vector, IB_POLL_SOFTIRQ);
 478         if (IS_ERR(queue->ib_cq)) {
 479                 ret = PTR_ERR(queue->ib_cq);
 480                 goto out_put_dev;
 481         }
 482
 483         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 484         if (ret)
 485                 goto out_destroy_ib_cq;
 486
 487         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 488                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 489         if (!queue->rsp_ring) {
 490                 ret = -ENOMEM;
 491                 goto out_destroy_qp;
 492         }
 493
 494         return 0;
 495
 496 out_destroy_qp:
 497         ib_destroy_qp(queue->qp);
 498 out_destroy_ib_cq:
 499         ib_free_cq(queue->ib_cq);
 500 out_put_dev:
 501         nvme_rdma_dev_put(queue->device);
 502         return ret;
 503 }
 504
 505 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 506                 int idx, size_t queue_size)
 507 {
 508         struct nvme_rdma_queue *queue;
 509         struct sockaddr *src_addr = NULL;
 510         int ret;
 511
 512         queue = &ctrl->queues[idx];
 513         queue->ctrl = ctrl;
 514         init_completion(&queue->cm_done);
 515
 516         if (idx > 0)
 517                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 518         else
 519                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 520
 521         queue->queue_size = queue_size;
 522         atomic_set(&queue->sig_count, 0);
 523
 524         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 525                         RDMA_PS_TCP, IB_QPT_RC);
 526         if (IS_ERR(queue->cm_id)) {
 527                 dev_info(ctrl->ctrl.device,
 528                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 529                 return PTR_ERR(queue->cm_id);
 530         }
 531
 532         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 533                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 534
 535         queue->cm_error = -ETIMEDOUT;
 536         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 537                         (struct sockaddr *)&ctrl->addr,
 538                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 539         if (ret) {
 540                 dev_info(ctrl->ctrl.device,
 541                         "rdma_resolve_addr failed (%d).\n", ret);
 542                 goto out_destroy_cm_id;
 543         }
 544
 545         ret = nvme_rdma_wait_for_cm(queue);
 546         if (ret) {
 547                 dev_info(ctrl->ctrl.device,
 548                         "rdma connection establishment failed (%d)\n", ret);
 549                 goto out_destroy_cm_id;
 550         }
 551
 552         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 553
 554         return 0;
 555
 556 out_destroy_cm_id:
 557         rdma_destroy_id(queue->cm_id);
 558         return ret;
 559 }
 560
 561 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 562 {
 563         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 564                 return;
 565
 566         rdma_disconnect(queue->cm_id);
 567         ib_drain_qp(queue->qp);
 568 }
 569
 570 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 571 {
 572         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 573                 return;
 574
 575         nvme_rdma_destroy_queue_ib(queue);
 576         rdma_destroy_id(queue->cm_id);
 577 }
 578
 579 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 580 {
 581         int i;
 582
 583         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 584                 nvme_rdma_free_queue(&ctrl->queues[i]);
 585 }
 586
 587 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 588 {
 589         int i;
 590
 591         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 592                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 593 }
 594
 595 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 596 {
 597         int ret;
 598
 599         if (idx)
 600                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 601         else
 602                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 603
 604         if (!ret)
 605                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
 606         else
 607                 dev_info(ctrl->ctrl.device,
 608                         "failed to connect queue: %d ret=%d\n", idx, ret);
 609         return ret;
 610 }
 611
 612 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 613 {
 614         int i, ret = 0;
 615
 616         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 617                 ret = nvme_rdma_start_queue(ctrl, i);
 618                 if (ret)
 619                         goto out_stop_queues;
 620         }
 621
 622         return 0;
 623
 624 out_stop_queues:
 625         for (i--; i >= 1; i--)
 626                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 627         return ret;
 628 }
 629
 630 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 631 {
 632         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 633         struct ib_device *ibdev = ctrl->device->dev;
 634         unsigned int nr_io_queues;
 635         int i, ret;
 636
 637         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 638
 639         /*
 640          * we map queues according to the device irq vectors for
 641          * optimal locality so we don't need more queues than
 642          * completion vectors.
 643          */
 644         nr_io_queues = min_t(unsigned int, nr_io_queues,
 645                                 ibdev->num_comp_vectors);
 646
 647         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 648         if (ret)
 649                 return ret;
 650
 651         ctrl->ctrl.queue_count = nr_io_queues + 1;
 652         if (ctrl->ctrl.queue_count < 2)
 653                 return 0;
 654
 655         dev_info(ctrl->ctrl.device,
 656                 "creating %d I/O queues.\n", nr_io_queues);
 657
 658         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 659                 ret = nvme_rdma_alloc_queue(ctrl, i,
 660                                 ctrl->ctrl.sqsize + 1);
 661                 if (ret)
 662                         goto out_free_queues;
 663         }
 664
 665         return 0;
 666
 667 out_free_queues:
 668         for (i--; i >= 1; i--)
 669                 nvme_rdma_free_queue(&ctrl->queues[i]);
 670
 671         return ret;
 672 }
 673
 674 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
 675                 struct blk_mq_tag_set *set)
 676 {
 677         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 678
 679         blk_mq_free_tag_set(set);
 680         nvme_rdma_dev_put(ctrl->device);
 681 }
 682
 683 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 684                 bool admin)
 685 {
 686         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 687         struct blk_mq_tag_set *set;
 688         int ret;
 689
 690         if (admin) {
 691                 set = &ctrl->admin_tag_set;
 692                 memset(set, 0, sizeof(*set));
 693                 set->ops = &nvme_rdma_admin_mq_ops;
 694                 set->queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
 695                 set->reserved_tags = 2; /* connect + keep-alive */
 696                 set->numa_node = NUMA_NO_NODE;
 697                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 698                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 699                 set->driver_data = ctrl;
 700                 set->nr_hw_queues = 1;
 701                 set->timeout = ADMIN_TIMEOUT;
 702                 set->flags = BLK_MQ_F_NO_SCHED;
 703         } else {
 704                 set = &ctrl->tag_set;
 705                 memset(set, 0, sizeof(*set));
 706                 set->ops = &nvme_rdma_mq_ops;
 707                 set->queue_depth = nctrl->opts->queue_size;
 708                 set->reserved_tags = 1; /* fabric connect */
 709                 set->numa_node = NUMA_NO_NODE;
 710                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 711                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 712                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 713                 set->driver_data = ctrl;
 714                 set->nr_hw_queues = nctrl->queue_count - 1;
 715                 set->timeout = NVME_IO_TIMEOUT;
 716         }
 717
 718         ret = blk_mq_alloc_tag_set(set);
 719         if (ret)
 720                 goto out;
 721
 722         /*
 723          * We need a reference on the device as long as the tag_set is alive,
 724          * as the MRs in the request structures need a valid ib_device.
 725          */
 726         ret = nvme_rdma_dev_get(ctrl->device);
 727         if (!ret) {
 728                 ret = -EINVAL;
 729                 goto out_free_tagset;
 730         }
 731
 732         return set;
 733
 734 out_free_tagset:
 735         blk_mq_free_tag_set(set);
 736 out:
 737         return ERR_PTR(ret);
 738 }
 739
 740 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 741                 bool remove)
 742 {
 743         nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
 744                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 745         nvme_rdma_stop_queue(&ctrl->queues[0]);
 746         if (remove) {
 747                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 748                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 749         }
 750         nvme_rdma_free_queue(&ctrl->queues[0]);
 751 }
 752
 753 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 754                 bool new)
 755 {
 756         int error;
 757
 758         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 759         if (error)
 760                 return error;
 761
 762         ctrl->device = ctrl->queues[0].device;
 763
 764         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
 765                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
 766
 767         if (new) {
 768                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 769                 if (IS_ERR(ctrl->ctrl.admin_tagset))
 770                         goto out_free_queue;
 771
 772                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 773                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 774                         error = PTR_ERR(ctrl->ctrl.admin_q);
 775                         goto out_free_tagset;
 776                 }
 777         } else {
 778                 error = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 779                 if (error)
 780                         goto out_free_queue;
 781         }
 782
 783         error = nvme_rdma_start_queue(ctrl, 0);
 784         if (error)
 785                 goto out_cleanup_queue;
 786
 787         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 788                         &ctrl->ctrl.cap);
 789         if (error) {
 790                 dev_err(ctrl->ctrl.device,
 791                         "prop_get NVME_REG_CAP failed\n");
 792                 goto out_cleanup_queue;
 793         }
 794
 795         ctrl->ctrl.sqsize =
 796                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 797
 798         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 799         if (error)
 800                 goto out_cleanup_queue;
 801
 802         ctrl->ctrl.max_hw_sectors =
 803                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 804
 805         error = nvme_init_identify(&ctrl->ctrl);
 806         if (error)
 807                 goto out_cleanup_queue;
 808
 809         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
 810                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
 811                         DMA_TO_DEVICE);
 812         if (error)
 813                 goto out_cleanup_queue;
 814
 815         return 0;
 816
 817 out_cleanup_queue:
 818         if (new)
 819                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 820 out_free_tagset:
 821         if (new)
 822                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 823 out_free_queue:
 824         nvme_rdma_free_queue(&ctrl->queues[0]);
 825         return error;
 826 }
 827
 828 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 829                 bool remove)
 830 {
 831         nvme_rdma_stop_io_queues(ctrl);
 832         if (remove) {
 833                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 834                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 835         }
 836         nvme_rdma_free_io_queues(ctrl);
 837 }
 838
 839 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 840 {
 841         int ret;
 842
 843         ret = nvme_rdma_alloc_io_queues(ctrl);
 844         if (ret)
 845                 return ret;
 846
 847         if (new) {
 848                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 849                 if (IS_ERR(ctrl->ctrl.tagset))
 850                         goto out_free_io_queues;
 851
 852                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 853                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 854                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 855                         goto out_free_tag_set;
 856                 }
 857         } else {
 858                 ret = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 859                 if (ret)
 860                         goto out_free_io_queues;
 861
 862                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 863                         ctrl->ctrl.queue_count - 1);
 864         }
 865
 866         ret = nvme_rdma_start_io_queues(ctrl);
 867         if (ret)
 868                 goto out_cleanup_connect_q;
 869
 870         return 0;
 871
 872 out_cleanup_connect_q:
 873         if (new)
 874                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 875 out_free_tag_set:
 876         if (new)
 877                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 878 out_free_io_queues:
 879         nvme_rdma_free_io_queues(ctrl);
 880         return ret;
 881 }
 882
 883 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 884 {
 885         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 886
 887         if (list_empty(&ctrl->list))
 888                 goto free_ctrl;
 889
 890         mutex_lock(&nvme_rdma_ctrl_mutex);
 891         list_del(&ctrl->list);
 892         mutex_unlock(&nvme_rdma_ctrl_mutex);
 893
 894         kfree(ctrl->queues);
 895         nvmf_free_options(nctrl->opts);
 896 free_ctrl:
 897         kfree(ctrl);
 898 }
 899
 900 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 901 {
 902         /* If we are resetting/deleting then do nothing */
 903         if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
 904                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 905                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 906                 return;
 907         }
 908
 909         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 910                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 911                         ctrl->ctrl.opts->reconnect_delay);
 912                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 913                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 914         } else {
 915                 dev_info(ctrl->ctrl.device, "Removing controller...\n");
 916                 queue_work(nvme_wq, &ctrl->delete_work);
 917         }
 918 }
 919
 920 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 921 {
 922         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 923                         struct nvme_rdma_ctrl, reconnect_work);
 924         bool changed;
 925         int ret;
 926
 927         ++ctrl->ctrl.nr_reconnects;
 928
 929         ret = nvme_rdma_configure_admin_queue(ctrl, false);
 930         if (ret)
 931                 goto requeue;
 932
 933         if (ctrl->ctrl.queue_count > 1) {
 934                 ret = nvme_rdma_configure_io_queues(ctrl, false);
 935                 if (ret)
 936                         goto destroy_admin;
 937         }
 938
 939         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 940         if (!changed) {
 941                 /* state change failure is ok if we're in DELETING state */
 942                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
 943                 return;
 944         }
 945
 946         nvme_start_ctrl(&ctrl->ctrl);
 947
 948         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
 949                         ctrl->ctrl.nr_reconnects);
 950
 951         ctrl->ctrl.nr_reconnects = 0;
 952
 953         return;
 954
 955 destroy_admin:
 956         nvme_rdma_destroy_admin_queue(ctrl, false);
 957 requeue:
 958         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
 959                         ctrl->ctrl.nr_reconnects);
 960         nvme_rdma_reconnect_or_remove(ctrl);
 961 }
 962
 963 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 964 {
 965         struct nvme_rdma_ctrl *ctrl = container_of(work,
 966                         struct nvme_rdma_ctrl, err_work);
 967
 968         nvme_stop_keep_alive(&ctrl->ctrl);
 969
 970         if (ctrl->ctrl.queue_count > 1) {
 971                 nvme_stop_queues(&ctrl->ctrl);
 972                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 973                                         nvme_cancel_request, &ctrl->ctrl);
 974                 nvme_rdma_destroy_io_queues(ctrl, false);
 975         }
 976
 977         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 978         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 979                                 nvme_cancel_request, &ctrl->ctrl);
 980         nvme_rdma_destroy_admin_queue(ctrl, false);
 981
 982         /*
 983          * queues are not a live anymore, so restart the queues to fail fast
 984          * new IO
 985          */
 986         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 987         nvme_start_queues(&ctrl->ctrl);
 988
 989         nvme_rdma_reconnect_or_remove(ctrl);
 990 }
 991
 992 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 993 {
 994         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 995                 return;
 996
 997         queue_work(nvme_wq, &ctrl->err_work);
 998 }
 999
1000 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1001                 const char *op)
1002 {
1003         struct nvme_rdma_queue *queue = cq->cq_context;
1004         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1005
1006         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1007                 dev_info(ctrl->ctrl.device,
1008                              "%s for CQE 0x%p failed with status %s (%d)\n",
1009                              op, wc->wr_cqe,
1010                              ib_wc_status_msg(wc->status), wc->status);
1011         nvme_rdma_error_recovery(ctrl);
1012 }
1013
1014 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1015 {
1016         if (unlikely(wc->status != IB_WC_SUCCESS))
1017                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1018 }
1019
1020 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1021 {
1022         if (unlikely(wc->status != IB_WC_SUCCESS))
1023                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1024 }
1025
1026 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1027                 struct nvme_rdma_request *req)
1028 {
1029         struct ib_send_wr *bad_wr;
1030         struct ib_send_wr wr = {
1031                 .opcode             = IB_WR_LOCAL_INV,
1032                 .next               = NULL,
1033                 .num_sge            = 0,
1034                 .send_flags         = 0,
1035                 .ex.invalidate_rkey = req->mr->rkey,
1036         };
1037
1038         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1039         wr.wr_cqe = &req->reg_cqe;
1040
1041         return ib_post_send(queue->qp, &wr, &bad_wr);
1042 }
1043
1044 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1045                 struct request *rq)
1046 {
1047         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1048         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1049         struct nvme_rdma_device *dev = queue->device;
1050         struct ib_device *ibdev = dev->dev;
1051         int res;
1052
1053         if (!blk_rq_bytes(rq))
1054                 return;
1055
1056         if (req->mr->need_inval && test_bit(NVME_RDMA_Q_LIVE, &req->queue->flags)) {
1057                 res = nvme_rdma_inv_rkey(queue, req);
1058                 if (unlikely(res < 0)) {
1059                         dev_err(ctrl->ctrl.device,
1060                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1061                                 req->mr->rkey, res);
1062                         nvme_rdma_error_recovery(queue->ctrl);
1063                 }
1064         }
1065
1066         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1067                         req->nents, rq_data_dir(rq) ==
1068                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1069
1070         nvme_cleanup_cmd(rq);
1071         sg_free_table_chained(&req->sg_table, true);
1072 }
1073
1074 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1075 {
1076         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1077
1078         sg->addr = 0;
1079         put_unaligned_le24(0, sg->length);
1080         put_unaligned_le32(0, sg->key);
1081         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1082         return 0;
1083 }
1084
1085 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1086                 struct nvme_rdma_request *req, struct nvme_command *c)
1087 {
1088         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1089
1090         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1091         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1092         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1093
1094         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1095         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1096         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1097
1098         req->inline_data = true;
1099         req->num_sge++;
1100         return 0;
1101 }
1102
1103 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1104                 struct nvme_rdma_request *req, struct nvme_command *c)
1105 {
1106         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1107
1108         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1109         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1110         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1111         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1112         return 0;
1113 }
1114
1115 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1116                 struct nvme_rdma_request *req, struct nvme_command *c,
1117                 int count)
1118 {
1119         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1120         int nr;
1121
1122         /*
1123          * Align the MR to a 4K page size to match the ctrl page size and
1124          * the block virtual boundary.
1125          */
1126         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1127         if (unlikely(nr < count)) {
1128                 if (nr < 0)
1129                         return nr;
1130                 return -EINVAL;
1131         }
1132
1133         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1134
1135         req->reg_cqe.done = nvme_rdma_memreg_done;
1136         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1137         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1138         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1139         req->reg_wr.wr.num_sge = 0;
1140         req->reg_wr.mr = req->mr;
1141         req->reg_wr.key = req->mr->rkey;
1142         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1143                              IB_ACCESS_REMOTE_READ |
1144                              IB_ACCESS_REMOTE_WRITE;
1145
1146         req->mr->need_inval = true;
1147
1148         sg->addr = cpu_to_le64(req->mr->iova);
1149         put_unaligned_le24(req->mr->length, sg->length);
1150         put_unaligned_le32(req->mr->rkey, sg->key);
1151         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1152                         NVME_SGL_FMT_INVALIDATE;
1153
1154         return 0;
1155 }
1156
1157 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1158                 struct request *rq, struct nvme_command *c)
1159 {
1160         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1161         struct nvme_rdma_device *dev = queue->device;
1162         struct ib_device *ibdev = dev->dev;
1163         int count, ret;
1164
1165         req->num_sge = 1;
1166         req->inline_data = false;
1167         req->mr->need_inval = false;
1168
1169         c->common.flags |= NVME_CMD_SGL_METABUF;
1170
1171         if (!blk_rq_bytes(rq))
1172                 return nvme_rdma_set_sg_null(c);
1173
1174         req->sg_table.sgl = req->first_sgl;
1175         ret = sg_alloc_table_chained(&req->sg_table,
1176                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1177         if (ret)
1178                 return -ENOMEM;
1179
1180         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1181
1182         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1183                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1184         if (unlikely(count <= 0)) {
1185                 sg_free_table_chained(&req->sg_table, true);
1186                 return -EIO;
1187         }
1188
1189         if (count == 1) {
1190                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1191                     blk_rq_payload_bytes(rq) <=
1192                                 nvme_rdma_inline_data_size(queue))
1193                         return nvme_rdma_map_sg_inline(queue, req, c);
1194
1195                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1196                         return nvme_rdma_map_sg_single(queue, req, c);
1197         }
1198
1199         return nvme_rdma_map_sg_fr(queue, req, c, count);
1200 }
1201
1202 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1203 {
1204         if (unlikely(wc->status != IB_WC_SUCCESS))
1205                 nvme_rdma_wr_error(cq, wc, "SEND");
1206 }
1207
1208 /*
1209  * We want to signal completion at least every queue depth/2.  This returns the
1210  * largest power of two that is not above half of (queue size + 1) to optimize
1211  * (avoid divisions).
1212  */
1213 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1214 {
1215         int limit = 1 << ilog2((queue->queue_size + 1) / 2);
1216
1217         return (atomic_inc_return(&queue->sig_count) & (limit - 1)) == 0;
1218 }
1219
1220 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1221                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1222                 struct ib_send_wr *first, bool flush)
1223 {
1224         struct ib_send_wr wr, *bad_wr;
1225         int ret;
1226
1227         sge->addr   = qe->dma;
1228         sge->length = sizeof(struct nvme_command),
1229         sge->lkey   = queue->device->pd->local_dma_lkey;
1230
1231         qe->cqe.done = nvme_rdma_send_done;
1232
1233         wr.next       = NULL;
1234         wr.wr_cqe     = &qe->cqe;
1235         wr.sg_list    = sge;
1236         wr.num_sge    = num_sge;
1237         wr.opcode     = IB_WR_SEND;
1238         wr.send_flags = 0;
1239
1240         /*
1241          * Unsignalled send completions are another giant desaster in the
1242          * IB Verbs spec:  If we don't regularly post signalled sends
1243          * the send queue will fill up and only a QP reset will rescue us.
1244          * Would have been way to obvious to handle this in hardware or
1245          * at least the RDMA stack..
1246          *
1247          * Always signal the flushes. The magic request used for the flush
1248          * sequencer is not allocated in our driver's tagset and it's
1249          * triggered to be freed by blk_cleanup_queue(). So we need to
1250          * always mark it as signaled to ensure that the "wr_cqe", which is
1251          * embedded in request's payload, is not freed when __ib_process_cq()
1252          * calls wr_cqe->done().
1253          */
1254         if (nvme_rdma_queue_sig_limit(queue) || flush)
1255                 wr.send_flags |= IB_SEND_SIGNALED;
1256
1257         if (first)
1258                 first->next = &wr;
1259         else
1260                 first = &wr;
1261
1262         ret = ib_post_send(queue->qp, first, &bad_wr);
1263         if (unlikely(ret)) {
1264                 dev_err(queue->ctrl->ctrl.device,
1265                              "%s failed with error code %d\n", __func__, ret);
1266         }
1267         return ret;
1268 }
1269
1270 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1271                 struct nvme_rdma_qe *qe)
1272 {
1273         struct ib_recv_wr wr, *bad_wr;
1274         struct ib_sge list;
1275         int ret;
1276
1277         list.addr   = qe->dma;
1278         list.length = sizeof(struct nvme_completion);
1279         list.lkey   = queue->device->pd->local_dma_lkey;
1280
1281         qe->cqe.done = nvme_rdma_recv_done;
1282
1283         wr.next     = NULL;
1284         wr.wr_cqe   = &qe->cqe;
1285         wr.sg_list  = &list;
1286         wr.num_sge  = 1;
1287
1288         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1289         if (unlikely(ret)) {
1290                 dev_err(queue->ctrl->ctrl.device,
1291                         "%s failed with error code %d\n", __func__, ret);
1292         }
1293         return ret;
1294 }
1295
1296 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1297 {
1298         u32 queue_idx = nvme_rdma_queue_idx(queue);
1299
1300         if (queue_idx == 0)
1301                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1302         return queue->ctrl->tag_set.tags[queue_idx - 1];
1303 }
1304
1305 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1306 {
1307         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1308         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1309         struct ib_device *dev = queue->device->dev;
1310         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1311         struct nvme_command *cmd = sqe->data;
1312         struct ib_sge sge;
1313         int ret;
1314
1315         if (WARN_ON_ONCE(aer_idx != 0))
1316                 return;
1317
1318         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1319
1320         memset(cmd, 0, sizeof(*cmd));
1321         cmd->common.opcode = nvme_admin_async_event;
1322         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1323         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1324         nvme_rdma_set_sg_null(cmd);
1325
1326         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1327                         DMA_TO_DEVICE);
1328
1329         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1330         WARN_ON_ONCE(ret);
1331 }
1332
1333 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1334                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1335 {
1336         struct request *rq;
1337         struct nvme_rdma_request *req;
1338         int ret = 0;
1339
1340         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1341         if (!rq) {
1342                 dev_err(queue->ctrl->ctrl.device,
1343                         "tag 0x%x on QP %#x not found\n",
1344                         cqe->command_id, queue->qp->qp_num);
1345                 nvme_rdma_error_recovery(queue->ctrl);
1346                 return ret;
1347         }
1348         req = blk_mq_rq_to_pdu(rq);
1349
1350         if (rq->tag == tag)
1351                 ret = 1;
1352
1353         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1354             wc->ex.invalidate_rkey == req->mr->rkey)
1355                 req->mr->need_inval = false;
1356
1357         nvme_end_request(rq, cqe->status, cqe->result);
1358         return ret;
1359 }
1360
1361 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1362 {
1363         struct nvme_rdma_qe *qe =
1364                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1365         struct nvme_rdma_queue *queue = cq->cq_context;
1366         struct ib_device *ibdev = queue->device->dev;
1367         struct nvme_completion *cqe = qe->data;
1368         const size_t len = sizeof(struct nvme_completion);
1369         int ret = 0;
1370
1371         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1372                 nvme_rdma_wr_error(cq, wc, "RECV");
1373                 return 0;
1374         }
1375
1376         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1377         /*
1378          * AEN requests are special as they don't time out and can
1379          * survive any kind of queue freeze and often don't respond to
1380          * aborts.  We don't even bother to allocate a struct request
1381          * for them but rather special case them here.
1382          */
1383         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1384                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1385                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1386                                 &cqe->result);
1387         else
1388                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1389         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1390
1391         nvme_rdma_post_recv(queue, qe);
1392         return ret;
1393 }
1394
1395 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1396 {
1397         __nvme_rdma_recv_done(cq, wc, -1);
1398 }
1399
1400 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1401 {
1402         int ret, i;
1403
1404         for (i = 0; i < queue->queue_size; i++) {
1405                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1406                 if (ret)
1407                         goto out_destroy_queue_ib;
1408         }
1409
1410         return 0;
1411
1412 out_destroy_queue_ib:
1413         nvme_rdma_destroy_queue_ib(queue);
1414         return ret;
1415 }
1416
1417 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1418                 struct rdma_cm_event *ev)
1419 {
1420         struct rdma_cm_id *cm_id = queue->cm_id;
1421         int status = ev->status;
1422         const char *rej_msg;
1423         const struct nvme_rdma_cm_rej *rej_data;
1424         u8 rej_data_len;
1425
1426         rej_msg = rdma_reject_msg(cm_id, status);
1427         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1428
1429         if (rej_data && rej_data_len >= sizeof(u16)) {
1430                 u16 sts = le16_to_cpu(rej_data->sts);
1431
1432                 dev_err(queue->ctrl->ctrl.device,
1433                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1434                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1435         } else {
1436                 dev_err(queue->ctrl->ctrl.device,
1437                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1438         }
1439
1440         return -ECONNRESET;
1441 }
1442
1443 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1444 {
1445         int ret;
1446
1447         ret = nvme_rdma_create_queue_ib(queue);
1448         if (ret)
1449                 return ret;
1450
1451         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1452         if (ret) {
1453                 dev_err(queue->ctrl->ctrl.device,
1454                         "rdma_resolve_route failed (%d).\n",
1455                         queue->cm_error);
1456                 goto out_destroy_queue;
1457         }
1458
1459         return 0;
1460
1461 out_destroy_queue:
1462         nvme_rdma_destroy_queue_ib(queue);
1463         return ret;
1464 }
1465
1466 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1467 {
1468         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1469         struct rdma_conn_param param = { };
1470         struct nvme_rdma_cm_req priv = { };
1471         int ret;
1472
1473         param.qp_num = queue->qp->qp_num;
1474         param.flow_control = 1;
1475
1476         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1477         /* maximum retry count */
1478         param.retry_count = 7;
1479         param.rnr_retry_count = 7;
1480         param.private_data = &priv;
1481         param.private_data_len = sizeof(priv);
1482
1483         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1484         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1485         /*
1486          * set the admin queue depth to the minimum size
1487          * specified by the Fabrics standard.
1488          */
1489         if (priv.qid == 0) {
1490                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1491                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1492         } else {
1493                 /*
1494                  * current interpretation of the fabrics spec
1495                  * is at minimum you make hrqsize sqsize+1, or a
1496                  * 1's based representation of sqsize.
1497                  */
1498                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1499                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1500         }
1501
1502         ret = rdma_connect(queue->cm_id, &param);
1503         if (ret) {
1504                 dev_err(ctrl->ctrl.device,
1505                         "rdma_connect failed (%d).\n", ret);
1506                 goto out_destroy_queue_ib;
1507         }
1508
1509         return 0;
1510
1511 out_destroy_queue_ib:
1512         nvme_rdma_destroy_queue_ib(queue);
1513         return ret;
1514 }
1515
1516 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1517                 struct rdma_cm_event *ev)
1518 {
1519         struct nvme_rdma_queue *queue = cm_id->context;
1520         int cm_error = 0;
1521
1522         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1523                 rdma_event_msg(ev->event), ev->event,
1524                 ev->status, cm_id);
1525
1526         switch (ev->event) {
1527         case RDMA_CM_EVENT_ADDR_RESOLVED:
1528                 cm_error = nvme_rdma_addr_resolved(queue);
1529                 break;
1530         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1531                 cm_error = nvme_rdma_route_resolved(queue);
1532                 break;
1533         case RDMA_CM_EVENT_ESTABLISHED:
1534                 queue->cm_error = nvme_rdma_conn_established(queue);
1535                 /* complete cm_done regardless of success/failure */
1536                 complete(&queue->cm_done);
1537                 return 0;
1538         case RDMA_CM_EVENT_REJECTED:
1539                 nvme_rdma_destroy_queue_ib(queue);
1540                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1541                 break;
1542         case RDMA_CM_EVENT_ROUTE_ERROR:
1543         case RDMA_CM_EVENT_CONNECT_ERROR:
1544         case RDMA_CM_EVENT_UNREACHABLE:
1545                 nvme_rdma_destroy_queue_ib(queue);
1546         case RDMA_CM_EVENT_ADDR_ERROR:
1547                 dev_dbg(queue->ctrl->ctrl.device,
1548                         "CM error event %d\n", ev->event);
1549                 cm_error = -ECONNRESET;
1550                 break;
1551         case RDMA_CM_EVENT_DISCONNECTED:
1552         case RDMA_CM_EVENT_ADDR_CHANGE:
1553         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1554                 dev_dbg(queue->ctrl->ctrl.device,
1555                         "disconnect received - connection closed\n");
1556                 nvme_rdma_error_recovery(queue->ctrl);
1557                 break;
1558         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1559                 /* device removal is handled via the ib_client API */
1560                 break;
1561         default:
1562                 dev_err(queue->ctrl->ctrl.device,
1563                         "Unexpected RDMA CM event (%d)\n", ev->event);
1564                 nvme_rdma_error_recovery(queue->ctrl);
1565                 break;
1566         }
1567
1568         if (cm_error) {
1569                 queue->cm_error = cm_error;
1570                 complete(&queue->cm_done);
1571         }
1572
1573         return 0;
1574 }
1575
1576 static enum blk_eh_timer_return
1577 nvme_rdma_timeout(struct request *rq, bool reserved)
1578 {
1579         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1580
1581         /* queue error recovery */
1582         nvme_rdma_error_recovery(req->queue->ctrl);
1583
1584         /* fail with DNR on cmd timeout */
1585         nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1586
1587         return BLK_EH_HANDLED;
1588 }
1589
1590 /*
1591  * We cannot accept any other command until the Connect command has completed.
1592  */
1593 static inline blk_status_t
1594 nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1595 {
1596         if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1597                 struct nvme_command *cmd = nvme_req(rq)->cmd;
1598
1599                 if (!blk_rq_is_passthrough(rq) ||
1600                     cmd->common.opcode != nvme_fabrics_command ||
1601                     cmd->fabrics.fctype != nvme_fabrics_type_connect) {
1602                         /*
1603                          * reconnecting state means transport disruption, which
1604                          * can take a long time and even might fail permanently,
1605                          * so we can't let incoming I/O be requeued forever.
1606                          * fail it fast to allow upper layers a chance to
1607                          * failover.
1608                          */
1609                         if (queue->ctrl->ctrl.state == NVME_CTRL_RECONNECTING)
1610                                 return BLK_STS_IOERR;
1611                         return BLK_STS_RESOURCE; /* try again later */
1612                 }
1613         }
1614
1615         return 0;
1616 }
1617
1618 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1619                 const struct blk_mq_queue_data *bd)
1620 {
1621         struct nvme_ns *ns = hctx->queue->queuedata;
1622         struct nvme_rdma_queue *queue = hctx->driver_data;
1623         struct request *rq = bd->rq;
1624         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1625         struct nvme_rdma_qe *sqe = &req->sqe;
1626         struct nvme_command *c = sqe->data;
1627         bool flush = false;
1628         struct ib_device *dev;
1629         blk_status_t ret;
1630         int err;
1631
1632         WARN_ON_ONCE(rq->tag < 0);
1633
1634         ret = nvme_rdma_queue_is_ready(queue, rq);
1635         if (unlikely(ret))
1636                 return ret;
1637
1638         dev = queue->device->dev;
1639         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1640                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1641
1642         ret = nvme_setup_cmd(ns, rq, c);
1643         if (ret)
1644                 return ret;
1645
1646         blk_mq_start_request(rq);
1647
1648         err = nvme_rdma_map_data(queue, rq, c);
1649         if (unlikely(err < 0)) {
1650                 dev_err(queue->ctrl->ctrl.device,
1651                              "Failed to map data (%d)\n", err);
1652                 nvme_cleanup_cmd(rq);
1653                 goto err;
1654         }
1655
1656         ib_dma_sync_single_for_device(dev, sqe->dma,
1657                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1658
1659         if (req_op(rq) == REQ_OP_FLUSH)
1660                 flush = true;
1661         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1662                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1663         if (unlikely(err)) {
1664                 nvme_rdma_unmap_data(queue, rq);
1665                 goto err;
1666         }
1667
1668         return BLK_STS_OK;
1669 err:
1670         if (err == -ENOMEM || err == -EAGAIN)
1671                 return BLK_STS_RESOURCE;
1672         return BLK_STS_IOERR;
1673 }
1674
1675 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1676 {
1677         struct nvme_rdma_queue *queue = hctx->driver_data;
1678         struct ib_cq *cq = queue->ib_cq;
1679         struct ib_wc wc;
1680         int found = 0;
1681
1682         while (ib_poll_cq(cq, 1, &wc) > 0) {
1683                 struct ib_cqe *cqe = wc.wr_cqe;
1684
1685                 if (cqe) {
1686                         if (cqe->done == nvme_rdma_recv_done)
1687                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1688                         else
1689                                 cqe->done(cq, &wc);
1690                 }
1691         }
1692
1693         return found;
1694 }
1695
1696 static void nvme_rdma_complete_rq(struct request *rq)
1697 {
1698         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1699
1700         nvme_rdma_unmap_data(req->queue, rq);
1701         nvme_complete_rq(rq);
1702 }
1703
1704 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1705 {
1706         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1707
1708         return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1709 }
1710
1711 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1712         .queue_rq       = nvme_rdma_queue_rq,
1713         .complete       = nvme_rdma_complete_rq,
1714         .init_request   = nvme_rdma_init_request,
1715         .exit_request   = nvme_rdma_exit_request,
1716         .init_hctx      = nvme_rdma_init_hctx,
1717         .poll           = nvme_rdma_poll,
1718         .timeout        = nvme_rdma_timeout,
1719         .map_queues     = nvme_rdma_map_queues,
1720 };
1721
1722 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1723         .queue_rq       = nvme_rdma_queue_rq,
1724         .complete       = nvme_rdma_complete_rq,
1725         .init_request   = nvme_rdma_init_request,
1726         .exit_request   = nvme_rdma_exit_request,
1727         .init_hctx      = nvme_rdma_init_admin_hctx,
1728         .timeout        = nvme_rdma_timeout,
1729 };
1730
1731 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1732 {
1733         cancel_work_sync(&ctrl->err_work);
1734         cancel_delayed_work_sync(&ctrl->reconnect_work);
1735
1736         if (ctrl->ctrl.queue_count > 1) {
1737                 nvme_stop_queues(&ctrl->ctrl);
1738                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1739                                         nvme_cancel_request, &ctrl->ctrl);
1740                 nvme_rdma_destroy_io_queues(ctrl, shutdown);
1741         }
1742
1743         if (shutdown)
1744                 nvme_shutdown_ctrl(&ctrl->ctrl);
1745         else
1746                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1747
1748         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1749         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1750                                 nvme_cancel_request, &ctrl->ctrl);
1751         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1752         nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1753 }
1754
1755 static void nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl)
1756 {
1757         nvme_remove_namespaces(&ctrl->ctrl);
1758         nvme_rdma_shutdown_ctrl(ctrl, true);
1759         nvme_uninit_ctrl(&ctrl->ctrl);
1760         nvme_put_ctrl(&ctrl->ctrl);
1761 }
1762
1763 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1764 {
1765         struct nvme_rdma_ctrl *ctrl = container_of(work,
1766                                 struct nvme_rdma_ctrl, delete_work);
1767
1768         nvme_stop_ctrl(&ctrl->ctrl);
1769         nvme_rdma_remove_ctrl(ctrl);
1770 }
1771
1772 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1773 {
1774         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1775                 return -EBUSY;
1776
1777         if (!queue_work(nvme_wq, &ctrl->delete_work))
1778                 return -EBUSY;
1779
1780         return 0;
1781 }
1782
1783 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1784 {
1785         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1786         int ret = 0;
1787
1788         /*
1789          * Keep a reference until all work is flushed since
1790          * __nvme_rdma_del_ctrl can free the ctrl mem
1791          */
1792         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1793                 return -EBUSY;
1794         ret = __nvme_rdma_del_ctrl(ctrl);
1795         if (!ret)
1796                 flush_work(&ctrl->delete_work);
1797         nvme_put_ctrl(&ctrl->ctrl);
1798         return ret;
1799 }
1800
1801 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1802 {
1803         struct nvme_rdma_ctrl *ctrl =
1804                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1805         int ret;
1806         bool changed;
1807
1808         nvme_stop_ctrl(&ctrl->ctrl);
1809         nvme_rdma_shutdown_ctrl(ctrl, false);
1810
1811         ret = nvme_rdma_configure_admin_queue(ctrl, false);
1812         if (ret)
1813                 goto out_fail;
1814
1815         if (ctrl->ctrl.queue_count > 1) {
1816                 ret = nvme_rdma_configure_io_queues(ctrl, false);
1817                 if (ret)
1818                         goto out_fail;
1819         }
1820
1821         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1822         if (!changed) {
1823                 /* state change failure is ok if we're in DELETING state */
1824                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1825                 return;
1826         }
1827
1828         nvme_start_ctrl(&ctrl->ctrl);
1829
1830         return;
1831
1832 out_fail:
1833         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1834         nvme_rdma_remove_ctrl(ctrl);
1835 }
1836
1837 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1838         .name                   = "rdma",
1839         .module                 = THIS_MODULE,
1840         .flags                  = NVME_F_FABRICS,
1841         .reg_read32             = nvmf_reg_read32,
1842         .reg_read64             = nvmf_reg_read64,
1843         .reg_write32            = nvmf_reg_write32,
1844         .free_ctrl              = nvme_rdma_free_ctrl,
1845         .submit_async_event     = nvme_rdma_submit_async_event,
1846         .delete_ctrl            = nvme_rdma_del_ctrl,
1847         .get_address            = nvmf_get_address,
1848         .reinit_request         = nvme_rdma_reinit_request,
1849 };
1850
1851 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1852                 struct nvmf_ctrl_options *opts)
1853 {
1854         struct nvme_rdma_ctrl *ctrl;
1855         int ret;
1856         bool changed;
1857         char *port;
1858
1859         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1860         if (!ctrl)
1861                 return ERR_PTR(-ENOMEM);
1862         ctrl->ctrl.opts = opts;
1863         INIT_LIST_HEAD(&ctrl->list);
1864
1865         if (opts->mask & NVMF_OPT_TRSVCID)
1866                 port = opts->trsvcid;
1867         else
1868                 port = __stringify(NVME_RDMA_IP_PORT);
1869
1870         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1871                         opts->traddr, port, &ctrl->addr);
1872         if (ret) {
1873                 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1874                 goto out_free_ctrl;
1875         }
1876
1877         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1878                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1879                         opts->host_traddr, NULL, &ctrl->src_addr);
1880                 if (ret) {
1881                         pr_err("malformed src address passed: %s\n",
1882                                opts->host_traddr);
1883                         goto out_free_ctrl;
1884                 }
1885         }
1886
1887         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1888                                 0 /* no quirks, we're perfect! */);
1889         if (ret)
1890                 goto out_free_ctrl;
1891
1892         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1893                         nvme_rdma_reconnect_ctrl_work);
1894         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1895         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1896         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1897
1898         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1899         ctrl->ctrl.sqsize = opts->queue_size - 1;
1900         ctrl->ctrl.kato = opts->kato;
1901
1902         ret = -ENOMEM;
1903         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1904                                 GFP_KERNEL);
1905         if (!ctrl->queues)
1906                 goto out_uninit_ctrl;
1907
1908         ret = nvme_rdma_configure_admin_queue(ctrl, true);
1909         if (ret)
1910                 goto out_kfree_queues;
1911
1912         /* sanity check icdoff */
1913         if (ctrl->ctrl.icdoff) {
1914                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1915                 ret = -EINVAL;
1916                 goto out_remove_admin_queue;
1917         }
1918
1919         /* sanity check keyed sgls */
1920         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1921                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1922                 ret = -EINVAL;
1923                 goto out_remove_admin_queue;
1924         }
1925
1926         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1927                 /* warn if maxcmd is lower than queue_size */
1928                 dev_warn(ctrl->ctrl.device,
1929                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1930                         opts->queue_size, ctrl->ctrl.maxcmd);
1931                 opts->queue_size = ctrl->ctrl.maxcmd;
1932         }
1933
1934         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1935                 /* warn if sqsize is lower than queue_size */
1936                 dev_warn(ctrl->ctrl.device,
1937                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1938                         opts->queue_size, ctrl->ctrl.sqsize + 1);
1939                 opts->queue_size = ctrl->ctrl.sqsize + 1;
1940         }
1941
1942         if (opts->nr_io_queues) {
1943                 ret = nvme_rdma_configure_io_queues(ctrl, true);
1944                 if (ret)
1945                         goto out_remove_admin_queue;
1946         }
1947
1948         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1949         WARN_ON_ONCE(!changed);
1950
1951         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1952                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1953
1954         kref_get(&ctrl->ctrl.kref);
1955
1956         mutex_lock(&nvme_rdma_ctrl_mutex);
1957         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1958         mutex_unlock(&nvme_rdma_ctrl_mutex);
1959
1960         nvme_start_ctrl(&ctrl->ctrl);
1961
1962         return &ctrl->ctrl;
1963
1964 out_remove_admin_queue:
1965         nvme_rdma_destroy_admin_queue(ctrl, true);
1966 out_kfree_queues:
1967         kfree(ctrl->queues);
1968 out_uninit_ctrl:
1969         nvme_uninit_ctrl(&ctrl->ctrl);
1970         nvme_put_ctrl(&ctrl->ctrl);
1971         if (ret > 0)
1972                 ret = -EIO;
1973         return ERR_PTR(ret);
1974 out_free_ctrl:
1975         kfree(ctrl);
1976         return ERR_PTR(ret);
1977 }
1978
1979 static struct nvmf_transport_ops nvme_rdma_transport = {
1980         .name           = "rdma",
1981         .required_opts  = NVMF_OPT_TRADDR,
1982         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1983                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1984         .create_ctrl    = nvme_rdma_create_ctrl,
1985 };
1986
1987 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1988 {
1989         struct nvme_rdma_ctrl *ctrl;
1990
1991         /* Delete all controllers using this device */
1992         mutex_lock(&nvme_rdma_ctrl_mutex);
1993         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1994                 if (ctrl->device->dev != ib_device)
1995                         continue;
1996                 dev_info(ctrl->ctrl.device,
1997                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
1998                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1999                 __nvme_rdma_del_ctrl(ctrl);
2000         }
2001         mutex_unlock(&nvme_rdma_ctrl_mutex);
2002
2003         flush_workqueue(nvme_wq);
2004 }
2005
2006 static struct ib_client nvme_rdma_ib_client = {
2007         .name   = "nvme_rdma",
2008         .remove = nvme_rdma_remove_one
2009 };
2010
2011 static int __init nvme_rdma_init_module(void)
2012 {
2013         int ret;
2014
2015         ret = ib_register_client(&nvme_rdma_ib_client);
2016         if (ret)
2017                 return ret;
2018
2019         ret = nvmf_register_transport(&nvme_rdma_transport);
2020         if (ret)
2021                 goto err_unreg_client;
2022
2023         return 0;
2024
2025 err_unreg_client:
2026         ib_unregister_client(&nvme_rdma_ib_client);
2027         return ret;
2028 }
2029
2030 static void __exit nvme_rdma_cleanup_module(void)
2031 {
2032         nvmf_unregister_transport(&nvme_rdma_transport);
2033         ib_unregister_client(&nvme_rdma_ib_client);
2034 }
2035
2036 module_init(nvme_rdma_init_module);
2037 module_exit(nvme_rdma_cleanup_module);
2038
2039 MODULE_LICENSE("GPL v2");