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 struct nvme_rdma_device {
  45         struct ib_device        *dev;
  46         struct ib_pd            *pd;
  47         struct kref             ref;
  48         struct list_head        entry;
  49 };
  50
  51 struct nvme_rdma_qe {
  52         struct ib_cqe           cqe;
  53         void                    *data;
  54         u64                     dma;
  55 };
  56
  57 struct nvme_rdma_queue;
  58 struct nvme_rdma_request {
  59         struct nvme_request     req;
  60         struct ib_mr            *mr;
  61         struct nvme_rdma_qe     sqe;
  62         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  63         u32                     num_sge;
  64         int                     nents;
  65         bool                    inline_data;
  66         struct ib_reg_wr        reg_wr;
  67         struct ib_cqe           reg_cqe;
  68         struct nvme_rdma_queue  *queue;
  69         struct sg_table         sg_table;
  70         struct scatterlist      first_sgl[];
  71 };
  72
  73 enum nvme_rdma_queue_flags {
  74         NVME_RDMA_Q_ALLOCATED           = 0,
  75         NVME_RDMA_Q_LIVE                = 1,
  76 };
  77
  78 struct nvme_rdma_queue {
  79         struct nvme_rdma_qe     *rsp_ring;
  80         atomic_t                sig_count;
  81         int                     queue_size;
  82         size_t                  cmnd_capsule_len;
  83         struct nvme_rdma_ctrl   *ctrl;
  84         struct nvme_rdma_device *device;
  85         struct ib_cq            *ib_cq;
  86         struct ib_qp            *qp;
  87
  88         unsigned long           flags;
  89         struct rdma_cm_id       *cm_id;
  90         int                     cm_error;
  91         struct completion       cm_done;
  92 };
  93
  94 struct nvme_rdma_ctrl {
  95         /* read only in the hot path */
  96         struct nvme_rdma_queue  *queues;
  97
  98         /* other member variables */
  99         struct blk_mq_tag_set   tag_set;
 100         struct work_struct      err_work;
 101
 102         struct nvme_rdma_qe     async_event_sqe;
 103
 104         struct delayed_work     reconnect_work;
 105
 106         struct list_head        list;
 107
 108         struct blk_mq_tag_set   admin_tag_set;
 109         struct nvme_rdma_device *device;
 110
 111         u32                     max_fr_pages;
 112
 113         struct sockaddr_storage addr;
 114         struct sockaddr_storage src_addr;
 115
 116         struct nvme_ctrl        ctrl;
 117 };
 118
 119 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 120 {
 121         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 122 }
 123
 124 static LIST_HEAD(device_list);
 125 static DEFINE_MUTEX(device_list_mutex);
 126
 127 static LIST_HEAD(nvme_rdma_ctrl_list);
 128 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 129
 130 /*
 131  * Disabling this option makes small I/O goes faster, but is fundamentally
 132  * unsafe.  With it turned off we will have to register a global rkey that
 133  * allows read and write access to all physical memory.
 134  */
 135 static bool register_always = true;
 136 module_param(register_always, bool, 0444);
 137 MODULE_PARM_DESC(register_always,
 138          "Use memory registration even for contiguous memory regions");
 139
 140 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 141                 struct rdma_cm_event *event);
 142 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 143
 144 static const struct blk_mq_ops nvme_rdma_mq_ops;
 145 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
 146
 147 /* XXX: really should move to a generic header sooner or later.. */
 148 static inline void put_unaligned_le24(u32 val, u8 *p)
 149 {
 150         *p++ = val;
 151         *p++ = val >> 8;
 152         *p++ = val >> 16;
 153 }
 154
 155 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 156 {
 157         return queue - queue->ctrl->queues;
 158 }
 159
 160 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 161 {
 162         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 163 }
 164
 165 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 166                 size_t capsule_size, enum dma_data_direction dir)
 167 {
 168         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 169         kfree(qe->data);
 170 }
 171
 172 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 173                 size_t capsule_size, enum dma_data_direction dir)
 174 {
 175         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 176         if (!qe->data)
 177                 return -ENOMEM;
 178
 179         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 180         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 181                 kfree(qe->data);
 182                 return -ENOMEM;
 183         }
 184
 185         return 0;
 186 }
 187
 188 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 189                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 190                 size_t capsule_size, enum dma_data_direction dir)
 191 {
 192         int i;
 193
 194         for (i = 0; i < ib_queue_size; i++)
 195                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 196         kfree(ring);
 197 }
 198
 199 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 200                 size_t ib_queue_size, size_t capsule_size,
 201                 enum dma_data_direction dir)
 202 {
 203         struct nvme_rdma_qe *ring;
 204         int i;
 205
 206         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 207         if (!ring)
 208                 return NULL;
 209
 210         for (i = 0; i < ib_queue_size; i++) {
 211                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 212                         goto out_free_ring;
 213         }
 214
 215         return ring;
 216
 217 out_free_ring:
 218         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 219         return NULL;
 220 }
 221
 222 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 223 {
 224         pr_debug("QP event %s (%d)\n",
 225                  ib_event_msg(event->event), event->event);
 226
 227 }
 228
 229 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 230 {
 231         wait_for_completion_interruptible_timeout(&queue->cm_done,
 232                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 233         return queue->cm_error;
 234 }
 235
 236 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 237 {
 238         struct nvme_rdma_device *dev = queue->device;
 239         struct ib_qp_init_attr init_attr;
 240         int ret;
 241
 242         memset(&init_attr, 0, sizeof(init_attr));
 243         init_attr.event_handler = nvme_rdma_qp_event;
 244         /* +1 for drain */
 245         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 246         /* +1 for drain */
 247         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 248         init_attr.cap.max_recv_sge = 1;
 249         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 250         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 251         init_attr.qp_type = IB_QPT_RC;
 252         init_attr.send_cq = queue->ib_cq;
 253         init_attr.recv_cq = queue->ib_cq;
 254
 255         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 256
 257         queue->qp = queue->cm_id->qp;
 258         return ret;
 259 }
 260
 261 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 262 {
 263         struct nvme_rdma_ctrl *ctrl = data;
 264         struct nvme_rdma_device *dev = ctrl->device;
 265         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 266         int ret = 0;
 267
 268         if (WARN_ON_ONCE(!req->mr))
 269                 return 0;
 270
 271         ib_dereg_mr(req->mr);
 272
 273         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 274                         ctrl->max_fr_pages);
 275         if (IS_ERR(req->mr)) {
 276                 ret = PTR_ERR(req->mr);
 277                 req->mr = NULL;
 278                 goto out;
 279         }
 280
 281         req->mr->need_inval = false;
 282
 283 out:
 284         return ret;
 285 }
 286
 287 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 288                 struct request *rq, unsigned int hctx_idx)
 289 {
 290         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 291         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 292         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 293         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 294         struct nvme_rdma_device *dev = queue->device;
 295
 296         if (req->mr)
 297                 ib_dereg_mr(req->mr);
 298
 299         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 300                         DMA_TO_DEVICE);
 301 }
 302
 303 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 304                 struct request *rq, unsigned int hctx_idx,
 305                 unsigned int numa_node)
 306 {
 307         struct nvme_rdma_ctrl *ctrl = set->driver_data;
 308         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 309         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 310         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 311         struct nvme_rdma_device *dev = queue->device;
 312         struct ib_device *ibdev = dev->dev;
 313         int ret;
 314
 315         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 316                         DMA_TO_DEVICE);
 317         if (ret)
 318                 return ret;
 319
 320         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 321                         ctrl->max_fr_pages);
 322         if (IS_ERR(req->mr)) {
 323                 ret = PTR_ERR(req->mr);
 324                 goto out_free_qe;
 325         }
 326
 327         req->queue = queue;
 328
 329         return 0;
 330
 331 out_free_qe:
 332         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 333                         DMA_TO_DEVICE);
 334         return -ENOMEM;
 335 }
 336
 337 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 338                 unsigned int hctx_idx)
 339 {
 340         struct nvme_rdma_ctrl *ctrl = data;
 341         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 342
 343         BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
 344
 345         hctx->driver_data = queue;
 346         return 0;
 347 }
 348
 349 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 350                 unsigned int hctx_idx)
 351 {
 352         struct nvme_rdma_ctrl *ctrl = data;
 353         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 354
 355         BUG_ON(hctx_idx != 0);
 356
 357         hctx->driver_data = queue;
 358         return 0;
 359 }
 360
 361 static void nvme_rdma_free_dev(struct kref *ref)
 362 {
 363         struct nvme_rdma_device *ndev =
 364                 container_of(ref, struct nvme_rdma_device, ref);
 365
 366         mutex_lock(&device_list_mutex);
 367         list_del(&ndev->entry);
 368         mutex_unlock(&device_list_mutex);
 369
 370         ib_dealloc_pd(ndev->pd);
 371         kfree(ndev);
 372 }
 373
 374 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 375 {
 376         kref_put(&dev->ref, nvme_rdma_free_dev);
 377 }
 378
 379 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 380 {
 381         return kref_get_unless_zero(&dev->ref);
 382 }
 383
 384 static struct nvme_rdma_device *
 385 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 386 {
 387         struct nvme_rdma_device *ndev;
 388
 389         mutex_lock(&device_list_mutex);
 390         list_for_each_entry(ndev, &device_list, entry) {
 391                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 392                     nvme_rdma_dev_get(ndev))
 393                         goto out_unlock;
 394         }
 395
 396         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 397         if (!ndev)
 398                 goto out_err;
 399
 400         ndev->dev = cm_id->device;
 401         kref_init(&ndev->ref);
 402
 403         ndev->pd = ib_alloc_pd(ndev->dev,
 404                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 405         if (IS_ERR(ndev->pd))
 406                 goto out_free_dev;
 407
 408         if (!(ndev->dev->attrs.device_cap_flags &
 409               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 410                 dev_err(&ndev->dev->dev,
 411                         "Memory registrations not supported.\n");
 412                 goto out_free_pd;
 413         }
 414
 415         list_add(&ndev->entry, &device_list);
 416 out_unlock:
 417         mutex_unlock(&device_list_mutex);
 418         return ndev;
 419
 420 out_free_pd:
 421         ib_dealloc_pd(ndev->pd);
 422 out_free_dev:
 423         kfree(ndev);
 424 out_err:
 425         mutex_unlock(&device_list_mutex);
 426         return NULL;
 427 }
 428
 429 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 430 {
 431         struct nvme_rdma_device *dev = queue->device;
 432         struct ib_device *ibdev = dev->dev;
 433
 434         rdma_destroy_qp(queue->cm_id);
 435         ib_free_cq(queue->ib_cq);
 436
 437         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 438                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 439
 440         nvme_rdma_dev_put(dev);
 441 }
 442
 443 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 444 {
 445         struct ib_device *ibdev;
 446         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 447         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 448         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 449         int ret;
 450
 451         queue->device = nvme_rdma_find_get_device(queue->cm_id);
 452         if (!queue->device) {
 453                 dev_err(queue->cm_id->device->dev.parent,
 454                         "no client data found!\n");
 455                 return -ECONNREFUSED;
 456         }
 457         ibdev = queue->device->dev;
 458
 459         /*
 460          * Spread I/O queues completion vectors according their queue index.
 461          * Admin queues can always go on completion vector 0.
 462          */
 463         comp_vector = idx == 0 ? idx : idx - 1;
 464
 465         /* +1 for ib_stop_cq */
 466         queue->ib_cq = ib_alloc_cq(ibdev, queue,
 467                                 cq_factor * queue->queue_size + 1,
 468                                 comp_vector, IB_POLL_SOFTIRQ);
 469         if (IS_ERR(queue->ib_cq)) {
 470                 ret = PTR_ERR(queue->ib_cq);
 471                 goto out_put_dev;
 472         }
 473
 474         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 475         if (ret)
 476                 goto out_destroy_ib_cq;
 477
 478         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 479                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 480         if (!queue->rsp_ring) {
 481                 ret = -ENOMEM;
 482                 goto out_destroy_qp;
 483         }
 484
 485         return 0;
 486
 487 out_destroy_qp:
 488         rdma_destroy_qp(queue->cm_id);
 489 out_destroy_ib_cq:
 490         ib_free_cq(queue->ib_cq);
 491 out_put_dev:
 492         nvme_rdma_dev_put(queue->device);
 493         return ret;
 494 }
 495
 496 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
 497                 int idx, size_t queue_size)
 498 {
 499         struct nvme_rdma_queue *queue;
 500         struct sockaddr *src_addr = NULL;
 501         int ret;
 502
 503         queue = &ctrl->queues[idx];
 504         queue->ctrl = ctrl;
 505         init_completion(&queue->cm_done);
 506
 507         if (idx > 0)
 508                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 509         else
 510                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 511
 512         queue->queue_size = queue_size;
 513         atomic_set(&queue->sig_count, 0);
 514
 515         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 516                         RDMA_PS_TCP, IB_QPT_RC);
 517         if (IS_ERR(queue->cm_id)) {
 518                 dev_info(ctrl->ctrl.device,
 519                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 520                 return PTR_ERR(queue->cm_id);
 521         }
 522
 523         if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
 524                 src_addr = (struct sockaddr *)&ctrl->src_addr;
 525
 526         queue->cm_error = -ETIMEDOUT;
 527         ret = rdma_resolve_addr(queue->cm_id, src_addr,
 528                         (struct sockaddr *)&ctrl->addr,
 529                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 530         if (ret) {
 531                 dev_info(ctrl->ctrl.device,
 532                         "rdma_resolve_addr failed (%d).\n", ret);
 533                 goto out_destroy_cm_id;
 534         }
 535
 536         ret = nvme_rdma_wait_for_cm(queue);
 537         if (ret) {
 538                 dev_info(ctrl->ctrl.device,
 539                         "rdma connection establishment failed (%d)\n", ret);
 540                 goto out_destroy_cm_id;
 541         }
 542
 543         set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
 544
 545         return 0;
 546
 547 out_destroy_cm_id:
 548         rdma_destroy_id(queue->cm_id);
 549         return ret;
 550 }
 551
 552 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 553 {
 554         if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
 555                 return;
 556
 557         rdma_disconnect(queue->cm_id);
 558         ib_drain_qp(queue->qp);
 559 }
 560
 561 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 562 {
 563         if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
 564                 return;
 565
 566         if (nvme_rdma_queue_idx(queue) == 0) {
 567                 nvme_rdma_free_qe(queue->device->dev,
 568                         &queue->ctrl->async_event_sqe,
 569                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 570         }
 571
 572         nvme_rdma_destroy_queue_ib(queue);
 573         rdma_destroy_id(queue->cm_id);
 574 }
 575
 576 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 577 {
 578         int i;
 579
 580         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 581                 nvme_rdma_free_queue(&ctrl->queues[i]);
 582 }
 583
 584 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
 585 {
 586         int i;
 587
 588         for (i = 1; i < ctrl->ctrl.queue_count; i++)
 589                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 590 }
 591
 592 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
 593 {
 594         int ret;
 595
 596         if (idx)
 597                 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
 598         else
 599                 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 600
 601         if (!ret)
 602                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
 603         else
 604                 dev_info(ctrl->ctrl.device,
 605                         "failed to connect queue: %d ret=%d\n", idx, ret);
 606         return ret;
 607 }
 608
 609 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
 610 {
 611         int i, ret = 0;
 612
 613         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 614                 ret = nvme_rdma_start_queue(ctrl, i);
 615                 if (ret)
 616                         goto out_stop_queues;
 617         }
 618
 619         return 0;
 620
 621 out_stop_queues:
 622         for (i--; i >= 1; i--)
 623                 nvme_rdma_stop_queue(&ctrl->queues[i]);
 624         return ret;
 625 }
 626
 627 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
 628 {
 629         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 630         struct ib_device *ibdev = ctrl->device->dev;
 631         unsigned int nr_io_queues;
 632         int i, ret;
 633
 634         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 635
 636         /*
 637          * we map queues according to the device irq vectors for
 638          * optimal locality so we don't need more queues than
 639          * completion vectors.
 640          */
 641         nr_io_queues = min_t(unsigned int, nr_io_queues,
 642                                 ibdev->num_comp_vectors);
 643
 644         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 645         if (ret)
 646                 return ret;
 647
 648         ctrl->ctrl.queue_count = nr_io_queues + 1;
 649         if (ctrl->ctrl.queue_count < 2)
 650                 return 0;
 651
 652         dev_info(ctrl->ctrl.device,
 653                 "creating %d I/O queues.\n", nr_io_queues);
 654
 655         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
 656                 ret = nvme_rdma_alloc_queue(ctrl, i,
 657                                 ctrl->ctrl.sqsize + 1);
 658                 if (ret)
 659                         goto out_free_queues;
 660         }
 661
 662         return 0;
 663
 664 out_free_queues:
 665         for (i--; i >= 1; i--)
 666                 nvme_rdma_free_queue(&ctrl->queues[i]);
 667
 668         return ret;
 669 }
 670
 671 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
 672                 struct blk_mq_tag_set *set)
 673 {
 674         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 675
 676         blk_mq_free_tag_set(set);
 677         nvme_rdma_dev_put(ctrl->device);
 678 }
 679
 680 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 681                 bool admin)
 682 {
 683         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 684         struct blk_mq_tag_set *set;
 685         int ret;
 686
 687         if (admin) {
 688                 set = &ctrl->admin_tag_set;
 689                 memset(set, 0, sizeof(*set));
 690                 set->ops = &nvme_rdma_admin_mq_ops;
 691                 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
 692                 set->reserved_tags = 2; /* connect + keep-alive */
 693                 set->numa_node = NUMA_NO_NODE;
 694                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 695                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 696                 set->driver_data = ctrl;
 697                 set->nr_hw_queues = 1;
 698                 set->timeout = ADMIN_TIMEOUT;
 699                 set->flags = BLK_MQ_F_NO_SCHED;
 700         } else {
 701                 set = &ctrl->tag_set;
 702                 memset(set, 0, sizeof(*set));
 703                 set->ops = &nvme_rdma_mq_ops;
 704                 set->queue_depth = nctrl->opts->queue_size;
 705                 set->reserved_tags = 1; /* fabric connect */
 706                 set->numa_node = NUMA_NO_NODE;
 707                 set->flags = BLK_MQ_F_SHOULD_MERGE;
 708                 set->cmd_size = sizeof(struct nvme_rdma_request) +
 709                         SG_CHUNK_SIZE * sizeof(struct scatterlist);
 710                 set->driver_data = ctrl;
 711                 set->nr_hw_queues = nctrl->queue_count - 1;
 712                 set->timeout = NVME_IO_TIMEOUT;
 713         }
 714
 715         ret = blk_mq_alloc_tag_set(set);
 716         if (ret)
 717                 goto out;
 718
 719         /*
 720          * We need a reference on the device as long as the tag_set is alive,
 721          * as the MRs in the request structures need a valid ib_device.
 722          */
 723         ret = nvme_rdma_dev_get(ctrl->device);
 724         if (!ret) {
 725                 ret = -EINVAL;
 726                 goto out_free_tagset;
 727         }
 728
 729         return set;
 730
 731 out_free_tagset:
 732         blk_mq_free_tag_set(set);
 733 out:
 734         return ERR_PTR(ret);
 735 }
 736
 737 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
 738                 bool remove)
 739 {
 740         nvme_rdma_stop_queue(&ctrl->queues[0]);
 741         if (remove) {
 742                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 743                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 744         }
 745         nvme_rdma_free_queue(&ctrl->queues[0]);
 746 }
 747
 748 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 749                 bool new)
 750 {
 751         int error;
 752
 753         error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
 754         if (error)
 755                 return error;
 756
 757         ctrl->device = ctrl->queues[0].device;
 758
 759         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
 760                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
 761
 762         if (new) {
 763                 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 764                 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
 765                         error = PTR_ERR(ctrl->ctrl.admin_tagset);
 766                         goto out_free_queue;
 767                 }
 768
 769                 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
 770                 if (IS_ERR(ctrl->ctrl.admin_q)) {
 771                         error = PTR_ERR(ctrl->ctrl.admin_q);
 772                         goto out_free_tagset;
 773                 }
 774         } else {
 775                 error = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 776                 if (error)
 777                         goto out_free_queue;
 778         }
 779
 780         error = nvme_rdma_start_queue(ctrl, 0);
 781         if (error)
 782                 goto out_cleanup_queue;
 783
 784         error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
 785                         &ctrl->ctrl.cap);
 786         if (error) {
 787                 dev_err(ctrl->ctrl.device,
 788                         "prop_get NVME_REG_CAP failed\n");
 789                 goto out_cleanup_queue;
 790         }
 791
 792         ctrl->ctrl.sqsize =
 793                 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
 794
 795         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
 796         if (error)
 797                 goto out_cleanup_queue;
 798
 799         ctrl->ctrl.max_hw_sectors =
 800                 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
 801
 802         error = nvme_init_identify(&ctrl->ctrl);
 803         if (error)
 804                 goto out_cleanup_queue;
 805
 806         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
 807                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
 808                         DMA_TO_DEVICE);
 809         if (error)
 810                 goto out_cleanup_queue;
 811
 812         return 0;
 813
 814 out_cleanup_queue:
 815         if (new)
 816                 blk_cleanup_queue(ctrl->ctrl.admin_q);
 817 out_free_tagset:
 818         if (new)
 819                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
 820 out_free_queue:
 821         nvme_rdma_free_queue(&ctrl->queues[0]);
 822         return error;
 823 }
 824
 825 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
 826                 bool remove)
 827 {
 828         nvme_rdma_stop_io_queues(ctrl);
 829         if (remove) {
 830                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 831                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 832         }
 833         nvme_rdma_free_io_queues(ctrl);
 834 }
 835
 836 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 837 {
 838         int ret;
 839
 840         ret = nvme_rdma_alloc_io_queues(ctrl);
 841         if (ret)
 842                 return ret;
 843
 844         if (new) {
 845                 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
 846                 if (IS_ERR(ctrl->ctrl.tagset)) {
 847                         ret = PTR_ERR(ctrl->ctrl.tagset);
 848                         goto out_free_io_queues;
 849                 }
 850
 851                 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
 852                 if (IS_ERR(ctrl->ctrl.connect_q)) {
 853                         ret = PTR_ERR(ctrl->ctrl.connect_q);
 854                         goto out_free_tag_set;
 855                 }
 856         } else {
 857                 ret = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 858                 if (ret)
 859                         goto out_free_io_queues;
 860
 861                 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 862                         ctrl->ctrl.queue_count - 1);
 863         }
 864
 865         ret = nvme_rdma_start_io_queues(ctrl);
 866         if (ret)
 867                 goto out_cleanup_connect_q;
 868
 869         return 0;
 870
 871 out_cleanup_connect_q:
 872         if (new)
 873                 blk_cleanup_queue(ctrl->ctrl.connect_q);
 874 out_free_tag_set:
 875         if (new)
 876                 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
 877 out_free_io_queues:
 878         nvme_rdma_free_io_queues(ctrl);
 879         return ret;
 880 }
 881
 882 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 883 {
 884         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 885
 886         if (list_empty(&ctrl->list))
 887                 goto free_ctrl;
 888
 889         mutex_lock(&nvme_rdma_ctrl_mutex);
 890         list_del(&ctrl->list);
 891         mutex_unlock(&nvme_rdma_ctrl_mutex);
 892
 893         kfree(ctrl->queues);
 894         nvmf_free_options(nctrl->opts);
 895 free_ctrl:
 896         kfree(ctrl);
 897 }
 898
 899 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
 900 {
 901         /* If we are resetting/deleting then do nothing */
 902         if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
 903                 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
 904                         ctrl->ctrl.state == NVME_CTRL_LIVE);
 905                 return;
 906         }
 907
 908         if (nvmf_should_reconnect(&ctrl->ctrl)) {
 909                 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
 910                         ctrl->ctrl.opts->reconnect_delay);
 911                 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
 912                                 ctrl->ctrl.opts->reconnect_delay * HZ);
 913         } else {
 914                 dev_info(ctrl->ctrl.device, "Removing controller...\n");
 915                 nvme_delete_ctrl(&ctrl->ctrl);
 916         }
 917 }
 918
 919 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 920 {
 921         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 922                         struct nvme_rdma_ctrl, reconnect_work);
 923         bool changed;
 924         int ret;
 925
 926         ++ctrl->ctrl.nr_reconnects;
 927
 928         ret = nvme_rdma_configure_admin_queue(ctrl, false);
 929         if (ret)
 930                 goto requeue;
 931
 932         if (ctrl->ctrl.queue_count > 1) {
 933                 ret = nvme_rdma_configure_io_queues(ctrl, false);
 934                 if (ret)
 935                         goto destroy_admin;
 936         }
 937
 938         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 939         if (!changed) {
 940                 /* state change failure is ok if we're in DELETING state */
 941                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
 942                 return;
 943         }
 944
 945         nvme_start_ctrl(&ctrl->ctrl);
 946
 947         dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
 948                         ctrl->ctrl.nr_reconnects);
 949
 950         ctrl->ctrl.nr_reconnects = 0;
 951
 952         return;
 953
 954 destroy_admin:
 955         nvme_rdma_destroy_admin_queue(ctrl, false);
 956 requeue:
 957         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
 958                         ctrl->ctrl.nr_reconnects);
 959         nvme_rdma_reconnect_or_remove(ctrl);
 960 }
 961
 962 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 963 {
 964         struct nvme_rdma_ctrl *ctrl = container_of(work,
 965                         struct nvme_rdma_ctrl, err_work);
 966
 967         nvme_stop_keep_alive(&ctrl->ctrl);
 968
 969         if (ctrl->ctrl.queue_count > 1) {
 970                 nvme_stop_queues(&ctrl->ctrl);
 971                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 972                                         nvme_cancel_request, &ctrl->ctrl);
 973                 nvme_rdma_destroy_io_queues(ctrl, false);
 974         }
 975
 976         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
 977         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 978                                 nvme_cancel_request, &ctrl->ctrl);
 979         nvme_rdma_destroy_admin_queue(ctrl, false);
 980
 981         /*
 982          * queues are not a live anymore, so restart the queues to fail fast
 983          * new IO
 984          */
 985         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
 986         nvme_start_queues(&ctrl->ctrl);
 987
 988         nvme_rdma_reconnect_or_remove(ctrl);
 989 }
 990
 991 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 992 {
 993         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 994                 return;
 995
 996         queue_work(nvme_wq, &ctrl->err_work);
 997 }
 998
 999 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1000                 const char *op)
1001 {
1002         struct nvme_rdma_queue *queue = cq->cq_context;
1003         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1004
1005         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1006                 dev_info(ctrl->ctrl.device,
1007                              "%s for CQE 0x%p failed with status %s (%d)\n",
1008                              op, wc->wr_cqe,
1009                              ib_wc_status_msg(wc->status), wc->status);
1010         nvme_rdma_error_recovery(ctrl);
1011 }
1012
1013 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1014 {
1015         if (unlikely(wc->status != IB_WC_SUCCESS))
1016                 nvme_rdma_wr_error(cq, wc, "MEMREG");
1017 }
1018
1019 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1020 {
1021         if (unlikely(wc->status != IB_WC_SUCCESS))
1022                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1023 }
1024
1025 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1026                 struct nvme_rdma_request *req)
1027 {
1028         struct ib_send_wr *bad_wr;
1029         struct ib_send_wr wr = {
1030                 .opcode             = IB_WR_LOCAL_INV,
1031                 .next               = NULL,
1032                 .num_sge            = 0,
1033                 .send_flags         = 0,
1034                 .ex.invalidate_rkey = req->mr->rkey,
1035         };
1036
1037         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1038         wr.wr_cqe = &req->reg_cqe;
1039
1040         return ib_post_send(queue->qp, &wr, &bad_wr);
1041 }
1042
1043 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1044                 struct request *rq)
1045 {
1046         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1047         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1048         struct nvme_rdma_device *dev = queue->device;
1049         struct ib_device *ibdev = dev->dev;
1050         int res;
1051
1052         if (!blk_rq_bytes(rq))
1053                 return;
1054
1055         if (req->mr->need_inval && test_bit(NVME_RDMA_Q_LIVE, &req->queue->flags)) {
1056                 res = nvme_rdma_inv_rkey(queue, req);
1057                 if (unlikely(res < 0)) {
1058                         dev_err(ctrl->ctrl.device,
1059                                 "Queueing INV WR for rkey %#x failed (%d)\n",
1060                                 req->mr->rkey, res);
1061                         nvme_rdma_error_recovery(queue->ctrl);
1062                 }
1063         }
1064
1065         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1066                         req->nents, rq_data_dir(rq) ==
1067                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1068
1069         nvme_cleanup_cmd(rq);
1070         sg_free_table_chained(&req->sg_table, true);
1071 }
1072
1073 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1074 {
1075         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1076
1077         sg->addr = 0;
1078         put_unaligned_le24(0, sg->length);
1079         put_unaligned_le32(0, sg->key);
1080         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1081         return 0;
1082 }
1083
1084 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1085                 struct nvme_rdma_request *req, struct nvme_command *c)
1086 {
1087         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1088
1089         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1090         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1091         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1092
1093         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1094         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1095         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1096
1097         req->inline_data = true;
1098         req->num_sge++;
1099         return 0;
1100 }
1101
1102 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1103                 struct nvme_rdma_request *req, struct nvme_command *c)
1104 {
1105         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1106
1107         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1108         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1109         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1110         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1111         return 0;
1112 }
1113
1114 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1115                 struct nvme_rdma_request *req, struct nvme_command *c,
1116                 int count)
1117 {
1118         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1119         int nr;
1120
1121         /*
1122          * Align the MR to a 4K page size to match the ctrl page size and
1123          * the block virtual boundary.
1124          */
1125         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1126         if (unlikely(nr < count)) {
1127                 if (nr < 0)
1128                         return nr;
1129                 return -EINVAL;
1130         }
1131
1132         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1133
1134         req->reg_cqe.done = nvme_rdma_memreg_done;
1135         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1136         req->reg_wr.wr.opcode = IB_WR_REG_MR;
1137         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1138         req->reg_wr.wr.num_sge = 0;
1139         req->reg_wr.mr = req->mr;
1140         req->reg_wr.key = req->mr->rkey;
1141         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1142                              IB_ACCESS_REMOTE_READ |
1143                              IB_ACCESS_REMOTE_WRITE;
1144
1145         req->mr->need_inval = true;
1146
1147         sg->addr = cpu_to_le64(req->mr->iova);
1148         put_unaligned_le24(req->mr->length, sg->length);
1149         put_unaligned_le32(req->mr->rkey, sg->key);
1150         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1151                         NVME_SGL_FMT_INVALIDATE;
1152
1153         return 0;
1154 }
1155
1156 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1157                 struct request *rq, struct nvme_command *c)
1158 {
1159         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1160         struct nvme_rdma_device *dev = queue->device;
1161         struct ib_device *ibdev = dev->dev;
1162         int count, ret;
1163
1164         req->num_sge = 1;
1165         req->inline_data = false;
1166         req->mr->need_inval = false;
1167
1168         c->common.flags |= NVME_CMD_SGL_METABUF;
1169
1170         if (!blk_rq_bytes(rq))
1171                 return nvme_rdma_set_sg_null(c);
1172
1173         req->sg_table.sgl = req->first_sgl;
1174         ret = sg_alloc_table_chained(&req->sg_table,
1175                         blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1176         if (ret)
1177                 return -ENOMEM;
1178
1179         req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1180
1181         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1182                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1183         if (unlikely(count <= 0)) {
1184                 sg_free_table_chained(&req->sg_table, true);
1185                 return -EIO;
1186         }
1187
1188         if (count == 1) {
1189                 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1190                     blk_rq_payload_bytes(rq) <=
1191                                 nvme_rdma_inline_data_size(queue))
1192                         return nvme_rdma_map_sg_inline(queue, req, c);
1193
1194                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1195                         return nvme_rdma_map_sg_single(queue, req, c);
1196         }
1197
1198         return nvme_rdma_map_sg_fr(queue, req, c, count);
1199 }
1200
1201 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1202 {
1203         if (unlikely(wc->status != IB_WC_SUCCESS))
1204                 nvme_rdma_wr_error(cq, wc, "SEND");
1205 }
1206
1207 /*
1208  * We want to signal completion at least every queue depth/2.  This returns the
1209  * largest power of two that is not above half of (queue size + 1) to optimize
1210  * (avoid divisions).
1211  */
1212 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1213 {
1214         int limit = 1 << ilog2((queue->queue_size + 1) / 2);
1215
1216         return (atomic_inc_return(&queue->sig_count) & (limit - 1)) == 0;
1217 }
1218
1219 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1220                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1221                 struct ib_send_wr *first, bool flush)
1222 {
1223         struct ib_send_wr wr, *bad_wr;
1224         int ret;
1225
1226         sge->addr   = qe->dma;
1227         sge->length = sizeof(struct nvme_command),
1228         sge->lkey   = queue->device->pd->local_dma_lkey;
1229
1230         qe->cqe.done = nvme_rdma_send_done;
1231
1232         wr.next       = NULL;
1233         wr.wr_cqe     = &qe->cqe;
1234         wr.sg_list    = sge;
1235         wr.num_sge    = num_sge;
1236         wr.opcode     = IB_WR_SEND;
1237         wr.send_flags = 0;
1238
1239         /*
1240          * Unsignalled send completions are another giant desaster in the
1241          * IB Verbs spec:  If we don't regularly post signalled sends
1242          * the send queue will fill up and only a QP reset will rescue us.
1243          * Would have been way to obvious to handle this in hardware or
1244          * at least the RDMA stack..
1245          *
1246          * Always signal the flushes. The magic request used for the flush
1247          * sequencer is not allocated in our driver's tagset and it's
1248          * triggered to be freed by blk_cleanup_queue(). So we need to
1249          * always mark it as signaled to ensure that the "wr_cqe", which is
1250          * embedded in request's payload, is not freed when __ib_process_cq()
1251          * calls wr_cqe->done().
1252          */
1253         if (nvme_rdma_queue_sig_limit(queue) || flush)
1254                 wr.send_flags |= IB_SEND_SIGNALED;
1255
1256         if (first)
1257                 first->next = &wr;
1258         else
1259                 first = &wr;
1260
1261         ret = ib_post_send(queue->qp, first, &bad_wr);
1262         if (unlikely(ret)) {
1263                 dev_err(queue->ctrl->ctrl.device,
1264                              "%s failed with error code %d\n", __func__, ret);
1265         }
1266         return ret;
1267 }
1268
1269 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1270                 struct nvme_rdma_qe *qe)
1271 {
1272         struct ib_recv_wr wr, *bad_wr;
1273         struct ib_sge list;
1274         int ret;
1275
1276         list.addr   = qe->dma;
1277         list.length = sizeof(struct nvme_completion);
1278         list.lkey   = queue->device->pd->local_dma_lkey;
1279
1280         qe->cqe.done = nvme_rdma_recv_done;
1281
1282         wr.next     = NULL;
1283         wr.wr_cqe   = &qe->cqe;
1284         wr.sg_list  = &list;
1285         wr.num_sge  = 1;
1286
1287         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1288         if (unlikely(ret)) {
1289                 dev_err(queue->ctrl->ctrl.device,
1290                         "%s failed with error code %d\n", __func__, ret);
1291         }
1292         return ret;
1293 }
1294
1295 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1296 {
1297         u32 queue_idx = nvme_rdma_queue_idx(queue);
1298
1299         if (queue_idx == 0)
1300                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1301         return queue->ctrl->tag_set.tags[queue_idx - 1];
1302 }
1303
1304 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1305 {
1306         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1307         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1308         struct ib_device *dev = queue->device->dev;
1309         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1310         struct nvme_command *cmd = sqe->data;
1311         struct ib_sge sge;
1312         int ret;
1313
1314         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1315
1316         memset(cmd, 0, sizeof(*cmd));
1317         cmd->common.opcode = nvme_admin_async_event;
1318         cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1319         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1320         nvme_rdma_set_sg_null(cmd);
1321
1322         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1323                         DMA_TO_DEVICE);
1324
1325         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1326         WARN_ON_ONCE(ret);
1327 }
1328
1329 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1330                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1331 {
1332         struct request *rq;
1333         struct nvme_rdma_request *req;
1334         int ret = 0;
1335
1336         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1337         if (!rq) {
1338                 dev_err(queue->ctrl->ctrl.device,
1339                         "tag 0x%x on QP %#x not found\n",
1340                         cqe->command_id, queue->qp->qp_num);
1341                 nvme_rdma_error_recovery(queue->ctrl);
1342                 return ret;
1343         }
1344         req = blk_mq_rq_to_pdu(rq);
1345
1346         if (rq->tag == tag)
1347                 ret = 1;
1348
1349         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1350             wc->ex.invalidate_rkey == req->mr->rkey)
1351                 req->mr->need_inval = false;
1352
1353         nvme_end_request(rq, cqe->status, cqe->result);
1354         return ret;
1355 }
1356
1357 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1358 {
1359         struct nvme_rdma_qe *qe =
1360                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1361         struct nvme_rdma_queue *queue = cq->cq_context;
1362         struct ib_device *ibdev = queue->device->dev;
1363         struct nvme_completion *cqe = qe->data;
1364         const size_t len = sizeof(struct nvme_completion);
1365         int ret = 0;
1366
1367         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1368                 nvme_rdma_wr_error(cq, wc, "RECV");
1369                 return 0;
1370         }
1371
1372         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1373         /*
1374          * AEN requests are special as they don't time out and can
1375          * survive any kind of queue freeze and often don't respond to
1376          * aborts.  We don't even bother to allocate a struct request
1377          * for them but rather special case them here.
1378          */
1379         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1380                         cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1381                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1382                                 &cqe->result);
1383         else
1384                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1385         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1386
1387         nvme_rdma_post_recv(queue, qe);
1388         return ret;
1389 }
1390
1391 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1392 {
1393         __nvme_rdma_recv_done(cq, wc, -1);
1394 }
1395
1396 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1397 {
1398         int ret, i;
1399
1400         for (i = 0; i < queue->queue_size; i++) {
1401                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1402                 if (ret)
1403                         goto out_destroy_queue_ib;
1404         }
1405
1406         return 0;
1407
1408 out_destroy_queue_ib:
1409         nvme_rdma_destroy_queue_ib(queue);
1410         return ret;
1411 }
1412
1413 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1414                 struct rdma_cm_event *ev)
1415 {
1416         struct rdma_cm_id *cm_id = queue->cm_id;
1417         int status = ev->status;
1418         const char *rej_msg;
1419         const struct nvme_rdma_cm_rej *rej_data;
1420         u8 rej_data_len;
1421
1422         rej_msg = rdma_reject_msg(cm_id, status);
1423         rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1424
1425         if (rej_data && rej_data_len >= sizeof(u16)) {
1426                 u16 sts = le16_to_cpu(rej_data->sts);
1427
1428                 dev_err(queue->ctrl->ctrl.device,
1429                       "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1430                       status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1431         } else {
1432                 dev_err(queue->ctrl->ctrl.device,
1433                         "Connect rejected: status %d (%s).\n", status, rej_msg);
1434         }
1435
1436         return -ECONNRESET;
1437 }
1438
1439 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1440 {
1441         int ret;
1442
1443         ret = nvme_rdma_create_queue_ib(queue);
1444         if (ret)
1445                 return ret;
1446
1447         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1448         if (ret) {
1449                 dev_err(queue->ctrl->ctrl.device,
1450                         "rdma_resolve_route failed (%d).\n",
1451                         queue->cm_error);
1452                 goto out_destroy_queue;
1453         }
1454
1455         return 0;
1456
1457 out_destroy_queue:
1458         nvme_rdma_destroy_queue_ib(queue);
1459         return ret;
1460 }
1461
1462 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1463 {
1464         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1465         struct rdma_conn_param param = { };
1466         struct nvme_rdma_cm_req priv = { };
1467         int ret;
1468
1469         param.qp_num = queue->qp->qp_num;
1470         param.flow_control = 1;
1471
1472         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1473         /* maximum retry count */
1474         param.retry_count = 7;
1475         param.rnr_retry_count = 7;
1476         param.private_data = &priv;
1477         param.private_data_len = sizeof(priv);
1478
1479         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1480         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1481         /*
1482          * set the admin queue depth to the minimum size
1483          * specified by the Fabrics standard.
1484          */
1485         if (priv.qid == 0) {
1486                 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1487                 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1488         } else {
1489                 /*
1490                  * current interpretation of the fabrics spec
1491                  * is at minimum you make hrqsize sqsize+1, or a
1492                  * 1's based representation of sqsize.
1493                  */
1494                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1495                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1496         }
1497
1498         ret = rdma_connect(queue->cm_id, &param);
1499         if (ret) {
1500                 dev_err(ctrl->ctrl.device,
1501                         "rdma_connect failed (%d).\n", ret);
1502                 goto out_destroy_queue_ib;
1503         }
1504
1505         return 0;
1506
1507 out_destroy_queue_ib:
1508         nvme_rdma_destroy_queue_ib(queue);
1509         return ret;
1510 }
1511
1512 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1513                 struct rdma_cm_event *ev)
1514 {
1515         struct nvme_rdma_queue *queue = cm_id->context;
1516         int cm_error = 0;
1517
1518         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1519                 rdma_event_msg(ev->event), ev->event,
1520                 ev->status, cm_id);
1521
1522         switch (ev->event) {
1523         case RDMA_CM_EVENT_ADDR_RESOLVED:
1524                 cm_error = nvme_rdma_addr_resolved(queue);
1525                 break;
1526         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1527                 cm_error = nvme_rdma_route_resolved(queue);
1528                 break;
1529         case RDMA_CM_EVENT_ESTABLISHED:
1530                 queue->cm_error = nvme_rdma_conn_established(queue);
1531                 /* complete cm_done regardless of success/failure */
1532                 complete(&queue->cm_done);
1533                 return 0;
1534         case RDMA_CM_EVENT_REJECTED:
1535                 nvme_rdma_destroy_queue_ib(queue);
1536                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1537                 break;
1538         case RDMA_CM_EVENT_ROUTE_ERROR:
1539         case RDMA_CM_EVENT_CONNECT_ERROR:
1540         case RDMA_CM_EVENT_UNREACHABLE:
1541                 nvme_rdma_destroy_queue_ib(queue);
1542         case RDMA_CM_EVENT_ADDR_ERROR:
1543                 dev_dbg(queue->ctrl->ctrl.device,
1544                         "CM error event %d\n", ev->event);
1545                 cm_error = -ECONNRESET;
1546                 break;
1547         case RDMA_CM_EVENT_DISCONNECTED:
1548         case RDMA_CM_EVENT_ADDR_CHANGE:
1549         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1550                 dev_dbg(queue->ctrl->ctrl.device,
1551                         "disconnect received - connection closed\n");
1552                 nvme_rdma_error_recovery(queue->ctrl);
1553                 break;
1554         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1555                 /* device removal is handled via the ib_client API */
1556                 break;
1557         default:
1558                 dev_err(queue->ctrl->ctrl.device,
1559                         "Unexpected RDMA CM event (%d)\n", ev->event);
1560                 nvme_rdma_error_recovery(queue->ctrl);
1561                 break;
1562         }
1563
1564         if (cm_error) {
1565                 queue->cm_error = cm_error;
1566                 complete(&queue->cm_done);
1567         }
1568
1569         return 0;
1570 }
1571
1572 static enum blk_eh_timer_return
1573 nvme_rdma_timeout(struct request *rq, bool reserved)
1574 {
1575         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1576
1577         dev_warn(req->queue->ctrl->ctrl.device,
1578                  "I/O %d QID %d timeout, reset controller\n",
1579                  rq->tag, nvme_rdma_queue_idx(req->queue));
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                          * fail fast to give upper layers a chance to failover.
1606                          * deleting state means that the ctrl will never accept
1607                          * commands again, fail it permanently.
1608                          */
1609                         if (queue->ctrl->ctrl.state == NVME_CTRL_RECONNECTING ||
1610                             queue->ctrl->ctrl.state == NVME_CTRL_DELETING) {
1611                                 nvme_req(rq)->status = NVME_SC_ABORT_REQ;
1612                                 return BLK_STS_IOERR;
1613                         }
1614                         return BLK_STS_RESOURCE; /* try again later */
1615                 }
1616         }
1617
1618         return 0;
1619 }
1620
1621 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1622                 const struct blk_mq_queue_data *bd)
1623 {
1624         struct nvme_ns *ns = hctx->queue->queuedata;
1625         struct nvme_rdma_queue *queue = hctx->driver_data;
1626         struct request *rq = bd->rq;
1627         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1628         struct nvme_rdma_qe *sqe = &req->sqe;
1629         struct nvme_command *c = sqe->data;
1630         bool flush = false;
1631         struct ib_device *dev;
1632         blk_status_t ret;
1633         int err;
1634
1635         WARN_ON_ONCE(rq->tag < 0);
1636
1637         ret = nvme_rdma_queue_is_ready(queue, rq);
1638         if (unlikely(ret))
1639                 return ret;
1640
1641         dev = queue->device->dev;
1642         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1643                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1644
1645         ret = nvme_setup_cmd(ns, rq, c);
1646         if (ret)
1647                 return ret;
1648
1649         blk_mq_start_request(rq);
1650
1651         err = nvme_rdma_map_data(queue, rq, c);
1652         if (unlikely(err < 0)) {
1653                 dev_err(queue->ctrl->ctrl.device,
1654                              "Failed to map data (%d)\n", err);
1655                 nvme_cleanup_cmd(rq);
1656                 goto err;
1657         }
1658
1659         ib_dma_sync_single_for_device(dev, sqe->dma,
1660                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1661
1662         if (req_op(rq) == REQ_OP_FLUSH)
1663                 flush = true;
1664         err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1665                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1666         if (unlikely(err)) {
1667                 nvme_rdma_unmap_data(queue, rq);
1668                 goto err;
1669         }
1670
1671         return BLK_STS_OK;
1672 err:
1673         if (err == -ENOMEM || err == -EAGAIN)
1674                 return BLK_STS_RESOURCE;
1675         return BLK_STS_IOERR;
1676 }
1677
1678 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1679 {
1680         struct nvme_rdma_queue *queue = hctx->driver_data;
1681         struct ib_cq *cq = queue->ib_cq;
1682         struct ib_wc wc;
1683         int found = 0;
1684
1685         while (ib_poll_cq(cq, 1, &wc) > 0) {
1686                 struct ib_cqe *cqe = wc.wr_cqe;
1687
1688                 if (cqe) {
1689                         if (cqe->done == nvme_rdma_recv_done)
1690                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1691                         else
1692                                 cqe->done(cq, &wc);
1693                 }
1694         }
1695
1696         return found;
1697 }
1698
1699 static void nvme_rdma_complete_rq(struct request *rq)
1700 {
1701         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1702
1703         nvme_rdma_unmap_data(req->queue, rq);
1704         nvme_complete_rq(rq);
1705 }
1706
1707 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1708 {
1709         struct nvme_rdma_ctrl *ctrl = set->driver_data;
1710
1711         return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1712 }
1713
1714 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1715         .queue_rq       = nvme_rdma_queue_rq,
1716         .complete       = nvme_rdma_complete_rq,
1717         .init_request   = nvme_rdma_init_request,
1718         .exit_request   = nvme_rdma_exit_request,
1719         .init_hctx      = nvme_rdma_init_hctx,
1720         .poll           = nvme_rdma_poll,
1721         .timeout        = nvme_rdma_timeout,
1722         .map_queues     = nvme_rdma_map_queues,
1723 };
1724
1725 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1726         .queue_rq       = nvme_rdma_queue_rq,
1727         .complete       = nvme_rdma_complete_rq,
1728         .init_request   = nvme_rdma_init_request,
1729         .exit_request   = nvme_rdma_exit_request,
1730         .init_hctx      = nvme_rdma_init_admin_hctx,
1731         .timeout        = nvme_rdma_timeout,
1732 };
1733
1734 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1735 {
1736         cancel_work_sync(&ctrl->err_work);
1737         cancel_delayed_work_sync(&ctrl->reconnect_work);
1738
1739         if (ctrl->ctrl.queue_count > 1) {
1740                 nvme_stop_queues(&ctrl->ctrl);
1741                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1742                                         nvme_cancel_request, &ctrl->ctrl);
1743                 nvme_rdma_destroy_io_queues(ctrl, shutdown);
1744         }
1745
1746         if (shutdown)
1747                 nvme_shutdown_ctrl(&ctrl->ctrl);
1748         else
1749                 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1750
1751         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1752         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1753                                 nvme_cancel_request, &ctrl->ctrl);
1754         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1755         nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1756 }
1757
1758 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1759 {
1760         nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1761 }
1762
1763 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1764 {
1765         struct nvme_rdma_ctrl *ctrl =
1766                 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1767         int ret;
1768         bool changed;
1769
1770         nvme_stop_ctrl(&ctrl->ctrl);
1771         nvme_rdma_shutdown_ctrl(ctrl, false);
1772
1773         ret = nvme_rdma_configure_admin_queue(ctrl, false);
1774         if (ret)
1775                 goto out_fail;
1776
1777         if (ctrl->ctrl.queue_count > 1) {
1778                 ret = nvme_rdma_configure_io_queues(ctrl, false);
1779                 if (ret)
1780                         goto out_fail;
1781         }
1782
1783         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1784         if (!changed) {
1785                 /* state change failure is ok if we're in DELETING state */
1786                 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1787                 return;
1788         }
1789
1790         nvme_start_ctrl(&ctrl->ctrl);
1791
1792         return;
1793
1794 out_fail:
1795         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1796         nvme_remove_namespaces(&ctrl->ctrl);
1797         nvme_rdma_shutdown_ctrl(ctrl, true);
1798         nvme_uninit_ctrl(&ctrl->ctrl);
1799         nvme_put_ctrl(&ctrl->ctrl);
1800 }
1801
1802 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1803         .name                   = "rdma",
1804         .module                 = THIS_MODULE,
1805         .flags                  = NVME_F_FABRICS,
1806         .reg_read32             = nvmf_reg_read32,
1807         .reg_read64             = nvmf_reg_read64,
1808         .reg_write32            = nvmf_reg_write32,
1809         .free_ctrl              = nvme_rdma_free_ctrl,
1810         .submit_async_event     = nvme_rdma_submit_async_event,
1811         .delete_ctrl            = nvme_rdma_delete_ctrl,
1812         .get_address            = nvmf_get_address,
1813         .reinit_request         = nvme_rdma_reinit_request,
1814 };
1815
1816 static inline bool
1817 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1818         struct nvmf_ctrl_options *opts)
1819 {
1820         char *stdport = __stringify(NVME_RDMA_IP_PORT);
1821
1822
1823         if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1824             strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1825                 return false;
1826
1827         if (opts->mask & NVMF_OPT_TRSVCID &&
1828             ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1829                 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1830                         return false;
1831         } else if (opts->mask & NVMF_OPT_TRSVCID) {
1832                 if (strcmp(opts->trsvcid, stdport))
1833                         return false;
1834         } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1835                 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1836                         return false;
1837         }
1838         /* else, it's a match as both have stdport. Fall to next checks */
1839
1840         /*
1841          * checking the local address is rough. In most cases, one
1842          * is not specified and the host port is selected by the stack.
1843          *
1844          * Assume no match if:
1845          *  local address is specified and address is not the same
1846          *  local address is not specified but remote is, or vice versa
1847          *    (admin using specific host_traddr when it matters).
1848          */
1849         if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1850             ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1851                 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1852                         return false;
1853         } else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1854                    ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1855                 return false;
1856         /*
1857          * if neither controller had an host port specified, assume it's
1858          * a match as everything else matched.
1859          */
1860
1861         return true;
1862 }
1863
1864 /*
1865  * Fails a connection request if it matches an existing controller
1866  * (association) with the same tuple:
1867  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1868  *
1869  * if local address is not specified in the request, it will match an
1870  * existing controller with all the other parameters the same and no
1871  * local port address specified as well.
1872  *
1873  * The ports don't need to be compared as they are intrinsically
1874  * already matched by the port pointers supplied.
1875  */
1876 static bool
1877 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1878 {
1879         struct nvme_rdma_ctrl *ctrl;
1880         bool found = false;
1881
1882         mutex_lock(&nvme_rdma_ctrl_mutex);
1883         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1884                 found = __nvme_rdma_options_match(ctrl, opts);
1885                 if (found)
1886                         break;
1887         }
1888         mutex_unlock(&nvme_rdma_ctrl_mutex);
1889
1890         return found;
1891 }
1892
1893 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1894                 struct nvmf_ctrl_options *opts)
1895 {
1896         struct nvme_rdma_ctrl *ctrl;
1897         int ret;
1898         bool changed;
1899         char *port;
1900
1901         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1902         if (!ctrl)
1903                 return ERR_PTR(-ENOMEM);
1904         ctrl->ctrl.opts = opts;
1905         INIT_LIST_HEAD(&ctrl->list);
1906
1907         if (opts->mask & NVMF_OPT_TRSVCID)
1908                 port = opts->trsvcid;
1909         else
1910                 port = __stringify(NVME_RDMA_IP_PORT);
1911
1912         ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1913                         opts->traddr, port, &ctrl->addr);
1914         if (ret) {
1915                 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1916                 goto out_free_ctrl;
1917         }
1918
1919         if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1920                 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1921                         opts->host_traddr, NULL, &ctrl->src_addr);
1922                 if (ret) {
1923                         pr_err("malformed src address passed: %s\n",
1924                                opts->host_traddr);
1925                         goto out_free_ctrl;
1926                 }
1927         }
1928
1929         if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1930                 ret = -EALREADY;
1931                 goto out_free_ctrl;
1932         }
1933
1934         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1935                                 0 /* no quirks, we're perfect! */);
1936         if (ret)
1937                 goto out_free_ctrl;
1938
1939         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1940                         nvme_rdma_reconnect_ctrl_work);
1941         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1942         INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1943
1944         ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1945         ctrl->ctrl.sqsize = opts->queue_size - 1;
1946         ctrl->ctrl.kato = opts->kato;
1947
1948         ret = -ENOMEM;
1949         ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1950                                 GFP_KERNEL);
1951         if (!ctrl->queues)
1952                 goto out_uninit_ctrl;
1953
1954         ret = nvme_rdma_configure_admin_queue(ctrl, true);
1955         if (ret)
1956                 goto out_kfree_queues;
1957
1958         /* sanity check icdoff */
1959         if (ctrl->ctrl.icdoff) {
1960                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1961                 ret = -EINVAL;
1962                 goto out_remove_admin_queue;
1963         }
1964
1965         /* sanity check keyed sgls */
1966         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1967                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1968                 ret = -EINVAL;
1969                 goto out_remove_admin_queue;
1970         }
1971
1972         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1973                 /* warn if maxcmd is lower than queue_size */
1974                 dev_warn(ctrl->ctrl.device,
1975                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1976                         opts->queue_size, ctrl->ctrl.maxcmd);
1977                 opts->queue_size = ctrl->ctrl.maxcmd;
1978         }
1979
1980         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1981                 /* warn if sqsize is lower than queue_size */
1982                 dev_warn(ctrl->ctrl.device,
1983                         "queue_size %zu > ctrl sqsize %u, clamping down\n",
1984                         opts->queue_size, ctrl->ctrl.sqsize + 1);
1985                 opts->queue_size = ctrl->ctrl.sqsize + 1;
1986         }
1987
1988         if (opts->nr_io_queues) {
1989                 ret = nvme_rdma_configure_io_queues(ctrl, true);
1990                 if (ret)
1991                         goto out_remove_admin_queue;
1992         }
1993
1994         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1995         WARN_ON_ONCE(!changed);
1996
1997         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1998                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1999
2000         nvme_get_ctrl(&ctrl->ctrl);
2001
2002         mutex_lock(&nvme_rdma_ctrl_mutex);
2003         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2004         mutex_unlock(&nvme_rdma_ctrl_mutex);
2005
2006         nvme_start_ctrl(&ctrl->ctrl);
2007
2008         return &ctrl->ctrl;
2009
2010 out_remove_admin_queue:
2011         nvme_rdma_destroy_admin_queue(ctrl, true);
2012 out_kfree_queues:
2013         kfree(ctrl->queues);
2014 out_uninit_ctrl:
2015         nvme_uninit_ctrl(&ctrl->ctrl);
2016         nvme_put_ctrl(&ctrl->ctrl);
2017         if (ret > 0)
2018                 ret = -EIO;
2019         return ERR_PTR(ret);
2020 out_free_ctrl:
2021         kfree(ctrl);
2022         return ERR_PTR(ret);
2023 }
2024
2025 static struct nvmf_transport_ops nvme_rdma_transport = {
2026         .name           = "rdma",
2027         .required_opts  = NVMF_OPT_TRADDR,
2028         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2029                           NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2030         .create_ctrl    = nvme_rdma_create_ctrl,
2031 };
2032
2033 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2034 {
2035         struct nvme_rdma_ctrl *ctrl;
2036
2037         /* Delete all controllers using this device */
2038         mutex_lock(&nvme_rdma_ctrl_mutex);
2039         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2040                 if (ctrl->device->dev != ib_device)
2041                         continue;
2042                 dev_info(ctrl->ctrl.device,
2043                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
2044                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2045                 nvme_delete_ctrl(&ctrl->ctrl);
2046         }
2047         mutex_unlock(&nvme_rdma_ctrl_mutex);
2048
2049         flush_workqueue(nvme_wq);
2050 }
2051
2052 static struct ib_client nvme_rdma_ib_client = {
2053         .name   = "nvme_rdma",
2054         .remove = nvme_rdma_remove_one
2055 };
2056
2057 static int __init nvme_rdma_init_module(void)
2058 {
2059         int ret;
2060
2061         ret = ib_register_client(&nvme_rdma_ib_client);
2062         if (ret)
2063                 return ret;
2064
2065         ret = nvmf_register_transport(&nvme_rdma_transport);
2066         if (ret)
2067                 goto err_unreg_client;
2068
2069         return 0;
2070
2071 err_unreg_client:
2072         ib_unregister_client(&nvme_rdma_ib_client);
2073         return ret;
2074 }
2075
2076 static void __exit nvme_rdma_cleanup_module(void)
2077 {
2078         nvmf_unregister_transport(&nvme_rdma_transport);
2079         ib_unregister_client(&nvme_rdma_ib_client);
2080 }
2081
2082 module_init(nvme_rdma_init_module);
2083 module_exit(nvme_rdma_cleanup_module);
2084
2085 MODULE_LICENSE("GPL v2");