drivers/nvme/host/multipath.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2017-2018 Christoph Hellwig.
   4  */
   5
   6 #include <linux/moduleparam.h>
   7 #include <trace/events/block.h>
   8 #include "nvme.h"
   9
  10 static bool multipath = true;
  11 module_param(multipath, bool, 0444);
  12 MODULE_PARM_DESC(multipath,
  13         "turn on native support for multiple controllers per subsystem");
  14
  15 inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
  16 {
  17         return multipath && ctrl->subsys && (ctrl->subsys->cmic & (1 << 3));
  18 }
  19
  20 /*
  21  * If multipathing is enabled we need to always use the subsystem instance
  22  * number for numbering our devices to avoid conflicts between subsystems that
  23  * have multiple controllers and thus use the multipath-aware subsystem node
  24  * and those that have a single controller and use the controller node
  25  * directly.
  26  */
  27 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
  28                         struct nvme_ctrl *ctrl, int *flags)
  29 {
  30         if (!multipath) {
  31                 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
  32         } else if (ns->head->disk) {
  33                 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
  34                                 ctrl->instance, ns->head->instance);
  35                 *flags = GENHD_FL_HIDDEN;
  36         } else {
  37                 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
  38                                 ns->head->instance);
  39         }
  40 }
  41
  42 void nvme_failover_req(struct request *req)
  43 {
  44         struct nvme_ns *ns = req->q->queuedata;
  45         u16 status = nvme_req(req)->status;
  46         unsigned long flags;
  47
  48         spin_lock_irqsave(&ns->head->requeue_lock, flags);
  49         blk_steal_bios(&ns->head->requeue_list, req);
  50         spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
  51         blk_mq_end_request(req, 0);
  52
  53         switch (status & 0x7ff) {
  54         case NVME_SC_ANA_TRANSITION:
  55         case NVME_SC_ANA_INACCESSIBLE:
  56         case NVME_SC_ANA_PERSISTENT_LOSS:
  57                 /*
  58                  * If we got back an ANA error we know the controller is alive,
  59                  * but not ready to serve this namespaces.  The spec suggests
  60                  * we should update our general state here, but due to the fact
  61                  * that the admin and I/O queues are not serialized that is
  62                  * fundamentally racy.  So instead just clear the current path,
  63                  * mark the the path as pending and kick of a re-read of the ANA
  64                  * log page ASAP.
  65                  */
  66                 nvme_mpath_clear_current_path(ns);
  67                 if (ns->ctrl->ana_log_buf) {
  68                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
  69                         queue_work(nvme_wq, &ns->ctrl->ana_work);
  70                 }
  71                 break;
  72         case NVME_SC_HOST_PATH_ERROR:
  73                 /*
  74                  * Temporary transport disruption in talking to the controller.
  75                  * Try to send on a new path.
  76                  */
  77                 nvme_mpath_clear_current_path(ns);
  78                 break;
  79         default:
  80                 /*
  81                  * Reset the controller for any non-ANA error as we don't know
  82                  * what caused the error.
  83                  */
  84                 nvme_reset_ctrl(ns->ctrl);
  85                 break;
  86         }
  87
  88         kblockd_schedule_work(&ns->head->requeue_work);
  89 }
  90
  91 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
  92 {
  93         struct nvme_ns *ns;
  94
  95         down_read(&ctrl->namespaces_rwsem);
  96         list_for_each_entry(ns, &ctrl->namespaces, list) {
  97                 if (ns->head->disk)
  98                         kblockd_schedule_work(&ns->head->requeue_work);
  99         }
 100         up_read(&ctrl->namespaces_rwsem);
 101 }
 102
 103 static const char *nvme_ana_state_names[] = {
 104         [0]                             = "invalid state",
 105         [NVME_ANA_OPTIMIZED]            = "optimized",
 106         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 107         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 108         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 109         [NVME_ANA_CHANGE]               = "change",
 110 };
 111
 112 void nvme_mpath_clear_current_path(struct nvme_ns *ns)
 113 {
 114         struct nvme_ns_head *head = ns->head;
 115         int node;
 116
 117         if (!head)
 118                 return;
 119
 120         for_each_node(node) {
 121                 if (ns == rcu_access_pointer(head->current_path[node]))
 122                         rcu_assign_pointer(head->current_path[node], NULL);
 123         }
 124 }
 125
 126 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 127 {
 128         return ns->ctrl->state != NVME_CTRL_LIVE ||
 129                 test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 130                 test_bit(NVME_NS_REMOVING, &ns->flags);
 131 }
 132
 133 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 134 {
 135         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 136         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 137
 138         list_for_each_entry_rcu(ns, &head->list, siblings) {
 139                 if (nvme_path_is_disabled(ns))
 140                         continue;
 141
 142                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 143                         distance = node_distance(node, ns->ctrl->numa_node);
 144                 else
 145                         distance = LOCAL_DISTANCE;
 146
 147                 switch (ns->ana_state) {
 148                 case NVME_ANA_OPTIMIZED:
 149                         if (distance < found_distance) {
 150                                 found_distance = distance;
 151                                 found = ns;
 152                         }
 153                         break;
 154                 case NVME_ANA_NONOPTIMIZED:
 155                         if (distance < fallback_distance) {
 156                                 fallback_distance = distance;
 157                                 fallback = ns;
 158                         }
 159                         break;
 160                 default:
 161                         break;
 162                 }
 163         }
 164
 165         if (!found)
 166                 found = fallback;
 167         if (found)
 168                 rcu_assign_pointer(head->current_path[node], found);
 169         return found;
 170 }
 171
 172 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 173                 struct nvme_ns *ns)
 174 {
 175         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 176                         siblings);
 177         if (ns)
 178                 return ns;
 179         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 180 }
 181
 182 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 183                 int node, struct nvme_ns *old)
 184 {
 185         struct nvme_ns *ns, *found, *fallback = NULL;
 186
 187         if (list_is_singular(&head->list)) {
 188                 if (nvme_path_is_disabled(old))
 189                         return NULL;
 190                 return old;
 191         }
 192
 193         for (ns = nvme_next_ns(head, old);
 194              ns != old;
 195              ns = nvme_next_ns(head, ns)) {
 196                 if (nvme_path_is_disabled(ns))
 197                         continue;
 198
 199                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 200                         found = ns;
 201                         goto out;
 202                 }
 203                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 204                         fallback = ns;
 205         }
 206
 207         if (!fallback)
 208                 return NULL;
 209         found = fallback;
 210 out:
 211         rcu_assign_pointer(head->current_path[node], found);
 212         return found;
 213 }
 214
 215 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 216 {
 217         return ns->ctrl->state == NVME_CTRL_LIVE &&
 218                 ns->ana_state == NVME_ANA_OPTIMIZED;
 219 }
 220
 221 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 222 {
 223         int node = numa_node_id();
 224         struct nvme_ns *ns;
 225
 226         ns = srcu_dereference(head->current_path[node], &head->srcu);
 227         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns)
 228                 ns = nvme_round_robin_path(head, node, ns);
 229         if (unlikely(!ns || !nvme_path_is_optimized(ns)))
 230                 ns = __nvme_find_path(head, node);
 231         return ns;
 232 }
 233
 234 static blk_qc_t nvme_ns_head_make_request(struct request_queue *q,
 235                 struct bio *bio)
 236 {
 237         struct nvme_ns_head *head = q->queuedata;
 238         struct device *dev = disk_to_dev(head->disk);
 239         struct nvme_ns *ns;
 240         blk_qc_t ret = BLK_QC_T_NONE;
 241         int srcu_idx;
 242
 243         /*
 244          * The namespace might be going away and the bio might
 245          * be moved to a different queue via blk_steal_bios(),
 246          * so we need to use the bio_split pool from the original
 247          * queue to allocate the bvecs from.
 248          */
 249         blk_queue_split(q, &bio);
 250
 251         srcu_idx = srcu_read_lock(&head->srcu);
 252         ns = nvme_find_path(head);
 253         if (likely(ns)) {
 254                 bio->bi_disk = ns->disk;
 255                 bio->bi_opf |= REQ_NVME_MPATH;
 256                 trace_block_bio_remap(bio->bi_disk->queue, bio,
 257                                       disk_devt(ns->head->disk),
 258                                       bio->bi_iter.bi_sector);
 259                 ret = direct_make_request(bio);
 260         } else if (!list_empty_careful(&head->list)) {
 261                 dev_warn_ratelimited(dev, "no path available - requeuing I/O\n");
 262
 263                 spin_lock_irq(&head->requeue_lock);
 264                 bio_list_add(&head->requeue_list, bio);
 265                 spin_unlock_irq(&head->requeue_lock);
 266         } else {
 267                 dev_warn_ratelimited(dev, "no path - failing I/O\n");
 268
 269                 bio->bi_status = BLK_STS_IOERR;
 270                 bio_endio(bio);
 271         }
 272
 273         srcu_read_unlock(&head->srcu, srcu_idx);
 274         return ret;
 275 }
 276
 277 static void nvme_requeue_work(struct work_struct *work)
 278 {
 279         struct nvme_ns_head *head =
 280                 container_of(work, struct nvme_ns_head, requeue_work);
 281         struct bio *bio, *next;
 282
 283         spin_lock_irq(&head->requeue_lock);
 284         next = bio_list_get(&head->requeue_list);
 285         spin_unlock_irq(&head->requeue_lock);
 286
 287         while ((bio = next) != NULL) {
 288                 next = bio->bi_next;
 289                 bio->bi_next = NULL;
 290
 291                 /*
 292                  * Reset disk to the mpath node and resubmit to select a new
 293                  * path.
 294                  */
 295                 bio->bi_disk = head->disk;
 296                 generic_make_request(bio);
 297         }
 298 }
 299
 300 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 301 {
 302         struct request_queue *q;
 303         bool vwc = false;
 304
 305         mutex_init(&head->lock);
 306         bio_list_init(&head->requeue_list);
 307         spin_lock_init(&head->requeue_lock);
 308         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 309
 310         /*
 311          * Add a multipath node if the subsystems supports multiple controllers.
 312          * We also do this for private namespaces as the namespace sharing data could
 313          * change after a rescan.
 314          */
 315         if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath)
 316                 return 0;
 317
 318         q = blk_alloc_queue_node(GFP_KERNEL, ctrl->numa_node);
 319         if (!q)
 320                 goto out;
 321         q->queuedata = head;
 322         blk_queue_make_request(q, nvme_ns_head_make_request);
 323         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 324         /* set to a default value for 512 until disk is validated */
 325         blk_queue_logical_block_size(q, 512);
 326         blk_set_stacking_limits(&q->limits);
 327
 328         /* we need to propagate up the VMC settings */
 329         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 330                 vwc = true;
 331         blk_queue_write_cache(q, vwc, vwc);
 332
 333         head->disk = alloc_disk(0);
 334         if (!head->disk)
 335                 goto out_cleanup_queue;
 336         head->disk->fops = &nvme_ns_head_ops;
 337         head->disk->private_data = head;
 338         head->disk->queue = q;
 339         head->disk->flags = GENHD_FL_EXT_DEVT;
 340         sprintf(head->disk->disk_name, "nvme%dn%d",
 341                         ctrl->subsys->instance, head->instance);
 342         return 0;
 343
 344 out_cleanup_queue:
 345         blk_cleanup_queue(q);
 346 out:
 347         return -ENOMEM;
 348 }
 349
 350 static void nvme_mpath_set_live(struct nvme_ns *ns)
 351 {
 352         struct nvme_ns_head *head = ns->head;
 353
 354         lockdep_assert_held(&ns->head->lock);
 355
 356         if (!head->disk)
 357                 return;
 358
 359         if (!(head->disk->flags & GENHD_FL_UP))
 360                 device_add_disk(&head->subsys->dev, head->disk,
 361                                 nvme_ns_id_attr_groups);
 362
 363         if (nvme_path_is_optimized(ns)) {
 364                 int node, srcu_idx;
 365
 366                 srcu_idx = srcu_read_lock(&head->srcu);
 367                 for_each_node(node)
 368                         __nvme_find_path(head, node);
 369                 srcu_read_unlock(&head->srcu, srcu_idx);
 370         }
 371
 372         kblockd_schedule_work(&ns->head->requeue_work);
 373 }
 374
 375 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 376                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 377                         void *))
 378 {
 379         void *base = ctrl->ana_log_buf;
 380         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 381         int error, i;
 382
 383         lockdep_assert_held(&ctrl->ana_lock);
 384
 385         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 386                 struct nvme_ana_group_desc *desc = base + offset;
 387                 u32 nr_nsids = le32_to_cpu(desc->nnsids);
 388                 size_t nsid_buf_size = nr_nsids * sizeof(__le32);
 389
 390                 if (WARN_ON_ONCE(desc->grpid == 0))
 391                         return -EINVAL;
 392                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 393                         return -EINVAL;
 394                 if (WARN_ON_ONCE(desc->state == 0))
 395                         return -EINVAL;
 396                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 397                         return -EINVAL;
 398
 399                 offset += sizeof(*desc);
 400                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 401                         return -EINVAL;
 402
 403                 error = cb(ctrl, desc, data);
 404                 if (error)
 405                         return error;
 406
 407                 offset += nsid_buf_size;
 408                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 409                         return -EINVAL;
 410         }
 411
 412         return 0;
 413 }
 414
 415 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 416 {
 417         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 418 }
 419
 420 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 421                 struct nvme_ns *ns)
 422 {
 423         mutex_lock(&ns->head->lock);
 424         ns->ana_grpid = le32_to_cpu(desc->grpid);
 425         ns->ana_state = desc->state;
 426         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 427
 428         if (nvme_state_is_live(ns->ana_state))
 429                 nvme_mpath_set_live(ns);
 430         mutex_unlock(&ns->head->lock);
 431 }
 432
 433 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 434                 struct nvme_ana_group_desc *desc, void *data)
 435 {
 436         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 437         unsigned *nr_change_groups = data;
 438         struct nvme_ns *ns;
 439
 440         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 441                         le32_to_cpu(desc->grpid),
 442                         nvme_ana_state_names[desc->state]);
 443
 444         if (desc->state == NVME_ANA_CHANGE)
 445                 (*nr_change_groups)++;
 446
 447         if (!nr_nsids)
 448                 return 0;
 449
 450         down_write(&ctrl->namespaces_rwsem);
 451         list_for_each_entry(ns, &ctrl->namespaces, list) {
 452                 if (ns->head->ns_id != le32_to_cpu(desc->nsids[n]))
 453                         continue;
 454                 nvme_update_ns_ana_state(desc, ns);
 455                 if (++n == nr_nsids)
 456                         break;
 457         }
 458         up_write(&ctrl->namespaces_rwsem);
 459         WARN_ON_ONCE(n < nr_nsids);
 460         return 0;
 461 }
 462
 463 static int nvme_read_ana_log(struct nvme_ctrl *ctrl, bool groups_only)
 464 {
 465         u32 nr_change_groups = 0;
 466         int error;
 467
 468         mutex_lock(&ctrl->ana_lock);
 469         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA,
 470                         groups_only ? NVME_ANA_LOG_RGO : 0,
 471                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 472         if (error) {
 473                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 474                 goto out_unlock;
 475         }
 476
 477         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 478                         nvme_update_ana_state);
 479         if (error)
 480                 goto out_unlock;
 481
 482         /*
 483          * In theory we should have an ANATT timer per group as they might enter
 484          * the change state at different times.  But that is a lot of overhead
 485          * just to protect against a target that keeps entering new changes
 486          * states while never finishing previous ones.  But we'll still
 487          * eventually time out once all groups are in change state, so this
 488          * isn't a big deal.
 489          *
 490          * We also double the ANATT value to provide some slack for transports
 491          * or AEN processing overhead.
 492          */
 493         if (nr_change_groups)
 494                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 495         else
 496                 del_timer_sync(&ctrl->anatt_timer);
 497 out_unlock:
 498         mutex_unlock(&ctrl->ana_lock);
 499         return error;
 500 }
 501
 502 static void nvme_ana_work(struct work_struct *work)
 503 {
 504         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 505
 506         nvme_read_ana_log(ctrl, false);
 507 }
 508
 509 static void nvme_anatt_timeout(struct timer_list *t)
 510 {
 511         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 512
 513         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 514         nvme_reset_ctrl(ctrl);
 515 }
 516
 517 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 518 {
 519         if (!nvme_ctrl_use_ana(ctrl))
 520                 return;
 521         del_timer_sync(&ctrl->anatt_timer);
 522         cancel_work_sync(&ctrl->ana_work);
 523 }
 524
 525 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 526         struct device_attribute subsys_attr_##_name =   \
 527                 __ATTR(_name, _mode, _show, _store)
 528
 529 static const char *nvme_iopolicy_names[] = {
 530         [NVME_IOPOLICY_NUMA]    = "numa",
 531         [NVME_IOPOLICY_RR]      = "round-robin",
 532 };
 533
 534 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 535                 struct device_attribute *attr, char *buf)
 536 {
 537         struct nvme_subsystem *subsys =
 538                 container_of(dev, struct nvme_subsystem, dev);
 539
 540         return sprintf(buf, "%s\n",
 541                         nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 542 }
 543
 544 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 545                 struct device_attribute *attr, const char *buf, size_t count)
 546 {
 547         struct nvme_subsystem *subsys =
 548                 container_of(dev, struct nvme_subsystem, dev);
 549         int i;
 550
 551         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 552                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 553                         WRITE_ONCE(subsys->iopolicy, i);
 554                         return count;
 555                 }
 556         }
 557
 558         return -EINVAL;
 559 }
 560 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 561                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 562
 563 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 564                 char *buf)
 565 {
 566         return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 567 }
 568 DEVICE_ATTR_RO(ana_grpid);
 569
 570 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 571                 char *buf)
 572 {
 573         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 574
 575         return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 576 }
 577 DEVICE_ATTR_RO(ana_state);
 578
 579 static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl,
 580                 struct nvme_ana_group_desc *desc, void *data)
 581 {
 582         struct nvme_ns *ns = data;
 583
 584         if (ns->ana_grpid == le32_to_cpu(desc->grpid)) {
 585                 nvme_update_ns_ana_state(desc, ns);
 586                 return -ENXIO; /* just break out of the loop */
 587         }
 588
 589         return 0;
 590 }
 591
 592 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 593 {
 594         if (nvme_ctrl_use_ana(ns->ctrl)) {
 595                 mutex_lock(&ns->ctrl->ana_lock);
 596                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 597                 nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state);
 598                 mutex_unlock(&ns->ctrl->ana_lock);
 599         } else {
 600                 mutex_lock(&ns->head->lock);
 601                 ns->ana_state = NVME_ANA_OPTIMIZED;
 602                 nvme_mpath_set_live(ns);
 603                 mutex_unlock(&ns->head->lock);
 604         }
 605 }
 606
 607 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 608 {
 609         if (!head->disk)
 610                 return;
 611         if (head->disk->flags & GENHD_FL_UP)
 612                 del_gendisk(head->disk);
 613         blk_set_queue_dying(head->disk->queue);
 614         /* make sure all pending bios are cleaned up */
 615         kblockd_schedule_work(&head->requeue_work);
 616         flush_work(&head->requeue_work);
 617         blk_cleanup_queue(head->disk->queue);
 618         put_disk(head->disk);
 619 }
 620
 621 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 622 {
 623         int error;
 624
 625         if (!nvme_ctrl_use_ana(ctrl))
 626                 return 0;
 627
 628         ctrl->anacap = id->anacap;
 629         ctrl->anatt = id->anatt;
 630         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 631         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 632
 633         mutex_init(&ctrl->ana_lock);
 634         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 635         ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 636                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
 637         ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
 638
 639         if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
 640                 dev_err(ctrl->device,
 641                         "ANA log page size (%zd) larger than MDTS (%d).\n",
 642                         ctrl->ana_log_size,
 643                         ctrl->max_hw_sectors << SECTOR_SHIFT);
 644                 dev_err(ctrl->device, "disabling ANA support.\n");
 645                 return 0;
 646         }
 647
 648         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 649         ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
 650         if (!ctrl->ana_log_buf) {
 651                 error = -ENOMEM;
 652                 goto out;
 653         }
 654
 655         error = nvme_read_ana_log(ctrl, true);
 656         if (error)
 657                 goto out_free_ana_log_buf;
 658         return 0;
 659 out_free_ana_log_buf:
 660         kfree(ctrl->ana_log_buf);
 661         ctrl->ana_log_buf = NULL;
 662 out:
 663         return error;
 664 }
 665
 666 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 667 {
 668         kfree(ctrl->ana_log_buf);
 669         ctrl->ana_log_buf = NULL;
 670 }
 671