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/backing-dev.h>
   7 #include <linux/moduleparam.h>
   8 #include <trace/events/block.h>
   9 #include "nvme.h"
  10
  11 static bool multipath = true;
  12 module_param(multipath, bool, 0444);
  13 MODULE_PARM_DESC(multipath,
  14         "turn on native support for multiple controllers per subsystem");
  15
  16 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
  17 {
  18         struct nvme_ns_head *h;
  19
  20         lockdep_assert_held(&subsys->lock);
  21         list_for_each_entry(h, &subsys->nsheads, entry)
  22                 if (h->disk)
  23                         blk_mq_unfreeze_queue(h->disk->queue);
  24 }
  25
  26 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
  27 {
  28         struct nvme_ns_head *h;
  29
  30         lockdep_assert_held(&subsys->lock);
  31         list_for_each_entry(h, &subsys->nsheads, entry)
  32                 if (h->disk)
  33                         blk_mq_freeze_queue_wait(h->disk->queue);
  34 }
  35
  36 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
  37 {
  38         struct nvme_ns_head *h;
  39
  40         lockdep_assert_held(&subsys->lock);
  41         list_for_each_entry(h, &subsys->nsheads, entry)
  42                 if (h->disk)
  43                         blk_freeze_queue_start(h->disk->queue);
  44 }
  45
  46 /*
  47  * If multipathing is enabled we need to always use the subsystem instance
  48  * number for numbering our devices to avoid conflicts between subsystems that
  49  * have multiple controllers and thus use the multipath-aware subsystem node
  50  * and those that have a single controller and use the controller node
  51  * directly.
  52  */
  53 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
  54                         struct nvme_ctrl *ctrl, int *flags)
  55 {
  56         if (!multipath) {
  57                 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
  58         } else if (ns->head->disk) {
  59                 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
  60                                 ctrl->instance, ns->head->instance);
  61                 *flags = GENHD_FL_HIDDEN;
  62         } else {
  63                 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
  64                                 ns->head->instance);
  65         }
  66 }
  67
  68 bool nvme_failover_req(struct request *req)
  69 {
  70         struct nvme_ns *ns = req->q->queuedata;
  71         u16 status = nvme_req(req)->status;
  72         unsigned long flags;
  73
  74         switch (status & 0x7ff) {
  75         case NVME_SC_ANA_TRANSITION:
  76         case NVME_SC_ANA_INACCESSIBLE:
  77         case NVME_SC_ANA_PERSISTENT_LOSS:
  78                 /*
  79                  * If we got back an ANA error we know the controller is alive,
  80                  * but not ready to serve this namespaces.  The spec suggests
  81                  * we should update our general state here, but due to the fact
  82                  * that the admin and I/O queues are not serialized that is
  83                  * fundamentally racy.  So instead just clear the current path,
  84                  * mark the the path as pending and kick of a re-read of the ANA
  85                  * log page ASAP.
  86                  */
  87                 nvme_mpath_clear_current_path(ns);
  88                 if (ns->ctrl->ana_log_buf) {
  89                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
  90                         queue_work(nvme_wq, &ns->ctrl->ana_work);
  91                 }
  92                 break;
  93         case NVME_SC_HOST_PATH_ERROR:
  94         case NVME_SC_HOST_ABORTED_CMD:
  95                 /*
  96                  * Temporary transport disruption in talking to the controller.
  97                  * Try to send on a new path.
  98                  */
  99                 nvme_mpath_clear_current_path(ns);
 100                 break;
 101         default:
 102                 /* This was a non-ANA error so follow the normal error path. */
 103                 return false;
 104         }
 105
 106         spin_lock_irqsave(&ns->head->requeue_lock, flags);
 107         blk_steal_bios(&ns->head->requeue_list, req);
 108         spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
 109         blk_mq_end_request(req, 0);
 110
 111         kblockd_schedule_work(&ns->head->requeue_work);
 112         return true;
 113 }
 114
 115 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 116 {
 117         struct nvme_ns *ns;
 118
 119         down_read(&ctrl->namespaces_rwsem);
 120         list_for_each_entry(ns, &ctrl->namespaces, list) {
 121                 if (ns->head->disk)
 122                         kblockd_schedule_work(&ns->head->requeue_work);
 123         }
 124         up_read(&ctrl->namespaces_rwsem);
 125 }
 126
 127 static const char *nvme_ana_state_names[] = {
 128         [0]                             = "invalid state",
 129         [NVME_ANA_OPTIMIZED]            = "optimized",
 130         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 131         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 132         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 133         [NVME_ANA_CHANGE]               = "change",
 134 };
 135
 136 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 137 {
 138         struct nvme_ns_head *head = ns->head;
 139         bool changed = false;
 140         int node;
 141
 142         if (!head)
 143                 goto out;
 144
 145         for_each_node(node) {
 146                 if (ns == rcu_access_pointer(head->current_path[node])) {
 147                         rcu_assign_pointer(head->current_path[node], NULL);
 148                         changed = true;
 149                 }
 150         }
 151 out:
 152         return changed;
 153 }
 154
 155 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 156 {
 157         struct nvme_ns *ns;
 158
 159         mutex_lock(&ctrl->scan_lock);
 160         down_read(&ctrl->namespaces_rwsem);
 161         list_for_each_entry(ns, &ctrl->namespaces, list)
 162                 if (nvme_mpath_clear_current_path(ns))
 163                         kblockd_schedule_work(&ns->head->requeue_work);
 164         up_read(&ctrl->namespaces_rwsem);
 165         mutex_unlock(&ctrl->scan_lock);
 166 }
 167
 168 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 169 {
 170         /*
 171          * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
 172          * still be able to complete assuming that the controller is connected.
 173          * Otherwise it will fail immediately and return to the requeue list.
 174          */
 175         if (ns->ctrl->state != NVME_CTRL_LIVE &&
 176             ns->ctrl->state != NVME_CTRL_DELETING)
 177                 return true;
 178         if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 179             test_bit(NVME_NS_REMOVING, &ns->flags))
 180                 return true;
 181         return false;
 182 }
 183
 184 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 185 {
 186         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 187         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 188
 189         list_for_each_entry_rcu(ns, &head->list, siblings) {
 190                 if (nvme_path_is_disabled(ns))
 191                         continue;
 192
 193                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 194                         distance = node_distance(node, ns->ctrl->numa_node);
 195                 else
 196                         distance = LOCAL_DISTANCE;
 197
 198                 switch (ns->ana_state) {
 199                 case NVME_ANA_OPTIMIZED:
 200                         if (distance < found_distance) {
 201                                 found_distance = distance;
 202                                 found = ns;
 203                         }
 204                         break;
 205                 case NVME_ANA_NONOPTIMIZED:
 206                         if (distance < fallback_distance) {
 207                                 fallback_distance = distance;
 208                                 fallback = ns;
 209                         }
 210                         break;
 211                 default:
 212                         break;
 213                 }
 214         }
 215
 216         if (!found)
 217                 found = fallback;
 218         if (found)
 219                 rcu_assign_pointer(head->current_path[node], found);
 220         return found;
 221 }
 222
 223 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 224                 struct nvme_ns *ns)
 225 {
 226         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 227                         siblings);
 228         if (ns)
 229                 return ns;
 230         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 231 }
 232
 233 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 234                 int node, struct nvme_ns *old)
 235 {
 236         struct nvme_ns *ns, *found, *fallback = NULL;
 237
 238         if (list_is_singular(&head->list)) {
 239                 if (nvme_path_is_disabled(old))
 240                         return NULL;
 241                 return old;
 242         }
 243
 244         for (ns = nvme_next_ns(head, old);
 245              ns != old;
 246              ns = nvme_next_ns(head, ns)) {
 247                 if (nvme_path_is_disabled(ns))
 248                         continue;
 249
 250                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 251                         found = ns;
 252                         goto out;
 253                 }
 254                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 255                         fallback = ns;
 256         }
 257
 258         /* No optimized path found, re-check the current path */
 259         if (!nvme_path_is_disabled(old) &&
 260             old->ana_state == NVME_ANA_OPTIMIZED) {
 261                 found = old;
 262                 goto out;
 263         }
 264         if (!fallback)
 265                 return NULL;
 266         found = fallback;
 267 out:
 268         rcu_assign_pointer(head->current_path[node], found);
 269         return found;
 270 }
 271
 272 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 273 {
 274         return ns->ctrl->state == NVME_CTRL_LIVE &&
 275                 ns->ana_state == NVME_ANA_OPTIMIZED;
 276 }
 277
 278 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 279 {
 280         int node = numa_node_id();
 281         struct nvme_ns *ns;
 282
 283         ns = srcu_dereference(head->current_path[node], &head->srcu);
 284         if (unlikely(!ns))
 285                 return __nvme_find_path(head, node);
 286
 287         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
 288                 return nvme_round_robin_path(head, node, ns);
 289         if (unlikely(!nvme_path_is_optimized(ns)))
 290                 return __nvme_find_path(head, node);
 291         return ns;
 292 }
 293
 294 static bool nvme_available_path(struct nvme_ns_head *head)
 295 {
 296         struct nvme_ns *ns;
 297
 298         list_for_each_entry_rcu(ns, &head->list, siblings) {
 299                 switch (ns->ctrl->state) {
 300                 case NVME_CTRL_LIVE:
 301                 case NVME_CTRL_RESETTING:
 302                 case NVME_CTRL_CONNECTING:
 303                         /* fallthru */
 304                         return true;
 305                 default:
 306                         break;
 307                 }
 308         }
 309         return false;
 310 }
 311
 312 blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
 313 {
 314         struct nvme_ns_head *head = bio->bi_disk->private_data;
 315         struct device *dev = disk_to_dev(head->disk);
 316         struct nvme_ns *ns;
 317         blk_qc_t ret = BLK_QC_T_NONE;
 318         int srcu_idx;
 319
 320         /*
 321          * The namespace might be going away and the bio might be moved to a
 322          * different queue via blk_steal_bios(), so we need to use the bio_split
 323          * pool from the original queue to allocate the bvecs from.
 324          */
 325         blk_queue_split(&bio);
 326
 327         srcu_idx = srcu_read_lock(&head->srcu);
 328         ns = nvme_find_path(head);
 329         if (likely(ns)) {
 330                 bio->bi_disk = ns->disk;
 331                 bio->bi_opf |= REQ_NVME_MPATH;
 332                 trace_block_bio_remap(bio->bi_disk->queue, bio,
 333                                       disk_devt(ns->head->disk),
 334                                       bio->bi_iter.bi_sector);
 335                 ret = submit_bio_noacct(bio);
 336         } else if (nvme_available_path(head)) {
 337                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 338
 339                 spin_lock_irq(&head->requeue_lock);
 340                 bio_list_add(&head->requeue_list, bio);
 341                 spin_unlock_irq(&head->requeue_lock);
 342         } else {
 343                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 344
 345                 bio->bi_status = BLK_STS_IOERR;
 346                 bio_endio(bio);
 347         }
 348
 349         srcu_read_unlock(&head->srcu, srcu_idx);
 350         return ret;
 351 }
 352
 353 static void nvme_requeue_work(struct work_struct *work)
 354 {
 355         struct nvme_ns_head *head =
 356                 container_of(work, struct nvme_ns_head, requeue_work);
 357         struct bio *bio, *next;
 358
 359         spin_lock_irq(&head->requeue_lock);
 360         next = bio_list_get(&head->requeue_list);
 361         spin_unlock_irq(&head->requeue_lock);
 362
 363         while ((bio = next) != NULL) {
 364                 next = bio->bi_next;
 365                 bio->bi_next = NULL;
 366
 367                 /*
 368                  * Reset disk to the mpath node and resubmit to select a new
 369                  * path.
 370                  */
 371                 bio->bi_disk = head->disk;
 372                 submit_bio_noacct(bio);
 373         }
 374 }
 375
 376 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 377 {
 378         struct request_queue *q;
 379         bool vwc = false;
 380
 381         mutex_init(&head->lock);
 382         bio_list_init(&head->requeue_list);
 383         spin_lock_init(&head->requeue_lock);
 384         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 385
 386         /*
 387          * Add a multipath node if the subsystems supports multiple controllers.
 388          * We also do this for private namespaces as the namespace sharing data could
 389          * change after a rescan.
 390          */
 391         if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) || !multipath)
 392                 return 0;
 393
 394         q = blk_alloc_queue(ctrl->numa_node);
 395         if (!q)
 396                 goto out;
 397         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 398         /* set to a default value for 512 until disk is validated */
 399         blk_queue_logical_block_size(q, 512);
 400         blk_set_stacking_limits(&q->limits);
 401
 402         /* we need to propagate up the VMC settings */
 403         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 404                 vwc = true;
 405         blk_queue_write_cache(q, vwc, vwc);
 406
 407         head->disk = alloc_disk(0);
 408         if (!head->disk)
 409                 goto out_cleanup_queue;
 410         head->disk->fops = &nvme_ns_head_ops;
 411         head->disk->private_data = head;
 412         head->disk->queue = q;
 413         head->disk->flags = GENHD_FL_EXT_DEVT;
 414         sprintf(head->disk->disk_name, "nvme%dn%d",
 415                         ctrl->subsys->instance, head->instance);
 416         return 0;
 417
 418 out_cleanup_queue:
 419         blk_cleanup_queue(q);
 420 out:
 421         return -ENOMEM;
 422 }
 423
 424 static void nvme_mpath_set_live(struct nvme_ns *ns)
 425 {
 426         struct nvme_ns_head *head = ns->head;
 427
 428         if (!head->disk)
 429                 return;
 430
 431         if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags))
 432                 device_add_disk(&head->subsys->dev, head->disk,
 433                                 nvme_ns_id_attr_groups);
 434
 435         mutex_lock(&head->lock);
 436         if (nvme_path_is_optimized(ns)) {
 437                 int node, srcu_idx;
 438
 439                 srcu_idx = srcu_read_lock(&head->srcu);
 440                 for_each_node(node)
 441                         __nvme_find_path(head, node);
 442                 srcu_read_unlock(&head->srcu, srcu_idx);
 443         }
 444         mutex_unlock(&head->lock);
 445
 446         synchronize_srcu(&head->srcu);
 447         kblockd_schedule_work(&head->requeue_work);
 448 }
 449
 450 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 451                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 452                         void *))
 453 {
 454         void *base = ctrl->ana_log_buf;
 455         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 456         int error, i;
 457
 458         lockdep_assert_held(&ctrl->ana_lock);
 459
 460         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 461                 struct nvme_ana_group_desc *desc = base + offset;
 462                 u32 nr_nsids;
 463                 size_t nsid_buf_size;
 464
 465                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 466                         return -EINVAL;
 467
 468                 nr_nsids = le32_to_cpu(desc->nnsids);
 469                 nsid_buf_size = nr_nsids * sizeof(__le32);
 470
 471                 if (WARN_ON_ONCE(desc->grpid == 0))
 472                         return -EINVAL;
 473                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 474                         return -EINVAL;
 475                 if (WARN_ON_ONCE(desc->state == 0))
 476                         return -EINVAL;
 477                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 478                         return -EINVAL;
 479
 480                 offset += sizeof(*desc);
 481                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 482                         return -EINVAL;
 483
 484                 error = cb(ctrl, desc, data);
 485                 if (error)
 486                         return error;
 487
 488                 offset += nsid_buf_size;
 489         }
 490
 491         return 0;
 492 }
 493
 494 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 495 {
 496         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 497 }
 498
 499 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 500                 struct nvme_ns *ns)
 501 {
 502         ns->ana_grpid = le32_to_cpu(desc->grpid);
 503         ns->ana_state = desc->state;
 504         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 505
 506         if (nvme_state_is_live(ns->ana_state))
 507                 nvme_mpath_set_live(ns);
 508 }
 509
 510 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 511                 struct nvme_ana_group_desc *desc, void *data)
 512 {
 513         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 514         unsigned *nr_change_groups = data;
 515         struct nvme_ns *ns;
 516
 517         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 518                         le32_to_cpu(desc->grpid),
 519                         nvme_ana_state_names[desc->state]);
 520
 521         if (desc->state == NVME_ANA_CHANGE)
 522                 (*nr_change_groups)++;
 523
 524         if (!nr_nsids)
 525                 return 0;
 526
 527         down_read(&ctrl->namespaces_rwsem);
 528         list_for_each_entry(ns, &ctrl->namespaces, list) {
 529                 unsigned nsid = le32_to_cpu(desc->nsids[n]);
 530
 531                 if (ns->head->ns_id < nsid)
 532                         continue;
 533                 if (ns->head->ns_id == nsid)
 534                         nvme_update_ns_ana_state(desc, ns);
 535                 if (++n == nr_nsids)
 536                         break;
 537         }
 538         up_read(&ctrl->namespaces_rwsem);
 539         return 0;
 540 }
 541
 542 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 543 {
 544         u32 nr_change_groups = 0;
 545         int error;
 546
 547         mutex_lock(&ctrl->ana_lock);
 548         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
 549                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 550         if (error) {
 551                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 552                 goto out_unlock;
 553         }
 554
 555         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 556                         nvme_update_ana_state);
 557         if (error)
 558                 goto out_unlock;
 559
 560         /*
 561          * In theory we should have an ANATT timer per group as they might enter
 562          * the change state at different times.  But that is a lot of overhead
 563          * just to protect against a target that keeps entering new changes
 564          * states while never finishing previous ones.  But we'll still
 565          * eventually time out once all groups are in change state, so this
 566          * isn't a big deal.
 567          *
 568          * We also double the ANATT value to provide some slack for transports
 569          * or AEN processing overhead.
 570          */
 571         if (nr_change_groups)
 572                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 573         else
 574                 del_timer_sync(&ctrl->anatt_timer);
 575 out_unlock:
 576         mutex_unlock(&ctrl->ana_lock);
 577         return error;
 578 }
 579
 580 static void nvme_ana_work(struct work_struct *work)
 581 {
 582         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 583
 584         if (ctrl->state != NVME_CTRL_LIVE)
 585                 return;
 586
 587         nvme_read_ana_log(ctrl);
 588 }
 589
 590 static void nvme_anatt_timeout(struct timer_list *t)
 591 {
 592         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 593
 594         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 595         nvme_reset_ctrl(ctrl);
 596 }
 597
 598 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 599 {
 600         if (!nvme_ctrl_use_ana(ctrl))
 601                 return;
 602         del_timer_sync(&ctrl->anatt_timer);
 603         cancel_work_sync(&ctrl->ana_work);
 604 }
 605
 606 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 607         struct device_attribute subsys_attr_##_name =   \
 608                 __ATTR(_name, _mode, _show, _store)
 609
 610 static const char *nvme_iopolicy_names[] = {
 611         [NVME_IOPOLICY_NUMA]    = "numa",
 612         [NVME_IOPOLICY_RR]      = "round-robin",
 613 };
 614
 615 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 616                 struct device_attribute *attr, char *buf)
 617 {
 618         struct nvme_subsystem *subsys =
 619                 container_of(dev, struct nvme_subsystem, dev);
 620
 621         return sprintf(buf, "%s\n",
 622                         nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 623 }
 624
 625 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 626                 struct device_attribute *attr, const char *buf, size_t count)
 627 {
 628         struct nvme_subsystem *subsys =
 629                 container_of(dev, struct nvme_subsystem, dev);
 630         int i;
 631
 632         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 633                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 634                         WRITE_ONCE(subsys->iopolicy, i);
 635                         return count;
 636                 }
 637         }
 638
 639         return -EINVAL;
 640 }
 641 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 642                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 643
 644 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 645                 char *buf)
 646 {
 647         return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 648 }
 649 DEVICE_ATTR_RO(ana_grpid);
 650
 651 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 652                 char *buf)
 653 {
 654         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 655
 656         return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 657 }
 658 DEVICE_ATTR_RO(ana_state);
 659
 660 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
 661                 struct nvme_ana_group_desc *desc, void *data)
 662 {
 663         struct nvme_ana_group_desc *dst = data;
 664
 665         if (desc->grpid != dst->grpid)
 666                 return 0;
 667
 668         *dst = *desc;
 669         return -ENXIO; /* just break out of the loop */
 670 }
 671
 672 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 673 {
 674         if (nvme_ctrl_use_ana(ns->ctrl)) {
 675                 struct nvme_ana_group_desc desc = {
 676                         .grpid = id->anagrpid,
 677                         .state = 0,
 678                 };
 679
 680                 mutex_lock(&ns->ctrl->ana_lock);
 681                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 682                 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
 683                 mutex_unlock(&ns->ctrl->ana_lock);
 684                 if (desc.state) {
 685                         /* found the group desc: update */
 686                         nvme_update_ns_ana_state(&desc, ns);
 687                 }
 688         } else {
 689                 ns->ana_state = NVME_ANA_OPTIMIZED;
 690                 nvme_mpath_set_live(ns);
 691         }
 692
 693         if (bdi_cap_stable_pages_required(ns->queue->backing_dev_info)) {
 694                 struct gendisk *disk = ns->head->disk;
 695
 696                 if (disk)
 697                         disk->queue->backing_dev_info->capabilities |=
 698                                         BDI_CAP_STABLE_WRITES;
 699         }
 700 }
 701
 702 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 703 {
 704         if (!head->disk)
 705                 return;
 706         if (head->disk->flags & GENHD_FL_UP)
 707                 del_gendisk(head->disk);
 708         blk_set_queue_dying(head->disk->queue);
 709         /* make sure all pending bios are cleaned up */
 710         kblockd_schedule_work(&head->requeue_work);
 711         flush_work(&head->requeue_work);
 712         blk_cleanup_queue(head->disk->queue);
 713         if (!test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 714                 /*
 715                  * if device_add_disk wasn't called, prevent
 716                  * disk release to put a bogus reference on the
 717                  * request queue
 718                  */
 719                 head->disk->queue = NULL;
 720         }
 721         put_disk(head->disk);
 722 }
 723
 724 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 725 {
 726         int error;
 727
 728         /* check if multipath is enabled and we have the capability */
 729         if (!multipath || !ctrl->subsys ||
 730             !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
 731                 return 0;
 732
 733         ctrl->anacap = id->anacap;
 734         ctrl->anatt = id->anatt;
 735         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 736         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 737
 738         mutex_init(&ctrl->ana_lock);
 739         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 740         ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 741                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
 742         ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
 743
 744         if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
 745                 dev_err(ctrl->device,
 746                         "ANA log page size (%zd) larger than MDTS (%d).\n",
 747                         ctrl->ana_log_size,
 748                         ctrl->max_hw_sectors << SECTOR_SHIFT);
 749                 dev_err(ctrl->device, "disabling ANA support.\n");
 750                 return 0;
 751         }
 752
 753         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 754         kfree(ctrl->ana_log_buf);
 755         ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
 756         if (!ctrl->ana_log_buf) {
 757                 error = -ENOMEM;
 758                 goto out;
 759         }
 760
 761         error = nvme_read_ana_log(ctrl);
 762         if (error)
 763                 goto out_free_ana_log_buf;
 764         return 0;
 765 out_free_ana_log_buf:
 766         kfree(ctrl->ana_log_buf);
 767         ctrl->ana_log_buf = NULL;
 768 out:
 769         return error;
 770 }
 771
 772 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 773 {
 774         kfree(ctrl->ana_log_buf);
 775         ctrl->ana_log_buf = NULL;
 776 }
 777