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 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
  16 {
  17         struct nvme_ns_head *h;
  18
  19         lockdep_assert_held(&subsys->lock);
  20         list_for_each_entry(h, &subsys->nsheads, entry)
  21                 if (h->disk)
  22                         blk_mq_unfreeze_queue(h->disk->queue);
  23 }
  24
  25 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
  26 {
  27         struct nvme_ns_head *h;
  28
  29         lockdep_assert_held(&subsys->lock);
  30         list_for_each_entry(h, &subsys->nsheads, entry)
  31                 if (h->disk)
  32                         blk_mq_freeze_queue_wait(h->disk->queue);
  33 }
  34
  35 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
  36 {
  37         struct nvme_ns_head *h;
  38
  39         lockdep_assert_held(&subsys->lock);
  40         list_for_each_entry(h, &subsys->nsheads, entry)
  41                 if (h->disk)
  42                         blk_freeze_queue_start(h->disk->queue);
  43 }
  44
  45 /*
  46  * If multipathing is enabled we need to always use the subsystem instance
  47  * number for numbering our devices to avoid conflicts between subsystems that
  48  * have multiple controllers and thus use the multipath-aware subsystem node
  49  * and those that have a single controller and use the controller node
  50  * directly.
  51  */
  52 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
  53                         struct nvme_ctrl *ctrl, int *flags)
  54 {
  55         if (!multipath) {
  56                 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
  57         } else if (ns->head->disk) {
  58                 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
  59                                 ctrl->instance, ns->head->instance);
  60                 *flags = GENHD_FL_HIDDEN;
  61         } else {
  62                 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
  63                                 ns->head->instance);
  64         }
  65 }
  66
  67 void nvme_failover_req(struct request *req)
  68 {
  69         struct nvme_ns *ns = req->q->queuedata;
  70         u16 status = nvme_req(req)->status;
  71         unsigned long flags;
  72
  73         spin_lock_irqsave(&ns->head->requeue_lock, flags);
  74         blk_steal_bios(&ns->head->requeue_list, req);
  75         spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
  76         blk_mq_end_request(req, 0);
  77
  78         switch (status & 0x7ff) {
  79         case NVME_SC_ANA_TRANSITION:
  80         case NVME_SC_ANA_INACCESSIBLE:
  81         case NVME_SC_ANA_PERSISTENT_LOSS:
  82                 /*
  83                  * If we got back an ANA error we know the controller is alive,
  84                  * but not ready to serve this namespaces.  The spec suggests
  85                  * we should update our general state here, but due to the fact
  86                  * that the admin and I/O queues are not serialized that is
  87                  * fundamentally racy.  So instead just clear the current path,
  88                  * mark the the path as pending and kick of a re-read of the ANA
  89                  * log page ASAP.
  90                  */
  91                 nvme_mpath_clear_current_path(ns);
  92                 if (ns->ctrl->ana_log_buf) {
  93                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
  94                         queue_work(nvme_wq, &ns->ctrl->ana_work);
  95                 }
  96                 break;
  97         case NVME_SC_HOST_PATH_ERROR:
  98                 /*
  99                  * Temporary transport disruption in talking to the controller.
 100                  * Try to send on a new path.
 101                  */
 102                 nvme_mpath_clear_current_path(ns);
 103                 break;
 104         default:
 105                 /*
 106                  * Reset the controller for any non-ANA error as we don't know
 107                  * what caused the error.
 108                  */
 109                 nvme_reset_ctrl(ns->ctrl);
 110                 break;
 111         }
 112
 113         kblockd_schedule_work(&ns->head->requeue_work);
 114 }
 115
 116 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 117 {
 118         struct nvme_ns *ns;
 119
 120         down_read(&ctrl->namespaces_rwsem);
 121         list_for_each_entry(ns, &ctrl->namespaces, list) {
 122                 if (ns->head->disk)
 123                         kblockd_schedule_work(&ns->head->requeue_work);
 124         }
 125         up_read(&ctrl->namespaces_rwsem);
 126 }
 127
 128 static const char *nvme_ana_state_names[] = {
 129         [0]                             = "invalid state",
 130         [NVME_ANA_OPTIMIZED]            = "optimized",
 131         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 132         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 133         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 134         [NVME_ANA_CHANGE]               = "change",
 135 };
 136
 137 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 138 {
 139         struct nvme_ns_head *head = ns->head;
 140         bool changed = false;
 141         int node;
 142
 143         if (!head)
 144                 goto out;
 145
 146         for_each_node(node) {
 147                 if (ns == rcu_access_pointer(head->current_path[node])) {
 148                         rcu_assign_pointer(head->current_path[node], NULL);
 149                         changed = true;
 150                 }
 151         }
 152 out:
 153         return changed;
 154 }
 155
 156 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 157 {
 158         struct nvme_ns *ns;
 159
 160         mutex_lock(&ctrl->scan_lock);
 161         down_read(&ctrl->namespaces_rwsem);
 162         list_for_each_entry(ns, &ctrl->namespaces, list)
 163                 if (nvme_mpath_clear_current_path(ns))
 164                         kblockd_schedule_work(&ns->head->requeue_work);
 165         up_read(&ctrl->namespaces_rwsem);
 166         mutex_unlock(&ctrl->scan_lock);
 167 }
 168
 169 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 170 {
 171         return ns->ctrl->state != NVME_CTRL_LIVE ||
 172                 test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 173                 test_bit(NVME_NS_REMOVING, &ns->flags);
 174 }
 175
 176 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 177 {
 178         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 179         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 180
 181         list_for_each_entry_rcu(ns, &head->list, siblings) {
 182                 if (nvme_path_is_disabled(ns))
 183                         continue;
 184
 185                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 186                         distance = node_distance(node, ns->ctrl->numa_node);
 187                 else
 188                         distance = LOCAL_DISTANCE;
 189
 190                 switch (ns->ana_state) {
 191                 case NVME_ANA_OPTIMIZED:
 192                         if (distance < found_distance) {
 193                                 found_distance = distance;
 194                                 found = ns;
 195                         }
 196                         break;
 197                 case NVME_ANA_NONOPTIMIZED:
 198                         if (distance < fallback_distance) {
 199                                 fallback_distance = distance;
 200                                 fallback = ns;
 201                         }
 202                         break;
 203                 default:
 204                         break;
 205                 }
 206         }
 207
 208         if (!found)
 209                 found = fallback;
 210         if (found)
 211                 rcu_assign_pointer(head->current_path[node], found);
 212         return found;
 213 }
 214
 215 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 216                 struct nvme_ns *ns)
 217 {
 218         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 219                         siblings);
 220         if (ns)
 221                 return ns;
 222         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 223 }
 224
 225 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 226                 int node, struct nvme_ns *old)
 227 {
 228         struct nvme_ns *ns, *found, *fallback = NULL;
 229
 230         if (list_is_singular(&head->list)) {
 231                 if (nvme_path_is_disabled(old))
 232                         return NULL;
 233                 return old;
 234         }
 235
 236         for (ns = nvme_next_ns(head, old);
 237              ns != old;
 238              ns = nvme_next_ns(head, ns)) {
 239                 if (nvme_path_is_disabled(ns))
 240                         continue;
 241
 242                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 243                         found = ns;
 244                         goto out;
 245                 }
 246                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 247                         fallback = ns;
 248         }
 249
 250         if (!fallback)
 251                 return NULL;
 252         found = fallback;
 253 out:
 254         rcu_assign_pointer(head->current_path[node], found);
 255         return found;
 256 }
 257
 258 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 259 {
 260         return ns->ctrl->state == NVME_CTRL_LIVE &&
 261                 ns->ana_state == NVME_ANA_OPTIMIZED;
 262 }
 263
 264 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 265 {
 266         int node = numa_node_id();
 267         struct nvme_ns *ns;
 268
 269         ns = srcu_dereference(head->current_path[node], &head->srcu);
 270         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns)
 271                 ns = nvme_round_robin_path(head, node, ns);
 272         if (unlikely(!ns || !nvme_path_is_optimized(ns)))
 273                 ns = __nvme_find_path(head, node);
 274         return ns;
 275 }
 276
 277 static bool nvme_available_path(struct nvme_ns_head *head)
 278 {
 279         struct nvme_ns *ns;
 280
 281         list_for_each_entry_rcu(ns, &head->list, siblings) {
 282                 switch (ns->ctrl->state) {
 283                 case NVME_CTRL_LIVE:
 284                 case NVME_CTRL_RESETTING:
 285                 case NVME_CTRL_CONNECTING:
 286                         /* fallthru */
 287                         return true;
 288                 default:
 289                         break;
 290                 }
 291         }
 292         return false;
 293 }
 294
 295 static blk_qc_t nvme_ns_head_make_request(struct request_queue *q,
 296                 struct bio *bio)
 297 {
 298         struct nvme_ns_head *head = q->queuedata;
 299         struct device *dev = disk_to_dev(head->disk);
 300         struct nvme_ns *ns;
 301         blk_qc_t ret = BLK_QC_T_NONE;
 302         int srcu_idx;
 303
 304         /*
 305          * The namespace might be going away and the bio might
 306          * be moved to a different queue via blk_steal_bios(),
 307          * so we need to use the bio_split pool from the original
 308          * queue to allocate the bvecs from.
 309          */
 310         blk_queue_split(q, &bio);
 311
 312         srcu_idx = srcu_read_lock(&head->srcu);
 313         ns = nvme_find_path(head);
 314         if (likely(ns)) {
 315                 bio->bi_disk = ns->disk;
 316                 bio->bi_opf |= REQ_NVME_MPATH;
 317                 trace_block_bio_remap(bio->bi_disk->queue, bio,
 318                                       disk_devt(ns->head->disk),
 319                                       bio->bi_iter.bi_sector);
 320                 ret = direct_make_request(bio);
 321         } else if (nvme_available_path(head)) {
 322                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 323
 324                 spin_lock_irq(&head->requeue_lock);
 325                 bio_list_add(&head->requeue_list, bio);
 326                 spin_unlock_irq(&head->requeue_lock);
 327         } else {
 328                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 329
 330                 bio->bi_status = BLK_STS_IOERR;
 331                 bio_endio(bio);
 332         }
 333
 334         srcu_read_unlock(&head->srcu, srcu_idx);
 335         return ret;
 336 }
 337
 338 static void nvme_requeue_work(struct work_struct *work)
 339 {
 340         struct nvme_ns_head *head =
 341                 container_of(work, struct nvme_ns_head, requeue_work);
 342         struct bio *bio, *next;
 343
 344         spin_lock_irq(&head->requeue_lock);
 345         next = bio_list_get(&head->requeue_list);
 346         spin_unlock_irq(&head->requeue_lock);
 347
 348         while ((bio = next) != NULL) {
 349                 next = bio->bi_next;
 350                 bio->bi_next = NULL;
 351
 352                 /*
 353                  * Reset disk to the mpath node and resubmit to select a new
 354                  * path.
 355                  */
 356                 bio->bi_disk = head->disk;
 357                 generic_make_request(bio);
 358         }
 359 }
 360
 361 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 362 {
 363         struct request_queue *q;
 364         bool vwc = false;
 365
 366         mutex_init(&head->lock);
 367         bio_list_init(&head->requeue_list);
 368         spin_lock_init(&head->requeue_lock);
 369         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 370
 371         /*
 372          * Add a multipath node if the subsystems supports multiple controllers.
 373          * We also do this for private namespaces as the namespace sharing data could
 374          * change after a rescan.
 375          */
 376         if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath)
 377                 return 0;
 378
 379         q = blk_alloc_queue_node(GFP_KERNEL, ctrl->numa_node);
 380         if (!q)
 381                 goto out;
 382         q->queuedata = head;
 383         blk_queue_make_request(q, nvme_ns_head_make_request);
 384         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 385         /* set to a default value for 512 until disk is validated */
 386         blk_queue_logical_block_size(q, 512);
 387         blk_set_stacking_limits(&q->limits);
 388
 389         /* we need to propagate up the VMC settings */
 390         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 391                 vwc = true;
 392         blk_queue_write_cache(q, vwc, vwc);
 393
 394         head->disk = alloc_disk(0);
 395         if (!head->disk)
 396                 goto out_cleanup_queue;
 397         head->disk->fops = &nvme_ns_head_ops;
 398         head->disk->private_data = head;
 399         head->disk->queue = q;
 400         head->disk->flags = GENHD_FL_EXT_DEVT;
 401         sprintf(head->disk->disk_name, "nvme%dn%d",
 402                         ctrl->subsys->instance, head->instance);
 403         return 0;
 404
 405 out_cleanup_queue:
 406         blk_cleanup_queue(q);
 407 out:
 408         return -ENOMEM;
 409 }
 410
 411 static void nvme_mpath_set_live(struct nvme_ns *ns)
 412 {
 413         struct nvme_ns_head *head = ns->head;
 414
 415         lockdep_assert_held(&ns->head->lock);
 416
 417         if (!head->disk)
 418                 return;
 419
 420         if (!(head->disk->flags & GENHD_FL_UP))
 421                 device_add_disk(&head->subsys->dev, head->disk,
 422                                 nvme_ns_id_attr_groups);
 423
 424         if (nvme_path_is_optimized(ns)) {
 425                 int node, srcu_idx;
 426
 427                 srcu_idx = srcu_read_lock(&head->srcu);
 428                 for_each_node(node)
 429                         __nvme_find_path(head, node);
 430                 srcu_read_unlock(&head->srcu, srcu_idx);
 431         }
 432
 433         synchronize_srcu(&ns->head->srcu);
 434         kblockd_schedule_work(&ns->head->requeue_work);
 435 }
 436
 437 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 438                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 439                         void *))
 440 {
 441         void *base = ctrl->ana_log_buf;
 442         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 443         int error, i;
 444
 445         lockdep_assert_held(&ctrl->ana_lock);
 446
 447         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 448                 struct nvme_ana_group_desc *desc = base + offset;
 449                 u32 nr_nsids = le32_to_cpu(desc->nnsids);
 450                 size_t nsid_buf_size = nr_nsids * sizeof(__le32);
 451
 452                 if (WARN_ON_ONCE(desc->grpid == 0))
 453                         return -EINVAL;
 454                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 455                         return -EINVAL;
 456                 if (WARN_ON_ONCE(desc->state == 0))
 457                         return -EINVAL;
 458                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 459                         return -EINVAL;
 460
 461                 offset += sizeof(*desc);
 462                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 463                         return -EINVAL;
 464
 465                 error = cb(ctrl, desc, data);
 466                 if (error)
 467                         return error;
 468
 469                 offset += nsid_buf_size;
 470                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 471                         return -EINVAL;
 472         }
 473
 474         return 0;
 475 }
 476
 477 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 478 {
 479         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 480 }
 481
 482 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 483                 struct nvme_ns *ns)
 484 {
 485         mutex_lock(&ns->head->lock);
 486         ns->ana_grpid = le32_to_cpu(desc->grpid);
 487         ns->ana_state = desc->state;
 488         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 489
 490         if (nvme_state_is_live(ns->ana_state))
 491                 nvme_mpath_set_live(ns);
 492         mutex_unlock(&ns->head->lock);
 493 }
 494
 495 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 496                 struct nvme_ana_group_desc *desc, void *data)
 497 {
 498         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 499         unsigned *nr_change_groups = data;
 500         struct nvme_ns *ns;
 501
 502         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 503                         le32_to_cpu(desc->grpid),
 504                         nvme_ana_state_names[desc->state]);
 505
 506         if (desc->state == NVME_ANA_CHANGE)
 507                 (*nr_change_groups)++;
 508
 509         if (!nr_nsids)
 510                 return 0;
 511
 512         down_write(&ctrl->namespaces_rwsem);
 513         list_for_each_entry(ns, &ctrl->namespaces, list) {
 514                 unsigned nsid = le32_to_cpu(desc->nsids[n]);
 515
 516                 if (ns->head->ns_id < nsid)
 517                         continue;
 518                 if (ns->head->ns_id == nsid)
 519                         nvme_update_ns_ana_state(desc, ns);
 520                 if (++n == nr_nsids)
 521                         break;
 522         }
 523         up_write(&ctrl->namespaces_rwsem);
 524         return 0;
 525 }
 526
 527 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 528 {
 529         u32 nr_change_groups = 0;
 530         int error;
 531
 532         mutex_lock(&ctrl->ana_lock);
 533         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0,
 534                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 535         if (error) {
 536                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 537                 goto out_unlock;
 538         }
 539
 540         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 541                         nvme_update_ana_state);
 542         if (error)
 543                 goto out_unlock;
 544
 545         /*
 546          * In theory we should have an ANATT timer per group as they might enter
 547          * the change state at different times.  But that is a lot of overhead
 548          * just to protect against a target that keeps entering new changes
 549          * states while never finishing previous ones.  But we'll still
 550          * eventually time out once all groups are in change state, so this
 551          * isn't a big deal.
 552          *
 553          * We also double the ANATT value to provide some slack for transports
 554          * or AEN processing overhead.
 555          */
 556         if (nr_change_groups)
 557                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 558         else
 559                 del_timer_sync(&ctrl->anatt_timer);
 560 out_unlock:
 561         mutex_unlock(&ctrl->ana_lock);
 562         return error;
 563 }
 564
 565 static void nvme_ana_work(struct work_struct *work)
 566 {
 567         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 568
 569         nvme_read_ana_log(ctrl);
 570 }
 571
 572 static void nvme_anatt_timeout(struct timer_list *t)
 573 {
 574         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 575
 576         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 577         nvme_reset_ctrl(ctrl);
 578 }
 579
 580 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 581 {
 582         if (!nvme_ctrl_use_ana(ctrl))
 583                 return;
 584         del_timer_sync(&ctrl->anatt_timer);
 585         cancel_work_sync(&ctrl->ana_work);
 586 }
 587
 588 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 589         struct device_attribute subsys_attr_##_name =   \
 590                 __ATTR(_name, _mode, _show, _store)
 591
 592 static const char *nvme_iopolicy_names[] = {
 593         [NVME_IOPOLICY_NUMA]    = "numa",
 594         [NVME_IOPOLICY_RR]      = "round-robin",
 595 };
 596
 597 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 598                 struct device_attribute *attr, char *buf)
 599 {
 600         struct nvme_subsystem *subsys =
 601                 container_of(dev, struct nvme_subsystem, dev);
 602
 603         return sprintf(buf, "%s\n",
 604                         nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 605 }
 606
 607 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 608                 struct device_attribute *attr, const char *buf, size_t count)
 609 {
 610         struct nvme_subsystem *subsys =
 611                 container_of(dev, struct nvme_subsystem, dev);
 612         int i;
 613
 614         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 615                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 616                         WRITE_ONCE(subsys->iopolicy, i);
 617                         return count;
 618                 }
 619         }
 620
 621         return -EINVAL;
 622 }
 623 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 624                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 625
 626 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 627                 char *buf)
 628 {
 629         return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 630 }
 631 DEVICE_ATTR_RO(ana_grpid);
 632
 633 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 634                 char *buf)
 635 {
 636         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 637
 638         return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 639 }
 640 DEVICE_ATTR_RO(ana_state);
 641
 642 static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl,
 643                 struct nvme_ana_group_desc *desc, void *data)
 644 {
 645         struct nvme_ns *ns = data;
 646
 647         if (ns->ana_grpid == le32_to_cpu(desc->grpid)) {
 648                 nvme_update_ns_ana_state(desc, ns);
 649                 return -ENXIO; /* just break out of the loop */
 650         }
 651
 652         return 0;
 653 }
 654
 655 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 656 {
 657         if (nvme_ctrl_use_ana(ns->ctrl)) {
 658                 mutex_lock(&ns->ctrl->ana_lock);
 659                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 660                 nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state);
 661                 mutex_unlock(&ns->ctrl->ana_lock);
 662         } else {
 663                 mutex_lock(&ns->head->lock);
 664                 ns->ana_state = NVME_ANA_OPTIMIZED;
 665                 nvme_mpath_set_live(ns);
 666                 mutex_unlock(&ns->head->lock);
 667         }
 668 }
 669
 670 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 671 {
 672         if (!head->disk)
 673                 return;
 674         if (head->disk->flags & GENHD_FL_UP)
 675                 del_gendisk(head->disk);
 676         blk_set_queue_dying(head->disk->queue);
 677         /* make sure all pending bios are cleaned up */
 678         kblockd_schedule_work(&head->requeue_work);
 679         flush_work(&head->requeue_work);
 680         blk_cleanup_queue(head->disk->queue);
 681         put_disk(head->disk);
 682 }
 683
 684 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 685 {
 686         int error;
 687
 688         /* check if multipath is enabled and we have the capability */
 689         if (!multipath || !ctrl->subsys || !(ctrl->subsys->cmic & (1 << 3)))
 690                 return 0;
 691
 692         ctrl->anacap = id->anacap;
 693         ctrl->anatt = id->anatt;
 694         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 695         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 696
 697         mutex_init(&ctrl->ana_lock);
 698         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 699         ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 700                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
 701         ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
 702
 703         if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
 704                 dev_err(ctrl->device,
 705                         "ANA log page size (%zd) larger than MDTS (%d).\n",
 706                         ctrl->ana_log_size,
 707                         ctrl->max_hw_sectors << SECTOR_SHIFT);
 708                 dev_err(ctrl->device, "disabling ANA support.\n");
 709                 return 0;
 710         }
 711
 712         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 713         ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
 714         if (!ctrl->ana_log_buf) {
 715                 error = -ENOMEM;
 716                 goto out;
 717         }
 718
 719         error = nvme_read_ana_log(ctrl);
 720         if (error)
 721                 goto out_free_ana_log_buf;
 722         return 0;
 723 out_free_ana_log_buf:
 724         kfree(ctrl->ana_log_buf);
 725         ctrl->ana_log_buf = NULL;
 726 out:
 727         return error;
 728 }
 729
 730 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 731 {
 732         kfree(ctrl->ana_log_buf);
 733         ctrl->ana_log_buf = NULL;
 734 }
 735