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