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         case NVME_SC_HOST_ABORTED_CMD:
  99                 /*
 100                  * Temporary transport disruption in talking to the controller.
 101                  * Try to send on a new path.
 102                  */
 103                 nvme_mpath_clear_current_path(ns);
 104                 break;
 105         default:
 106                 /*
 107                  * Reset the controller for any non-ANA error as we don't know
 108                  * what caused the error.
 109                  */
 110                 nvme_reset_ctrl(ns->ctrl);
 111                 break;
 112         }
 113
 114         kblockd_schedule_work(&ns->head->requeue_work);
 115 }
 116
 117 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 118 {
 119         struct nvme_ns *ns;
 120
 121         down_read(&ctrl->namespaces_rwsem);
 122         list_for_each_entry(ns, &ctrl->namespaces, list) {
 123                 if (ns->head->disk)
 124                         kblockd_schedule_work(&ns->head->requeue_work);
 125         }
 126         up_read(&ctrl->namespaces_rwsem);
 127 }
 128
 129 static const char *nvme_ana_state_names[] = {
 130         [0]                             = "invalid state",
 131         [NVME_ANA_OPTIMIZED]            = "optimized",
 132         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 133         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 134         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 135         [NVME_ANA_CHANGE]               = "change",
 136 };
 137
 138 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 139 {
 140         struct nvme_ns_head *head = ns->head;
 141         bool changed = false;
 142         int node;
 143
 144         if (!head)
 145                 goto out;
 146
 147         for_each_node(node) {
 148                 if (ns == rcu_access_pointer(head->current_path[node])) {
 149                         rcu_assign_pointer(head->current_path[node], NULL);
 150                         changed = true;
 151                 }
 152         }
 153 out:
 154         return changed;
 155 }
 156
 157 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 158 {
 159         struct nvme_ns *ns;
 160
 161         mutex_lock(&ctrl->scan_lock);
 162         down_read(&ctrl->namespaces_rwsem);
 163         list_for_each_entry(ns, &ctrl->namespaces, list)
 164                 if (nvme_mpath_clear_current_path(ns))
 165                         kblockd_schedule_work(&ns->head->requeue_work);
 166         up_read(&ctrl->namespaces_rwsem);
 167         mutex_unlock(&ctrl->scan_lock);
 168 }
 169
 170 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 171 {
 172         return ns->ctrl->state != NVME_CTRL_LIVE ||
 173                 test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 174                 test_bit(NVME_NS_REMOVING, &ns->flags);
 175 }
 176
 177 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 178 {
 179         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 180         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 181
 182         list_for_each_entry_rcu(ns, &head->list, siblings) {
 183                 if (nvme_path_is_disabled(ns))
 184                         continue;
 185
 186                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 187                         distance = node_distance(node, ns->ctrl->numa_node);
 188                 else
 189                         distance = LOCAL_DISTANCE;
 190
 191                 switch (ns->ana_state) {
 192                 case NVME_ANA_OPTIMIZED:
 193                         if (distance < found_distance) {
 194                                 found_distance = distance;
 195                                 found = ns;
 196                         }
 197                         break;
 198                 case NVME_ANA_NONOPTIMIZED:
 199                         if (distance < fallback_distance) {
 200                                 fallback_distance = distance;
 201                                 fallback = ns;
 202                         }
 203                         break;
 204                 default:
 205                         break;
 206                 }
 207         }
 208
 209         if (!found)
 210                 found = fallback;
 211         if (found)
 212                 rcu_assign_pointer(head->current_path[node], found);
 213         return found;
 214 }
 215
 216 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 217                 struct nvme_ns *ns)
 218 {
 219         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 220                         siblings);
 221         if (ns)
 222                 return ns;
 223         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 224 }
 225
 226 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 227                 int node, struct nvme_ns *old)
 228 {
 229         struct nvme_ns *ns, *found, *fallback = NULL;
 230
 231         if (list_is_singular(&head->list)) {
 232                 if (nvme_path_is_disabled(old))
 233                         return NULL;
 234                 return old;
 235         }
 236
 237         for (ns = nvme_next_ns(head, old);
 238              ns != old;
 239              ns = nvme_next_ns(head, ns)) {
 240                 if (nvme_path_is_disabled(ns))
 241                         continue;
 242
 243                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 244                         found = ns;
 245                         goto out;
 246                 }
 247                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 248                         fallback = ns;
 249         }
 250
 251         if (!fallback)
 252                 return NULL;
 253         found = fallback;
 254 out:
 255         rcu_assign_pointer(head->current_path[node], found);
 256         return found;
 257 }
 258
 259 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 260 {
 261         return ns->ctrl->state == NVME_CTRL_LIVE &&
 262                 ns->ana_state == NVME_ANA_OPTIMIZED;
 263 }
 264
 265 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 266 {
 267         int node = numa_node_id();
 268         struct nvme_ns *ns;
 269
 270         ns = srcu_dereference(head->current_path[node], &head->srcu);
 271         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns)
 272                 ns = nvme_round_robin_path(head, node, ns);
 273         if (unlikely(!ns || !nvme_path_is_optimized(ns)))
 274                 ns = __nvme_find_path(head, node);
 275         return ns;
 276 }
 277
 278 static bool nvme_available_path(struct nvme_ns_head *head)
 279 {
 280         struct nvme_ns *ns;
 281
 282         list_for_each_entry_rcu(ns, &head->list, siblings) {
 283                 switch (ns->ctrl->state) {
 284                 case NVME_CTRL_LIVE:
 285                 case NVME_CTRL_RESETTING:
 286                 case NVME_CTRL_CONNECTING:
 287                         /* fallthru */
 288                         return true;
 289                 default:
 290                         break;
 291                 }
 292         }
 293         return false;
 294 }
 295
 296 static blk_qc_t nvme_ns_head_make_request(struct request_queue *q,
 297                 struct bio *bio)
 298 {
 299         struct nvme_ns_head *head = q->queuedata;
 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
 307          * be moved to a different queue via blk_steal_bios(),
 308          * so we need to use the bio_split pool from the original
 309          * queue to allocate the bvecs from.
 310          */
 311         blk_queue_split(q, &bio);
 312
 313         srcu_idx = srcu_read_lock(&head->srcu);
 314         ns = nvme_find_path(head);
 315         if (likely(ns)) {
 316                 bio->bi_disk = ns->disk;
 317                 bio->bi_opf |= REQ_NVME_MPATH;
 318                 trace_block_bio_remap(bio->bi_disk->queue, bio,
 319                                       disk_devt(ns->head->disk),
 320                                       bio->bi_iter.bi_sector);
 321                 ret = direct_make_request(bio);
 322         } else if (nvme_available_path(head)) {
 323                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 324
 325                 spin_lock_irq(&head->requeue_lock);
 326                 bio_list_add(&head->requeue_list, bio);
 327                 spin_unlock_irq(&head->requeue_lock);
 328         } else {
 329                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 330
 331                 bio->bi_status = BLK_STS_IOERR;
 332                 bio_endio(bio);
 333         }
 334
 335         srcu_read_unlock(&head->srcu, srcu_idx);
 336         return ret;
 337 }
 338
 339 static void nvme_requeue_work(struct work_struct *work)
 340 {
 341         struct nvme_ns_head *head =
 342                 container_of(work, struct nvme_ns_head, requeue_work);
 343         struct bio *bio, *next;
 344
 345         spin_lock_irq(&head->requeue_lock);
 346         next = bio_list_get(&head->requeue_list);
 347         spin_unlock_irq(&head->requeue_lock);
 348
 349         while ((bio = next) != NULL) {
 350                 next = bio->bi_next;
 351                 bio->bi_next = NULL;
 352
 353                 /*
 354                  * Reset disk to the mpath node and resubmit to select a new
 355                  * path.
 356                  */
 357                 bio->bi_disk = head->disk;
 358                 generic_make_request(bio);
 359         }
 360 }
 361
 362 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 363 {
 364         struct request_queue *q;
 365         bool vwc = false;
 366
 367         mutex_init(&head->lock);
 368         bio_list_init(&head->requeue_list);
 369         spin_lock_init(&head->requeue_lock);
 370         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 371
 372         /*
 373          * Add a multipath node if the subsystems supports multiple controllers.
 374          * We also do this for private namespaces as the namespace sharing data could
 375          * change after a rescan.
 376          */
 377         if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath)
 378                 return 0;
 379
 380         q = blk_alloc_queue_node(GFP_KERNEL, ctrl->numa_node);
 381         if (!q)
 382                 goto out;
 383         q->queuedata = head;
 384         blk_queue_make_request(q, nvme_ns_head_make_request);
 385         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 386         /* set to a default value for 512 until disk is validated */
 387         blk_queue_logical_block_size(q, 512);
 388         blk_set_stacking_limits(&q->limits);
 389
 390         /* we need to propagate up the VMC settings */
 391         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 392                 vwc = true;
 393         blk_queue_write_cache(q, vwc, vwc);
 394
 395         head->disk = alloc_disk(0);
 396         if (!head->disk)
 397                 goto out_cleanup_queue;
 398         head->disk->fops = &nvme_ns_head_ops;
 399         head->disk->private_data = head;
 400         head->disk->queue = q;
 401         head->disk->flags = GENHD_FL_EXT_DEVT;
 402         sprintf(head->disk->disk_name, "nvme%dn%d",
 403                         ctrl->subsys->instance, head->instance);
 404         return 0;
 405
 406 out_cleanup_queue:
 407         blk_cleanup_queue(q);
 408 out:
 409         return -ENOMEM;
 410 }
 411
 412 static void nvme_mpath_set_live(struct nvme_ns *ns)
 413 {
 414         struct nvme_ns_head *head = ns->head;
 415
 416         lockdep_assert_held(&ns->head->lock);
 417
 418         if (!head->disk)
 419                 return;
 420
 421         if (!(head->disk->flags & GENHD_FL_UP))
 422                 device_add_disk(&head->subsys->dev, head->disk,
 423                                 nvme_ns_id_attr_groups);
 424
 425         if (nvme_path_is_optimized(ns)) {
 426                 int node, srcu_idx;
 427
 428                 srcu_idx = srcu_read_lock(&head->srcu);
 429                 for_each_node(node)
 430                         __nvme_find_path(head, node);
 431                 srcu_read_unlock(&head->srcu, srcu_idx);
 432         }
 433
 434         synchronize_srcu(&ns->head->srcu);
 435         kblockd_schedule_work(&ns->head->requeue_work);
 436 }
 437
 438 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 439                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 440                         void *))
 441 {
 442         void *base = ctrl->ana_log_buf;
 443         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 444         int error, i;
 445
 446         lockdep_assert_held(&ctrl->ana_lock);
 447
 448         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 449                 struct nvme_ana_group_desc *desc = base + offset;
 450                 u32 nr_nsids;
 451                 size_t nsid_buf_size;
 452
 453                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 454                         return -EINVAL;
 455
 456                 nr_nsids = le32_to_cpu(desc->nnsids);
 457                 nsid_buf_size = nr_nsids * sizeof(__le32);
 458
 459                 if (WARN_ON_ONCE(desc->grpid == 0))
 460                         return -EINVAL;
 461                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 462                         return -EINVAL;
 463                 if (WARN_ON_ONCE(desc->state == 0))
 464                         return -EINVAL;
 465                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 466                         return -EINVAL;
 467
 468                 offset += sizeof(*desc);
 469                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 470                         return -EINVAL;
 471
 472                 error = cb(ctrl, desc, data);
 473                 if (error)
 474                         return error;
 475
 476                 offset += nsid_buf_size;
 477         }
 478
 479         return 0;
 480 }
 481
 482 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 483 {
 484         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 485 }
 486
 487 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 488                 struct nvme_ns *ns)
 489 {
 490         mutex_lock(&ns->head->lock);
 491         ns->ana_grpid = le32_to_cpu(desc->grpid);
 492         ns->ana_state = desc->state;
 493         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 494
 495         if (nvme_state_is_live(ns->ana_state))
 496                 nvme_mpath_set_live(ns);
 497         mutex_unlock(&ns->head->lock);
 498 }
 499
 500 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 501                 struct nvme_ana_group_desc *desc, void *data)
 502 {
 503         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 504         unsigned *nr_change_groups = data;
 505         struct nvme_ns *ns;
 506
 507         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 508                         le32_to_cpu(desc->grpid),
 509                         nvme_ana_state_names[desc->state]);
 510
 511         if (desc->state == NVME_ANA_CHANGE)
 512                 (*nr_change_groups)++;
 513
 514         if (!nr_nsids)
 515                 return 0;
 516
 517         down_write(&ctrl->namespaces_rwsem);
 518         list_for_each_entry(ns, &ctrl->namespaces, list) {
 519                 unsigned nsid = le32_to_cpu(desc->nsids[n]);
 520
 521                 if (ns->head->ns_id < nsid)
 522                         continue;
 523                 if (ns->head->ns_id == nsid)
 524                         nvme_update_ns_ana_state(desc, ns);
 525                 if (++n == nr_nsids)
 526                         break;
 527         }
 528         up_write(&ctrl->namespaces_rwsem);
 529         return 0;
 530 }
 531
 532 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 533 {
 534         u32 nr_change_groups = 0;
 535         int error;
 536
 537         mutex_lock(&ctrl->ana_lock);
 538         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0,
 539                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 540         if (error) {
 541                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 542                 goto out_unlock;
 543         }
 544
 545         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 546                         nvme_update_ana_state);
 547         if (error)
 548                 goto out_unlock;
 549
 550         /*
 551          * In theory we should have an ANATT timer per group as they might enter
 552          * the change state at different times.  But that is a lot of overhead
 553          * just to protect against a target that keeps entering new changes
 554          * states while never finishing previous ones.  But we'll still
 555          * eventually time out once all groups are in change state, so this
 556          * isn't a big deal.
 557          *
 558          * We also double the ANATT value to provide some slack for transports
 559          * or AEN processing overhead.
 560          */
 561         if (nr_change_groups)
 562                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 563         else
 564                 del_timer_sync(&ctrl->anatt_timer);
 565 out_unlock:
 566         mutex_unlock(&ctrl->ana_lock);
 567         return error;
 568 }
 569
 570 static void nvme_ana_work(struct work_struct *work)
 571 {
 572         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 573
 574         nvme_read_ana_log(ctrl);
 575 }
 576
 577 static void nvme_anatt_timeout(struct timer_list *t)
 578 {
 579         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 580
 581         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 582         nvme_reset_ctrl(ctrl);
 583 }
 584
 585 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 586 {
 587         if (!nvme_ctrl_use_ana(ctrl))
 588                 return;
 589         del_timer_sync(&ctrl->anatt_timer);
 590         cancel_work_sync(&ctrl->ana_work);
 591 }
 592
 593 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 594         struct device_attribute subsys_attr_##_name =   \
 595                 __ATTR(_name, _mode, _show, _store)
 596
 597 static const char *nvme_iopolicy_names[] = {
 598         [NVME_IOPOLICY_NUMA]    = "numa",
 599         [NVME_IOPOLICY_RR]      = "round-robin",
 600 };
 601
 602 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 603                 struct device_attribute *attr, char *buf)
 604 {
 605         struct nvme_subsystem *subsys =
 606                 container_of(dev, struct nvme_subsystem, dev);
 607
 608         return sprintf(buf, "%s\n",
 609                         nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 610 }
 611
 612 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 613                 struct device_attribute *attr, const char *buf, size_t count)
 614 {
 615         struct nvme_subsystem *subsys =
 616                 container_of(dev, struct nvme_subsystem, dev);
 617         int i;
 618
 619         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 620                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 621                         WRITE_ONCE(subsys->iopolicy, i);
 622                         return count;
 623                 }
 624         }
 625
 626         return -EINVAL;
 627 }
 628 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 629                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 630
 631 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 632                 char *buf)
 633 {
 634         return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 635 }
 636 DEVICE_ATTR_RO(ana_grpid);
 637
 638 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 639                 char *buf)
 640 {
 641         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 642
 643         return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 644 }
 645 DEVICE_ATTR_RO(ana_state);
 646
 647 static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl,
 648                 struct nvme_ana_group_desc *desc, void *data)
 649 {
 650         struct nvme_ns *ns = data;
 651
 652         if (ns->ana_grpid == le32_to_cpu(desc->grpid)) {
 653                 nvme_update_ns_ana_state(desc, ns);
 654                 return -ENXIO; /* just break out of the loop */
 655         }
 656
 657         return 0;
 658 }
 659
 660 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 661 {
 662         if (nvme_ctrl_use_ana(ns->ctrl)) {
 663                 mutex_lock(&ns->ctrl->ana_lock);
 664                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 665                 nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state);
 666                 mutex_unlock(&ns->ctrl->ana_lock);
 667         } else {
 668                 mutex_lock(&ns->head->lock);
 669                 ns->ana_state = NVME_ANA_OPTIMIZED;
 670                 nvme_mpath_set_live(ns);
 671                 mutex_unlock(&ns->head->lock);
 672         }
 673 }
 674
 675 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 676 {
 677         if (!head->disk)
 678                 return;
 679         if (head->disk->flags & GENHD_FL_UP)
 680                 del_gendisk(head->disk);
 681         blk_set_queue_dying(head->disk->queue);
 682         /* make sure all pending bios are cleaned up */
 683         kblockd_schedule_work(&head->requeue_work);
 684         flush_work(&head->requeue_work);
 685         blk_cleanup_queue(head->disk->queue);
 686         put_disk(head->disk);
 687 }
 688
 689 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 690 {
 691         int error;
 692
 693         /* check if multipath is enabled and we have the capability */
 694         if (!multipath || !ctrl->subsys || !(ctrl->subsys->cmic & (1 << 3)))
 695                 return 0;
 696
 697         ctrl->anacap = id->anacap;
 698         ctrl->anatt = id->anatt;
 699         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 700         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 701
 702         mutex_init(&ctrl->ana_lock);
 703         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 704         ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 705                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
 706         ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
 707
 708         if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
 709                 dev_err(ctrl->device,
 710                         "ANA log page size (%zd) larger than MDTS (%d).\n",
 711                         ctrl->ana_log_size,
 712                         ctrl->max_hw_sectors << SECTOR_SHIFT);
 713                 dev_err(ctrl->device, "disabling ANA support.\n");
 714                 return 0;
 715         }
 716
 717         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 718         ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
 719         if (!ctrl->ana_log_buf) {
 720                 error = -ENOMEM;
 721                 goto out;
 722         }
 723
 724         error = nvme_read_ana_log(ctrl);
 725         if (error)
 726                 goto out_free_ana_log_buf;
 727         return 0;
 728 out_free_ana_log_buf:
 729         kfree(ctrl->ana_log_buf);
 730         ctrl->ana_log_buf = NULL;
 731 out:
 732         return error;
 733 }
 734
 735 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 736 {
 737         kfree(ctrl->ana_log_buf);
 738         ctrl->ana_log_buf = NULL;
 739 }
 740