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         return ns->ctrl->state != NVME_CTRL_LIVE ||
 171                 test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 172                 test_bit(NVME_NS_REMOVING, &ns->flags);
 173 }
 174
 175 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 176 {
 177         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 178         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 179
 180         list_for_each_entry_rcu(ns, &head->list, siblings) {
 181                 if (nvme_path_is_disabled(ns))
 182                         continue;
 183
 184                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 185                         distance = node_distance(node, ns->ctrl->numa_node);
 186                 else
 187                         distance = LOCAL_DISTANCE;
 188
 189                 switch (ns->ana_state) {
 190                 case NVME_ANA_OPTIMIZED:
 191                         if (distance < found_distance) {
 192                                 found_distance = distance;
 193                                 found = ns;
 194                         }
 195                         break;
 196                 case NVME_ANA_NONOPTIMIZED:
 197                         if (distance < fallback_distance) {
 198                                 fallback_distance = distance;
 199                                 fallback = ns;
 200                         }
 201                         break;
 202                 default:
 203                         break;
 204                 }
 205         }
 206
 207         if (!found)
 208                 found = fallback;
 209         if (found)
 210                 rcu_assign_pointer(head->current_path[node], found);
 211         return found;
 212 }
 213
 214 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 215                 struct nvme_ns *ns)
 216 {
 217         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 218                         siblings);
 219         if (ns)
 220                 return ns;
 221         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 222 }
 223
 224 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 225                 int node, struct nvme_ns *old)
 226 {
 227         struct nvme_ns *ns, *found, *fallback = NULL;
 228
 229         if (list_is_singular(&head->list)) {
 230                 if (nvme_path_is_disabled(old))
 231                         return NULL;
 232                 return old;
 233         }
 234
 235         for (ns = nvme_next_ns(head, old);
 236              ns != old;
 237              ns = nvme_next_ns(head, ns)) {
 238                 if (nvme_path_is_disabled(ns))
 239                         continue;
 240
 241                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 242                         found = ns;
 243                         goto out;
 244                 }
 245                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 246                         fallback = ns;
 247         }
 248
 249         if (!fallback)
 250                 return NULL;
 251         found = fallback;
 252 out:
 253         rcu_assign_pointer(head->current_path[node], found);
 254         return found;
 255 }
 256
 257 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 258 {
 259         return ns->ctrl->state == NVME_CTRL_LIVE &&
 260                 ns->ana_state == NVME_ANA_OPTIMIZED;
 261 }
 262
 263 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 264 {
 265         int node = numa_node_id();
 266         struct nvme_ns *ns;
 267
 268         ns = srcu_dereference(head->current_path[node], &head->srcu);
 269         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns)
 270                 ns = nvme_round_robin_path(head, node, ns);
 271         if (unlikely(!ns || !nvme_path_is_optimized(ns)))
 272                 ns = __nvme_find_path(head, node);
 273         return ns;
 274 }
 275
 276 static bool nvme_available_path(struct nvme_ns_head *head)
 277 {
 278         struct nvme_ns *ns;
 279
 280         list_for_each_entry_rcu(ns, &head->list, siblings) {
 281                 switch (ns->ctrl->state) {
 282                 case NVME_CTRL_LIVE:
 283                 case NVME_CTRL_RESETTING:
 284                 case NVME_CTRL_CONNECTING:
 285                         /* fallthru */
 286                         return true;
 287                 default:
 288                         break;
 289                 }
 290         }
 291         return false;
 292 }
 293
 294 blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
 295 {
 296         struct nvme_ns_head *head = bio->bi_disk->private_data;
 297         struct device *dev = disk_to_dev(head->disk);
 298         struct nvme_ns *ns;
 299         blk_qc_t ret = BLK_QC_T_NONE;
 300         int srcu_idx;
 301
 302         /*
 303          * The namespace might be going away and the bio might be moved to a
 304          * different queue via blk_steal_bios(), so we need to use the bio_split
 305          * pool from the original queue to allocate the bvecs from.
 306          */
 307         blk_queue_split(&bio);
 308
 309         srcu_idx = srcu_read_lock(&head->srcu);
 310         ns = nvme_find_path(head);
 311         if (likely(ns)) {
 312                 bio->bi_disk = ns->disk;
 313                 bio->bi_opf |= REQ_NVME_MPATH;
 314                 trace_block_bio_remap(bio->bi_disk->queue, bio,
 315                                       disk_devt(ns->head->disk),
 316                                       bio->bi_iter.bi_sector);
 317                 ret = submit_bio_noacct(bio);
 318         } else if (nvme_available_path(head)) {
 319                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 320
 321                 spin_lock_irq(&head->requeue_lock);
 322                 bio_list_add(&head->requeue_list, bio);
 323                 spin_unlock_irq(&head->requeue_lock);
 324         } else {
 325                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 326
 327                 bio->bi_status = BLK_STS_IOERR;
 328                 bio_endio(bio);
 329         }
 330
 331         srcu_read_unlock(&head->srcu, srcu_idx);
 332         return ret;
 333 }
 334
 335 static void nvme_requeue_work(struct work_struct *work)
 336 {
 337         struct nvme_ns_head *head =
 338                 container_of(work, struct nvme_ns_head, requeue_work);
 339         struct bio *bio, *next;
 340
 341         spin_lock_irq(&head->requeue_lock);
 342         next = bio_list_get(&head->requeue_list);
 343         spin_unlock_irq(&head->requeue_lock);
 344
 345         while ((bio = next) != NULL) {
 346                 next = bio->bi_next;
 347                 bio->bi_next = NULL;
 348
 349                 /*
 350                  * Reset disk to the mpath node and resubmit to select a new
 351                  * path.
 352                  */
 353                 bio->bi_disk = head->disk;
 354                 submit_bio_noacct(bio);
 355         }
 356 }
 357
 358 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 359 {
 360         struct request_queue *q;
 361         bool vwc = false;
 362
 363         mutex_init(&head->lock);
 364         bio_list_init(&head->requeue_list);
 365         spin_lock_init(&head->requeue_lock);
 366         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 367
 368         /*
 369          * Add a multipath node if the subsystems supports multiple controllers.
 370          * We also do this for private namespaces as the namespace sharing data could
 371          * change after a rescan.
 372          */
 373         if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) || !multipath)
 374                 return 0;
 375
 376         q = blk_alloc_queue(ctrl->numa_node);
 377         if (!q)
 378                 goto out;
 379         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 380         /* set to a default value for 512 until disk is validated */
 381         blk_queue_logical_block_size(q, 512);
 382         blk_set_stacking_limits(&q->limits);
 383
 384         /* we need to propagate up the VMC settings */
 385         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 386                 vwc = true;
 387         blk_queue_write_cache(q, vwc, vwc);
 388
 389         head->disk = alloc_disk(0);
 390         if (!head->disk)
 391                 goto out_cleanup_queue;
 392         head->disk->fops = &nvme_ns_head_ops;
 393         head->disk->private_data = head;
 394         head->disk->queue = q;
 395         head->disk->flags = GENHD_FL_EXT_DEVT;
 396         sprintf(head->disk->disk_name, "nvme%dn%d",
 397                         ctrl->subsys->instance, head->instance);
 398         return 0;
 399
 400 out_cleanup_queue:
 401         blk_cleanup_queue(q);
 402 out:
 403         return -ENOMEM;
 404 }
 405
 406 static void nvme_mpath_set_live(struct nvme_ns *ns)
 407 {
 408         struct nvme_ns_head *head = ns->head;
 409
 410         if (!head->disk)
 411                 return;
 412
 413         if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags))
 414                 device_add_disk(&head->subsys->dev, head->disk,
 415                                 nvme_ns_id_attr_groups);
 416
 417         mutex_lock(&head->lock);
 418         if (nvme_path_is_optimized(ns)) {
 419                 int node, srcu_idx;
 420
 421                 srcu_idx = srcu_read_lock(&head->srcu);
 422                 for_each_node(node)
 423                         __nvme_find_path(head, node);
 424                 srcu_read_unlock(&head->srcu, srcu_idx);
 425         }
 426         mutex_unlock(&head->lock);
 427
 428         synchronize_srcu(&head->srcu);
 429         kblockd_schedule_work(&head->requeue_work);
 430 }
 431
 432 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 433                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 434                         void *))
 435 {
 436         void *base = ctrl->ana_log_buf;
 437         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 438         int error, i;
 439
 440         lockdep_assert_held(&ctrl->ana_lock);
 441
 442         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 443                 struct nvme_ana_group_desc *desc = base + offset;
 444                 u32 nr_nsids;
 445                 size_t nsid_buf_size;
 446
 447                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 448                         return -EINVAL;
 449
 450                 nr_nsids = le32_to_cpu(desc->nnsids);
 451                 nsid_buf_size = nr_nsids * sizeof(__le32);
 452
 453                 if (WARN_ON_ONCE(desc->grpid == 0))
 454                         return -EINVAL;
 455                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 456                         return -EINVAL;
 457                 if (WARN_ON_ONCE(desc->state == 0))
 458                         return -EINVAL;
 459                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 460                         return -EINVAL;
 461
 462                 offset += sizeof(*desc);
 463                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 464                         return -EINVAL;
 465
 466                 error = cb(ctrl, desc, data);
 467                 if (error)
 468                         return error;
 469
 470                 offset += nsid_buf_size;
 471         }
 472
 473         return 0;
 474 }
 475
 476 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 477 {
 478         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 479 }
 480
 481 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 482                 struct nvme_ns *ns)
 483 {
 484         ns->ana_grpid = le32_to_cpu(desc->grpid);
 485         ns->ana_state = desc->state;
 486         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 487
 488         if (nvme_state_is_live(ns->ana_state))
 489                 nvme_mpath_set_live(ns);
 490 }
 491
 492 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 493                 struct nvme_ana_group_desc *desc, void *data)
 494 {
 495         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 496         unsigned *nr_change_groups = data;
 497         struct nvme_ns *ns;
 498
 499         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 500                         le32_to_cpu(desc->grpid),
 501                         nvme_ana_state_names[desc->state]);
 502
 503         if (desc->state == NVME_ANA_CHANGE)
 504                 (*nr_change_groups)++;
 505
 506         if (!nr_nsids)
 507                 return 0;
 508
 509         down_read(&ctrl->namespaces_rwsem);
 510         list_for_each_entry(ns, &ctrl->namespaces, list) {
 511                 unsigned nsid = le32_to_cpu(desc->nsids[n]);
 512
 513                 if (ns->head->ns_id < nsid)
 514                         continue;
 515                 if (ns->head->ns_id == nsid)
 516                         nvme_update_ns_ana_state(desc, ns);
 517                 if (++n == nr_nsids)
 518                         break;
 519         }
 520         up_read(&ctrl->namespaces_rwsem);
 521         return 0;
 522 }
 523
 524 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 525 {
 526         u32 nr_change_groups = 0;
 527         int error;
 528
 529         mutex_lock(&ctrl->ana_lock);
 530         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0,
 531                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 532         if (error) {
 533                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 534                 goto out_unlock;
 535         }
 536
 537         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 538                         nvme_update_ana_state);
 539         if (error)
 540                 goto out_unlock;
 541
 542         /*
 543          * In theory we should have an ANATT timer per group as they might enter
 544          * the change state at different times.  But that is a lot of overhead
 545          * just to protect against a target that keeps entering new changes
 546          * states while never finishing previous ones.  But we'll still
 547          * eventually time out once all groups are in change state, so this
 548          * isn't a big deal.
 549          *
 550          * We also double the ANATT value to provide some slack for transports
 551          * or AEN processing overhead.
 552          */
 553         if (nr_change_groups)
 554                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 555         else
 556                 del_timer_sync(&ctrl->anatt_timer);
 557 out_unlock:
 558         mutex_unlock(&ctrl->ana_lock);
 559         return error;
 560 }
 561
 562 static void nvme_ana_work(struct work_struct *work)
 563 {
 564         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 565
 566         nvme_read_ana_log(ctrl);
 567 }
 568
 569 static void nvme_anatt_timeout(struct timer_list *t)
 570 {
 571         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 572
 573         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 574         nvme_reset_ctrl(ctrl);
 575 }
 576
 577 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 578 {
 579         if (!nvme_ctrl_use_ana(ctrl))
 580                 return;
 581         del_timer_sync(&ctrl->anatt_timer);
 582         cancel_work_sync(&ctrl->ana_work);
 583 }
 584
 585 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 586         struct device_attribute subsys_attr_##_name =   \
 587                 __ATTR(_name, _mode, _show, _store)
 588
 589 static const char *nvme_iopolicy_names[] = {
 590         [NVME_IOPOLICY_NUMA]    = "numa",
 591         [NVME_IOPOLICY_RR]      = "round-robin",
 592 };
 593
 594 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 595                 struct device_attribute *attr, char *buf)
 596 {
 597         struct nvme_subsystem *subsys =
 598                 container_of(dev, struct nvme_subsystem, dev);
 599
 600         return sprintf(buf, "%s\n",
 601                         nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 602 }
 603
 604 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 605                 struct device_attribute *attr, const char *buf, size_t count)
 606 {
 607         struct nvme_subsystem *subsys =
 608                 container_of(dev, struct nvme_subsystem, dev);
 609         int i;
 610
 611         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 612                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 613                         WRITE_ONCE(subsys->iopolicy, i);
 614                         return count;
 615                 }
 616         }
 617
 618         return -EINVAL;
 619 }
 620 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 621                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 622
 623 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 624                 char *buf)
 625 {
 626         return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 627 }
 628 DEVICE_ATTR_RO(ana_grpid);
 629
 630 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 631                 char *buf)
 632 {
 633         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 634
 635         return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 636 }
 637 DEVICE_ATTR_RO(ana_state);
 638
 639 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
 640                 struct nvme_ana_group_desc *desc, void *data)
 641 {
 642         struct nvme_ana_group_desc *dst = data;
 643
 644         if (desc->grpid != dst->grpid)
 645                 return 0;
 646
 647         *dst = *desc;
 648         return -ENXIO; /* just break out of the loop */
 649 }
 650
 651 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 652 {
 653         if (nvme_ctrl_use_ana(ns->ctrl)) {
 654                 struct nvme_ana_group_desc desc = {
 655                         .grpid = id->anagrpid,
 656                         .state = 0,
 657                 };
 658
 659                 mutex_lock(&ns->ctrl->ana_lock);
 660                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 661                 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
 662                 mutex_unlock(&ns->ctrl->ana_lock);
 663                 if (desc.state) {
 664                         /* found the group desc: update */
 665                         nvme_update_ns_ana_state(&desc, ns);
 666                 }
 667         } else {
 668                 ns->ana_state = NVME_ANA_OPTIMIZED;
 669                 nvme_mpath_set_live(ns);
 670         }
 671
 672         if (bdi_cap_stable_pages_required(ns->queue->backing_dev_info)) {
 673                 struct gendisk *disk = ns->head->disk;
 674
 675                 if (disk)
 676                         disk->queue->backing_dev_info->capabilities |=
 677                                         BDI_CAP_STABLE_WRITES;
 678         }
 679 }
 680
 681 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 682 {
 683         if (!head->disk)
 684                 return;
 685         if (head->disk->flags & GENHD_FL_UP)
 686                 del_gendisk(head->disk);
 687         blk_set_queue_dying(head->disk->queue);
 688         /* make sure all pending bios are cleaned up */
 689         kblockd_schedule_work(&head->requeue_work);
 690         flush_work(&head->requeue_work);
 691         blk_cleanup_queue(head->disk->queue);
 692         if (!test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 693                 /*
 694                  * if device_add_disk wasn't called, prevent
 695                  * disk release to put a bogus reference on the
 696                  * request queue
 697                  */
 698                 head->disk->queue = NULL;
 699         }
 700         put_disk(head->disk);
 701 }
 702
 703 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 704 {
 705         int error;
 706
 707         /* check if multipath is enabled and we have the capability */
 708         if (!multipath || !ctrl->subsys ||
 709             !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
 710                 return 0;
 711
 712         ctrl->anacap = id->anacap;
 713         ctrl->anatt = id->anatt;
 714         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 715         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 716
 717         mutex_init(&ctrl->ana_lock);
 718         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 719         ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 720                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
 721         ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
 722
 723         if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
 724                 dev_err(ctrl->device,
 725                         "ANA log page size (%zd) larger than MDTS (%d).\n",
 726                         ctrl->ana_log_size,
 727                         ctrl->max_hw_sectors << SECTOR_SHIFT);
 728                 dev_err(ctrl->device, "disabling ANA support.\n");
 729                 return 0;
 730         }
 731
 732         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 733         kfree(ctrl->ana_log_buf);
 734         ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
 735         if (!ctrl->ana_log_buf) {
 736                 error = -ENOMEM;
 737                 goto out;
 738         }
 739
 740         error = nvme_read_ana_log(ctrl);
 741         if (error)
 742                 goto out_free_ana_log_buf;
 743         return 0;
 744 out_free_ana_log_buf:
 745         kfree(ctrl->ana_log_buf);
 746         ctrl->ana_log_buf = NULL;
 747 out:
 748         return error;
 749 }
 750
 751 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 752 {
 753         kfree(ctrl->ana_log_buf);
 754         ctrl->ana_log_buf = NULL;
 755 }
 756