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 bool nvme_mpath_set_disk_name(struct nvme_ns *ns, char *disk_name, int *flags)
  54 {
  55         if (!multipath)
  56                 return false;
  57         if (!ns->head->disk) {
  58                 sprintf(disk_name, "nvme%dn%d", ns->ctrl->subsys->instance,
  59                         ns->head->instance);
  60                 return true;
  61         }
  62         sprintf(disk_name, "nvme%dc%dn%d", ns->ctrl->subsys->instance,
  63                 ns->ctrl->instance, ns->head->instance);
  64         *flags = GENHD_FL_HIDDEN;
  65         return true;
  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         struct bio *bio;
  74
  75         nvme_mpath_clear_current_path(ns);
  76
  77         /*
  78          * If we got back an ANA error, we know the controller is alive but not
  79          * ready to serve this namespace.  Kick of a re-read of the ANA
  80          * information page, and just try any other available path for now.
  81          */
  82         if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) {
  83                 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
  84                 queue_work(nvme_wq, &ns->ctrl->ana_work);
  85         }
  86
  87         spin_lock_irqsave(&ns->head->requeue_lock, flags);
  88         for (bio = req->bio; bio; bio = bio->bi_next) {
  89                 bio_set_dev(bio, ns->head->disk->part0);
  90                 if (bio->bi_opf & REQ_POLLED) {
  91                         bio->bi_opf &= ~REQ_POLLED;
  92                         bio->bi_cookie = BLK_QC_T_NONE;
  93                 }
  94         }
  95         blk_steal_bios(&ns->head->requeue_list, req);
  96         spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
  97
  98         blk_mq_end_request(req, 0);
  99         kblockd_schedule_work(&ns->head->requeue_work);
 100 }
 101
 102 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 103 {
 104         struct nvme_ns *ns;
 105
 106         down_read(&ctrl->namespaces_rwsem);
 107         list_for_each_entry(ns, &ctrl->namespaces, list) {
 108                 if (!ns->head->disk)
 109                         continue;
 110                 kblockd_schedule_work(&ns->head->requeue_work);
 111                 if (ctrl->state == NVME_CTRL_LIVE)
 112                         disk_uevent(ns->head->disk, KOBJ_CHANGE);
 113         }
 114         up_read(&ctrl->namespaces_rwsem);
 115 }
 116
 117 static const char *nvme_ana_state_names[] = {
 118         [0]                             = "invalid state",
 119         [NVME_ANA_OPTIMIZED]            = "optimized",
 120         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 121         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 122         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 123         [NVME_ANA_CHANGE]               = "change",
 124 };
 125
 126 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 127 {
 128         struct nvme_ns_head *head = ns->head;
 129         bool changed = false;
 130         int node;
 131
 132         if (!head)
 133                 goto out;
 134
 135         for_each_node(node) {
 136                 if (ns == rcu_access_pointer(head->current_path[node])) {
 137                         rcu_assign_pointer(head->current_path[node], NULL);
 138                         changed = true;
 139                 }
 140         }
 141 out:
 142         return changed;
 143 }
 144
 145 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 146 {
 147         struct nvme_ns *ns;
 148
 149         down_read(&ctrl->namespaces_rwsem);
 150         list_for_each_entry(ns, &ctrl->namespaces, list) {
 151                 nvme_mpath_clear_current_path(ns);
 152                 kblockd_schedule_work(&ns->head->requeue_work);
 153         }
 154         up_read(&ctrl->namespaces_rwsem);
 155 }
 156
 157 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
 158 {
 159         struct nvme_ns_head *head = ns->head;
 160         sector_t capacity = get_capacity(head->disk);
 161         int node;
 162
 163         list_for_each_entry_rcu(ns, &head->list, siblings) {
 164                 if (capacity != get_capacity(ns->disk))
 165                         clear_bit(NVME_NS_READY, &ns->flags);
 166         }
 167
 168         for_each_node(node)
 169                 rcu_assign_pointer(head->current_path[node], NULL);
 170 }
 171
 172 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 173 {
 174         /*
 175          * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
 176          * still be able to complete assuming that the controller is connected.
 177          * Otherwise it will fail immediately and return to the requeue list.
 178          */
 179         if (ns->ctrl->state != NVME_CTRL_LIVE &&
 180             ns->ctrl->state != NVME_CTRL_DELETING)
 181                 return true;
 182         if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 183             !test_bit(NVME_NS_READY, &ns->flags))
 184                 return true;
 185         return false;
 186 }
 187
 188 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 189 {
 190         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 191         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 192
 193         list_for_each_entry_rcu(ns, &head->list, siblings) {
 194                 if (nvme_path_is_disabled(ns))
 195                         continue;
 196
 197                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 198                         distance = node_distance(node, ns->ctrl->numa_node);
 199                 else
 200                         distance = LOCAL_DISTANCE;
 201
 202                 switch (ns->ana_state) {
 203                 case NVME_ANA_OPTIMIZED:
 204                         if (distance < found_distance) {
 205                                 found_distance = distance;
 206                                 found = ns;
 207                         }
 208                         break;
 209                 case NVME_ANA_NONOPTIMIZED:
 210                         if (distance < fallback_distance) {
 211                                 fallback_distance = distance;
 212                                 fallback = ns;
 213                         }
 214                         break;
 215                 default:
 216                         break;
 217                 }
 218         }
 219
 220         if (!found)
 221                 found = fallback;
 222         if (found)
 223                 rcu_assign_pointer(head->current_path[node], found);
 224         return found;
 225 }
 226
 227 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 228                 struct nvme_ns *ns)
 229 {
 230         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 231                         siblings);
 232         if (ns)
 233                 return ns;
 234         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 235 }
 236
 237 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 238                 int node, struct nvme_ns *old)
 239 {
 240         struct nvme_ns *ns, *found = NULL;
 241
 242         if (list_is_singular(&head->list)) {
 243                 if (nvme_path_is_disabled(old))
 244                         return NULL;
 245                 return old;
 246         }
 247
 248         for (ns = nvme_next_ns(head, old);
 249              ns && ns != old;
 250              ns = nvme_next_ns(head, ns)) {
 251                 if (nvme_path_is_disabled(ns))
 252                         continue;
 253
 254                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 255                         found = ns;
 256                         goto out;
 257                 }
 258                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 259                         found = ns;
 260         }
 261
 262         /*
 263          * The loop above skips the current path for round-robin semantics.
 264          * Fall back to the current path if either:
 265          *  - no other optimized path found and current is optimized,
 266          *  - no other usable path found and current is usable.
 267          */
 268         if (!nvme_path_is_disabled(old) &&
 269             (old->ana_state == NVME_ANA_OPTIMIZED ||
 270              (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
 271                 return old;
 272
 273         if (!found)
 274                 return NULL;
 275 out:
 276         rcu_assign_pointer(head->current_path[node], found);
 277         return found;
 278 }
 279
 280 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 281 {
 282         return ns->ctrl->state == NVME_CTRL_LIVE &&
 283                 ns->ana_state == NVME_ANA_OPTIMIZED;
 284 }
 285
 286 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 287 {
 288         int node = numa_node_id();
 289         struct nvme_ns *ns;
 290
 291         ns = srcu_dereference(head->current_path[node], &head->srcu);
 292         if (unlikely(!ns))
 293                 return __nvme_find_path(head, node);
 294
 295         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
 296                 return nvme_round_robin_path(head, node, ns);
 297         if (unlikely(!nvme_path_is_optimized(ns)))
 298                 return __nvme_find_path(head, node);
 299         return ns;
 300 }
 301
 302 static bool nvme_available_path(struct nvme_ns_head *head)
 303 {
 304         struct nvme_ns *ns;
 305
 306         list_for_each_entry_rcu(ns, &head->list, siblings) {
 307                 if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
 308                         continue;
 309                 switch (ns->ctrl->state) {
 310                 case NVME_CTRL_LIVE:
 311                 case NVME_CTRL_RESETTING:
 312                 case NVME_CTRL_CONNECTING:
 313                         /* fallthru */
 314                         return true;
 315                 default:
 316                         break;
 317                 }
 318         }
 319         return false;
 320 }
 321
 322 static void nvme_ns_head_submit_bio(struct bio *bio)
 323 {
 324         struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
 325         struct device *dev = disk_to_dev(head->disk);
 326         struct nvme_ns *ns;
 327         int srcu_idx;
 328
 329         /*
 330          * The namespace might be going away and the bio might be moved to a
 331          * different queue via blk_steal_bios(), so we need to use the bio_split
 332          * pool from the original queue to allocate the bvecs from.
 333          */
 334         blk_queue_split(&bio);
 335
 336         srcu_idx = srcu_read_lock(&head->srcu);
 337         ns = nvme_find_path(head);
 338         if (likely(ns)) {
 339                 bio_set_dev(bio, ns->disk->part0);
 340                 bio->bi_opf |= REQ_NVME_MPATH;
 341                 trace_block_bio_remap(bio, disk_devt(ns->head->disk),
 342                                       bio->bi_iter.bi_sector);
 343                 submit_bio_noacct(bio);
 344         } else if (nvme_available_path(head)) {
 345                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 346
 347                 spin_lock_irq(&head->requeue_lock);
 348                 bio_list_add(&head->requeue_list, bio);
 349                 spin_unlock_irq(&head->requeue_lock);
 350         } else {
 351                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 352
 353                 bio->bi_status = BLK_STS_IOERR;
 354                 bio_endio(bio);
 355         }
 356
 357         srcu_read_unlock(&head->srcu, srcu_idx);
 358 }
 359
 360 static int nvme_ns_head_open(struct block_device *bdev, fmode_t mode)
 361 {
 362         if (!nvme_tryget_ns_head(bdev->bd_disk->private_data))
 363                 return -ENXIO;
 364         return 0;
 365 }
 366
 367 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode)
 368 {
 369         nvme_put_ns_head(disk->private_data);
 370 }
 371
 372 #ifdef CONFIG_BLK_DEV_ZONED
 373 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
 374                 unsigned int nr_zones, report_zones_cb cb, void *data)
 375 {
 376         struct nvme_ns_head *head = disk->private_data;
 377         struct nvme_ns *ns;
 378         int srcu_idx, ret = -EWOULDBLOCK;
 379
 380         srcu_idx = srcu_read_lock(&head->srcu);
 381         ns = nvme_find_path(head);
 382         if (ns)
 383                 ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
 384         srcu_read_unlock(&head->srcu, srcu_idx);
 385         return ret;
 386 }
 387 #else
 388 #define nvme_ns_head_report_zones       NULL
 389 #endif /* CONFIG_BLK_DEV_ZONED */
 390
 391 const struct block_device_operations nvme_ns_head_ops = {
 392         .owner          = THIS_MODULE,
 393         .submit_bio     = nvme_ns_head_submit_bio,
 394         .open           = nvme_ns_head_open,
 395         .release        = nvme_ns_head_release,
 396         .ioctl          = nvme_ns_head_ioctl,
 397         .getgeo         = nvme_getgeo,
 398         .report_zones   = nvme_ns_head_report_zones,
 399         .pr_ops         = &nvme_pr_ops,
 400 };
 401
 402 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
 403 {
 404         return container_of(cdev, struct nvme_ns_head, cdev);
 405 }
 406
 407 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
 408 {
 409         if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
 410                 return -ENXIO;
 411         return 0;
 412 }
 413
 414 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
 415 {
 416         nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
 417         return 0;
 418 }
 419
 420 static const struct file_operations nvme_ns_head_chr_fops = {
 421         .owner          = THIS_MODULE,
 422         .open           = nvme_ns_head_chr_open,
 423         .release        = nvme_ns_head_chr_release,
 424         .unlocked_ioctl = nvme_ns_head_chr_ioctl,
 425         .compat_ioctl   = compat_ptr_ioctl,
 426 };
 427
 428 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
 429 {
 430         int ret;
 431
 432         head->cdev_device.parent = &head->subsys->dev;
 433         ret = dev_set_name(&head->cdev_device, "ng%dn%d",
 434                            head->subsys->instance, head->instance);
 435         if (ret)
 436                 return ret;
 437         ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
 438                             &nvme_ns_head_chr_fops, THIS_MODULE);
 439         return ret;
 440 }
 441
 442 static void nvme_requeue_work(struct work_struct *work)
 443 {
 444         struct nvme_ns_head *head =
 445                 container_of(work, struct nvme_ns_head, requeue_work);
 446         struct bio *bio, *next;
 447
 448         spin_lock_irq(&head->requeue_lock);
 449         next = bio_list_get(&head->requeue_list);
 450         spin_unlock_irq(&head->requeue_lock);
 451
 452         while ((bio = next) != NULL) {
 453                 next = bio->bi_next;
 454                 bio->bi_next = NULL;
 455
 456                 submit_bio_noacct(bio);
 457         }
 458 }
 459
 460 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 461 {
 462         bool vwc = false;
 463
 464         mutex_init(&head->lock);
 465         bio_list_init(&head->requeue_list);
 466         spin_lock_init(&head->requeue_lock);
 467         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 468
 469         /*
 470          * Add a multipath node if the subsystems supports multiple controllers.
 471          * We also do this for private namespaces as the namespace sharing data could
 472          * change after a rescan.
 473          */
 474         if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) || !multipath)
 475                 return 0;
 476
 477         head->disk = blk_alloc_disk(ctrl->numa_node);
 478         if (!head->disk)
 479                 return -ENOMEM;
 480         head->disk->fops = &nvme_ns_head_ops;
 481         head->disk->private_data = head;
 482         sprintf(head->disk->disk_name, "nvme%dn%d",
 483                         ctrl->subsys->instance, head->instance);
 484
 485         blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
 486         blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
 487         /*
 488          * This assumes all controllers that refer to a namespace either
 489          * support poll queues or not.  That is not a strict guarantee,
 490          * but if the assumption is wrong the effect is only suboptimal
 491          * performance but not correctness problem.
 492          */
 493         if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL &&
 494             ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues)
 495                 blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue);
 496
 497         /* set to a default value of 512 until the disk is validated */
 498         blk_queue_logical_block_size(head->disk->queue, 512);
 499         blk_set_stacking_limits(&head->disk->queue->limits);
 500
 501         /* we need to propagate up the VMC settings */
 502         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 503                 vwc = true;
 504         blk_queue_write_cache(head->disk->queue, vwc, vwc);
 505         return 0;
 506 }
 507
 508 static void nvme_mpath_set_live(struct nvme_ns *ns)
 509 {
 510         struct nvme_ns_head *head = ns->head;
 511         int rc;
 512
 513         if (!head->disk)
 514                 return;
 515
 516         /*
 517          * test_and_set_bit() is used because it is protecting against two nvme
 518          * paths simultaneously calling device_add_disk() on the same namespace
 519          * head.
 520          */
 521         if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 522                 rc = device_add_disk(&head->subsys->dev, head->disk,
 523                                      nvme_ns_id_attr_groups);
 524                 if (rc) {
 525                         clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
 526                         return;
 527                 }
 528                 nvme_add_ns_head_cdev(head);
 529         }
 530
 531         mutex_lock(&head->lock);
 532         if (nvme_path_is_optimized(ns)) {
 533                 int node, srcu_idx;
 534
 535                 srcu_idx = srcu_read_lock(&head->srcu);
 536                 for_each_node(node)
 537                         __nvme_find_path(head, node);
 538                 srcu_read_unlock(&head->srcu, srcu_idx);
 539         }
 540         mutex_unlock(&head->lock);
 541
 542         synchronize_srcu(&head->srcu);
 543         kblockd_schedule_work(&head->requeue_work);
 544 }
 545
 546 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 547                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 548                         void *))
 549 {
 550         void *base = ctrl->ana_log_buf;
 551         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 552         int error, i;
 553
 554         lockdep_assert_held(&ctrl->ana_lock);
 555
 556         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 557                 struct nvme_ana_group_desc *desc = base + offset;
 558                 u32 nr_nsids;
 559                 size_t nsid_buf_size;
 560
 561                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 562                         return -EINVAL;
 563
 564                 nr_nsids = le32_to_cpu(desc->nnsids);
 565                 nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
 566
 567                 if (WARN_ON_ONCE(desc->grpid == 0))
 568                         return -EINVAL;
 569                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 570                         return -EINVAL;
 571                 if (WARN_ON_ONCE(desc->state == 0))
 572                         return -EINVAL;
 573                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 574                         return -EINVAL;
 575
 576                 offset += sizeof(*desc);
 577                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 578                         return -EINVAL;
 579
 580                 error = cb(ctrl, desc, data);
 581                 if (error)
 582                         return error;
 583
 584                 offset += nsid_buf_size;
 585         }
 586
 587         return 0;
 588 }
 589
 590 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 591 {
 592         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 593 }
 594
 595 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 596                 struct nvme_ns *ns)
 597 {
 598         ns->ana_grpid = le32_to_cpu(desc->grpid);
 599         ns->ana_state = desc->state;
 600         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 601
 602         if (nvme_state_is_live(ns->ana_state))
 603                 nvme_mpath_set_live(ns);
 604 }
 605
 606 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 607                 struct nvme_ana_group_desc *desc, void *data)
 608 {
 609         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 610         unsigned *nr_change_groups = data;
 611         struct nvme_ns *ns;
 612
 613         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 614                         le32_to_cpu(desc->grpid),
 615                         nvme_ana_state_names[desc->state]);
 616
 617         if (desc->state == NVME_ANA_CHANGE)
 618                 (*nr_change_groups)++;
 619
 620         if (!nr_nsids)
 621                 return 0;
 622
 623         down_read(&ctrl->namespaces_rwsem);
 624         list_for_each_entry(ns, &ctrl->namespaces, list) {
 625                 unsigned nsid;
 626 again:
 627                 nsid = le32_to_cpu(desc->nsids[n]);
 628                 if (ns->head->ns_id < nsid)
 629                         continue;
 630                 if (ns->head->ns_id == nsid)
 631                         nvme_update_ns_ana_state(desc, ns);
 632                 if (++n == nr_nsids)
 633                         break;
 634                 if (ns->head->ns_id > nsid)
 635                         goto again;
 636         }
 637         up_read(&ctrl->namespaces_rwsem);
 638         return 0;
 639 }
 640
 641 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 642 {
 643         u32 nr_change_groups = 0;
 644         int error;
 645
 646         mutex_lock(&ctrl->ana_lock);
 647         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
 648                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 649         if (error) {
 650                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 651                 goto out_unlock;
 652         }
 653
 654         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 655                         nvme_update_ana_state);
 656         if (error)
 657                 goto out_unlock;
 658
 659         /*
 660          * In theory we should have an ANATT timer per group as they might enter
 661          * the change state at different times.  But that is a lot of overhead
 662          * just to protect against a target that keeps entering new changes
 663          * states while never finishing previous ones.  But we'll still
 664          * eventually time out once all groups are in change state, so this
 665          * isn't a big deal.
 666          *
 667          * We also double the ANATT value to provide some slack for transports
 668          * or AEN processing overhead.
 669          */
 670         if (nr_change_groups)
 671                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 672         else
 673                 del_timer_sync(&ctrl->anatt_timer);
 674 out_unlock:
 675         mutex_unlock(&ctrl->ana_lock);
 676         return error;
 677 }
 678
 679 static void nvme_ana_work(struct work_struct *work)
 680 {
 681         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 682
 683         if (ctrl->state != NVME_CTRL_LIVE)
 684                 return;
 685
 686         nvme_read_ana_log(ctrl);
 687 }
 688
 689 static void nvme_anatt_timeout(struct timer_list *t)
 690 {
 691         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 692
 693         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 694         nvme_reset_ctrl(ctrl);
 695 }
 696
 697 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 698 {
 699         if (!nvme_ctrl_use_ana(ctrl))
 700                 return;
 701         del_timer_sync(&ctrl->anatt_timer);
 702         cancel_work_sync(&ctrl->ana_work);
 703 }
 704
 705 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 706         struct device_attribute subsys_attr_##_name =   \
 707                 __ATTR(_name, _mode, _show, _store)
 708
 709 static const char *nvme_iopolicy_names[] = {
 710         [NVME_IOPOLICY_NUMA]    = "numa",
 711         [NVME_IOPOLICY_RR]      = "round-robin",
 712 };
 713
 714 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 715                 struct device_attribute *attr, char *buf)
 716 {
 717         struct nvme_subsystem *subsys =
 718                 container_of(dev, struct nvme_subsystem, dev);
 719
 720         return sysfs_emit(buf, "%s\n",
 721                           nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 722 }
 723
 724 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 725                 struct device_attribute *attr, const char *buf, size_t count)
 726 {
 727         struct nvme_subsystem *subsys =
 728                 container_of(dev, struct nvme_subsystem, dev);
 729         int i;
 730
 731         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 732                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 733                         WRITE_ONCE(subsys->iopolicy, i);
 734                         return count;
 735                 }
 736         }
 737
 738         return -EINVAL;
 739 }
 740 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 741                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 742
 743 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 744                 char *buf)
 745 {
 746         return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 747 }
 748 DEVICE_ATTR_RO(ana_grpid);
 749
 750 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 751                 char *buf)
 752 {
 753         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 754
 755         return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 756 }
 757 DEVICE_ATTR_RO(ana_state);
 758
 759 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
 760                 struct nvme_ana_group_desc *desc, void *data)
 761 {
 762         struct nvme_ana_group_desc *dst = data;
 763
 764         if (desc->grpid != dst->grpid)
 765                 return 0;
 766
 767         *dst = *desc;
 768         return -ENXIO; /* just break out of the loop */
 769 }
 770
 771 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 772 {
 773         if (nvme_ctrl_use_ana(ns->ctrl)) {
 774                 struct nvme_ana_group_desc desc = {
 775                         .grpid = id->anagrpid,
 776                         .state = 0,
 777                 };
 778
 779                 mutex_lock(&ns->ctrl->ana_lock);
 780                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 781                 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
 782                 mutex_unlock(&ns->ctrl->ana_lock);
 783                 if (desc.state) {
 784                         /* found the group desc: update */
 785                         nvme_update_ns_ana_state(&desc, ns);
 786                 } else {
 787                         /* group desc not found: trigger a re-read */
 788                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
 789                         queue_work(nvme_wq, &ns->ctrl->ana_work);
 790                 }
 791         } else {
 792                 ns->ana_state = NVME_ANA_OPTIMIZED;
 793                 nvme_mpath_set_live(ns);
 794         }
 795
 796         if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
 797                 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
 798                                    ns->head->disk->queue);
 799 #ifdef CONFIG_BLK_DEV_ZONED
 800         if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
 801                 ns->head->disk->queue->nr_zones = ns->queue->nr_zones;
 802 #endif
 803 }
 804
 805 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
 806 {
 807         if (!head->disk)
 808                 return;
 809         kblockd_schedule_work(&head->requeue_work);
 810         if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 811                 nvme_cdev_del(&head->cdev, &head->cdev_device);
 812                 del_gendisk(head->disk);
 813         }
 814 }
 815
 816 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 817 {
 818         if (!head->disk)
 819                 return;
 820         blk_set_queue_dying(head->disk->queue);
 821         /* make sure all pending bios are cleaned up */
 822         kblockd_schedule_work(&head->requeue_work);
 823         flush_work(&head->requeue_work);
 824         blk_cleanup_disk(head->disk);
 825 }
 826
 827 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
 828 {
 829         mutex_init(&ctrl->ana_lock);
 830         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 831         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 832 }
 833
 834 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 835 {
 836         size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
 837         size_t ana_log_size;
 838         int error = 0;
 839
 840         /* check if multipath is enabled and we have the capability */
 841         if (!multipath || !ctrl->subsys ||
 842             !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
 843                 return 0;
 844
 845         if (!ctrl->max_namespaces ||
 846             ctrl->max_namespaces > le32_to_cpu(id->nn)) {
 847                 dev_err(ctrl->device,
 848                         "Invalid MNAN value %u\n", ctrl->max_namespaces);
 849                 return -EINVAL;
 850         }
 851
 852         ctrl->anacap = id->anacap;
 853         ctrl->anatt = id->anatt;
 854         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 855         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 856
 857         ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 858                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
 859                 ctrl->max_namespaces * sizeof(__le32);
 860         if (ana_log_size > max_transfer_size) {
 861                 dev_err(ctrl->device,
 862                         "ANA log page size (%zd) larger than MDTS (%zd).\n",
 863                         ana_log_size, max_transfer_size);
 864                 dev_err(ctrl->device, "disabling ANA support.\n");
 865                 goto out_uninit;
 866         }
 867         if (ana_log_size > ctrl->ana_log_size) {
 868                 nvme_mpath_stop(ctrl);
 869                 kfree(ctrl->ana_log_buf);
 870                 ctrl->ana_log_buf = kmalloc(ana_log_size, GFP_KERNEL);
 871                 if (!ctrl->ana_log_buf)
 872                         return -ENOMEM;
 873         }
 874         ctrl->ana_log_size = ana_log_size;
 875         error = nvme_read_ana_log(ctrl);
 876         if (error)
 877                 goto out_uninit;
 878         return 0;
 879
 880 out_uninit:
 881         nvme_mpath_uninit(ctrl);
 882         return error;
 883 }
 884
 885 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 886 {
 887         kfree(ctrl->ana_log_buf);
 888         ctrl->ana_log_buf = NULL;
 889 }