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
  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 static blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
 298 {
 299         struct nvme_ns_head *head = bio->bi_bdev->bd_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_set_dev(bio, ns->disk->part0);
 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 int nvme_ns_head_open(struct block_device *bdev, fmode_t mode)
 338 {
 339         if (!nvme_tryget_ns_head(bdev->bd_disk->private_data))
 340                 return -ENXIO;
 341         return 0;
 342 }
 343
 344 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode)
 345 {
 346         nvme_put_ns_head(disk->private_data);
 347 }
 348
 349 const struct block_device_operations nvme_ns_head_ops = {
 350         .owner          = THIS_MODULE,
 351         .submit_bio     = nvme_ns_head_submit_bio,
 352         .open           = nvme_ns_head_open,
 353         .release        = nvme_ns_head_release,
 354         .ioctl          = nvme_ns_head_ioctl,
 355         .getgeo         = nvme_getgeo,
 356         .report_zones   = nvme_report_zones,
 357         .pr_ops         = &nvme_pr_ops,
 358 };
 359
 360 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
 361 {
 362         return container_of(cdev, struct nvme_ns_head, cdev);
 363 }
 364
 365 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
 366 {
 367         if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
 368                 return -ENXIO;
 369         return 0;
 370 }
 371
 372 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
 373 {
 374         nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
 375         return 0;
 376 }
 377
 378 static const struct file_operations nvme_ns_head_chr_fops = {
 379         .owner          = THIS_MODULE,
 380         .open           = nvme_ns_head_chr_open,
 381         .release        = nvme_ns_head_chr_release,
 382         .unlocked_ioctl = nvme_ns_head_chr_ioctl,
 383         .compat_ioctl   = compat_ptr_ioctl,
 384 };
 385
 386 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
 387 {
 388         int ret;
 389
 390         head->cdev_device.parent = &head->subsys->dev;
 391         ret = dev_set_name(&head->cdev_device, "ng%dn%d",
 392                            head->subsys->instance, head->instance);
 393         if (ret)
 394                 return ret;
 395         ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
 396                             &nvme_ns_head_chr_fops, THIS_MODULE);
 397         if (ret)
 398                 kfree_const(head->cdev_device.kobj.name);
 399         return ret;
 400 }
 401
 402 static void nvme_requeue_work(struct work_struct *work)
 403 {
 404         struct nvme_ns_head *head =
 405                 container_of(work, struct nvme_ns_head, requeue_work);
 406         struct bio *bio, *next;
 407
 408         spin_lock_irq(&head->requeue_lock);
 409         next = bio_list_get(&head->requeue_list);
 410         spin_unlock_irq(&head->requeue_lock);
 411
 412         while ((bio = next) != NULL) {
 413                 next = bio->bi_next;
 414                 bio->bi_next = NULL;
 415
 416                 /*
 417                  * Reset disk to the mpath node and resubmit to select a new
 418                  * path.
 419                  */
 420                 bio_set_dev(bio, head->disk->part0);
 421                 submit_bio_noacct(bio);
 422         }
 423 }
 424
 425 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 426 {
 427         struct request_queue *q;
 428         bool vwc = false;
 429
 430         mutex_init(&head->lock);
 431         bio_list_init(&head->requeue_list);
 432         spin_lock_init(&head->requeue_lock);
 433         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 434
 435         /*
 436          * Add a multipath node if the subsystems supports multiple controllers.
 437          * We also do this for private namespaces as the namespace sharing data could
 438          * change after a rescan.
 439          */
 440         if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) || !multipath)
 441                 return 0;
 442
 443         q = blk_alloc_queue(ctrl->numa_node);
 444         if (!q)
 445                 goto out;
 446         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 447         /* set to a default value for 512 until disk is validated */
 448         blk_queue_logical_block_size(q, 512);
 449         blk_set_stacking_limits(&q->limits);
 450
 451         /* we need to propagate up the VMC settings */
 452         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 453                 vwc = true;
 454         blk_queue_write_cache(q, vwc, vwc);
 455
 456         head->disk = alloc_disk(0);
 457         if (!head->disk)
 458                 goto out_cleanup_queue;
 459         head->disk->fops = &nvme_ns_head_ops;
 460         head->disk->private_data = head;
 461         head->disk->queue = q;
 462         head->disk->flags = GENHD_FL_EXT_DEVT;
 463         sprintf(head->disk->disk_name, "nvme%dn%d",
 464                         ctrl->subsys->instance, head->instance);
 465         return 0;
 466
 467 out_cleanup_queue:
 468         blk_cleanup_queue(q);
 469 out:
 470         return -ENOMEM;
 471 }
 472
 473 static void nvme_mpath_set_live(struct nvme_ns *ns)
 474 {
 475         struct nvme_ns_head *head = ns->head;
 476
 477         if (!head->disk)
 478                 return;
 479
 480         if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 481                 device_add_disk(&head->subsys->dev, head->disk,
 482                                 nvme_ns_id_attr_groups);
 483                 nvme_add_ns_head_cdev(head);
 484         }
 485
 486         mutex_lock(&head->lock);
 487         if (nvme_path_is_optimized(ns)) {
 488                 int node, srcu_idx;
 489
 490                 srcu_idx = srcu_read_lock(&head->srcu);
 491                 for_each_node(node)
 492                         __nvme_find_path(head, node);
 493                 srcu_read_unlock(&head->srcu, srcu_idx);
 494         }
 495         mutex_unlock(&head->lock);
 496
 497         synchronize_srcu(&head->srcu);
 498         kblockd_schedule_work(&head->requeue_work);
 499 }
 500
 501 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 502                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 503                         void *))
 504 {
 505         void *base = ctrl->ana_log_buf;
 506         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 507         int error, i;
 508
 509         lockdep_assert_held(&ctrl->ana_lock);
 510
 511         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 512                 struct nvme_ana_group_desc *desc = base + offset;
 513                 u32 nr_nsids;
 514                 size_t nsid_buf_size;
 515
 516                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 517                         return -EINVAL;
 518
 519                 nr_nsids = le32_to_cpu(desc->nnsids);
 520                 nsid_buf_size = nr_nsids * sizeof(__le32);
 521
 522                 if (WARN_ON_ONCE(desc->grpid == 0))
 523                         return -EINVAL;
 524                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 525                         return -EINVAL;
 526                 if (WARN_ON_ONCE(desc->state == 0))
 527                         return -EINVAL;
 528                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 529                         return -EINVAL;
 530
 531                 offset += sizeof(*desc);
 532                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 533                         return -EINVAL;
 534
 535                 error = cb(ctrl, desc, data);
 536                 if (error)
 537                         return error;
 538
 539                 offset += nsid_buf_size;
 540         }
 541
 542         return 0;
 543 }
 544
 545 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 546 {
 547         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 548 }
 549
 550 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 551                 struct nvme_ns *ns)
 552 {
 553         ns->ana_grpid = le32_to_cpu(desc->grpid);
 554         ns->ana_state = desc->state;
 555         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 556
 557         if (nvme_state_is_live(ns->ana_state))
 558                 nvme_mpath_set_live(ns);
 559 }
 560
 561 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 562                 struct nvme_ana_group_desc *desc, void *data)
 563 {
 564         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 565         unsigned *nr_change_groups = data;
 566         struct nvme_ns *ns;
 567
 568         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 569                         le32_to_cpu(desc->grpid),
 570                         nvme_ana_state_names[desc->state]);
 571
 572         if (desc->state == NVME_ANA_CHANGE)
 573                 (*nr_change_groups)++;
 574
 575         if (!nr_nsids)
 576                 return 0;
 577
 578         down_read(&ctrl->namespaces_rwsem);
 579         list_for_each_entry(ns, &ctrl->namespaces, list) {
 580                 unsigned nsid = le32_to_cpu(desc->nsids[n]);
 581
 582                 if (ns->head->ns_id < nsid)
 583                         continue;
 584                 if (ns->head->ns_id == nsid)
 585                         nvme_update_ns_ana_state(desc, ns);
 586                 if (++n == nr_nsids)
 587                         break;
 588         }
 589         up_read(&ctrl->namespaces_rwsem);
 590         return 0;
 591 }
 592
 593 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 594 {
 595         u32 nr_change_groups = 0;
 596         int error;
 597
 598         mutex_lock(&ctrl->ana_lock);
 599         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
 600                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 601         if (error) {
 602                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 603                 goto out_unlock;
 604         }
 605
 606         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 607                         nvme_update_ana_state);
 608         if (error)
 609                 goto out_unlock;
 610
 611         /*
 612          * In theory we should have an ANATT timer per group as they might enter
 613          * the change state at different times.  But that is a lot of overhead
 614          * just to protect against a target that keeps entering new changes
 615          * states while never finishing previous ones.  But we'll still
 616          * eventually time out once all groups are in change state, so this
 617          * isn't a big deal.
 618          *
 619          * We also double the ANATT value to provide some slack for transports
 620          * or AEN processing overhead.
 621          */
 622         if (nr_change_groups)
 623                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 624         else
 625                 del_timer_sync(&ctrl->anatt_timer);
 626 out_unlock:
 627         mutex_unlock(&ctrl->ana_lock);
 628         return error;
 629 }
 630
 631 static void nvme_ana_work(struct work_struct *work)
 632 {
 633         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 634
 635         if (ctrl->state != NVME_CTRL_LIVE)
 636                 return;
 637
 638         nvme_read_ana_log(ctrl);
 639 }
 640
 641 static void nvme_anatt_timeout(struct timer_list *t)
 642 {
 643         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 644
 645         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 646         nvme_reset_ctrl(ctrl);
 647 }
 648
 649 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 650 {
 651         if (!nvme_ctrl_use_ana(ctrl))
 652                 return;
 653         del_timer_sync(&ctrl->anatt_timer);
 654         cancel_work_sync(&ctrl->ana_work);
 655 }
 656
 657 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 658         struct device_attribute subsys_attr_##_name =   \
 659                 __ATTR(_name, _mode, _show, _store)
 660
 661 static const char *nvme_iopolicy_names[] = {
 662         [NVME_IOPOLICY_NUMA]    = "numa",
 663         [NVME_IOPOLICY_RR]      = "round-robin",
 664 };
 665
 666 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 667                 struct device_attribute *attr, char *buf)
 668 {
 669         struct nvme_subsystem *subsys =
 670                 container_of(dev, struct nvme_subsystem, dev);
 671
 672         return sysfs_emit(buf, "%s\n",
 673                           nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 674 }
 675
 676 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 677                 struct device_attribute *attr, const char *buf, size_t count)
 678 {
 679         struct nvme_subsystem *subsys =
 680                 container_of(dev, struct nvme_subsystem, dev);
 681         int i;
 682
 683         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 684                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 685                         WRITE_ONCE(subsys->iopolicy, i);
 686                         return count;
 687                 }
 688         }
 689
 690         return -EINVAL;
 691 }
 692 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 693                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 694
 695 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 696                 char *buf)
 697 {
 698         return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 699 }
 700 DEVICE_ATTR_RO(ana_grpid);
 701
 702 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 703                 char *buf)
 704 {
 705         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 706
 707         return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 708 }
 709 DEVICE_ATTR_RO(ana_state);
 710
 711 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
 712                 struct nvme_ana_group_desc *desc, void *data)
 713 {
 714         struct nvme_ana_group_desc *dst = data;
 715
 716         if (desc->grpid != dst->grpid)
 717                 return 0;
 718
 719         *dst = *desc;
 720         return -ENXIO; /* just break out of the loop */
 721 }
 722
 723 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 724 {
 725         if (nvme_ctrl_use_ana(ns->ctrl)) {
 726                 struct nvme_ana_group_desc desc = {
 727                         .grpid = id->anagrpid,
 728                         .state = 0,
 729                 };
 730
 731                 mutex_lock(&ns->ctrl->ana_lock);
 732                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 733                 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
 734                 mutex_unlock(&ns->ctrl->ana_lock);
 735                 if (desc.state) {
 736                         /* found the group desc: update */
 737                         nvme_update_ns_ana_state(&desc, ns);
 738                 } else {
 739                         /* group desc not found: trigger a re-read */
 740                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
 741                         queue_work(nvme_wq, &ns->ctrl->ana_work);
 742                 }
 743         } else {
 744                 ns->ana_state = NVME_ANA_OPTIMIZED;
 745                 nvme_mpath_set_live(ns);
 746         }
 747
 748         if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
 749                 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
 750                                    ns->head->disk->queue);
 751 #ifdef CONFIG_BLK_DEV_ZONED
 752         if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
 753                 ns->head->disk->queue->nr_zones = ns->queue->nr_zones;
 754 #endif
 755 }
 756
 757 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 758 {
 759         if (!head->disk)
 760                 return;
 761         if (head->disk->flags & GENHD_FL_UP) {
 762                 nvme_cdev_del(&head->cdev, &head->cdev_device);
 763                 del_gendisk(head->disk);
 764         }
 765         blk_set_queue_dying(head->disk->queue);
 766         /* make sure all pending bios are cleaned up */
 767         kblockd_schedule_work(&head->requeue_work);
 768         flush_work(&head->requeue_work);
 769         blk_cleanup_queue(head->disk->queue);
 770         if (!test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 771                 /*
 772                  * if device_add_disk wasn't called, prevent
 773                  * disk release to put a bogus reference on the
 774                  * request queue
 775                  */
 776                 head->disk->queue = NULL;
 777         }
 778         put_disk(head->disk);
 779 }
 780
 781 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 782 {
 783         int error;
 784
 785         /* check if multipath is enabled and we have the capability */
 786         if (!multipath || !ctrl->subsys ||
 787             !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
 788                 return 0;
 789
 790         ctrl->anacap = id->anacap;
 791         ctrl->anatt = id->anatt;
 792         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 793         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 794
 795         mutex_init(&ctrl->ana_lock);
 796         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 797         ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 798                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
 799         ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
 800
 801         if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
 802                 dev_err(ctrl->device,
 803                         "ANA log page size (%zd) larger than MDTS (%d).\n",
 804                         ctrl->ana_log_size,
 805                         ctrl->max_hw_sectors << SECTOR_SHIFT);
 806                 dev_err(ctrl->device, "disabling ANA support.\n");
 807                 return 0;
 808         }
 809
 810         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 811         kfree(ctrl->ana_log_buf);
 812         ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
 813         if (!ctrl->ana_log_buf) {
 814                 error = -ENOMEM;
 815                 goto out;
 816         }
 817
 818         error = nvme_read_ana_log(ctrl);
 819         if (error)
 820                 goto out_free_ana_log_buf;
 821         return 0;
 822 out_free_ana_log_buf:
 823         kfree(ctrl->ana_log_buf);
 824         ctrl->ana_log_buf = NULL;
 825 out:
 826         return error;
 827 }
 828
 829 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 830 {
 831         kfree(ctrl->ana_log_buf);
 832         ctrl->ana_log_buf = NULL;
 833 }