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 <linux/vmalloc.h>
   9 #include <trace/events/block.h>
  10 #include "nvme.h"
  11
  12 bool multipath = true;
  13 module_param(multipath, bool, 0444);
  14 MODULE_PARM_DESC(multipath,
  15         "turn on native support for multiple controllers per subsystem");
  16
  17 static const char *nvme_iopolicy_names[] = {
  18         [NVME_IOPOLICY_NUMA]    = "numa",
  19         [NVME_IOPOLICY_RR]      = "round-robin",
  20 };
  21
  22 static int iopolicy = NVME_IOPOLICY_NUMA;
  23
  24 static int nvme_set_iopolicy(const char *val, const struct kernel_param *kp)
  25 {
  26         if (!val)
  27                 return -EINVAL;
  28         if (!strncmp(val, "numa", 4))
  29                 iopolicy = NVME_IOPOLICY_NUMA;
  30         else if (!strncmp(val, "round-robin", 11))
  31                 iopolicy = NVME_IOPOLICY_RR;
  32         else
  33                 return -EINVAL;
  34
  35         return 0;
  36 }
  37
  38 static int nvme_get_iopolicy(char *buf, const struct kernel_param *kp)
  39 {
  40         return sprintf(buf, "%s\n", nvme_iopolicy_names[iopolicy]);
  41 }
  42
  43 module_param_call(iopolicy, nvme_set_iopolicy, nvme_get_iopolicy,
  44         &iopolicy, 0644);
  45 MODULE_PARM_DESC(iopolicy,
  46         "Default multipath I/O policy; 'numa' (default) or 'round-robin'");
  47
  48 void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys)
  49 {
  50         subsys->iopolicy = iopolicy;
  51 }
  52
  53 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
  54 {
  55         struct nvme_ns_head *h;
  56
  57         lockdep_assert_held(&subsys->lock);
  58         list_for_each_entry(h, &subsys->nsheads, entry)
  59                 if (h->disk)
  60                         blk_mq_unfreeze_queue(h->disk->queue);
  61 }
  62
  63 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
  64 {
  65         struct nvme_ns_head *h;
  66
  67         lockdep_assert_held(&subsys->lock);
  68         list_for_each_entry(h, &subsys->nsheads, entry)
  69                 if (h->disk)
  70                         blk_mq_freeze_queue_wait(h->disk->queue);
  71 }
  72
  73 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
  74 {
  75         struct nvme_ns_head *h;
  76
  77         lockdep_assert_held(&subsys->lock);
  78         list_for_each_entry(h, &subsys->nsheads, entry)
  79                 if (h->disk)
  80                         blk_freeze_queue_start(h->disk->queue);
  81 }
  82
  83 void nvme_failover_req(struct request *req)
  84 {
  85         struct nvme_ns *ns = req->q->queuedata;
  86         u16 status = nvme_req(req)->status & 0x7ff;
  87         unsigned long flags;
  88         struct bio *bio;
  89
  90         nvme_mpath_clear_current_path(ns);
  91
  92         /*
  93          * If we got back an ANA error, we know the controller is alive but not
  94          * ready to serve this namespace.  Kick of a re-read of the ANA
  95          * information page, and just try any other available path for now.
  96          */
  97         if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) {
  98                 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
  99                 queue_work(nvme_wq, &ns->ctrl->ana_work);
 100         }
 101
 102         spin_lock_irqsave(&ns->head->requeue_lock, flags);
 103         for (bio = req->bio; bio; bio = bio->bi_next) {
 104                 bio_set_dev(bio, ns->head->disk->part0);
 105                 if (bio->bi_opf & REQ_POLLED) {
 106                         bio->bi_opf &= ~REQ_POLLED;
 107                         bio->bi_cookie = BLK_QC_T_NONE;
 108                 }
 109         }
 110         blk_steal_bios(&ns->head->requeue_list, req);
 111         spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
 112
 113         blk_mq_end_request(req, 0);
 114         kblockd_schedule_work(&ns->head->requeue_work);
 115 }
 116
 117 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 118 {
 119         struct nvme_ns *ns;
 120
 121         down_read(&ctrl->namespaces_rwsem);
 122         list_for_each_entry(ns, &ctrl->namespaces, list) {
 123                 if (!ns->head->disk)
 124                         continue;
 125                 kblockd_schedule_work(&ns->head->requeue_work);
 126                 if (ctrl->state == NVME_CTRL_LIVE)
 127                         disk_uevent(ns->head->disk, KOBJ_CHANGE);
 128         }
 129         up_read(&ctrl->namespaces_rwsem);
 130 }
 131
 132 static const char *nvme_ana_state_names[] = {
 133         [0]                             = "invalid state",
 134         [NVME_ANA_OPTIMIZED]            = "optimized",
 135         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 136         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 137         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 138         [NVME_ANA_CHANGE]               = "change",
 139 };
 140
 141 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 142 {
 143         struct nvme_ns_head *head = ns->head;
 144         bool changed = false;
 145         int node;
 146
 147         if (!head)
 148                 goto out;
 149
 150         for_each_node(node) {
 151                 if (ns == rcu_access_pointer(head->current_path[node])) {
 152                         rcu_assign_pointer(head->current_path[node], NULL);
 153                         changed = true;
 154                 }
 155         }
 156 out:
 157         return changed;
 158 }
 159
 160 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 161 {
 162         struct nvme_ns *ns;
 163
 164         down_read(&ctrl->namespaces_rwsem);
 165         list_for_each_entry(ns, &ctrl->namespaces, list) {
 166                 nvme_mpath_clear_current_path(ns);
 167                 kblockd_schedule_work(&ns->head->requeue_work);
 168         }
 169         up_read(&ctrl->namespaces_rwsem);
 170 }
 171
 172 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
 173 {
 174         struct nvme_ns_head *head = ns->head;
 175         sector_t capacity = get_capacity(head->disk);
 176         int node;
 177
 178         list_for_each_entry_rcu(ns, &head->list, siblings) {
 179                 if (capacity != get_capacity(ns->disk))
 180                         clear_bit(NVME_NS_READY, &ns->flags);
 181         }
 182
 183         for_each_node(node)
 184                 rcu_assign_pointer(head->current_path[node], NULL);
 185 }
 186
 187 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 188 {
 189         /*
 190          * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
 191          * still be able to complete assuming that the controller is connected.
 192          * Otherwise it will fail immediately and return to the requeue list.
 193          */
 194         if (ns->ctrl->state != NVME_CTRL_LIVE &&
 195             ns->ctrl->state != NVME_CTRL_DELETING)
 196                 return true;
 197         if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 198             !test_bit(NVME_NS_READY, &ns->flags))
 199                 return true;
 200         return false;
 201 }
 202
 203 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 204 {
 205         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 206         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 207
 208         list_for_each_entry_rcu(ns, &head->list, siblings) {
 209                 if (nvme_path_is_disabled(ns))
 210                         continue;
 211
 212                 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 213                         distance = node_distance(node, ns->ctrl->numa_node);
 214                 else
 215                         distance = LOCAL_DISTANCE;
 216
 217                 switch (ns->ana_state) {
 218                 case NVME_ANA_OPTIMIZED:
 219                         if (distance < found_distance) {
 220                                 found_distance = distance;
 221                                 found = ns;
 222                         }
 223                         break;
 224                 case NVME_ANA_NONOPTIMIZED:
 225                         if (distance < fallback_distance) {
 226                                 fallback_distance = distance;
 227                                 fallback = ns;
 228                         }
 229                         break;
 230                 default:
 231                         break;
 232                 }
 233         }
 234
 235         if (!found)
 236                 found = fallback;
 237         if (found)
 238                 rcu_assign_pointer(head->current_path[node], found);
 239         return found;
 240 }
 241
 242 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 243                 struct nvme_ns *ns)
 244 {
 245         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 246                         siblings);
 247         if (ns)
 248                 return ns;
 249         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 250 }
 251
 252 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
 253                 int node, struct nvme_ns *old)
 254 {
 255         struct nvme_ns *ns, *found = NULL;
 256
 257         if (list_is_singular(&head->list)) {
 258                 if (nvme_path_is_disabled(old))
 259                         return NULL;
 260                 return old;
 261         }
 262
 263         for (ns = nvme_next_ns(head, old);
 264              ns && ns != old;
 265              ns = nvme_next_ns(head, ns)) {
 266                 if (nvme_path_is_disabled(ns))
 267                         continue;
 268
 269                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 270                         found = ns;
 271                         goto out;
 272                 }
 273                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 274                         found = ns;
 275         }
 276
 277         /*
 278          * The loop above skips the current path for round-robin semantics.
 279          * Fall back to the current path if either:
 280          *  - no other optimized path found and current is optimized,
 281          *  - no other usable path found and current is usable.
 282          */
 283         if (!nvme_path_is_disabled(old) &&
 284             (old->ana_state == NVME_ANA_OPTIMIZED ||
 285              (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
 286                 return old;
 287
 288         if (!found)
 289                 return NULL;
 290 out:
 291         rcu_assign_pointer(head->current_path[node], found);
 292         return found;
 293 }
 294
 295 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 296 {
 297         return ns->ctrl->state == NVME_CTRL_LIVE &&
 298                 ns->ana_state == NVME_ANA_OPTIMIZED;
 299 }
 300
 301 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 302 {
 303         int node = numa_node_id();
 304         struct nvme_ns *ns;
 305
 306         ns = srcu_dereference(head->current_path[node], &head->srcu);
 307         if (unlikely(!ns))
 308                 return __nvme_find_path(head, node);
 309
 310         if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
 311                 return nvme_round_robin_path(head, node, ns);
 312         if (unlikely(!nvme_path_is_optimized(ns)))
 313                 return __nvme_find_path(head, node);
 314         return ns;
 315 }
 316
 317 static bool nvme_available_path(struct nvme_ns_head *head)
 318 {
 319         struct nvme_ns *ns;
 320
 321         list_for_each_entry_rcu(ns, &head->list, siblings) {
 322                 if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
 323                         continue;
 324                 switch (ns->ctrl->state) {
 325                 case NVME_CTRL_LIVE:
 326                 case NVME_CTRL_RESETTING:
 327                 case NVME_CTRL_CONNECTING:
 328                         /* fallthru */
 329                         return true;
 330                 default:
 331                         break;
 332                 }
 333         }
 334         return false;
 335 }
 336
 337 static void nvme_ns_head_submit_bio(struct bio *bio)
 338 {
 339         struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
 340         struct device *dev = disk_to_dev(head->disk);
 341         struct nvme_ns *ns;
 342         int srcu_idx;
 343
 344         /*
 345          * The namespace might be going away and the bio might be moved to a
 346          * different queue via blk_steal_bios(), so we need to use the bio_split
 347          * pool from the original queue to allocate the bvecs from.
 348          */
 349         blk_queue_split(&bio);
 350
 351         srcu_idx = srcu_read_lock(&head->srcu);
 352         ns = nvme_find_path(head);
 353         if (likely(ns)) {
 354                 bio_set_dev(bio, ns->disk->part0);
 355                 bio->bi_opf |= REQ_NVME_MPATH;
 356                 trace_block_bio_remap(bio, disk_devt(ns->head->disk),
 357                                       bio->bi_iter.bi_sector);
 358                 submit_bio_noacct(bio);
 359         } else if (nvme_available_path(head)) {
 360                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 361
 362                 spin_lock_irq(&head->requeue_lock);
 363                 bio_list_add(&head->requeue_list, bio);
 364                 spin_unlock_irq(&head->requeue_lock);
 365         } else {
 366                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 367
 368                 bio_io_error(bio);
 369         }
 370
 371         srcu_read_unlock(&head->srcu, srcu_idx);
 372 }
 373
 374 static int nvme_ns_head_open(struct block_device *bdev, fmode_t mode)
 375 {
 376         if (!nvme_tryget_ns_head(bdev->bd_disk->private_data))
 377                 return -ENXIO;
 378         return 0;
 379 }
 380
 381 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode)
 382 {
 383         nvme_put_ns_head(disk->private_data);
 384 }
 385
 386 #ifdef CONFIG_BLK_DEV_ZONED
 387 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
 388                 unsigned int nr_zones, report_zones_cb cb, void *data)
 389 {
 390         struct nvme_ns_head *head = disk->private_data;
 391         struct nvme_ns *ns;
 392         int srcu_idx, ret = -EWOULDBLOCK;
 393
 394         srcu_idx = srcu_read_lock(&head->srcu);
 395         ns = nvme_find_path(head);
 396         if (ns)
 397                 ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
 398         srcu_read_unlock(&head->srcu, srcu_idx);
 399         return ret;
 400 }
 401 #else
 402 #define nvme_ns_head_report_zones       NULL
 403 #endif /* CONFIG_BLK_DEV_ZONED */
 404
 405 const struct block_device_operations nvme_ns_head_ops = {
 406         .owner          = THIS_MODULE,
 407         .submit_bio     = nvme_ns_head_submit_bio,
 408         .open           = nvme_ns_head_open,
 409         .release        = nvme_ns_head_release,
 410         .ioctl          = nvme_ns_head_ioctl,
 411         .getgeo         = nvme_getgeo,
 412         .report_zones   = nvme_ns_head_report_zones,
 413         .pr_ops         = &nvme_pr_ops,
 414 };
 415
 416 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
 417 {
 418         return container_of(cdev, struct nvme_ns_head, cdev);
 419 }
 420
 421 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
 422 {
 423         if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
 424                 return -ENXIO;
 425         return 0;
 426 }
 427
 428 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
 429 {
 430         nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
 431         return 0;
 432 }
 433
 434 static const struct file_operations nvme_ns_head_chr_fops = {
 435         .owner          = THIS_MODULE,
 436         .open           = nvme_ns_head_chr_open,
 437         .release        = nvme_ns_head_chr_release,
 438         .unlocked_ioctl = nvme_ns_head_chr_ioctl,
 439         .compat_ioctl   = compat_ptr_ioctl,
 440         .uring_cmd      = nvme_ns_head_chr_uring_cmd,
 441 };
 442
 443 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
 444 {
 445         int ret;
 446
 447         head->cdev_device.parent = &head->subsys->dev;
 448         ret = dev_set_name(&head->cdev_device, "ng%dn%d",
 449                            head->subsys->instance, head->instance);
 450         if (ret)
 451                 return ret;
 452         ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
 453                             &nvme_ns_head_chr_fops, THIS_MODULE);
 454         return ret;
 455 }
 456
 457 static void nvme_requeue_work(struct work_struct *work)
 458 {
 459         struct nvme_ns_head *head =
 460                 container_of(work, struct nvme_ns_head, requeue_work);
 461         struct bio *bio, *next;
 462
 463         spin_lock_irq(&head->requeue_lock);
 464         next = bio_list_get(&head->requeue_list);
 465         spin_unlock_irq(&head->requeue_lock);
 466
 467         while ((bio = next) != NULL) {
 468                 next = bio->bi_next;
 469                 bio->bi_next = NULL;
 470
 471                 submit_bio_noacct(bio);
 472         }
 473 }
 474
 475 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 476 {
 477         bool vwc = false;
 478
 479         mutex_init(&head->lock);
 480         bio_list_init(&head->requeue_list);
 481         spin_lock_init(&head->requeue_lock);
 482         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 483
 484         /*
 485          * Add a multipath node if the subsystems supports multiple controllers.
 486          * We also do this for private namespaces as the namespace sharing flag
 487          * could change after a rescan.
 488          */
 489         if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
 490             !nvme_is_unique_nsid(ctrl, head) || !multipath)
 491                 return 0;
 492
 493         head->disk = blk_alloc_disk(ctrl->numa_node);
 494         if (!head->disk)
 495                 return -ENOMEM;
 496         head->disk->fops = &nvme_ns_head_ops;
 497         head->disk->private_data = head;
 498         sprintf(head->disk->disk_name, "nvme%dn%d",
 499                         ctrl->subsys->instance, head->instance);
 500
 501         blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
 502         blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
 503         /*
 504          * This assumes all controllers that refer to a namespace either
 505          * support poll queues or not.  That is not a strict guarantee,
 506          * but if the assumption is wrong the effect is only suboptimal
 507          * performance but not correctness problem.
 508          */
 509         if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL &&
 510             ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues)
 511                 blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue);
 512
 513         /* set to a default value of 512 until the disk is validated */
 514         blk_queue_logical_block_size(head->disk->queue, 512);
 515         blk_set_stacking_limits(&head->disk->queue->limits);
 516
 517         /* we need to propagate up the VMC settings */
 518         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
 519                 vwc = true;
 520         blk_queue_write_cache(head->disk->queue, vwc, vwc);
 521         return 0;
 522 }
 523
 524 static void nvme_mpath_set_live(struct nvme_ns *ns)
 525 {
 526         struct nvme_ns_head *head = ns->head;
 527         int rc;
 528
 529         if (!head->disk)
 530                 return;
 531
 532         /*
 533          * test_and_set_bit() is used because it is protecting against two nvme
 534          * paths simultaneously calling device_add_disk() on the same namespace
 535          * head.
 536          */
 537         if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 538                 rc = device_add_disk(&head->subsys->dev, head->disk,
 539                                      nvme_ns_id_attr_groups);
 540                 if (rc) {
 541                         clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
 542                         return;
 543                 }
 544                 nvme_add_ns_head_cdev(head);
 545         }
 546
 547         mutex_lock(&head->lock);
 548         if (nvme_path_is_optimized(ns)) {
 549                 int node, srcu_idx;
 550
 551                 srcu_idx = srcu_read_lock(&head->srcu);
 552                 for_each_node(node)
 553                         __nvme_find_path(head, node);
 554                 srcu_read_unlock(&head->srcu, srcu_idx);
 555         }
 556         mutex_unlock(&head->lock);
 557
 558         synchronize_srcu(&head->srcu);
 559         kblockd_schedule_work(&head->requeue_work);
 560 }
 561
 562 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 563                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 564                         void *))
 565 {
 566         void *base = ctrl->ana_log_buf;
 567         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 568         int error, i;
 569
 570         lockdep_assert_held(&ctrl->ana_lock);
 571
 572         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 573                 struct nvme_ana_group_desc *desc = base + offset;
 574                 u32 nr_nsids;
 575                 size_t nsid_buf_size;
 576
 577                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 578                         return -EINVAL;
 579
 580                 nr_nsids = le32_to_cpu(desc->nnsids);
 581                 nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
 582
 583                 if (WARN_ON_ONCE(desc->grpid == 0))
 584                         return -EINVAL;
 585                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 586                         return -EINVAL;
 587                 if (WARN_ON_ONCE(desc->state == 0))
 588                         return -EINVAL;
 589                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 590                         return -EINVAL;
 591
 592                 offset += sizeof(*desc);
 593                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 594                         return -EINVAL;
 595
 596                 error = cb(ctrl, desc, data);
 597                 if (error)
 598                         return error;
 599
 600                 offset += nsid_buf_size;
 601         }
 602
 603         return 0;
 604 }
 605
 606 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 607 {
 608         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 609 }
 610
 611 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 612                 struct nvme_ns *ns)
 613 {
 614         ns->ana_grpid = le32_to_cpu(desc->grpid);
 615         ns->ana_state = desc->state;
 616         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 617         /*
 618          * nvme_mpath_set_live() will trigger I/O to the multipath path device
 619          * and in turn to this path device.  However we cannot accept this I/O
 620          * if the controller is not live.  This may deadlock if called from
 621          * nvme_mpath_init_identify() and the ctrl will never complete
 622          * initialization, preventing I/O from completing.  For this case we
 623          * will reprocess the ANA log page in nvme_mpath_update() once the
 624          * controller is ready.
 625          */
 626         if (nvme_state_is_live(ns->ana_state) &&
 627             ns->ctrl->state == NVME_CTRL_LIVE)
 628                 nvme_mpath_set_live(ns);
 629 }
 630
 631 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 632                 struct nvme_ana_group_desc *desc, void *data)
 633 {
 634         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 635         unsigned *nr_change_groups = data;
 636         struct nvme_ns *ns;
 637
 638         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 639                         le32_to_cpu(desc->grpid),
 640                         nvme_ana_state_names[desc->state]);
 641
 642         if (desc->state == NVME_ANA_CHANGE)
 643                 (*nr_change_groups)++;
 644
 645         if (!nr_nsids)
 646                 return 0;
 647
 648         down_read(&ctrl->namespaces_rwsem);
 649         list_for_each_entry(ns, &ctrl->namespaces, list) {
 650                 unsigned nsid;
 651 again:
 652                 nsid = le32_to_cpu(desc->nsids[n]);
 653                 if (ns->head->ns_id < nsid)
 654                         continue;
 655                 if (ns->head->ns_id == nsid)
 656                         nvme_update_ns_ana_state(desc, ns);
 657                 if (++n == nr_nsids)
 658                         break;
 659                 if (ns->head->ns_id > nsid)
 660                         goto again;
 661         }
 662         up_read(&ctrl->namespaces_rwsem);
 663         return 0;
 664 }
 665
 666 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 667 {
 668         u32 nr_change_groups = 0;
 669         int error;
 670
 671         mutex_lock(&ctrl->ana_lock);
 672         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
 673                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 674         if (error) {
 675                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 676                 goto out_unlock;
 677         }
 678
 679         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 680                         nvme_update_ana_state);
 681         if (error)
 682                 goto out_unlock;
 683
 684         /*
 685          * In theory we should have an ANATT timer per group as they might enter
 686          * the change state at different times.  But that is a lot of overhead
 687          * just to protect against a target that keeps entering new changes
 688          * states while never finishing previous ones.  But we'll still
 689          * eventually time out once all groups are in change state, so this
 690          * isn't a big deal.
 691          *
 692          * We also double the ANATT value to provide some slack for transports
 693          * or AEN processing overhead.
 694          */
 695         if (nr_change_groups)
 696                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 697         else
 698                 del_timer_sync(&ctrl->anatt_timer);
 699 out_unlock:
 700         mutex_unlock(&ctrl->ana_lock);
 701         return error;
 702 }
 703
 704 static void nvme_ana_work(struct work_struct *work)
 705 {
 706         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 707
 708         if (ctrl->state != NVME_CTRL_LIVE)
 709                 return;
 710
 711         nvme_read_ana_log(ctrl);
 712 }
 713
 714 void nvme_mpath_update(struct nvme_ctrl *ctrl)
 715 {
 716         u32 nr_change_groups = 0;
 717
 718         if (!ctrl->ana_log_buf)
 719                 return;
 720
 721         mutex_lock(&ctrl->ana_lock);
 722         nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
 723         mutex_unlock(&ctrl->ana_lock);
 724 }
 725
 726 static void nvme_anatt_timeout(struct timer_list *t)
 727 {
 728         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 729
 730         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 731         nvme_reset_ctrl(ctrl);
 732 }
 733
 734 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 735 {
 736         if (!nvme_ctrl_use_ana(ctrl))
 737                 return;
 738         del_timer_sync(&ctrl->anatt_timer);
 739         cancel_work_sync(&ctrl->ana_work);
 740 }
 741
 742 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 743         struct device_attribute subsys_attr_##_name =   \
 744                 __ATTR(_name, _mode, _show, _store)
 745
 746 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 747                 struct device_attribute *attr, char *buf)
 748 {
 749         struct nvme_subsystem *subsys =
 750                 container_of(dev, struct nvme_subsystem, dev);
 751
 752         return sysfs_emit(buf, "%s\n",
 753                           nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 754 }
 755
 756 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 757                 struct device_attribute *attr, const char *buf, size_t count)
 758 {
 759         struct nvme_subsystem *subsys =
 760                 container_of(dev, struct nvme_subsystem, dev);
 761         int i;
 762
 763         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 764                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 765                         WRITE_ONCE(subsys->iopolicy, i);
 766                         return count;
 767                 }
 768         }
 769
 770         return -EINVAL;
 771 }
 772 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 773                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 774
 775 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 776                 char *buf)
 777 {
 778         return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 779 }
 780 DEVICE_ATTR_RO(ana_grpid);
 781
 782 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 783                 char *buf)
 784 {
 785         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 786
 787         return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 788 }
 789 DEVICE_ATTR_RO(ana_state);
 790
 791 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
 792                 struct nvme_ana_group_desc *desc, void *data)
 793 {
 794         struct nvme_ana_group_desc *dst = data;
 795
 796         if (desc->grpid != dst->grpid)
 797                 return 0;
 798
 799         *dst = *desc;
 800         return -ENXIO; /* just break out of the loop */
 801 }
 802
 803 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
 804 {
 805         if (nvme_ctrl_use_ana(ns->ctrl)) {
 806                 struct nvme_ana_group_desc desc = {
 807                         .grpid = id->anagrpid,
 808                         .state = 0,
 809                 };
 810
 811                 mutex_lock(&ns->ctrl->ana_lock);
 812                 ns->ana_grpid = le32_to_cpu(id->anagrpid);
 813                 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
 814                 mutex_unlock(&ns->ctrl->ana_lock);
 815                 if (desc.state) {
 816                         /* found the group desc: update */
 817                         nvme_update_ns_ana_state(&desc, ns);
 818                 } else {
 819                         /* group desc not found: trigger a re-read */
 820                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
 821                         queue_work(nvme_wq, &ns->ctrl->ana_work);
 822                 }
 823         } else {
 824                 ns->ana_state = NVME_ANA_OPTIMIZED;
 825                 nvme_mpath_set_live(ns);
 826         }
 827
 828         if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
 829                 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
 830                                    ns->head->disk->queue);
 831 #ifdef CONFIG_BLK_DEV_ZONED
 832         if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
 833                 ns->head->disk->queue->nr_zones = ns->queue->nr_zones;
 834 #endif
 835 }
 836
 837 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
 838 {
 839         if (!head->disk)
 840                 return;
 841         kblockd_schedule_work(&head->requeue_work);
 842         if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 843                 nvme_cdev_del(&head->cdev, &head->cdev_device);
 844                 del_gendisk(head->disk);
 845         }
 846 }
 847
 848 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 849 {
 850         if (!head->disk)
 851                 return;
 852         blk_mark_disk_dead(head->disk);
 853         /* make sure all pending bios are cleaned up */
 854         kblockd_schedule_work(&head->requeue_work);
 855         flush_work(&head->requeue_work);
 856         blk_cleanup_disk(head->disk);
 857 }
 858
 859 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
 860 {
 861         mutex_init(&ctrl->ana_lock);
 862         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 863         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 864 }
 865
 866 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 867 {
 868         size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
 869         size_t ana_log_size;
 870         int error = 0;
 871
 872         /* check if multipath is enabled and we have the capability */
 873         if (!multipath || !ctrl->subsys ||
 874             !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
 875                 return 0;
 876
 877         if (!ctrl->max_namespaces ||
 878             ctrl->max_namespaces > le32_to_cpu(id->nn)) {
 879                 dev_err(ctrl->device,
 880                         "Invalid MNAN value %u\n", ctrl->max_namespaces);
 881                 return -EINVAL;
 882         }
 883
 884         ctrl->anacap = id->anacap;
 885         ctrl->anatt = id->anatt;
 886         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
 887         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
 888
 889         ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
 890                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
 891                 ctrl->max_namespaces * sizeof(__le32);
 892         if (ana_log_size > max_transfer_size) {
 893                 dev_err(ctrl->device,
 894                         "ANA log page size (%zd) larger than MDTS (%zd).\n",
 895                         ana_log_size, max_transfer_size);
 896                 dev_err(ctrl->device, "disabling ANA support.\n");
 897                 goto out_uninit;
 898         }
 899         if (ana_log_size > ctrl->ana_log_size) {
 900                 nvme_mpath_stop(ctrl);
 901                 nvme_mpath_uninit(ctrl);
 902                 ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
 903                 if (!ctrl->ana_log_buf)
 904                         return -ENOMEM;
 905         }
 906         ctrl->ana_log_size = ana_log_size;
 907         error = nvme_read_ana_log(ctrl);
 908         if (error)
 909                 goto out_uninit;
 910         return 0;
 911
 912 out_uninit:
 913         nvme_mpath_uninit(ctrl);
 914         return error;
 915 }
 916
 917 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 918 {
 919         kvfree(ctrl->ana_log_buf);
 920         ctrl->ana_log_buf = NULL;
 921         ctrl->ana_log_size = 0;
 922 }