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