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         [NVME_IOPOLICY_QD]      = "queue-depth",
  21 };
  22
  23 static int iopolicy = NVME_IOPOLICY_NUMA;
  24
  25 static int nvme_set_iopolicy(const char *val, const struct kernel_param *kp)
  26 {
  27         if (!val)
  28                 return -EINVAL;
  29         if (!strncmp(val, "numa", 4))
  30                 iopolicy = NVME_IOPOLICY_NUMA;
  31         else if (!strncmp(val, "round-robin", 11))
  32                 iopolicy = NVME_IOPOLICY_RR;
  33         else if (!strncmp(val, "queue-depth", 11))
  34                 iopolicy = NVME_IOPOLICY_QD;
  35         else
  36                 return -EINVAL;
  37
  38         return 0;
  39 }
  40
  41 static int nvme_get_iopolicy(char *buf, const struct kernel_param *kp)
  42 {
  43         return sprintf(buf, "%s\n", nvme_iopolicy_names[iopolicy]);
  44 }
  45
  46 module_param_call(iopolicy, nvme_set_iopolicy, nvme_get_iopolicy,
  47         &iopolicy, 0644);
  48 MODULE_PARM_DESC(iopolicy,
  49         "Default multipath I/O policy; 'numa' (default), 'round-robin' or 'queue-depth'");
  50
  51 void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys)
  52 {
  53         subsys->iopolicy = iopolicy;
  54 }
  55
  56 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
  57 {
  58         struct nvme_ns_head *h;
  59
  60         lockdep_assert_held(&subsys->lock);
  61         list_for_each_entry(h, &subsys->nsheads, entry)
  62                 if (h->disk)
  63                         blk_mq_unfreeze_queue(h->disk->queue);
  64 }
  65
  66 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
  67 {
  68         struct nvme_ns_head *h;
  69
  70         lockdep_assert_held(&subsys->lock);
  71         list_for_each_entry(h, &subsys->nsheads, entry)
  72                 if (h->disk)
  73                         blk_mq_freeze_queue_wait(h->disk->queue);
  74 }
  75
  76 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
  77 {
  78         struct nvme_ns_head *h;
  79
  80         lockdep_assert_held(&subsys->lock);
  81         list_for_each_entry(h, &subsys->nsheads, entry)
  82                 if (h->disk)
  83                         blk_freeze_queue_start(h->disk->queue);
  84 }
  85
  86 void nvme_failover_req(struct request *req)
  87 {
  88         struct nvme_ns *ns = req->q->queuedata;
  89         u16 status = nvme_req(req)->status & NVME_SCT_SC_MASK;
  90         unsigned long flags;
  91         struct bio *bio;
  92
  93         nvme_mpath_clear_current_path(ns);
  94
  95         /*
  96          * If we got back an ANA error, we know the controller is alive but not
  97          * ready to serve this namespace.  Kick of a re-read of the ANA
  98          * information page, and just try any other available path for now.
  99          */
 100         if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) {
 101                 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
 102                 queue_work(nvme_wq, &ns->ctrl->ana_work);
 103         }
 104
 105         spin_lock_irqsave(&ns->head->requeue_lock, flags);
 106         for (bio = req->bio; bio; bio = bio->bi_next) {
 107                 bio_set_dev(bio, ns->head->disk->part0);
 108                 if (bio->bi_opf & REQ_POLLED) {
 109                         bio->bi_opf &= ~REQ_POLLED;
 110                         bio->bi_cookie = BLK_QC_T_NONE;
 111                 }
 112                 /*
 113                  * The alternate request queue that we may end up submitting
 114                  * the bio to may be frozen temporarily, in this case REQ_NOWAIT
 115                  * will fail the I/O immediately with EAGAIN to the issuer.
 116                  * We are not in the issuer context which cannot block. Clear
 117                  * the flag to avoid spurious EAGAIN I/O failures.
 118                  */
 119                 bio->bi_opf &= ~REQ_NOWAIT;
 120         }
 121         blk_steal_bios(&ns->head->requeue_list, req);
 122         spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
 123
 124         nvme_req(req)->status = 0;
 125         nvme_end_req(req);
 126         kblockd_schedule_work(&ns->head->requeue_work);
 127 }
 128
 129 void nvme_mpath_start_request(struct request *rq)
 130 {
 131         struct nvme_ns *ns = rq->q->queuedata;
 132         struct gendisk *disk = ns->head->disk;
 133
 134         if (READ_ONCE(ns->head->subsys->iopolicy) == NVME_IOPOLICY_QD) {
 135                 atomic_inc(&ns->ctrl->nr_active);
 136                 nvme_req(rq)->flags |= NVME_MPATH_CNT_ACTIVE;
 137         }
 138
 139         if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq))
 140                 return;
 141
 142         nvme_req(rq)->flags |= NVME_MPATH_IO_STATS;
 143         nvme_req(rq)->start_time = bdev_start_io_acct(disk->part0, req_op(rq),
 144                                                       jiffies);
 145 }
 146 EXPORT_SYMBOL_GPL(nvme_mpath_start_request);
 147
 148 void nvme_mpath_end_request(struct request *rq)
 149 {
 150         struct nvme_ns *ns = rq->q->queuedata;
 151
 152         if (nvme_req(rq)->flags & NVME_MPATH_CNT_ACTIVE)
 153                 atomic_dec_if_positive(&ns->ctrl->nr_active);
 154
 155         if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
 156                 return;
 157         bdev_end_io_acct(ns->head->disk->part0, req_op(rq),
 158                          blk_rq_bytes(rq) >> SECTOR_SHIFT,
 159                          nvme_req(rq)->start_time);
 160 }
 161
 162 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 163 {
 164         struct nvme_ns *ns;
 165         int srcu_idx;
 166
 167         srcu_idx = srcu_read_lock(&ctrl->srcu);
 168         list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
 169                 if (!ns->head->disk)
 170                         continue;
 171                 kblockd_schedule_work(&ns->head->requeue_work);
 172                 if (nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
 173                         disk_uevent(ns->head->disk, KOBJ_CHANGE);
 174         }
 175         srcu_read_unlock(&ctrl->srcu, srcu_idx);
 176 }
 177
 178 static const char *nvme_ana_state_names[] = {
 179         [0]                             = "invalid state",
 180         [NVME_ANA_OPTIMIZED]            = "optimized",
 181         [NVME_ANA_NONOPTIMIZED]         = "non-optimized",
 182         [NVME_ANA_INACCESSIBLE]         = "inaccessible",
 183         [NVME_ANA_PERSISTENT_LOSS]      = "persistent-loss",
 184         [NVME_ANA_CHANGE]               = "change",
 185 };
 186
 187 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 188 {
 189         struct nvme_ns_head *head = ns->head;
 190         bool changed = false;
 191         int node;
 192
 193         if (!head)
 194                 goto out;
 195
 196         for_each_node(node) {
 197                 if (ns == rcu_access_pointer(head->current_path[node])) {
 198                         rcu_assign_pointer(head->current_path[node], NULL);
 199                         changed = true;
 200                 }
 201         }
 202 out:
 203         return changed;
 204 }
 205
 206 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 207 {
 208         struct nvme_ns *ns;
 209         int srcu_idx;
 210
 211         srcu_idx = srcu_read_lock(&ctrl->srcu);
 212         list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
 213                 nvme_mpath_clear_current_path(ns);
 214                 kblockd_schedule_work(&ns->head->requeue_work);
 215         }
 216         srcu_read_unlock(&ctrl->srcu, srcu_idx);
 217 }
 218
 219 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
 220 {
 221         struct nvme_ns_head *head = ns->head;
 222         sector_t capacity = get_capacity(head->disk);
 223         int node;
 224         int srcu_idx;
 225
 226         srcu_idx = srcu_read_lock(&head->srcu);
 227         list_for_each_entry_rcu(ns, &head->list, siblings) {
 228                 if (capacity != get_capacity(ns->disk))
 229                         clear_bit(NVME_NS_READY, &ns->flags);
 230         }
 231         srcu_read_unlock(&head->srcu, srcu_idx);
 232
 233         for_each_node(node)
 234                 rcu_assign_pointer(head->current_path[node], NULL);
 235         kblockd_schedule_work(&head->requeue_work);
 236 }
 237
 238 static bool nvme_path_is_disabled(struct nvme_ns *ns)
 239 {
 240         enum nvme_ctrl_state state = nvme_ctrl_state(ns->ctrl);
 241
 242         /*
 243          * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
 244          * still be able to complete assuming that the controller is connected.
 245          * Otherwise it will fail immediately and return to the requeue list.
 246          */
 247         if (state != NVME_CTRL_LIVE && state != NVME_CTRL_DELETING)
 248                 return true;
 249         if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
 250             !test_bit(NVME_NS_READY, &ns->flags))
 251                 return true;
 252         return false;
 253 }
 254
 255 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
 256 {
 257         int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
 258         struct nvme_ns *found = NULL, *fallback = NULL, *ns;
 259
 260         list_for_each_entry_rcu(ns, &head->list, siblings) {
 261                 if (nvme_path_is_disabled(ns))
 262                         continue;
 263
 264                 if (ns->ctrl->numa_node != NUMA_NO_NODE &&
 265                     READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
 266                         distance = node_distance(node, ns->ctrl->numa_node);
 267                 else
 268                         distance = LOCAL_DISTANCE;
 269
 270                 switch (ns->ana_state) {
 271                 case NVME_ANA_OPTIMIZED:
 272                         if (distance < found_distance) {
 273                                 found_distance = distance;
 274                                 found = ns;
 275                         }
 276                         break;
 277                 case NVME_ANA_NONOPTIMIZED:
 278                         if (distance < fallback_distance) {
 279                                 fallback_distance = distance;
 280                                 fallback = ns;
 281                         }
 282                         break;
 283                 default:
 284                         break;
 285                 }
 286         }
 287
 288         if (!found)
 289                 found = fallback;
 290         if (found)
 291                 rcu_assign_pointer(head->current_path[node], found);
 292         return found;
 293 }
 294
 295 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
 296                 struct nvme_ns *ns)
 297 {
 298         ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
 299                         siblings);
 300         if (ns)
 301                 return ns;
 302         return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
 303 }
 304
 305 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head)
 306 {
 307         struct nvme_ns *ns, *found = NULL;
 308         int node = numa_node_id();
 309         struct nvme_ns *old = srcu_dereference(head->current_path[node],
 310                                                &head->srcu);
 311
 312         if (unlikely(!old))
 313                 return __nvme_find_path(head, node);
 314
 315         if (list_is_singular(&head->list)) {
 316                 if (nvme_path_is_disabled(old))
 317                         return NULL;
 318                 return old;
 319         }
 320
 321         for (ns = nvme_next_ns(head, old);
 322              ns && ns != old;
 323              ns = nvme_next_ns(head, ns)) {
 324                 if (nvme_path_is_disabled(ns))
 325                         continue;
 326
 327                 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
 328                         found = ns;
 329                         goto out;
 330                 }
 331                 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
 332                         found = ns;
 333         }
 334
 335         /*
 336          * The loop above skips the current path for round-robin semantics.
 337          * Fall back to the current path if either:
 338          *  - no other optimized path found and current is optimized,
 339          *  - no other usable path found and current is usable.
 340          */
 341         if (!nvme_path_is_disabled(old) &&
 342             (old->ana_state == NVME_ANA_OPTIMIZED ||
 343              (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
 344                 return old;
 345
 346         if (!found)
 347                 return NULL;
 348 out:
 349         rcu_assign_pointer(head->current_path[node], found);
 350         return found;
 351 }
 352
 353 static struct nvme_ns *nvme_queue_depth_path(struct nvme_ns_head *head)
 354 {
 355         struct nvme_ns *best_opt = NULL, *best_nonopt = NULL, *ns;
 356         unsigned int min_depth_opt = UINT_MAX, min_depth_nonopt = UINT_MAX;
 357         unsigned int depth;
 358
 359         list_for_each_entry_rcu(ns, &head->list, siblings) {
 360                 if (nvme_path_is_disabled(ns))
 361                         continue;
 362
 363                 depth = atomic_read(&ns->ctrl->nr_active);
 364
 365                 switch (ns->ana_state) {
 366                 case NVME_ANA_OPTIMIZED:
 367                         if (depth < min_depth_opt) {
 368                                 min_depth_opt = depth;
 369                                 best_opt = ns;
 370                         }
 371                         break;
 372                 case NVME_ANA_NONOPTIMIZED:
 373                         if (depth < min_depth_nonopt) {
 374                                 min_depth_nonopt = depth;
 375                                 best_nonopt = ns;
 376                         }
 377                         break;
 378                 default:
 379                         break;
 380                 }
 381
 382                 if (min_depth_opt == 0)
 383                         return best_opt;
 384         }
 385
 386         return best_opt ? best_opt : best_nonopt;
 387 }
 388
 389 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 390 {
 391         return nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE &&
 392                 ns->ana_state == NVME_ANA_OPTIMIZED;
 393 }
 394
 395 static struct nvme_ns *nvme_numa_path(struct nvme_ns_head *head)
 396 {
 397         int node = numa_node_id();
 398         struct nvme_ns *ns;
 399
 400         ns = srcu_dereference(head->current_path[node], &head->srcu);
 401         if (unlikely(!ns))
 402                 return __nvme_find_path(head, node);
 403         if (unlikely(!nvme_path_is_optimized(ns)))
 404                 return __nvme_find_path(head, node);
 405         return ns;
 406 }
 407
 408 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
 409 {
 410         switch (READ_ONCE(head->subsys->iopolicy)) {
 411         case NVME_IOPOLICY_QD:
 412                 return nvme_queue_depth_path(head);
 413         case NVME_IOPOLICY_RR:
 414                 return nvme_round_robin_path(head);
 415         default:
 416                 return nvme_numa_path(head);
 417         }
 418 }
 419
 420 static bool nvme_available_path(struct nvme_ns_head *head)
 421 {
 422         struct nvme_ns *ns;
 423
 424         list_for_each_entry_rcu(ns, &head->list, siblings) {
 425                 if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
 426                         continue;
 427                 switch (nvme_ctrl_state(ns->ctrl)) {
 428                 case NVME_CTRL_LIVE:
 429                 case NVME_CTRL_RESETTING:
 430                 case NVME_CTRL_CONNECTING:
 431                         /* fallthru */
 432                         return true;
 433                 default:
 434                         break;
 435                 }
 436         }
 437         return false;
 438 }
 439
 440 static void nvme_ns_head_submit_bio(struct bio *bio)
 441 {
 442         struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
 443         struct device *dev = disk_to_dev(head->disk);
 444         struct nvme_ns *ns;
 445         int srcu_idx;
 446
 447         /*
 448          * The namespace might be going away and the bio might be moved to a
 449          * different queue via blk_steal_bios(), so we need to use the bio_split
 450          * pool from the original queue to allocate the bvecs from.
 451          */
 452         bio = bio_split_to_limits(bio);
 453         if (!bio)
 454                 return;
 455
 456         srcu_idx = srcu_read_lock(&head->srcu);
 457         ns = nvme_find_path(head);
 458         if (likely(ns)) {
 459                 bio_set_dev(bio, ns->disk->part0);
 460                 bio->bi_opf |= REQ_NVME_MPATH;
 461                 trace_block_bio_remap(bio, disk_devt(ns->head->disk),
 462                                       bio->bi_iter.bi_sector);
 463                 submit_bio_noacct(bio);
 464         } else if (nvme_available_path(head)) {
 465                 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
 466
 467                 spin_lock_irq(&head->requeue_lock);
 468                 bio_list_add(&head->requeue_list, bio);
 469                 spin_unlock_irq(&head->requeue_lock);
 470         } else {
 471                 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
 472
 473                 bio_io_error(bio);
 474         }
 475
 476         srcu_read_unlock(&head->srcu, srcu_idx);
 477 }
 478
 479 static int nvme_ns_head_open(struct gendisk *disk, blk_mode_t mode)
 480 {
 481         if (!nvme_tryget_ns_head(disk->private_data))
 482                 return -ENXIO;
 483         return 0;
 484 }
 485
 486 static void nvme_ns_head_release(struct gendisk *disk)
 487 {
 488         nvme_put_ns_head(disk->private_data);
 489 }
 490
 491 static int nvme_ns_head_get_unique_id(struct gendisk *disk, u8 id[16],
 492                 enum blk_unique_id type)
 493 {
 494         struct nvme_ns_head *head = disk->private_data;
 495         struct nvme_ns *ns;
 496         int srcu_idx, ret = -EWOULDBLOCK;
 497
 498         srcu_idx = srcu_read_lock(&head->srcu);
 499         ns = nvme_find_path(head);
 500         if (ns)
 501                 ret = nvme_ns_get_unique_id(ns, id, type);
 502         srcu_read_unlock(&head->srcu, srcu_idx);
 503         return ret;
 504 }
 505
 506 #ifdef CONFIG_BLK_DEV_ZONED
 507 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
 508                 unsigned int nr_zones, report_zones_cb cb, void *data)
 509 {
 510         struct nvme_ns_head *head = disk->private_data;
 511         struct nvme_ns *ns;
 512         int srcu_idx, ret = -EWOULDBLOCK;
 513
 514         srcu_idx = srcu_read_lock(&head->srcu);
 515         ns = nvme_find_path(head);
 516         if (ns)
 517                 ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
 518         srcu_read_unlock(&head->srcu, srcu_idx);
 519         return ret;
 520 }
 521 #else
 522 #define nvme_ns_head_report_zones       NULL
 523 #endif /* CONFIG_BLK_DEV_ZONED */
 524
 525 const struct block_device_operations nvme_ns_head_ops = {
 526         .owner          = THIS_MODULE,
 527         .submit_bio     = nvme_ns_head_submit_bio,
 528         .open           = nvme_ns_head_open,
 529         .release        = nvme_ns_head_release,
 530         .ioctl          = nvme_ns_head_ioctl,
 531         .compat_ioctl   = blkdev_compat_ptr_ioctl,
 532         .getgeo         = nvme_getgeo,
 533         .get_unique_id  = nvme_ns_head_get_unique_id,
 534         .report_zones   = nvme_ns_head_report_zones,
 535         .pr_ops         = &nvme_pr_ops,
 536 };
 537
 538 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
 539 {
 540         return container_of(cdev, struct nvme_ns_head, cdev);
 541 }
 542
 543 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
 544 {
 545         if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
 546                 return -ENXIO;
 547         return 0;
 548 }
 549
 550 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
 551 {
 552         nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
 553         return 0;
 554 }
 555
 556 static const struct file_operations nvme_ns_head_chr_fops = {
 557         .owner          = THIS_MODULE,
 558         .open           = nvme_ns_head_chr_open,
 559         .release        = nvme_ns_head_chr_release,
 560         .unlocked_ioctl = nvme_ns_head_chr_ioctl,
 561         .compat_ioctl   = compat_ptr_ioctl,
 562         .uring_cmd      = nvme_ns_head_chr_uring_cmd,
 563         .uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
 564 };
 565
 566 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
 567 {
 568         int ret;
 569
 570         head->cdev_device.parent = &head->subsys->dev;
 571         ret = dev_set_name(&head->cdev_device, "ng%dn%d",
 572                            head->subsys->instance, head->instance);
 573         if (ret)
 574                 return ret;
 575         ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
 576                             &nvme_ns_head_chr_fops, THIS_MODULE);
 577         return ret;
 578 }
 579
 580 static void nvme_requeue_work(struct work_struct *work)
 581 {
 582         struct nvme_ns_head *head =
 583                 container_of(work, struct nvme_ns_head, requeue_work);
 584         struct bio *bio, *next;
 585
 586         spin_lock_irq(&head->requeue_lock);
 587         next = bio_list_get(&head->requeue_list);
 588         spin_unlock_irq(&head->requeue_lock);
 589
 590         while ((bio = next) != NULL) {
 591                 next = bio->bi_next;
 592                 bio->bi_next = NULL;
 593
 594                 submit_bio_noacct(bio);
 595         }
 596 }
 597
 598 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 599 {
 600         struct queue_limits lim;
 601
 602         mutex_init(&head->lock);
 603         bio_list_init(&head->requeue_list);
 604         spin_lock_init(&head->requeue_lock);
 605         INIT_WORK(&head->requeue_work, nvme_requeue_work);
 606
 607         /*
 608          * Add a multipath node if the subsystems supports multiple controllers.
 609          * We also do this for private namespaces as the namespace sharing flag
 610          * could change after a rescan.
 611          */
 612         if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
 613             !nvme_is_unique_nsid(ctrl, head) || !multipath)
 614                 return 0;
 615
 616         blk_set_stacking_limits(&lim);
 617         lim.dma_alignment = 3;
 618         lim.features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT | BLK_FEAT_POLL;
 619         if (head->ids.csi != NVME_CSI_ZNS)
 620                 lim.max_zone_append_sectors = 0;
 621
 622         head->disk = blk_alloc_disk(&lim, ctrl->numa_node);
 623         if (IS_ERR(head->disk))
 624                 return PTR_ERR(head->disk);
 625         head->disk->fops = &nvme_ns_head_ops;
 626         head->disk->private_data = head;
 627         sprintf(head->disk->disk_name, "nvme%dn%d",
 628                         ctrl->subsys->instance, head->instance);
 629         return 0;
 630 }
 631
 632 static void nvme_mpath_set_live(struct nvme_ns *ns)
 633 {
 634         struct nvme_ns_head *head = ns->head;
 635         int rc;
 636
 637         if (!head->disk)
 638                 return;
 639
 640         /*
 641          * test_and_set_bit() is used because it is protecting against two nvme
 642          * paths simultaneously calling device_add_disk() on the same namespace
 643          * head.
 644          */
 645         if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 646                 rc = device_add_disk(&head->subsys->dev, head->disk,
 647                                      nvme_ns_attr_groups);
 648                 if (rc) {
 649                         clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
 650                         return;
 651                 }
 652                 nvme_add_ns_head_cdev(head);
 653         }
 654
 655         mutex_lock(&head->lock);
 656         if (nvme_path_is_optimized(ns)) {
 657                 int node, srcu_idx;
 658
 659                 srcu_idx = srcu_read_lock(&head->srcu);
 660                 for_each_online_node(node)
 661                         __nvme_find_path(head, node);
 662                 srcu_read_unlock(&head->srcu, srcu_idx);
 663         }
 664         mutex_unlock(&head->lock);
 665
 666         synchronize_srcu(&head->srcu);
 667         kblockd_schedule_work(&head->requeue_work);
 668 }
 669
 670 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
 671                 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
 672                         void *))
 673 {
 674         void *base = ctrl->ana_log_buf;
 675         size_t offset = sizeof(struct nvme_ana_rsp_hdr);
 676         int error, i;
 677
 678         lockdep_assert_held(&ctrl->ana_lock);
 679
 680         for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
 681                 struct nvme_ana_group_desc *desc = base + offset;
 682                 u32 nr_nsids;
 683                 size_t nsid_buf_size;
 684
 685                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
 686                         return -EINVAL;
 687
 688                 nr_nsids = le32_to_cpu(desc->nnsids);
 689                 nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
 690
 691                 if (WARN_ON_ONCE(desc->grpid == 0))
 692                         return -EINVAL;
 693                 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
 694                         return -EINVAL;
 695                 if (WARN_ON_ONCE(desc->state == 0))
 696                         return -EINVAL;
 697                 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
 698                         return -EINVAL;
 699
 700                 offset += sizeof(*desc);
 701                 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
 702                         return -EINVAL;
 703
 704                 error = cb(ctrl, desc, data);
 705                 if (error)
 706                         return error;
 707
 708                 offset += nsid_buf_size;
 709         }
 710
 711         return 0;
 712 }
 713
 714 static inline bool nvme_state_is_live(enum nvme_ana_state state)
 715 {
 716         return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
 717 }
 718
 719 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
 720                 struct nvme_ns *ns)
 721 {
 722         ns->ana_grpid = le32_to_cpu(desc->grpid);
 723         ns->ana_state = desc->state;
 724         clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
 725         /*
 726          * nvme_mpath_set_live() will trigger I/O to the multipath path device
 727          * and in turn to this path device.  However we cannot accept this I/O
 728          * if the controller is not live.  This may deadlock if called from
 729          * nvme_mpath_init_identify() and the ctrl will never complete
 730          * initialization, preventing I/O from completing.  For this case we
 731          * will reprocess the ANA log page in nvme_mpath_update() once the
 732          * controller is ready.
 733          */
 734         if (nvme_state_is_live(ns->ana_state) &&
 735             nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
 736                 nvme_mpath_set_live(ns);
 737 }
 738
 739 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
 740                 struct nvme_ana_group_desc *desc, void *data)
 741 {
 742         u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
 743         unsigned *nr_change_groups = data;
 744         struct nvme_ns *ns;
 745         int srcu_idx;
 746
 747         dev_dbg(ctrl->device, "ANA group %d: %s.\n",
 748                         le32_to_cpu(desc->grpid),
 749                         nvme_ana_state_names[desc->state]);
 750
 751         if (desc->state == NVME_ANA_CHANGE)
 752                 (*nr_change_groups)++;
 753
 754         if (!nr_nsids)
 755                 return 0;
 756
 757         srcu_idx = srcu_read_lock(&ctrl->srcu);
 758         list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
 759                 unsigned nsid;
 760 again:
 761                 nsid = le32_to_cpu(desc->nsids[n]);
 762                 if (ns->head->ns_id < nsid)
 763                         continue;
 764                 if (ns->head->ns_id == nsid)
 765                         nvme_update_ns_ana_state(desc, ns);
 766                 if (++n == nr_nsids)
 767                         break;
 768                 if (ns->head->ns_id > nsid)
 769                         goto again;
 770         }
 771         srcu_read_unlock(&ctrl->srcu, srcu_idx);
 772         return 0;
 773 }
 774
 775 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
 776 {
 777         u32 nr_change_groups = 0;
 778         int error;
 779
 780         mutex_lock(&ctrl->ana_lock);
 781         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
 782                         ctrl->ana_log_buf, ctrl->ana_log_size, 0);
 783         if (error) {
 784                 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
 785                 goto out_unlock;
 786         }
 787
 788         error = nvme_parse_ana_log(ctrl, &nr_change_groups,
 789                         nvme_update_ana_state);
 790         if (error)
 791                 goto out_unlock;
 792
 793         /*
 794          * In theory we should have an ANATT timer per group as they might enter
 795          * the change state at different times.  But that is a lot of overhead
 796          * just to protect against a target that keeps entering new changes
 797          * states while never finishing previous ones.  But we'll still
 798          * eventually time out once all groups are in change state, so this
 799          * isn't a big deal.
 800          *
 801          * We also double the ANATT value to provide some slack for transports
 802          * or AEN processing overhead.
 803          */
 804         if (nr_change_groups)
 805                 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
 806         else
 807                 del_timer_sync(&ctrl->anatt_timer);
 808 out_unlock:
 809         mutex_unlock(&ctrl->ana_lock);
 810         return error;
 811 }
 812
 813 static void nvme_ana_work(struct work_struct *work)
 814 {
 815         struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
 816
 817         if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
 818                 return;
 819
 820         nvme_read_ana_log(ctrl);
 821 }
 822
 823 void nvme_mpath_update(struct nvme_ctrl *ctrl)
 824 {
 825         u32 nr_change_groups = 0;
 826
 827         if (!ctrl->ana_log_buf)
 828                 return;
 829
 830         mutex_lock(&ctrl->ana_lock);
 831         nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
 832         mutex_unlock(&ctrl->ana_lock);
 833 }
 834
 835 static void nvme_anatt_timeout(struct timer_list *t)
 836 {
 837         struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
 838
 839         dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
 840         nvme_reset_ctrl(ctrl);
 841 }
 842
 843 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 844 {
 845         if (!nvme_ctrl_use_ana(ctrl))
 846                 return;
 847         del_timer_sync(&ctrl->anatt_timer);
 848         cancel_work_sync(&ctrl->ana_work);
 849 }
 850
 851 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
 852         struct device_attribute subsys_attr_##_name =   \
 853                 __ATTR(_name, _mode, _show, _store)
 854
 855 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
 856                 struct device_attribute *attr, char *buf)
 857 {
 858         struct nvme_subsystem *subsys =
 859                 container_of(dev, struct nvme_subsystem, dev);
 860
 861         return sysfs_emit(buf, "%s\n",
 862                           nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
 863 }
 864
 865 static void nvme_subsys_iopolicy_update(struct nvme_subsystem *subsys,
 866                 int iopolicy)
 867 {
 868         struct nvme_ctrl *ctrl;
 869         int old_iopolicy = READ_ONCE(subsys->iopolicy);
 870
 871         if (old_iopolicy == iopolicy)
 872                 return;
 873
 874         WRITE_ONCE(subsys->iopolicy, iopolicy);
 875
 876         /* iopolicy changes clear the mpath by design */
 877         mutex_lock(&nvme_subsystems_lock);
 878         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 879                 nvme_mpath_clear_ctrl_paths(ctrl);
 880         mutex_unlock(&nvme_subsystems_lock);
 881
 882         pr_notice("subsysnqn %s iopolicy changed from %s to %s\n",
 883                         subsys->subnqn,
 884                         nvme_iopolicy_names[old_iopolicy],
 885                         nvme_iopolicy_names[iopolicy]);
 886 }
 887
 888 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
 889                 struct device_attribute *attr, const char *buf, size_t count)
 890 {
 891         struct nvme_subsystem *subsys =
 892                 container_of(dev, struct nvme_subsystem, dev);
 893         int i;
 894
 895         for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
 896                 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
 897                         nvme_subsys_iopolicy_update(subsys, i);
 898                         return count;
 899                 }
 900         }
 901
 902         return -EINVAL;
 903 }
 904 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
 905                       nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
 906
 907 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
 908                 char *buf)
 909 {
 910         return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
 911 }
 912 DEVICE_ATTR_RO(ana_grpid);
 913
 914 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
 915                 char *buf)
 916 {
 917         struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
 918
 919         return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
 920 }
 921 DEVICE_ATTR_RO(ana_state);
 922
 923 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
 924                 struct nvme_ana_group_desc *desc, void *data)
 925 {
 926         struct nvme_ana_group_desc *dst = data;
 927
 928         if (desc->grpid != dst->grpid)
 929                 return 0;
 930
 931         *dst = *desc;
 932         return -ENXIO; /* just break out of the loop */
 933 }
 934
 935 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
 936 {
 937         if (nvme_ctrl_use_ana(ns->ctrl)) {
 938                 struct nvme_ana_group_desc desc = {
 939                         .grpid = anagrpid,
 940                         .state = 0,
 941                 };
 942
 943                 mutex_lock(&ns->ctrl->ana_lock);
 944                 ns->ana_grpid = le32_to_cpu(anagrpid);
 945                 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
 946                 mutex_unlock(&ns->ctrl->ana_lock);
 947                 if (desc.state) {
 948                         /* found the group desc: update */
 949                         nvme_update_ns_ana_state(&desc, ns);
 950                 } else {
 951                         /* group desc not found: trigger a re-read */
 952                         set_bit(NVME_NS_ANA_PENDING, &ns->flags);
 953                         queue_work(nvme_wq, &ns->ctrl->ana_work);
 954                 }
 955         } else {
 956                 ns->ana_state = NVME_ANA_OPTIMIZED;
 957                 nvme_mpath_set_live(ns);
 958         }
 959
 960 #ifdef CONFIG_BLK_DEV_ZONED
 961         if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
 962                 ns->head->disk->nr_zones = ns->disk->nr_zones;
 963 #endif
 964 }
 965
 966 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
 967 {
 968         if (!head->disk)
 969                 return;
 970         kblockd_schedule_work(&head->requeue_work);
 971         if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
 972                 nvme_cdev_del(&head->cdev, &head->cdev_device);
 973                 del_gendisk(head->disk);
 974         }
 975 }
 976
 977 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 978 {
 979         if (!head->disk)
 980                 return;
 981         /* make sure all pending bios are cleaned up */
 982         kblockd_schedule_work(&head->requeue_work);
 983         flush_work(&head->requeue_work);
 984         put_disk(head->disk);
 985 }
 986
 987 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
 988 {
 989         mutex_init(&ctrl->ana_lock);
 990         timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
 991         INIT_WORK(&ctrl->ana_work, nvme_ana_work);
 992 }
 993
 994 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
 995 {
 996         size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
 997         size_t ana_log_size;
 998         int error = 0;
 999
1000         /* check if multipath is enabled and we have the capability */
1001         if (!multipath || !ctrl->subsys ||
1002             !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
1003                 return 0;
1004
1005         /* initialize this in the identify path to cover controller resets */
1006         atomic_set(&ctrl->nr_active, 0);
1007
1008         if (!ctrl->max_namespaces ||
1009             ctrl->max_namespaces > le32_to_cpu(id->nn)) {
1010                 dev_err(ctrl->device,
1011                         "Invalid MNAN value %u\n", ctrl->max_namespaces);
1012                 return -EINVAL;
1013         }
1014
1015         ctrl->anacap = id->anacap;
1016         ctrl->anatt = id->anatt;
1017         ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
1018         ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
1019
1020         ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
1021                 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
1022                 ctrl->max_namespaces * sizeof(__le32);
1023         if (ana_log_size > max_transfer_size) {
1024                 dev_err(ctrl->device,
1025                         "ANA log page size (%zd) larger than MDTS (%zd).\n",
1026                         ana_log_size, max_transfer_size);
1027                 dev_err(ctrl->device, "disabling ANA support.\n");
1028                 goto out_uninit;
1029         }
1030         if (ana_log_size > ctrl->ana_log_size) {
1031                 nvme_mpath_stop(ctrl);
1032                 nvme_mpath_uninit(ctrl);
1033                 ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
1034                 if (!ctrl->ana_log_buf)
1035                         return -ENOMEM;
1036         }
1037         ctrl->ana_log_size = ana_log_size;
1038         error = nvme_read_ana_log(ctrl);
1039         if (error)
1040                 goto out_uninit;
1041         return 0;
1042
1043 out_uninit:
1044         nvme_mpath_uninit(ctrl);
1045         return error;
1046 }
1047
1048 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
1049 {
1050         kvfree(ctrl->ana_log_buf);
1051         ctrl->ana_log_buf = NULL;
1052         ctrl->ana_log_size = 0;
1053 }