1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2017-2018 Christoph Hellwig.
6 #include <linux/backing-dev.h>
7 #include <linux/moduleparam.h>
8 #include <trace/events/block.h>
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");
16 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
18 struct nvme_ns_head *h;
20 lockdep_assert_held(&subsys->lock);
21 list_for_each_entry(h, &subsys->nsheads, entry)
23 blk_mq_unfreeze_queue(h->disk->queue);
26 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
28 struct nvme_ns_head *h;
30 lockdep_assert_held(&subsys->lock);
31 list_for_each_entry(h, &subsys->nsheads, entry)
33 blk_mq_freeze_queue_wait(h->disk->queue);
36 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
38 struct nvme_ns_head *h;
40 lockdep_assert_held(&subsys->lock);
41 list_for_each_entry(h, &subsys->nsheads, entry)
43 blk_freeze_queue_start(h->disk->queue);
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
53 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
54 struct nvme_ctrl *ctrl, int *flags)
57 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
58 } else if (ns->head->disk) {
59 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
60 ctrl->instance, ns->head->instance);
61 *flags = GENHD_FL_HIDDEN;
63 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
68 void nvme_failover_req(struct request *req)
70 struct nvme_ns *ns = req->q->queuedata;
71 u16 status = nvme_req(req)->status & 0x7ff;
74 nvme_mpath_clear_current_path(ns);
77 * If we got back an ANA error, we know the controller is alive but not
78 * ready to serve this namespace. Kick of a re-read of the ANA
79 * information page, and just try any other available path for now.
81 if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) {
82 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
83 queue_work(nvme_wq, &ns->ctrl->ana_work);
86 spin_lock_irqsave(&ns->head->requeue_lock, flags);
87 blk_steal_bios(&ns->head->requeue_list, req);
88 spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
90 blk_mq_end_request(req, 0);
91 kblockd_schedule_work(&ns->head->requeue_work);
94 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
98 down_read(&ctrl->namespaces_rwsem);
99 list_for_each_entry(ns, &ctrl->namespaces, list) {
101 kblockd_schedule_work(&ns->head->requeue_work);
103 up_read(&ctrl->namespaces_rwsem);
106 static const char *nvme_ana_state_names[] = {
107 [0] = "invalid state",
108 [NVME_ANA_OPTIMIZED] = "optimized",
109 [NVME_ANA_NONOPTIMIZED] = "non-optimized",
110 [NVME_ANA_INACCESSIBLE] = "inaccessible",
111 [NVME_ANA_PERSISTENT_LOSS] = "persistent-loss",
112 [NVME_ANA_CHANGE] = "change",
115 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
117 struct nvme_ns_head *head = ns->head;
118 bool changed = false;
124 for_each_node(node) {
125 if (ns == rcu_access_pointer(head->current_path[node])) {
126 rcu_assign_pointer(head->current_path[node], NULL);
134 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
138 mutex_lock(&ctrl->scan_lock);
139 down_read(&ctrl->namespaces_rwsem);
140 list_for_each_entry(ns, &ctrl->namespaces, list)
141 if (nvme_mpath_clear_current_path(ns))
142 kblockd_schedule_work(&ns->head->requeue_work);
143 up_read(&ctrl->namespaces_rwsem);
144 mutex_unlock(&ctrl->scan_lock);
147 static bool nvme_path_is_disabled(struct nvme_ns *ns)
150 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
151 * still be able to complete assuming that the controller is connected.
152 * Otherwise it will fail immediately and return to the requeue list.
154 if (ns->ctrl->state != NVME_CTRL_LIVE &&
155 ns->ctrl->state != NVME_CTRL_DELETING)
157 if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
158 test_bit(NVME_NS_REMOVING, &ns->flags))
163 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
165 int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
166 struct nvme_ns *found = NULL, *fallback = NULL, *ns;
168 list_for_each_entry_rcu(ns, &head->list, siblings) {
169 if (nvme_path_is_disabled(ns))
172 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
173 distance = node_distance(node, ns->ctrl->numa_node);
175 distance = LOCAL_DISTANCE;
177 switch (ns->ana_state) {
178 case NVME_ANA_OPTIMIZED:
179 if (distance < found_distance) {
180 found_distance = distance;
184 case NVME_ANA_NONOPTIMIZED:
185 if (distance < fallback_distance) {
186 fallback_distance = distance;
198 rcu_assign_pointer(head->current_path[node], found);
202 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
205 ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
209 return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
212 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
213 int node, struct nvme_ns *old)
215 struct nvme_ns *ns, *found = NULL;
217 if (list_is_singular(&head->list)) {
218 if (nvme_path_is_disabled(old))
223 for (ns = nvme_next_ns(head, old);
225 ns = nvme_next_ns(head, ns)) {
226 if (nvme_path_is_disabled(ns))
229 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
233 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
238 * The loop above skips the current path for round-robin semantics.
239 * Fall back to the current path if either:
240 * - no other optimized path found and current is optimized,
241 * - no other usable path found and current is usable.
243 if (!nvme_path_is_disabled(old) &&
244 (old->ana_state == NVME_ANA_OPTIMIZED ||
245 (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
251 rcu_assign_pointer(head->current_path[node], found);
255 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
257 return ns->ctrl->state == NVME_CTRL_LIVE &&
258 ns->ana_state == NVME_ANA_OPTIMIZED;
261 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
263 int node = numa_node_id();
266 ns = srcu_dereference(head->current_path[node], &head->srcu);
268 return __nvme_find_path(head, node);
270 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
271 return nvme_round_robin_path(head, node, ns);
272 if (unlikely(!nvme_path_is_optimized(ns)))
273 return __nvme_find_path(head, node);
277 static bool nvme_available_path(struct nvme_ns_head *head)
281 list_for_each_entry_rcu(ns, &head->list, siblings) {
282 switch (ns->ctrl->state) {
284 case NVME_CTRL_RESETTING:
285 case NVME_CTRL_CONNECTING:
295 blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
297 struct nvme_ns_head *head = bio->bi_disk->private_data;
298 struct device *dev = disk_to_dev(head->disk);
300 blk_qc_t ret = BLK_QC_T_NONE;
304 * The namespace might be going away and the bio might be moved to a
305 * different queue via blk_steal_bios(), so we need to use the bio_split
306 * pool from the original queue to allocate the bvecs from.
308 blk_queue_split(&bio);
310 srcu_idx = srcu_read_lock(&head->srcu);
311 ns = nvme_find_path(head);
313 bio->bi_disk = ns->disk;
314 bio->bi_opf |= REQ_NVME_MPATH;
315 trace_block_bio_remap(bio->bi_disk->queue, bio,
316 disk_devt(ns->head->disk),
317 bio->bi_iter.bi_sector);
318 ret = submit_bio_noacct(bio);
319 } else if (nvme_available_path(head)) {
320 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
322 spin_lock_irq(&head->requeue_lock);
323 bio_list_add(&head->requeue_list, bio);
324 spin_unlock_irq(&head->requeue_lock);
326 dev_warn_ratelimited(dev, "no available path - failing I/O\n");
328 bio->bi_status = BLK_STS_IOERR;
332 srcu_read_unlock(&head->srcu, srcu_idx);
336 static void nvme_requeue_work(struct work_struct *work)
338 struct nvme_ns_head *head =
339 container_of(work, struct nvme_ns_head, requeue_work);
340 struct bio *bio, *next;
342 spin_lock_irq(&head->requeue_lock);
343 next = bio_list_get(&head->requeue_list);
344 spin_unlock_irq(&head->requeue_lock);
346 while ((bio = next) != NULL) {
351 * Reset disk to the mpath node and resubmit to select a new
354 bio->bi_disk = head->disk;
355 submit_bio_noacct(bio);
359 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
361 struct request_queue *q;
364 mutex_init(&head->lock);
365 bio_list_init(&head->requeue_list);
366 spin_lock_init(&head->requeue_lock);
367 INIT_WORK(&head->requeue_work, nvme_requeue_work);
370 * Add a multipath node if the subsystems supports multiple controllers.
371 * We also do this for private namespaces as the namespace sharing data could
372 * change after a rescan.
374 if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) || !multipath)
377 q = blk_alloc_queue(ctrl->numa_node);
380 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
381 /* set to a default value for 512 until disk is validated */
382 blk_queue_logical_block_size(q, 512);
383 blk_set_stacking_limits(&q->limits);
385 /* we need to propagate up the VMC settings */
386 if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
388 blk_queue_write_cache(q, vwc, vwc);
390 head->disk = alloc_disk(0);
392 goto out_cleanup_queue;
393 head->disk->fops = &nvme_ns_head_ops;
394 head->disk->private_data = head;
395 head->disk->queue = q;
396 head->disk->flags = GENHD_FL_EXT_DEVT;
397 sprintf(head->disk->disk_name, "nvme%dn%d",
398 ctrl->subsys->instance, head->instance);
402 blk_cleanup_queue(q);
407 static void nvme_mpath_set_live(struct nvme_ns *ns)
409 struct nvme_ns_head *head = ns->head;
414 if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags))
415 device_add_disk(&head->subsys->dev, head->disk,
416 nvme_ns_id_attr_groups);
418 mutex_lock(&head->lock);
419 if (nvme_path_is_optimized(ns)) {
422 srcu_idx = srcu_read_lock(&head->srcu);
424 __nvme_find_path(head, node);
425 srcu_read_unlock(&head->srcu, srcu_idx);
427 mutex_unlock(&head->lock);
429 synchronize_srcu(&head->srcu);
430 kblockd_schedule_work(&head->requeue_work);
433 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
434 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
437 void *base = ctrl->ana_log_buf;
438 size_t offset = sizeof(struct nvme_ana_rsp_hdr);
441 lockdep_assert_held(&ctrl->ana_lock);
443 for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
444 struct nvme_ana_group_desc *desc = base + offset;
446 size_t nsid_buf_size;
448 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
451 nr_nsids = le32_to_cpu(desc->nnsids);
452 nsid_buf_size = nr_nsids * sizeof(__le32);
454 if (WARN_ON_ONCE(desc->grpid == 0))
456 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
458 if (WARN_ON_ONCE(desc->state == 0))
460 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
463 offset += sizeof(*desc);
464 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
467 error = cb(ctrl, desc, data);
471 offset += nsid_buf_size;
477 static inline bool nvme_state_is_live(enum nvme_ana_state state)
479 return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
482 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
485 ns->ana_grpid = le32_to_cpu(desc->grpid);
486 ns->ana_state = desc->state;
487 clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
489 if (nvme_state_is_live(ns->ana_state))
490 nvme_mpath_set_live(ns);
493 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
494 struct nvme_ana_group_desc *desc, void *data)
496 u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
497 unsigned *nr_change_groups = data;
500 dev_dbg(ctrl->device, "ANA group %d: %s.\n",
501 le32_to_cpu(desc->grpid),
502 nvme_ana_state_names[desc->state]);
504 if (desc->state == NVME_ANA_CHANGE)
505 (*nr_change_groups)++;
510 down_read(&ctrl->namespaces_rwsem);
511 list_for_each_entry(ns, &ctrl->namespaces, list) {
512 unsigned nsid = le32_to_cpu(desc->nsids[n]);
514 if (ns->head->ns_id < nsid)
516 if (ns->head->ns_id == nsid)
517 nvme_update_ns_ana_state(desc, ns);
521 up_read(&ctrl->namespaces_rwsem);
525 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
527 u32 nr_change_groups = 0;
530 mutex_lock(&ctrl->ana_lock);
531 error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
532 ctrl->ana_log_buf, ctrl->ana_log_size, 0);
534 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
538 error = nvme_parse_ana_log(ctrl, &nr_change_groups,
539 nvme_update_ana_state);
544 * In theory we should have an ANATT timer per group as they might enter
545 * the change state at different times. But that is a lot of overhead
546 * just to protect against a target that keeps entering new changes
547 * states while never finishing previous ones. But we'll still
548 * eventually time out once all groups are in change state, so this
551 * We also double the ANATT value to provide some slack for transports
552 * or AEN processing overhead.
554 if (nr_change_groups)
555 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
557 del_timer_sync(&ctrl->anatt_timer);
559 mutex_unlock(&ctrl->ana_lock);
563 static void nvme_ana_work(struct work_struct *work)
565 struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
567 if (ctrl->state != NVME_CTRL_LIVE)
570 nvme_read_ana_log(ctrl);
573 static void nvme_anatt_timeout(struct timer_list *t)
575 struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
577 dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
578 nvme_reset_ctrl(ctrl);
581 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
583 if (!nvme_ctrl_use_ana(ctrl))
585 del_timer_sync(&ctrl->anatt_timer);
586 cancel_work_sync(&ctrl->ana_work);
589 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store) \
590 struct device_attribute subsys_attr_##_name = \
591 __ATTR(_name, _mode, _show, _store)
593 static const char *nvme_iopolicy_names[] = {
594 [NVME_IOPOLICY_NUMA] = "numa",
595 [NVME_IOPOLICY_RR] = "round-robin",
598 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
599 struct device_attribute *attr, char *buf)
601 struct nvme_subsystem *subsys =
602 container_of(dev, struct nvme_subsystem, dev);
604 return sprintf(buf, "%s\n",
605 nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
608 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
609 struct device_attribute *attr, const char *buf, size_t count)
611 struct nvme_subsystem *subsys =
612 container_of(dev, struct nvme_subsystem, dev);
615 for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
616 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
617 WRITE_ONCE(subsys->iopolicy, i);
624 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
625 nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
627 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
630 return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
632 DEVICE_ATTR_RO(ana_grpid);
634 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
637 struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
639 return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
641 DEVICE_ATTR_RO(ana_state);
643 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
644 struct nvme_ana_group_desc *desc, void *data)
646 struct nvme_ana_group_desc *dst = data;
648 if (desc->grpid != dst->grpid)
652 return -ENXIO; /* just break out of the loop */
655 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
657 if (nvme_ctrl_use_ana(ns->ctrl)) {
658 struct nvme_ana_group_desc desc = {
659 .grpid = id->anagrpid,
663 mutex_lock(&ns->ctrl->ana_lock);
664 ns->ana_grpid = le32_to_cpu(id->anagrpid);
665 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
666 mutex_unlock(&ns->ctrl->ana_lock);
668 /* found the group desc: update */
669 nvme_update_ns_ana_state(&desc, ns);
672 ns->ana_state = NVME_ANA_OPTIMIZED;
673 nvme_mpath_set_live(ns);
676 if (bdi_cap_stable_pages_required(ns->queue->backing_dev_info)) {
677 struct gendisk *disk = ns->head->disk;
680 disk->queue->backing_dev_info->capabilities |=
681 BDI_CAP_STABLE_WRITES;
685 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
689 if (head->disk->flags & GENHD_FL_UP)
690 del_gendisk(head->disk);
691 blk_set_queue_dying(head->disk->queue);
692 /* make sure all pending bios are cleaned up */
693 kblockd_schedule_work(&head->requeue_work);
694 flush_work(&head->requeue_work);
695 blk_cleanup_queue(head->disk->queue);
696 if (!test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
698 * if device_add_disk wasn't called, prevent
699 * disk release to put a bogus reference on the
702 head->disk->queue = NULL;
704 put_disk(head->disk);
707 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
711 /* check if multipath is enabled and we have the capability */
712 if (!multipath || !ctrl->subsys ||
713 !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
716 ctrl->anacap = id->anacap;
717 ctrl->anatt = id->anatt;
718 ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
719 ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
721 mutex_init(&ctrl->ana_lock);
722 timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
723 ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
724 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
725 ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
727 if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
728 dev_err(ctrl->device,
729 "ANA log page size (%zd) larger than MDTS (%d).\n",
731 ctrl->max_hw_sectors << SECTOR_SHIFT);
732 dev_err(ctrl->device, "disabling ANA support.\n");
736 INIT_WORK(&ctrl->ana_work, nvme_ana_work);
737 kfree(ctrl->ana_log_buf);
738 ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
739 if (!ctrl->ana_log_buf) {
744 error = nvme_read_ana_log(ctrl);
746 goto out_free_ana_log_buf;
748 out_free_ana_log_buf:
749 kfree(ctrl->ana_log_buf);
750 ctrl->ana_log_buf = NULL;
755 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
757 kfree(ctrl->ana_log_buf);
758 ctrl->ana_log_buf = NULL;