1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2017-2018 Christoph Hellwig.
6 #include <linux/moduleparam.h>
7 #include <trace/events/block.h>
10 static bool multipath = true;
11 module_param(multipath, bool, 0444);
12 MODULE_PARM_DESC(multipath,
13 "turn on native support for multiple controllers per subsystem");
15 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
17 struct nvme_ns_head *h;
19 lockdep_assert_held(&subsys->lock);
20 list_for_each_entry(h, &subsys->nsheads, entry)
22 blk_mq_unfreeze_queue(h->disk->queue);
25 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
27 struct nvme_ns_head *h;
29 lockdep_assert_held(&subsys->lock);
30 list_for_each_entry(h, &subsys->nsheads, entry)
32 blk_mq_freeze_queue_wait(h->disk->queue);
35 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
37 struct nvme_ns_head *h;
39 lockdep_assert_held(&subsys->lock);
40 list_for_each_entry(h, &subsys->nsheads, entry)
42 blk_freeze_queue_start(h->disk->queue);
46 * If multipathing is enabled we need to always use the subsystem instance
47 * number for numbering our devices to avoid conflicts between subsystems that
48 * have multiple controllers and thus use the multipath-aware subsystem node
49 * and those that have a single controller and use the controller node
52 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
53 struct nvme_ctrl *ctrl, int *flags)
56 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
57 } else if (ns->head->disk) {
58 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
59 ctrl->instance, ns->head->instance);
60 *flags = GENHD_FL_HIDDEN;
62 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
67 void nvme_failover_req(struct request *req)
69 struct nvme_ns *ns = req->q->queuedata;
70 u16 status = nvme_req(req)->status;
73 spin_lock_irqsave(&ns->head->requeue_lock, flags);
74 blk_steal_bios(&ns->head->requeue_list, req);
75 spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
76 blk_mq_end_request(req, 0);
78 switch (status & 0x7ff) {
79 case NVME_SC_ANA_TRANSITION:
80 case NVME_SC_ANA_INACCESSIBLE:
81 case NVME_SC_ANA_PERSISTENT_LOSS:
83 * If we got back an ANA error we know the controller is alive,
84 * but not ready to serve this namespaces. The spec suggests
85 * we should update our general state here, but due to the fact
86 * that the admin and I/O queues are not serialized that is
87 * fundamentally racy. So instead just clear the current path,
88 * mark the the path as pending and kick of a re-read of the ANA
91 nvme_mpath_clear_current_path(ns);
92 if (ns->ctrl->ana_log_buf) {
93 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
94 queue_work(nvme_wq, &ns->ctrl->ana_work);
97 case NVME_SC_HOST_PATH_ERROR:
99 * Temporary transport disruption in talking to the controller.
100 * Try to send on a new path.
102 nvme_mpath_clear_current_path(ns);
106 * Reset the controller for any non-ANA error as we don't know
107 * what caused the error.
109 nvme_reset_ctrl(ns->ctrl);
113 kblockd_schedule_work(&ns->head->requeue_work);
116 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
120 down_read(&ctrl->namespaces_rwsem);
121 list_for_each_entry(ns, &ctrl->namespaces, list) {
123 kblockd_schedule_work(&ns->head->requeue_work);
125 up_read(&ctrl->namespaces_rwsem);
128 static const char *nvme_ana_state_names[] = {
129 [0] = "invalid state",
130 [NVME_ANA_OPTIMIZED] = "optimized",
131 [NVME_ANA_NONOPTIMIZED] = "non-optimized",
132 [NVME_ANA_INACCESSIBLE] = "inaccessible",
133 [NVME_ANA_PERSISTENT_LOSS] = "persistent-loss",
134 [NVME_ANA_CHANGE] = "change",
137 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
139 struct nvme_ns_head *head = ns->head;
140 bool changed = false;
146 for_each_node(node) {
147 if (ns == rcu_access_pointer(head->current_path[node])) {
148 rcu_assign_pointer(head->current_path[node], NULL);
156 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
160 mutex_lock(&ctrl->scan_lock);
161 list_for_each_entry(ns, &ctrl->namespaces, list)
162 if (nvme_mpath_clear_current_path(ns))
163 kblockd_schedule_work(&ns->head->requeue_work);
164 mutex_unlock(&ctrl->scan_lock);
167 static bool nvme_path_is_disabled(struct nvme_ns *ns)
169 return ns->ctrl->state != NVME_CTRL_LIVE ||
170 test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
171 test_bit(NVME_NS_REMOVING, &ns->flags);
174 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
176 int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
177 struct nvme_ns *found = NULL, *fallback = NULL, *ns;
179 list_for_each_entry_rcu(ns, &head->list, siblings) {
180 if (nvme_path_is_disabled(ns))
183 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
184 distance = node_distance(node, ns->ctrl->numa_node);
186 distance = LOCAL_DISTANCE;
188 switch (ns->ana_state) {
189 case NVME_ANA_OPTIMIZED:
190 if (distance < found_distance) {
191 found_distance = distance;
195 case NVME_ANA_NONOPTIMIZED:
196 if (distance < fallback_distance) {
197 fallback_distance = distance;
209 rcu_assign_pointer(head->current_path[node], found);
213 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
216 ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
220 return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
223 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
224 int node, struct nvme_ns *old)
226 struct nvme_ns *ns, *found, *fallback = NULL;
228 if (list_is_singular(&head->list)) {
229 if (nvme_path_is_disabled(old))
234 for (ns = nvme_next_ns(head, old);
236 ns = nvme_next_ns(head, ns)) {
237 if (nvme_path_is_disabled(ns))
240 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
244 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
252 rcu_assign_pointer(head->current_path[node], found);
256 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
258 return ns->ctrl->state == NVME_CTRL_LIVE &&
259 ns->ana_state == NVME_ANA_OPTIMIZED;
262 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
264 int node = numa_node_id();
267 ns = srcu_dereference(head->current_path[node], &head->srcu);
268 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns)
269 ns = nvme_round_robin_path(head, node, ns);
270 if (unlikely(!ns || !nvme_path_is_optimized(ns)))
271 ns = __nvme_find_path(head, node);
275 static bool nvme_available_path(struct nvme_ns_head *head)
279 list_for_each_entry_rcu(ns, &head->list, siblings) {
280 switch (ns->ctrl->state) {
282 case NVME_CTRL_RESETTING:
283 case NVME_CTRL_CONNECTING:
293 static blk_qc_t nvme_ns_head_make_request(struct request_queue *q,
296 struct nvme_ns_head *head = q->queuedata;
297 struct device *dev = disk_to_dev(head->disk);
299 blk_qc_t ret = BLK_QC_T_NONE;
303 * The namespace might be going away and the bio might
304 * be moved to a different queue via blk_steal_bios(),
305 * so we need to use the bio_split pool from the original
306 * queue to allocate the bvecs from.
308 blk_queue_split(q, &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 = direct_make_request(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 generic_make_request(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 & (1 << 1)) || !multipath)
377 q = blk_alloc_queue_node(GFP_KERNEL, ctrl->numa_node);
381 blk_queue_make_request(q, nvme_ns_head_make_request);
382 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
383 /* set to a default value for 512 until disk is validated */
384 blk_queue_logical_block_size(q, 512);
385 blk_set_stacking_limits(&q->limits);
387 /* we need to propagate up the VMC settings */
388 if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
390 blk_queue_write_cache(q, vwc, vwc);
392 head->disk = alloc_disk(0);
394 goto out_cleanup_queue;
395 head->disk->fops = &nvme_ns_head_ops;
396 head->disk->private_data = head;
397 head->disk->queue = q;
398 head->disk->flags = GENHD_FL_EXT_DEVT;
399 sprintf(head->disk->disk_name, "nvme%dn%d",
400 ctrl->subsys->instance, head->instance);
404 blk_cleanup_queue(q);
409 static void nvme_mpath_set_live(struct nvme_ns *ns)
411 struct nvme_ns_head *head = ns->head;
413 lockdep_assert_held(&ns->head->lock);
418 if (!(head->disk->flags & GENHD_FL_UP))
419 device_add_disk(&head->subsys->dev, head->disk,
420 nvme_ns_id_attr_groups);
422 if (nvme_path_is_optimized(ns)) {
425 srcu_idx = srcu_read_lock(&head->srcu);
427 __nvme_find_path(head, node);
428 srcu_read_unlock(&head->srcu, srcu_idx);
431 synchronize_srcu(&ns->head->srcu);
432 kblockd_schedule_work(&ns->head->requeue_work);
435 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
436 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
439 void *base = ctrl->ana_log_buf;
440 size_t offset = sizeof(struct nvme_ana_rsp_hdr);
443 lockdep_assert_held(&ctrl->ana_lock);
445 for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
446 struct nvme_ana_group_desc *desc = base + offset;
447 u32 nr_nsids = le32_to_cpu(desc->nnsids);
448 size_t nsid_buf_size = nr_nsids * sizeof(__le32);
450 if (WARN_ON_ONCE(desc->grpid == 0))
452 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
454 if (WARN_ON_ONCE(desc->state == 0))
456 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
459 offset += sizeof(*desc);
460 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
463 error = cb(ctrl, desc, data);
467 offset += nsid_buf_size;
468 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
475 static inline bool nvme_state_is_live(enum nvme_ana_state state)
477 return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
480 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
483 mutex_lock(&ns->head->lock);
484 ns->ana_grpid = le32_to_cpu(desc->grpid);
485 ns->ana_state = desc->state;
486 clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
488 if (nvme_state_is_live(ns->ana_state))
489 nvme_mpath_set_live(ns);
490 mutex_unlock(&ns->head->lock);
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_write(&ctrl->namespaces_rwsem);
511 list_for_each_entry(ns, &ctrl->namespaces, list) {
512 if (ns->head->ns_id != le32_to_cpu(desc->nsids[n]))
514 nvme_update_ns_ana_state(desc, ns);
518 up_write(&ctrl->namespaces_rwsem);
519 WARN_ON_ONCE(n < nr_nsids);
523 static int nvme_read_ana_log(struct nvme_ctrl *ctrl, bool groups_only)
525 u32 nr_change_groups = 0;
528 mutex_lock(&ctrl->ana_lock);
529 error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA,
530 groups_only ? NVME_ANA_LOG_RGO : 0,
531 ctrl->ana_log_buf, ctrl->ana_log_size, 0);
533 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
537 error = nvme_parse_ana_log(ctrl, &nr_change_groups,
538 nvme_update_ana_state);
543 * In theory we should have an ANATT timer per group as they might enter
544 * the change state at different times. But that is a lot of overhead
545 * just to protect against a target that keeps entering new changes
546 * states while never finishing previous ones. But we'll still
547 * eventually time out once all groups are in change state, so this
550 * We also double the ANATT value to provide some slack for transports
551 * or AEN processing overhead.
553 if (nr_change_groups)
554 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
556 del_timer_sync(&ctrl->anatt_timer);
558 mutex_unlock(&ctrl->ana_lock);
562 static void nvme_ana_work(struct work_struct *work)
564 struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
566 nvme_read_ana_log(ctrl, false);
569 static void nvme_anatt_timeout(struct timer_list *t)
571 struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
573 dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
574 nvme_reset_ctrl(ctrl);
577 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
579 if (!nvme_ctrl_use_ana(ctrl))
581 del_timer_sync(&ctrl->anatt_timer);
582 cancel_work_sync(&ctrl->ana_work);
585 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store) \
586 struct device_attribute subsys_attr_##_name = \
587 __ATTR(_name, _mode, _show, _store)
589 static const char *nvme_iopolicy_names[] = {
590 [NVME_IOPOLICY_NUMA] = "numa",
591 [NVME_IOPOLICY_RR] = "round-robin",
594 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
595 struct device_attribute *attr, char *buf)
597 struct nvme_subsystem *subsys =
598 container_of(dev, struct nvme_subsystem, dev);
600 return sprintf(buf, "%s\n",
601 nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
604 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
605 struct device_attribute *attr, const char *buf, size_t count)
607 struct nvme_subsystem *subsys =
608 container_of(dev, struct nvme_subsystem, dev);
611 for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
612 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
613 WRITE_ONCE(subsys->iopolicy, i);
620 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
621 nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
623 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
626 return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
628 DEVICE_ATTR_RO(ana_grpid);
630 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
633 struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
635 return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
637 DEVICE_ATTR_RO(ana_state);
639 static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl,
640 struct nvme_ana_group_desc *desc, void *data)
642 struct nvme_ns *ns = data;
644 if (ns->ana_grpid == le32_to_cpu(desc->grpid)) {
645 nvme_update_ns_ana_state(desc, ns);
646 return -ENXIO; /* just break out of the loop */
652 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
654 if (nvme_ctrl_use_ana(ns->ctrl)) {
655 mutex_lock(&ns->ctrl->ana_lock);
656 ns->ana_grpid = le32_to_cpu(id->anagrpid);
657 nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state);
658 mutex_unlock(&ns->ctrl->ana_lock);
660 mutex_lock(&ns->head->lock);
661 ns->ana_state = NVME_ANA_OPTIMIZED;
662 nvme_mpath_set_live(ns);
663 mutex_unlock(&ns->head->lock);
667 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
671 if (head->disk->flags & GENHD_FL_UP)
672 del_gendisk(head->disk);
673 blk_set_queue_dying(head->disk->queue);
674 /* make sure all pending bios are cleaned up */
675 kblockd_schedule_work(&head->requeue_work);
676 flush_work(&head->requeue_work);
677 blk_cleanup_queue(head->disk->queue);
678 put_disk(head->disk);
681 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
685 /* check if multipath is enabled and we have the capability */
686 if (!multipath || !ctrl->subsys || !(ctrl->subsys->cmic & (1 << 3)))
689 ctrl->anacap = id->anacap;
690 ctrl->anatt = id->anatt;
691 ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
692 ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
694 mutex_init(&ctrl->ana_lock);
695 timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
696 ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
697 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
698 ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
700 if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
701 dev_err(ctrl->device,
702 "ANA log page size (%zd) larger than MDTS (%d).\n",
704 ctrl->max_hw_sectors << SECTOR_SHIFT);
705 dev_err(ctrl->device, "disabling ANA support.\n");
709 INIT_WORK(&ctrl->ana_work, nvme_ana_work);
710 ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
711 if (!ctrl->ana_log_buf) {
716 error = nvme_read_ana_log(ctrl, true);
718 goto out_free_ana_log_buf;
720 out_free_ana_log_buf:
721 kfree(ctrl->ana_log_buf);
722 ctrl->ana_log_buf = NULL;
727 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
729 kfree(ctrl->ana_log_buf);
730 ctrl->ana_log_buf = NULL;