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 inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
17 return multipath && ctrl->subsys && (ctrl->subsys->cmic & (1 << 3));
21 * If multipathing is enabled we need to always use the subsystem instance
22 * number for numbering our devices to avoid conflicts between subsystems that
23 * have multiple controllers and thus use the multipath-aware subsystem node
24 * and those that have a single controller and use the controller node
27 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
28 struct nvme_ctrl *ctrl, int *flags)
31 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
32 } else if (ns->head->disk) {
33 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
34 ctrl->instance, ns->head->instance);
35 *flags = GENHD_FL_HIDDEN;
37 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
42 void nvme_failover_req(struct request *req)
44 struct nvme_ns *ns = req->q->queuedata;
45 u16 status = nvme_req(req)->status;
48 spin_lock_irqsave(&ns->head->requeue_lock, flags);
49 blk_steal_bios(&ns->head->requeue_list, req);
50 spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
51 blk_mq_end_request(req, 0);
53 switch (status & 0x7ff) {
54 case NVME_SC_ANA_TRANSITION:
55 case NVME_SC_ANA_INACCESSIBLE:
56 case NVME_SC_ANA_PERSISTENT_LOSS:
58 * If we got back an ANA error we know the controller is alive,
59 * but not ready to serve this namespaces. The spec suggests
60 * we should update our general state here, but due to the fact
61 * that the admin and I/O queues are not serialized that is
62 * fundamentally racy. So instead just clear the current path,
63 * mark the the path as pending and kick of a re-read of the ANA
66 nvme_mpath_clear_current_path(ns);
67 if (ns->ctrl->ana_log_buf) {
68 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
69 queue_work(nvme_wq, &ns->ctrl->ana_work);
72 case NVME_SC_HOST_PATH_ERROR:
74 * Temporary transport disruption in talking to the controller.
75 * Try to send on a new path.
77 nvme_mpath_clear_current_path(ns);
81 * Reset the controller for any non-ANA error as we don't know
82 * what caused the error.
84 nvme_reset_ctrl(ns->ctrl);
88 kblockd_schedule_work(&ns->head->requeue_work);
91 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
95 down_read(&ctrl->namespaces_rwsem);
96 list_for_each_entry(ns, &ctrl->namespaces, list) {
98 kblockd_schedule_work(&ns->head->requeue_work);
100 up_read(&ctrl->namespaces_rwsem);
103 static const char *nvme_ana_state_names[] = {
104 [0] = "invalid state",
105 [NVME_ANA_OPTIMIZED] = "optimized",
106 [NVME_ANA_NONOPTIMIZED] = "non-optimized",
107 [NVME_ANA_INACCESSIBLE] = "inaccessible",
108 [NVME_ANA_PERSISTENT_LOSS] = "persistent-loss",
109 [NVME_ANA_CHANGE] = "change",
112 void nvme_mpath_clear_current_path(struct nvme_ns *ns)
114 struct nvme_ns_head *head = ns->head;
120 for_each_node(node) {
121 if (ns == rcu_access_pointer(head->current_path[node]))
122 rcu_assign_pointer(head->current_path[node], NULL);
126 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
128 int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
129 struct nvme_ns *found = NULL, *fallback = NULL, *ns;
131 list_for_each_entry_rcu(ns, &head->list, siblings) {
132 if (ns->ctrl->state != NVME_CTRL_LIVE ||
133 test_bit(NVME_NS_ANA_PENDING, &ns->flags))
136 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
137 distance = node_distance(node, ns->ctrl->numa_node);
139 distance = LOCAL_DISTANCE;
141 switch (ns->ana_state) {
142 case NVME_ANA_OPTIMIZED:
143 if (distance < found_distance) {
144 found_distance = distance;
148 case NVME_ANA_NONOPTIMIZED:
149 if (distance < fallback_distance) {
150 fallback_distance = distance;
162 rcu_assign_pointer(head->current_path[node], found);
166 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
169 ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
173 return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
176 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
177 int node, struct nvme_ns *old)
179 struct nvme_ns *ns, *found, *fallback = NULL;
181 if (list_is_singular(&head->list))
184 for (ns = nvme_next_ns(head, old);
186 ns = nvme_next_ns(head, ns)) {
187 if (ns->ctrl->state != NVME_CTRL_LIVE ||
188 test_bit(NVME_NS_ANA_PENDING, &ns->flags))
191 if (ns->ana_state == NVME_ANA_OPTIMIZED) {
195 if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
203 rcu_assign_pointer(head->current_path[node], found);
207 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
209 return ns->ctrl->state == NVME_CTRL_LIVE &&
210 ns->ana_state == NVME_ANA_OPTIMIZED;
213 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
215 int node = numa_node_id();
218 ns = srcu_dereference(head->current_path[node], &head->srcu);
219 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns)
220 ns = nvme_round_robin_path(head, node, ns);
221 if (unlikely(!ns || !nvme_path_is_optimized(ns)))
222 ns = __nvme_find_path(head, node);
226 static blk_qc_t nvme_ns_head_make_request(struct request_queue *q,
229 struct nvme_ns_head *head = q->queuedata;
230 struct device *dev = disk_to_dev(head->disk);
232 blk_qc_t ret = BLK_QC_T_NONE;
236 * The namespace might be going away and the bio might
237 * be moved to a different queue via blk_steal_bios(),
238 * so we need to use the bio_split pool from the original
239 * queue to allocate the bvecs from.
241 blk_queue_split(q, &bio);
243 srcu_idx = srcu_read_lock(&head->srcu);
244 ns = nvme_find_path(head);
246 bio->bi_disk = ns->disk;
247 bio->bi_opf |= REQ_NVME_MPATH;
248 trace_block_bio_remap(bio->bi_disk->queue, bio,
249 disk_devt(ns->head->disk),
250 bio->bi_iter.bi_sector);
251 ret = direct_make_request(bio);
252 } else if (!list_empty_careful(&head->list)) {
253 dev_warn_ratelimited(dev, "no path available - requeuing I/O\n");
255 spin_lock_irq(&head->requeue_lock);
256 bio_list_add(&head->requeue_list, bio);
257 spin_unlock_irq(&head->requeue_lock);
259 dev_warn_ratelimited(dev, "no path - failing I/O\n");
261 bio->bi_status = BLK_STS_IOERR;
265 srcu_read_unlock(&head->srcu, srcu_idx);
269 static void nvme_requeue_work(struct work_struct *work)
271 struct nvme_ns_head *head =
272 container_of(work, struct nvme_ns_head, requeue_work);
273 struct bio *bio, *next;
275 spin_lock_irq(&head->requeue_lock);
276 next = bio_list_get(&head->requeue_list);
277 spin_unlock_irq(&head->requeue_lock);
279 while ((bio = next) != NULL) {
284 * Reset disk to the mpath node and resubmit to select a new
287 bio->bi_disk = head->disk;
288 generic_make_request(bio);
292 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
294 struct request_queue *q;
297 mutex_init(&head->lock);
298 bio_list_init(&head->requeue_list);
299 spin_lock_init(&head->requeue_lock);
300 INIT_WORK(&head->requeue_work, nvme_requeue_work);
303 * Add a multipath node if the subsystems supports multiple controllers.
304 * We also do this for private namespaces as the namespace sharing data could
305 * change after a rescan.
307 if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath)
310 q = blk_alloc_queue_node(GFP_KERNEL, ctrl->numa_node);
314 blk_queue_make_request(q, nvme_ns_head_make_request);
315 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
316 /* set to a default value for 512 until disk is validated */
317 blk_queue_logical_block_size(q, 512);
318 blk_set_stacking_limits(&q->limits);
320 /* we need to propagate up the VMC settings */
321 if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
323 blk_queue_write_cache(q, vwc, vwc);
325 head->disk = alloc_disk(0);
327 goto out_cleanup_queue;
328 head->disk->fops = &nvme_ns_head_ops;
329 head->disk->private_data = head;
330 head->disk->queue = q;
331 head->disk->flags = GENHD_FL_EXT_DEVT;
332 sprintf(head->disk->disk_name, "nvme%dn%d",
333 ctrl->subsys->instance, head->instance);
337 blk_cleanup_queue(q);
342 static void nvme_mpath_set_live(struct nvme_ns *ns)
344 struct nvme_ns_head *head = ns->head;
346 lockdep_assert_held(&ns->head->lock);
351 if (!(head->disk->flags & GENHD_FL_UP))
352 device_add_disk(&head->subsys->dev, head->disk,
353 nvme_ns_id_attr_groups);
355 if (nvme_path_is_optimized(ns)) {
358 srcu_idx = srcu_read_lock(&head->srcu);
360 __nvme_find_path(head, node);
361 srcu_read_unlock(&head->srcu, srcu_idx);
364 kblockd_schedule_work(&ns->head->requeue_work);
367 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
368 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
371 void *base = ctrl->ana_log_buf;
372 size_t offset = sizeof(struct nvme_ana_rsp_hdr);
375 lockdep_assert_held(&ctrl->ana_lock);
377 for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
378 struct nvme_ana_group_desc *desc = base + offset;
379 u32 nr_nsids = le32_to_cpu(desc->nnsids);
380 size_t nsid_buf_size = nr_nsids * sizeof(__le32);
382 if (WARN_ON_ONCE(desc->grpid == 0))
384 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
386 if (WARN_ON_ONCE(desc->state == 0))
388 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
391 offset += sizeof(*desc);
392 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
395 error = cb(ctrl, desc, data);
399 offset += nsid_buf_size;
400 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
407 static inline bool nvme_state_is_live(enum nvme_ana_state state)
409 return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
412 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
415 mutex_lock(&ns->head->lock);
416 ns->ana_grpid = le32_to_cpu(desc->grpid);
417 ns->ana_state = desc->state;
418 clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
420 if (nvme_state_is_live(ns->ana_state))
421 nvme_mpath_set_live(ns);
422 mutex_unlock(&ns->head->lock);
425 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
426 struct nvme_ana_group_desc *desc, void *data)
428 u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
429 unsigned *nr_change_groups = data;
432 dev_dbg(ctrl->device, "ANA group %d: %s.\n",
433 le32_to_cpu(desc->grpid),
434 nvme_ana_state_names[desc->state]);
436 if (desc->state == NVME_ANA_CHANGE)
437 (*nr_change_groups)++;
442 down_write(&ctrl->namespaces_rwsem);
443 list_for_each_entry(ns, &ctrl->namespaces, list) {
444 if (ns->head->ns_id != le32_to_cpu(desc->nsids[n]))
446 nvme_update_ns_ana_state(desc, ns);
450 up_write(&ctrl->namespaces_rwsem);
451 WARN_ON_ONCE(n < nr_nsids);
455 static int nvme_read_ana_log(struct nvme_ctrl *ctrl, bool groups_only)
457 u32 nr_change_groups = 0;
460 mutex_lock(&ctrl->ana_lock);
461 error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA,
462 groups_only ? NVME_ANA_LOG_RGO : 0,
463 ctrl->ana_log_buf, ctrl->ana_log_size, 0);
465 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
469 error = nvme_parse_ana_log(ctrl, &nr_change_groups,
470 nvme_update_ana_state);
475 * In theory we should have an ANATT timer per group as they might enter
476 * the change state at different times. But that is a lot of overhead
477 * just to protect against a target that keeps entering new changes
478 * states while never finishing previous ones. But we'll still
479 * eventually time out once all groups are in change state, so this
482 * We also double the ANATT value to provide some slack for transports
483 * or AEN processing overhead.
485 if (nr_change_groups)
486 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
488 del_timer_sync(&ctrl->anatt_timer);
490 mutex_unlock(&ctrl->ana_lock);
494 static void nvme_ana_work(struct work_struct *work)
496 struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
498 nvme_read_ana_log(ctrl, false);
501 static void nvme_anatt_timeout(struct timer_list *t)
503 struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
505 dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
506 nvme_reset_ctrl(ctrl);
509 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
511 if (!nvme_ctrl_use_ana(ctrl))
513 del_timer_sync(&ctrl->anatt_timer);
514 cancel_work_sync(&ctrl->ana_work);
517 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store) \
518 struct device_attribute subsys_attr_##_name = \
519 __ATTR(_name, _mode, _show, _store)
521 static const char *nvme_iopolicy_names[] = {
522 [NVME_IOPOLICY_NUMA] = "numa",
523 [NVME_IOPOLICY_RR] = "round-robin",
526 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
527 struct device_attribute *attr, char *buf)
529 struct nvme_subsystem *subsys =
530 container_of(dev, struct nvme_subsystem, dev);
532 return sprintf(buf, "%s\n",
533 nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
536 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
537 struct device_attribute *attr, const char *buf, size_t count)
539 struct nvme_subsystem *subsys =
540 container_of(dev, struct nvme_subsystem, dev);
543 for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
544 if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
545 WRITE_ONCE(subsys->iopolicy, i);
552 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
553 nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
555 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
558 return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
560 DEVICE_ATTR_RO(ana_grpid);
562 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
565 struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
567 return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
569 DEVICE_ATTR_RO(ana_state);
571 static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl,
572 struct nvme_ana_group_desc *desc, void *data)
574 struct nvme_ns *ns = data;
576 if (ns->ana_grpid == le32_to_cpu(desc->grpid)) {
577 nvme_update_ns_ana_state(desc, ns);
578 return -ENXIO; /* just break out of the loop */
584 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
586 if (nvme_ctrl_use_ana(ns->ctrl)) {
587 mutex_lock(&ns->ctrl->ana_lock);
588 ns->ana_grpid = le32_to_cpu(id->anagrpid);
589 nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state);
590 mutex_unlock(&ns->ctrl->ana_lock);
592 mutex_lock(&ns->head->lock);
593 ns->ana_state = NVME_ANA_OPTIMIZED;
594 nvme_mpath_set_live(ns);
595 mutex_unlock(&ns->head->lock);
599 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
603 if (head->disk->flags & GENHD_FL_UP)
604 del_gendisk(head->disk);
605 blk_set_queue_dying(head->disk->queue);
606 /* make sure all pending bios are cleaned up */
607 kblockd_schedule_work(&head->requeue_work);
608 flush_work(&head->requeue_work);
609 blk_cleanup_queue(head->disk->queue);
610 put_disk(head->disk);
613 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
617 if (!nvme_ctrl_use_ana(ctrl))
620 ctrl->anacap = id->anacap;
621 ctrl->anatt = id->anatt;
622 ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
623 ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
625 mutex_init(&ctrl->ana_lock);
626 timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
627 ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
628 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
629 ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
631 if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
632 dev_err(ctrl->device,
633 "ANA log page size (%zd) larger than MDTS (%d).\n",
635 ctrl->max_hw_sectors << SECTOR_SHIFT);
636 dev_err(ctrl->device, "disabling ANA support.\n");
640 INIT_WORK(&ctrl->ana_work, nvme_ana_work);
641 ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
642 if (!ctrl->ana_log_buf) {
647 error = nvme_read_ana_log(ctrl, true);
649 goto out_free_ana_log_buf;
651 out_free_ana_log_buf:
652 kfree(ctrl->ana_log_buf);
653 ctrl->ana_log_buf = NULL;
658 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
660 kfree(ctrl->ana_log_buf);
661 ctrl->ana_log_buf = NULL;