2 * Copyright (c) 2017-2018 Christoph Hellwig.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #include <linux/moduleparam.h>
15 #include <trace/events/block.h>
18 static bool multipath = true;
19 module_param(multipath, bool, 0444);
20 MODULE_PARM_DESC(multipath,
21 "turn on native support for multiple controllers per subsystem");
23 inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
25 return multipath && ctrl->subsys && (ctrl->subsys->cmic & (1 << 3));
29 * If multipathing is enabled we need to always use the subsystem instance
30 * number for numbering our devices to avoid conflicts between subsystems that
31 * have multiple controllers and thus use the multipath-aware subsystem node
32 * and those that have a single controller and use the controller node
35 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
36 struct nvme_ctrl *ctrl, int *flags)
39 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
40 } else if (ns->head->disk) {
41 sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
42 ctrl->cntlid, ns->head->instance);
43 *flags = GENHD_FL_HIDDEN;
45 sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance,
50 void nvme_failover_req(struct request *req)
52 struct nvme_ns *ns = req->q->queuedata;
53 u16 status = nvme_req(req)->status;
56 spin_lock_irqsave(&ns->head->requeue_lock, flags);
57 blk_steal_bios(&ns->head->requeue_list, req);
58 spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
59 blk_mq_end_request(req, 0);
61 switch (status & 0x7ff) {
62 case NVME_SC_ANA_TRANSITION:
63 case NVME_SC_ANA_INACCESSIBLE:
64 case NVME_SC_ANA_PERSISTENT_LOSS:
66 * If we got back an ANA error we know the controller is alive,
67 * but not ready to serve this namespaces. The spec suggests
68 * we should update our general state here, but due to the fact
69 * that the admin and I/O queues are not serialized that is
70 * fundamentally racy. So instead just clear the current path,
71 * mark the the path as pending and kick of a re-read of the ANA
74 nvme_mpath_clear_current_path(ns);
75 if (ns->ctrl->ana_log_buf) {
76 set_bit(NVME_NS_ANA_PENDING, &ns->flags);
77 queue_work(nvme_wq, &ns->ctrl->ana_work);
80 case NVME_SC_HOST_PATH_ERROR:
82 * Temporary transport disruption in talking to the controller.
83 * Try to send on a new path.
85 nvme_mpath_clear_current_path(ns);
89 * Reset the controller for any non-ANA error as we don't know
90 * what caused the error.
92 nvme_reset_ctrl(ns->ctrl);
96 kblockd_schedule_work(&ns->head->requeue_work);
99 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
103 down_read(&ctrl->namespaces_rwsem);
104 list_for_each_entry(ns, &ctrl->namespaces, list) {
106 kblockd_schedule_work(&ns->head->requeue_work);
108 up_read(&ctrl->namespaces_rwsem);
111 static const char *nvme_ana_state_names[] = {
112 [0] = "invalid state",
113 [NVME_ANA_OPTIMIZED] = "optimized",
114 [NVME_ANA_NONOPTIMIZED] = "non-optimized",
115 [NVME_ANA_INACCESSIBLE] = "inaccessible",
116 [NVME_ANA_PERSISTENT_LOSS] = "persistent-loss",
117 [NVME_ANA_CHANGE] = "change",
120 void nvme_mpath_clear_current_path(struct nvme_ns *ns)
122 struct nvme_ns_head *head = ns->head;
128 for_each_node(node) {
129 if (ns == rcu_access_pointer(head->current_path[node]))
130 rcu_assign_pointer(head->current_path[node], NULL);
134 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
136 int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
137 struct nvme_ns *found = NULL, *fallback = NULL, *ns;
139 list_for_each_entry_rcu(ns, &head->list, siblings) {
140 if (ns->ctrl->state != NVME_CTRL_LIVE ||
141 test_bit(NVME_NS_ANA_PENDING, &ns->flags))
144 distance = node_distance(node, dev_to_node(ns->ctrl->dev));
146 switch (ns->ana_state) {
147 case NVME_ANA_OPTIMIZED:
148 if (distance < found_distance) {
149 found_distance = distance;
153 case NVME_ANA_NONOPTIMIZED:
154 if (distance < fallback_distance) {
155 fallback_distance = distance;
167 rcu_assign_pointer(head->current_path[node], found);
171 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
173 return ns->ctrl->state == NVME_CTRL_LIVE &&
174 ns->ana_state == NVME_ANA_OPTIMIZED;
177 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
179 int node = numa_node_id();
182 ns = srcu_dereference(head->current_path[node], &head->srcu);
183 if (unlikely(!ns || !nvme_path_is_optimized(ns)))
184 ns = __nvme_find_path(head, node);
188 static blk_qc_t nvme_ns_head_make_request(struct request_queue *q,
191 struct nvme_ns_head *head = q->queuedata;
192 struct device *dev = disk_to_dev(head->disk);
194 blk_qc_t ret = BLK_QC_T_NONE;
197 srcu_idx = srcu_read_lock(&head->srcu);
198 ns = nvme_find_path(head);
200 bio->bi_disk = ns->disk;
201 bio->bi_opf |= REQ_NVME_MPATH;
202 trace_block_bio_remap(bio->bi_disk->queue, bio,
203 disk_devt(ns->head->disk),
204 bio->bi_iter.bi_sector);
205 ret = direct_make_request(bio);
206 } else if (!list_empty_careful(&head->list)) {
207 dev_warn_ratelimited(dev, "no path available - requeuing I/O\n");
209 spin_lock_irq(&head->requeue_lock);
210 bio_list_add(&head->requeue_list, bio);
211 spin_unlock_irq(&head->requeue_lock);
213 dev_warn_ratelimited(dev, "no path - failing I/O\n");
215 bio->bi_status = BLK_STS_IOERR;
219 srcu_read_unlock(&head->srcu, srcu_idx);
223 static bool nvme_ns_head_poll(struct request_queue *q, blk_qc_t qc)
225 struct nvme_ns_head *head = q->queuedata;
230 srcu_idx = srcu_read_lock(&head->srcu);
231 ns = srcu_dereference(head->current_path[numa_node_id()], &head->srcu);
232 if (likely(ns && nvme_path_is_optimized(ns)))
233 found = ns->queue->poll_fn(q, qc);
234 srcu_read_unlock(&head->srcu, srcu_idx);
238 static void nvme_requeue_work(struct work_struct *work)
240 struct nvme_ns_head *head =
241 container_of(work, struct nvme_ns_head, requeue_work);
242 struct bio *bio, *next;
244 spin_lock_irq(&head->requeue_lock);
245 next = bio_list_get(&head->requeue_list);
246 spin_unlock_irq(&head->requeue_lock);
248 while ((bio = next) != NULL) {
253 * Reset disk to the mpath node and resubmit to select a new
256 bio->bi_disk = head->disk;
257 generic_make_request(bio);
261 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
263 struct request_queue *q;
266 mutex_init(&head->lock);
267 bio_list_init(&head->requeue_list);
268 spin_lock_init(&head->requeue_lock);
269 INIT_WORK(&head->requeue_work, nvme_requeue_work);
272 * Add a multipath node if the subsystems supports multiple controllers.
273 * We also do this for private namespaces as the namespace sharing data could
274 * change after a rescan.
276 if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath)
279 q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, NULL);
283 blk_queue_make_request(q, nvme_ns_head_make_request);
284 q->poll_fn = nvme_ns_head_poll;
285 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
286 /* set to a default value for 512 until disk is validated */
287 blk_queue_logical_block_size(q, 512);
288 blk_set_stacking_limits(&q->limits);
290 /* we need to propagate up the VMC settings */
291 if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
293 blk_queue_write_cache(q, vwc, vwc);
295 head->disk = alloc_disk(0);
297 goto out_cleanup_queue;
298 head->disk->fops = &nvme_ns_head_ops;
299 head->disk->private_data = head;
300 head->disk->queue = q;
301 head->disk->flags = GENHD_FL_EXT_DEVT;
302 sprintf(head->disk->disk_name, "nvme%dn%d",
303 ctrl->subsys->instance, head->instance);
307 blk_cleanup_queue(q);
312 static void nvme_mpath_set_live(struct nvme_ns *ns)
314 struct nvme_ns_head *head = ns->head;
316 lockdep_assert_held(&ns->head->lock);
321 if (!(head->disk->flags & GENHD_FL_UP))
322 device_add_disk(&head->subsys->dev, head->disk,
323 nvme_ns_id_attr_groups);
325 if (nvme_path_is_optimized(ns)) {
328 srcu_idx = srcu_read_lock(&head->srcu);
330 __nvme_find_path(head, node);
331 srcu_read_unlock(&head->srcu, srcu_idx);
334 kblockd_schedule_work(&ns->head->requeue_work);
337 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
338 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
341 void *base = ctrl->ana_log_buf;
342 size_t offset = sizeof(struct nvme_ana_rsp_hdr);
345 lockdep_assert_held(&ctrl->ana_lock);
347 for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
348 struct nvme_ana_group_desc *desc = base + offset;
349 u32 nr_nsids = le32_to_cpu(desc->nnsids);
350 size_t nsid_buf_size = nr_nsids * sizeof(__le32);
352 if (WARN_ON_ONCE(desc->grpid == 0))
354 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
356 if (WARN_ON_ONCE(desc->state == 0))
358 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
361 offset += sizeof(*desc);
362 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
365 error = cb(ctrl, desc, data);
369 offset += nsid_buf_size;
370 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
377 static inline bool nvme_state_is_live(enum nvme_ana_state state)
379 return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
382 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
385 enum nvme_ana_state old;
387 mutex_lock(&ns->head->lock);
389 ns->ana_grpid = le32_to_cpu(desc->grpid);
390 ns->ana_state = desc->state;
391 clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
393 if (nvme_state_is_live(ns->ana_state) && !nvme_state_is_live(old))
394 nvme_mpath_set_live(ns);
395 mutex_unlock(&ns->head->lock);
398 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
399 struct nvme_ana_group_desc *desc, void *data)
401 u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
402 unsigned *nr_change_groups = data;
405 dev_info(ctrl->device, "ANA group %d: %s.\n",
406 le32_to_cpu(desc->grpid),
407 nvme_ana_state_names[desc->state]);
409 if (desc->state == NVME_ANA_CHANGE)
410 (*nr_change_groups)++;
415 down_write(&ctrl->namespaces_rwsem);
416 list_for_each_entry(ns, &ctrl->namespaces, list) {
417 if (ns->head->ns_id != le32_to_cpu(desc->nsids[n]))
419 nvme_update_ns_ana_state(desc, ns);
423 up_write(&ctrl->namespaces_rwsem);
424 WARN_ON_ONCE(n < nr_nsids);
428 static int nvme_read_ana_log(struct nvme_ctrl *ctrl, bool groups_only)
430 u32 nr_change_groups = 0;
433 mutex_lock(&ctrl->ana_lock);
434 error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA,
435 groups_only ? NVME_ANA_LOG_RGO : 0,
436 ctrl->ana_log_buf, ctrl->ana_log_size, 0);
438 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
442 error = nvme_parse_ana_log(ctrl, &nr_change_groups,
443 nvme_update_ana_state);
448 * In theory we should have an ANATT timer per group as they might enter
449 * the change state at different times. But that is a lot of overhead
450 * just to protect against a target that keeps entering new changes
451 * states while never finishing previous ones. But we'll still
452 * eventually time out once all groups are in change state, so this
455 * We also double the ANATT value to provide some slack for transports
456 * or AEN processing overhead.
458 if (nr_change_groups)
459 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
461 del_timer_sync(&ctrl->anatt_timer);
463 mutex_unlock(&ctrl->ana_lock);
467 static void nvme_ana_work(struct work_struct *work)
469 struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
471 nvme_read_ana_log(ctrl, false);
474 static void nvme_anatt_timeout(struct timer_list *t)
476 struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
478 dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
479 nvme_reset_ctrl(ctrl);
482 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
484 if (!nvme_ctrl_use_ana(ctrl))
486 del_timer_sync(&ctrl->anatt_timer);
487 cancel_work_sync(&ctrl->ana_work);
490 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
493 return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
495 DEVICE_ATTR_RO(ana_grpid);
497 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
500 struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
502 return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
504 DEVICE_ATTR_RO(ana_state);
506 static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl,
507 struct nvme_ana_group_desc *desc, void *data)
509 struct nvme_ns *ns = data;
511 if (ns->ana_grpid == le32_to_cpu(desc->grpid)) {
512 nvme_update_ns_ana_state(desc, ns);
513 return -ENXIO; /* just break out of the loop */
519 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
521 if (nvme_ctrl_use_ana(ns->ctrl)) {
522 mutex_lock(&ns->ctrl->ana_lock);
523 ns->ana_grpid = le32_to_cpu(id->anagrpid);
524 nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state);
525 mutex_unlock(&ns->ctrl->ana_lock);
527 mutex_lock(&ns->head->lock);
528 ns->ana_state = NVME_ANA_OPTIMIZED;
529 nvme_mpath_set_live(ns);
530 mutex_unlock(&ns->head->lock);
534 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
538 if (head->disk->flags & GENHD_FL_UP)
539 del_gendisk(head->disk);
540 blk_set_queue_dying(head->disk->queue);
541 /* make sure all pending bios are cleaned up */
542 kblockd_schedule_work(&head->requeue_work);
543 flush_work(&head->requeue_work);
544 blk_cleanup_queue(head->disk->queue);
545 put_disk(head->disk);
548 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
552 if (!nvme_ctrl_use_ana(ctrl))
555 ctrl->anacap = id->anacap;
556 ctrl->anatt = id->anatt;
557 ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
558 ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
560 mutex_init(&ctrl->ana_lock);
561 timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
562 ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
563 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc);
564 if (!(ctrl->anacap & (1 << 6)))
565 ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32);
567 if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) {
568 dev_err(ctrl->device,
569 "ANA log page size (%zd) larger than MDTS (%d).\n",
571 ctrl->max_hw_sectors << SECTOR_SHIFT);
572 dev_err(ctrl->device, "disabling ANA support.\n");
576 INIT_WORK(&ctrl->ana_work, nvme_ana_work);
577 ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL);
578 if (!ctrl->ana_log_buf) {
583 error = nvme_read_ana_log(ctrl, true);
585 goto out_free_ana_log_buf;
587 out_free_ana_log_buf:
588 kfree(ctrl->ana_log_buf);
593 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
595 kfree(ctrl->ana_log_buf);