xref: /linux/drivers/nvme/host/multipath.c (revision 55d0969c451159cff86949b38c39171cab962069)
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 	if (!test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags))
425 		return NULL;
426 
427 	list_for_each_entry_rcu(ns, &head->list, siblings) {
428 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
429 			continue;
430 		switch (nvme_ctrl_state(ns->ctrl)) {
431 		case NVME_CTRL_LIVE:
432 		case NVME_CTRL_RESETTING:
433 		case NVME_CTRL_CONNECTING:
434 			return true;
435 		default:
436 			break;
437 		}
438 	}
439 	return false;
440 }
441 
442 static void nvme_ns_head_submit_bio(struct bio *bio)
443 {
444 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
445 	struct device *dev = disk_to_dev(head->disk);
446 	struct nvme_ns *ns;
447 	int srcu_idx;
448 
449 	/*
450 	 * The namespace might be going away and the bio might be moved to a
451 	 * different queue via blk_steal_bios(), so we need to use the bio_split
452 	 * pool from the original queue to allocate the bvecs from.
453 	 */
454 	bio = bio_split_to_limits(bio);
455 	if (!bio)
456 		return;
457 
458 	srcu_idx = srcu_read_lock(&head->srcu);
459 	ns = nvme_find_path(head);
460 	if (likely(ns)) {
461 		bio_set_dev(bio, ns->disk->part0);
462 		bio->bi_opf |= REQ_NVME_MPATH;
463 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
464 				      bio->bi_iter.bi_sector);
465 		submit_bio_noacct(bio);
466 	} else if (nvme_available_path(head)) {
467 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
468 
469 		spin_lock_irq(&head->requeue_lock);
470 		bio_list_add(&head->requeue_list, bio);
471 		spin_unlock_irq(&head->requeue_lock);
472 	} else {
473 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
474 
475 		bio_io_error(bio);
476 	}
477 
478 	srcu_read_unlock(&head->srcu, srcu_idx);
479 }
480 
481 static int nvme_ns_head_open(struct gendisk *disk, blk_mode_t mode)
482 {
483 	if (!nvme_tryget_ns_head(disk->private_data))
484 		return -ENXIO;
485 	return 0;
486 }
487 
488 static void nvme_ns_head_release(struct gendisk *disk)
489 {
490 	nvme_put_ns_head(disk->private_data);
491 }
492 
493 static int nvme_ns_head_get_unique_id(struct gendisk *disk, u8 id[16],
494 		enum blk_unique_id type)
495 {
496 	struct nvme_ns_head *head = disk->private_data;
497 	struct nvme_ns *ns;
498 	int srcu_idx, ret = -EWOULDBLOCK;
499 
500 	srcu_idx = srcu_read_lock(&head->srcu);
501 	ns = nvme_find_path(head);
502 	if (ns)
503 		ret = nvme_ns_get_unique_id(ns, id, type);
504 	srcu_read_unlock(&head->srcu, srcu_idx);
505 	return ret;
506 }
507 
508 #ifdef CONFIG_BLK_DEV_ZONED
509 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
510 		unsigned int nr_zones, report_zones_cb cb, void *data)
511 {
512 	struct nvme_ns_head *head = disk->private_data;
513 	struct nvme_ns *ns;
514 	int srcu_idx, ret = -EWOULDBLOCK;
515 
516 	srcu_idx = srcu_read_lock(&head->srcu);
517 	ns = nvme_find_path(head);
518 	if (ns)
519 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
520 	srcu_read_unlock(&head->srcu, srcu_idx);
521 	return ret;
522 }
523 #else
524 #define nvme_ns_head_report_zones	NULL
525 #endif /* CONFIG_BLK_DEV_ZONED */
526 
527 const struct block_device_operations nvme_ns_head_ops = {
528 	.owner		= THIS_MODULE,
529 	.submit_bio	= nvme_ns_head_submit_bio,
530 	.open		= nvme_ns_head_open,
531 	.release	= nvme_ns_head_release,
532 	.ioctl		= nvme_ns_head_ioctl,
533 	.compat_ioctl	= blkdev_compat_ptr_ioctl,
534 	.getgeo		= nvme_getgeo,
535 	.get_unique_id	= nvme_ns_head_get_unique_id,
536 	.report_zones	= nvme_ns_head_report_zones,
537 	.pr_ops		= &nvme_pr_ops,
538 };
539 
540 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
541 {
542 	return container_of(cdev, struct nvme_ns_head, cdev);
543 }
544 
545 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
546 {
547 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
548 		return -ENXIO;
549 	return 0;
550 }
551 
552 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
553 {
554 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
555 	return 0;
556 }
557 
558 static const struct file_operations nvme_ns_head_chr_fops = {
559 	.owner		= THIS_MODULE,
560 	.open		= nvme_ns_head_chr_open,
561 	.release	= nvme_ns_head_chr_release,
562 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
563 	.compat_ioctl	= compat_ptr_ioctl,
564 	.uring_cmd	= nvme_ns_head_chr_uring_cmd,
565 	.uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
566 };
567 
568 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
569 {
570 	int ret;
571 
572 	head->cdev_device.parent = &head->subsys->dev;
573 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
574 			   head->subsys->instance, head->instance);
575 	if (ret)
576 		return ret;
577 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
578 			    &nvme_ns_head_chr_fops, THIS_MODULE);
579 	return ret;
580 }
581 
582 static void nvme_partition_scan_work(struct work_struct *work)
583 {
584 	struct nvme_ns_head *head =
585 		container_of(work, struct nvme_ns_head, partition_scan_work);
586 
587 	if (WARN_ON_ONCE(!test_and_clear_bit(GD_SUPPRESS_PART_SCAN,
588 					     &head->disk->state)))
589 		return;
590 
591 	mutex_lock(&head->disk->open_mutex);
592 	bdev_disk_changed(head->disk, false);
593 	mutex_unlock(&head->disk->open_mutex);
594 }
595 
596 static void nvme_requeue_work(struct work_struct *work)
597 {
598 	struct nvme_ns_head *head =
599 		container_of(work, struct nvme_ns_head, requeue_work);
600 	struct bio *bio, *next;
601 
602 	spin_lock_irq(&head->requeue_lock);
603 	next = bio_list_get(&head->requeue_list);
604 	spin_unlock_irq(&head->requeue_lock);
605 
606 	while ((bio = next) != NULL) {
607 		next = bio->bi_next;
608 		bio->bi_next = NULL;
609 
610 		submit_bio_noacct(bio);
611 	}
612 }
613 
614 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
615 {
616 	struct queue_limits lim;
617 
618 	mutex_init(&head->lock);
619 	bio_list_init(&head->requeue_list);
620 	spin_lock_init(&head->requeue_lock);
621 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
622 	INIT_WORK(&head->partition_scan_work, nvme_partition_scan_work);
623 
624 	/*
625 	 * Add a multipath node if the subsystems supports multiple controllers.
626 	 * We also do this for private namespaces as the namespace sharing flag
627 	 * could change after a rescan.
628 	 */
629 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
630 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
631 		return 0;
632 
633 	blk_set_stacking_limits(&lim);
634 	lim.dma_alignment = 3;
635 	lim.features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT | BLK_FEAT_POLL;
636 	if (head->ids.csi == NVME_CSI_ZNS)
637 		lim.features |= BLK_FEAT_ZONED;
638 	else
639 		lim.max_zone_append_sectors = 0;
640 
641 	head->disk = blk_alloc_disk(&lim, ctrl->numa_node);
642 	if (IS_ERR(head->disk))
643 		return PTR_ERR(head->disk);
644 	head->disk->fops = &nvme_ns_head_ops;
645 	head->disk->private_data = head;
646 
647 	/*
648 	 * We need to suppress the partition scan from occuring within the
649 	 * controller's scan_work context. If a path error occurs here, the IO
650 	 * will wait until a path becomes available or all paths are torn down,
651 	 * but that action also occurs within scan_work, so it would deadlock.
652 	 * Defer the partion scan to a different context that does not block
653 	 * scan_work.
654 	 */
655 	set_bit(GD_SUPPRESS_PART_SCAN, &head->disk->state);
656 	sprintf(head->disk->disk_name, "nvme%dn%d",
657 			ctrl->subsys->instance, head->instance);
658 	return 0;
659 }
660 
661 static void nvme_mpath_set_live(struct nvme_ns *ns)
662 {
663 	struct nvme_ns_head *head = ns->head;
664 	int rc;
665 
666 	if (!head->disk)
667 		return;
668 
669 	/*
670 	 * test_and_set_bit() is used because it is protecting against two nvme
671 	 * paths simultaneously calling device_add_disk() on the same namespace
672 	 * head.
673 	 */
674 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
675 		rc = device_add_disk(&head->subsys->dev, head->disk,
676 				     nvme_ns_attr_groups);
677 		if (rc) {
678 			clear_bit(NVME_NSHEAD_DISK_LIVE, &head->flags);
679 			return;
680 		}
681 		nvme_add_ns_head_cdev(head);
682 		kblockd_schedule_work(&head->partition_scan_work);
683 	}
684 
685 	mutex_lock(&head->lock);
686 	if (nvme_path_is_optimized(ns)) {
687 		int node, srcu_idx;
688 
689 		srcu_idx = srcu_read_lock(&head->srcu);
690 		for_each_online_node(node)
691 			__nvme_find_path(head, node);
692 		srcu_read_unlock(&head->srcu, srcu_idx);
693 	}
694 	mutex_unlock(&head->lock);
695 
696 	synchronize_srcu(&head->srcu);
697 	kblockd_schedule_work(&head->requeue_work);
698 }
699 
700 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
701 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
702 			void *))
703 {
704 	void *base = ctrl->ana_log_buf;
705 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
706 	int error, i;
707 
708 	lockdep_assert_held(&ctrl->ana_lock);
709 
710 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
711 		struct nvme_ana_group_desc *desc = base + offset;
712 		u32 nr_nsids;
713 		size_t nsid_buf_size;
714 
715 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
716 			return -EINVAL;
717 
718 		nr_nsids = le32_to_cpu(desc->nnsids);
719 		nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
720 
721 		if (WARN_ON_ONCE(desc->grpid == 0))
722 			return -EINVAL;
723 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
724 			return -EINVAL;
725 		if (WARN_ON_ONCE(desc->state == 0))
726 			return -EINVAL;
727 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
728 			return -EINVAL;
729 
730 		offset += sizeof(*desc);
731 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
732 			return -EINVAL;
733 
734 		error = cb(ctrl, desc, data);
735 		if (error)
736 			return error;
737 
738 		offset += nsid_buf_size;
739 	}
740 
741 	return 0;
742 }
743 
744 static inline bool nvme_state_is_live(enum nvme_ana_state state)
745 {
746 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
747 }
748 
749 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
750 		struct nvme_ns *ns)
751 {
752 	ns->ana_grpid = le32_to_cpu(desc->grpid);
753 	ns->ana_state = desc->state;
754 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
755 	/*
756 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
757 	 * and in turn to this path device.  However we cannot accept this I/O
758 	 * if the controller is not live.  This may deadlock if called from
759 	 * nvme_mpath_init_identify() and the ctrl will never complete
760 	 * initialization, preventing I/O from completing.  For this case we
761 	 * will reprocess the ANA log page in nvme_mpath_update() once the
762 	 * controller is ready.
763 	 */
764 	if (nvme_state_is_live(ns->ana_state) &&
765 	    nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
766 		nvme_mpath_set_live(ns);
767 }
768 
769 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
770 		struct nvme_ana_group_desc *desc, void *data)
771 {
772 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
773 	unsigned *nr_change_groups = data;
774 	struct nvme_ns *ns;
775 	int srcu_idx;
776 
777 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
778 			le32_to_cpu(desc->grpid),
779 			nvme_ana_state_names[desc->state]);
780 
781 	if (desc->state == NVME_ANA_CHANGE)
782 		(*nr_change_groups)++;
783 
784 	if (!nr_nsids)
785 		return 0;
786 
787 	srcu_idx = srcu_read_lock(&ctrl->srcu);
788 	list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
789 		unsigned nsid;
790 again:
791 		nsid = le32_to_cpu(desc->nsids[n]);
792 		if (ns->head->ns_id < nsid)
793 			continue;
794 		if (ns->head->ns_id == nsid)
795 			nvme_update_ns_ana_state(desc, ns);
796 		if (++n == nr_nsids)
797 			break;
798 		if (ns->head->ns_id > nsid)
799 			goto again;
800 	}
801 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
802 	return 0;
803 }
804 
805 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
806 {
807 	u32 nr_change_groups = 0;
808 	int error;
809 
810 	mutex_lock(&ctrl->ana_lock);
811 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
812 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
813 	if (error) {
814 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
815 		goto out_unlock;
816 	}
817 
818 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
819 			nvme_update_ana_state);
820 	if (error)
821 		goto out_unlock;
822 
823 	/*
824 	 * In theory we should have an ANATT timer per group as they might enter
825 	 * the change state at different times.  But that is a lot of overhead
826 	 * just to protect against a target that keeps entering new changes
827 	 * states while never finishing previous ones.  But we'll still
828 	 * eventually time out once all groups are in change state, so this
829 	 * isn't a big deal.
830 	 *
831 	 * We also double the ANATT value to provide some slack for transports
832 	 * or AEN processing overhead.
833 	 */
834 	if (nr_change_groups)
835 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
836 	else
837 		del_timer_sync(&ctrl->anatt_timer);
838 out_unlock:
839 	mutex_unlock(&ctrl->ana_lock);
840 	return error;
841 }
842 
843 static void nvme_ana_work(struct work_struct *work)
844 {
845 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
846 
847 	if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
848 		return;
849 
850 	nvme_read_ana_log(ctrl);
851 }
852 
853 void nvme_mpath_update(struct nvme_ctrl *ctrl)
854 {
855 	u32 nr_change_groups = 0;
856 
857 	if (!ctrl->ana_log_buf)
858 		return;
859 
860 	mutex_lock(&ctrl->ana_lock);
861 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
862 	mutex_unlock(&ctrl->ana_lock);
863 }
864 
865 static void nvme_anatt_timeout(struct timer_list *t)
866 {
867 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
868 
869 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
870 	nvme_reset_ctrl(ctrl);
871 }
872 
873 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
874 {
875 	if (!nvme_ctrl_use_ana(ctrl))
876 		return;
877 	del_timer_sync(&ctrl->anatt_timer);
878 	cancel_work_sync(&ctrl->ana_work);
879 }
880 
881 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
882 	struct device_attribute subsys_attr_##_name =	\
883 		__ATTR(_name, _mode, _show, _store)
884 
885 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
886 		struct device_attribute *attr, char *buf)
887 {
888 	struct nvme_subsystem *subsys =
889 		container_of(dev, struct nvme_subsystem, dev);
890 
891 	return sysfs_emit(buf, "%s\n",
892 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
893 }
894 
895 static void nvme_subsys_iopolicy_update(struct nvme_subsystem *subsys,
896 		int iopolicy)
897 {
898 	struct nvme_ctrl *ctrl;
899 	int old_iopolicy = READ_ONCE(subsys->iopolicy);
900 
901 	if (old_iopolicy == iopolicy)
902 		return;
903 
904 	WRITE_ONCE(subsys->iopolicy, iopolicy);
905 
906 	/* iopolicy changes clear the mpath by design */
907 	mutex_lock(&nvme_subsystems_lock);
908 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
909 		nvme_mpath_clear_ctrl_paths(ctrl);
910 	mutex_unlock(&nvme_subsystems_lock);
911 
912 	pr_notice("subsysnqn %s iopolicy changed from %s to %s\n",
913 			subsys->subnqn,
914 			nvme_iopolicy_names[old_iopolicy],
915 			nvme_iopolicy_names[iopolicy]);
916 }
917 
918 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
919 		struct device_attribute *attr, const char *buf, size_t count)
920 {
921 	struct nvme_subsystem *subsys =
922 		container_of(dev, struct nvme_subsystem, dev);
923 	int i;
924 
925 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
926 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
927 			nvme_subsys_iopolicy_update(subsys, i);
928 			return count;
929 		}
930 	}
931 
932 	return -EINVAL;
933 }
934 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
935 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
936 
937 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
938 		char *buf)
939 {
940 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
941 }
942 DEVICE_ATTR_RO(ana_grpid);
943 
944 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
945 		char *buf)
946 {
947 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
948 
949 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
950 }
951 DEVICE_ATTR_RO(ana_state);
952 
953 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
954 		struct nvme_ana_group_desc *desc, void *data)
955 {
956 	struct nvme_ana_group_desc *dst = data;
957 
958 	if (desc->grpid != dst->grpid)
959 		return 0;
960 
961 	*dst = *desc;
962 	return -ENXIO; /* just break out of the loop */
963 }
964 
965 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
966 {
967 	if (nvme_ctrl_use_ana(ns->ctrl)) {
968 		struct nvme_ana_group_desc desc = {
969 			.grpid = anagrpid,
970 			.state = 0,
971 		};
972 
973 		mutex_lock(&ns->ctrl->ana_lock);
974 		ns->ana_grpid = le32_to_cpu(anagrpid);
975 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
976 		mutex_unlock(&ns->ctrl->ana_lock);
977 		if (desc.state) {
978 			/* found the group desc: update */
979 			nvme_update_ns_ana_state(&desc, ns);
980 		} else {
981 			/* group desc not found: trigger a re-read */
982 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
983 			queue_work(nvme_wq, &ns->ctrl->ana_work);
984 		}
985 	} else {
986 		ns->ana_state = NVME_ANA_OPTIMIZED;
987 		nvme_mpath_set_live(ns);
988 	}
989 
990 #ifdef CONFIG_BLK_DEV_ZONED
991 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
992 		ns->head->disk->nr_zones = ns->disk->nr_zones;
993 #endif
994 }
995 
996 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
997 {
998 	if (!head->disk)
999 		return;
1000 	if (test_and_clear_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
1001 		nvme_cdev_del(&head->cdev, &head->cdev_device);
1002 		/*
1003 		 * requeue I/O after NVME_NSHEAD_DISK_LIVE has been cleared
1004 		 * to allow multipath to fail all I/O.
1005 		 */
1006 		synchronize_srcu(&head->srcu);
1007 		kblockd_schedule_work(&head->requeue_work);
1008 		del_gendisk(head->disk);
1009 	}
1010 }
1011 
1012 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
1013 {
1014 	if (!head->disk)
1015 		return;
1016 	/* make sure all pending bios are cleaned up */
1017 	kblockd_schedule_work(&head->requeue_work);
1018 	flush_work(&head->requeue_work);
1019 	flush_work(&head->partition_scan_work);
1020 	put_disk(head->disk);
1021 }
1022 
1023 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
1024 {
1025 	mutex_init(&ctrl->ana_lock);
1026 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
1027 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
1028 }
1029 
1030 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
1031 {
1032 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
1033 	size_t ana_log_size;
1034 	int error = 0;
1035 
1036 	/* check if multipath is enabled and we have the capability */
1037 	if (!multipath || !ctrl->subsys ||
1038 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
1039 		return 0;
1040 
1041 	/* initialize this in the identify path to cover controller resets */
1042 	atomic_set(&ctrl->nr_active, 0);
1043 
1044 	if (!ctrl->max_namespaces ||
1045 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
1046 		dev_err(ctrl->device,
1047 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
1048 		return -EINVAL;
1049 	}
1050 
1051 	ctrl->anacap = id->anacap;
1052 	ctrl->anatt = id->anatt;
1053 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
1054 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
1055 
1056 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
1057 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
1058 		ctrl->max_namespaces * sizeof(__le32);
1059 	if (ana_log_size > max_transfer_size) {
1060 		dev_err(ctrl->device,
1061 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
1062 			ana_log_size, max_transfer_size);
1063 		dev_err(ctrl->device, "disabling ANA support.\n");
1064 		goto out_uninit;
1065 	}
1066 	if (ana_log_size > ctrl->ana_log_size) {
1067 		nvme_mpath_stop(ctrl);
1068 		nvme_mpath_uninit(ctrl);
1069 		ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
1070 		if (!ctrl->ana_log_buf)
1071 			return -ENOMEM;
1072 	}
1073 	ctrl->ana_log_size = ana_log_size;
1074 	error = nvme_read_ana_log(ctrl);
1075 	if (error)
1076 		goto out_uninit;
1077 	return 0;
1078 
1079 out_uninit:
1080 	nvme_mpath_uninit(ctrl);
1081 	return error;
1082 }
1083 
1084 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
1085 {
1086 	kvfree(ctrl->ana_log_buf);
1087 	ctrl->ana_log_buf = NULL;
1088 	ctrl->ana_log_size = 0;
1089 }
1090