xref: /linux/drivers/nvme/host/multipath.c (revision 5dfe7a7e52ccdf60dfd11ccbe509e4365ea721ca)
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 };
21 
22 static int iopolicy = NVME_IOPOLICY_NUMA;
23 
24 static int nvme_set_iopolicy(const char *val, const struct kernel_param *kp)
25 {
26 	if (!val)
27 		return -EINVAL;
28 	if (!strncmp(val, "numa", 4))
29 		iopolicy = NVME_IOPOLICY_NUMA;
30 	else if (!strncmp(val, "round-robin", 11))
31 		iopolicy = NVME_IOPOLICY_RR;
32 	else
33 		return -EINVAL;
34 
35 	return 0;
36 }
37 
38 static int nvme_get_iopolicy(char *buf, const struct kernel_param *kp)
39 {
40 	return sprintf(buf, "%s\n", nvme_iopolicy_names[iopolicy]);
41 }
42 
43 module_param_call(iopolicy, nvme_set_iopolicy, nvme_get_iopolicy,
44 	&iopolicy, 0644);
45 MODULE_PARM_DESC(iopolicy,
46 	"Default multipath I/O policy; 'numa' (default) or 'round-robin'");
47 
48 void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys)
49 {
50 	subsys->iopolicy = iopolicy;
51 }
52 
53 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
54 {
55 	struct nvme_ns_head *h;
56 
57 	lockdep_assert_held(&subsys->lock);
58 	list_for_each_entry(h, &subsys->nsheads, entry)
59 		if (h->disk)
60 			blk_mq_unfreeze_queue(h->disk->queue);
61 }
62 
63 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
64 {
65 	struct nvme_ns_head *h;
66 
67 	lockdep_assert_held(&subsys->lock);
68 	list_for_each_entry(h, &subsys->nsheads, entry)
69 		if (h->disk)
70 			blk_mq_freeze_queue_wait(h->disk->queue);
71 }
72 
73 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
74 {
75 	struct nvme_ns_head *h;
76 
77 	lockdep_assert_held(&subsys->lock);
78 	list_for_each_entry(h, &subsys->nsheads, entry)
79 		if (h->disk)
80 			blk_freeze_queue_start(h->disk->queue);
81 }
82 
83 void nvme_failover_req(struct request *req)
84 {
85 	struct nvme_ns *ns = req->q->queuedata;
86 	u16 status = nvme_req(req)->status & 0x7ff;
87 	unsigned long flags;
88 	struct bio *bio;
89 
90 	nvme_mpath_clear_current_path(ns);
91 
92 	/*
93 	 * If we got back an ANA error, we know the controller is alive but not
94 	 * ready to serve this namespace.  Kick of a re-read of the ANA
95 	 * information page, and just try any other available path for now.
96 	 */
97 	if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) {
98 		set_bit(NVME_NS_ANA_PENDING, &ns->flags);
99 		queue_work(nvme_wq, &ns->ctrl->ana_work);
100 	}
101 
102 	spin_lock_irqsave(&ns->head->requeue_lock, flags);
103 	for (bio = req->bio; bio; bio = bio->bi_next) {
104 		bio_set_dev(bio, ns->head->disk->part0);
105 		if (bio->bi_opf & REQ_POLLED) {
106 			bio->bi_opf &= ~REQ_POLLED;
107 			bio->bi_cookie = BLK_QC_T_NONE;
108 		}
109 	}
110 	blk_steal_bios(&ns->head->requeue_list, req);
111 	spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
112 
113 	blk_mq_end_request(req, 0);
114 	kblockd_schedule_work(&ns->head->requeue_work);
115 }
116 
117 void nvme_mpath_start_request(struct request *rq)
118 {
119 	struct nvme_ns *ns = rq->q->queuedata;
120 	struct gendisk *disk = ns->head->disk;
121 
122 	if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq))
123 		return;
124 
125 	nvme_req(rq)->flags |= NVME_MPATH_IO_STATS;
126 	nvme_req(rq)->start_time = bdev_start_io_acct(disk->part0, req_op(rq),
127 						      jiffies);
128 }
129 EXPORT_SYMBOL_GPL(nvme_mpath_start_request);
130 
131 void nvme_mpath_end_request(struct request *rq)
132 {
133 	struct nvme_ns *ns = rq->q->queuedata;
134 
135 	if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
136 		return;
137 	bdev_end_io_acct(ns->head->disk->part0, req_op(rq),
138 			 blk_rq_bytes(rq) >> SECTOR_SHIFT,
139 			 nvme_req(rq)->start_time);
140 }
141 
142 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
143 {
144 	struct nvme_ns *ns;
145 
146 	down_read(&ctrl->namespaces_rwsem);
147 	list_for_each_entry(ns, &ctrl->namespaces, list) {
148 		if (!ns->head->disk)
149 			continue;
150 		kblockd_schedule_work(&ns->head->requeue_work);
151 		if (ctrl->state == NVME_CTRL_LIVE)
152 			disk_uevent(ns->head->disk, KOBJ_CHANGE);
153 	}
154 	up_read(&ctrl->namespaces_rwsem);
155 }
156 
157 static const char *nvme_ana_state_names[] = {
158 	[0]				= "invalid state",
159 	[NVME_ANA_OPTIMIZED]		= "optimized",
160 	[NVME_ANA_NONOPTIMIZED]		= "non-optimized",
161 	[NVME_ANA_INACCESSIBLE]		= "inaccessible",
162 	[NVME_ANA_PERSISTENT_LOSS]	= "persistent-loss",
163 	[NVME_ANA_CHANGE]		= "change",
164 };
165 
166 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
167 {
168 	struct nvme_ns_head *head = ns->head;
169 	bool changed = false;
170 	int node;
171 
172 	if (!head)
173 		goto out;
174 
175 	for_each_node(node) {
176 		if (ns == rcu_access_pointer(head->current_path[node])) {
177 			rcu_assign_pointer(head->current_path[node], NULL);
178 			changed = true;
179 		}
180 	}
181 out:
182 	return changed;
183 }
184 
185 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
186 {
187 	struct nvme_ns *ns;
188 
189 	down_read(&ctrl->namespaces_rwsem);
190 	list_for_each_entry(ns, &ctrl->namespaces, list) {
191 		nvme_mpath_clear_current_path(ns);
192 		kblockd_schedule_work(&ns->head->requeue_work);
193 	}
194 	up_read(&ctrl->namespaces_rwsem);
195 }
196 
197 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
198 {
199 	struct nvme_ns_head *head = ns->head;
200 	sector_t capacity = get_capacity(head->disk);
201 	int node;
202 	int srcu_idx;
203 
204 	srcu_idx = srcu_read_lock(&head->srcu);
205 	list_for_each_entry_rcu(ns, &head->list, siblings) {
206 		if (capacity != get_capacity(ns->disk))
207 			clear_bit(NVME_NS_READY, &ns->flags);
208 	}
209 	srcu_read_unlock(&head->srcu, srcu_idx);
210 
211 	for_each_node(node)
212 		rcu_assign_pointer(head->current_path[node], NULL);
213 	kblockd_schedule_work(&head->requeue_work);
214 }
215 
216 static bool nvme_path_is_disabled(struct nvme_ns *ns)
217 {
218 	/*
219 	 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
220 	 * still be able to complete assuming that the controller is connected.
221 	 * Otherwise it will fail immediately and return to the requeue list.
222 	 */
223 	if (ns->ctrl->state != NVME_CTRL_LIVE &&
224 	    ns->ctrl->state != NVME_CTRL_DELETING)
225 		return true;
226 	if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
227 	    !test_bit(NVME_NS_READY, &ns->flags))
228 		return true;
229 	return false;
230 }
231 
232 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
233 {
234 	int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
235 	struct nvme_ns *found = NULL, *fallback = NULL, *ns;
236 
237 	list_for_each_entry_rcu(ns, &head->list, siblings) {
238 		if (nvme_path_is_disabled(ns))
239 			continue;
240 
241 		if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
242 			distance = node_distance(node, ns->ctrl->numa_node);
243 		else
244 			distance = LOCAL_DISTANCE;
245 
246 		switch (ns->ana_state) {
247 		case NVME_ANA_OPTIMIZED:
248 			if (distance < found_distance) {
249 				found_distance = distance;
250 				found = ns;
251 			}
252 			break;
253 		case NVME_ANA_NONOPTIMIZED:
254 			if (distance < fallback_distance) {
255 				fallback_distance = distance;
256 				fallback = ns;
257 			}
258 			break;
259 		default:
260 			break;
261 		}
262 	}
263 
264 	if (!found)
265 		found = fallback;
266 	if (found)
267 		rcu_assign_pointer(head->current_path[node], found);
268 	return found;
269 }
270 
271 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
272 		struct nvme_ns *ns)
273 {
274 	ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
275 			siblings);
276 	if (ns)
277 		return ns;
278 	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
279 }
280 
281 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
282 		int node, struct nvme_ns *old)
283 {
284 	struct nvme_ns *ns, *found = NULL;
285 
286 	if (list_is_singular(&head->list)) {
287 		if (nvme_path_is_disabled(old))
288 			return NULL;
289 		return old;
290 	}
291 
292 	for (ns = nvme_next_ns(head, old);
293 	     ns && ns != old;
294 	     ns = nvme_next_ns(head, ns)) {
295 		if (nvme_path_is_disabled(ns))
296 			continue;
297 
298 		if (ns->ana_state == NVME_ANA_OPTIMIZED) {
299 			found = ns;
300 			goto out;
301 		}
302 		if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
303 			found = ns;
304 	}
305 
306 	/*
307 	 * The loop above skips the current path for round-robin semantics.
308 	 * Fall back to the current path if either:
309 	 *  - no other optimized path found and current is optimized,
310 	 *  - no other usable path found and current is usable.
311 	 */
312 	if (!nvme_path_is_disabled(old) &&
313 	    (old->ana_state == NVME_ANA_OPTIMIZED ||
314 	     (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
315 		return old;
316 
317 	if (!found)
318 		return NULL;
319 out:
320 	rcu_assign_pointer(head->current_path[node], found);
321 	return found;
322 }
323 
324 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
325 {
326 	return ns->ctrl->state == NVME_CTRL_LIVE &&
327 		ns->ana_state == NVME_ANA_OPTIMIZED;
328 }
329 
330 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
331 {
332 	int node = numa_node_id();
333 	struct nvme_ns *ns;
334 
335 	ns = srcu_dereference(head->current_path[node], &head->srcu);
336 	if (unlikely(!ns))
337 		return __nvme_find_path(head, node);
338 
339 	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
340 		return nvme_round_robin_path(head, node, ns);
341 	if (unlikely(!nvme_path_is_optimized(ns)))
342 		return __nvme_find_path(head, node);
343 	return ns;
344 }
345 
346 static bool nvme_available_path(struct nvme_ns_head *head)
347 {
348 	struct nvme_ns *ns;
349 
350 	list_for_each_entry_rcu(ns, &head->list, siblings) {
351 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
352 			continue;
353 		switch (ns->ctrl->state) {
354 		case NVME_CTRL_LIVE:
355 		case NVME_CTRL_RESETTING:
356 		case NVME_CTRL_CONNECTING:
357 			/* fallthru */
358 			return true;
359 		default:
360 			break;
361 		}
362 	}
363 	return false;
364 }
365 
366 static void nvme_ns_head_submit_bio(struct bio *bio)
367 {
368 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
369 	struct device *dev = disk_to_dev(head->disk);
370 	struct nvme_ns *ns;
371 	int srcu_idx;
372 
373 	/*
374 	 * The namespace might be going away and the bio might be moved to a
375 	 * different queue via blk_steal_bios(), so we need to use the bio_split
376 	 * pool from the original queue to allocate the bvecs from.
377 	 */
378 	bio = bio_split_to_limits(bio);
379 	if (!bio)
380 		return;
381 
382 	srcu_idx = srcu_read_lock(&head->srcu);
383 	ns = nvme_find_path(head);
384 	if (likely(ns)) {
385 		bio_set_dev(bio, ns->disk->part0);
386 		bio->bi_opf |= REQ_NVME_MPATH;
387 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
388 				      bio->bi_iter.bi_sector);
389 		submit_bio_noacct(bio);
390 	} else if (nvme_available_path(head)) {
391 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
392 
393 		spin_lock_irq(&head->requeue_lock);
394 		bio_list_add(&head->requeue_list, bio);
395 		spin_unlock_irq(&head->requeue_lock);
396 	} else {
397 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
398 
399 		bio_io_error(bio);
400 	}
401 
402 	srcu_read_unlock(&head->srcu, srcu_idx);
403 }
404 
405 static int nvme_ns_head_open(struct gendisk *disk, blk_mode_t mode)
406 {
407 	if (!nvme_tryget_ns_head(disk->private_data))
408 		return -ENXIO;
409 	return 0;
410 }
411 
412 static void nvme_ns_head_release(struct gendisk *disk)
413 {
414 	nvme_put_ns_head(disk->private_data);
415 }
416 
417 #ifdef CONFIG_BLK_DEV_ZONED
418 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
419 		unsigned int nr_zones, report_zones_cb cb, void *data)
420 {
421 	struct nvme_ns_head *head = disk->private_data;
422 	struct nvme_ns *ns;
423 	int srcu_idx, ret = -EWOULDBLOCK;
424 
425 	srcu_idx = srcu_read_lock(&head->srcu);
426 	ns = nvme_find_path(head);
427 	if (ns)
428 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
429 	srcu_read_unlock(&head->srcu, srcu_idx);
430 	return ret;
431 }
432 #else
433 #define nvme_ns_head_report_zones	NULL
434 #endif /* CONFIG_BLK_DEV_ZONED */
435 
436 const struct block_device_operations nvme_ns_head_ops = {
437 	.owner		= THIS_MODULE,
438 	.submit_bio	= nvme_ns_head_submit_bio,
439 	.open		= nvme_ns_head_open,
440 	.release	= nvme_ns_head_release,
441 	.ioctl		= nvme_ns_head_ioctl,
442 	.compat_ioctl	= blkdev_compat_ptr_ioctl,
443 	.getgeo		= nvme_getgeo,
444 	.report_zones	= nvme_ns_head_report_zones,
445 	.pr_ops		= &nvme_pr_ops,
446 };
447 
448 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
449 {
450 	return container_of(cdev, struct nvme_ns_head, cdev);
451 }
452 
453 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
454 {
455 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
456 		return -ENXIO;
457 	return 0;
458 }
459 
460 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
461 {
462 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
463 	return 0;
464 }
465 
466 static const struct file_operations nvme_ns_head_chr_fops = {
467 	.owner		= THIS_MODULE,
468 	.open		= nvme_ns_head_chr_open,
469 	.release	= nvme_ns_head_chr_release,
470 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
471 	.compat_ioctl	= compat_ptr_ioctl,
472 	.uring_cmd	= nvme_ns_head_chr_uring_cmd,
473 	.uring_cmd_iopoll = nvme_ns_head_chr_uring_cmd_iopoll,
474 };
475 
476 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
477 {
478 	int ret;
479 
480 	head->cdev_device.parent = &head->subsys->dev;
481 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
482 			   head->subsys->instance, head->instance);
483 	if (ret)
484 		return ret;
485 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
486 			    &nvme_ns_head_chr_fops, THIS_MODULE);
487 	return ret;
488 }
489 
490 static void nvme_requeue_work(struct work_struct *work)
491 {
492 	struct nvme_ns_head *head =
493 		container_of(work, struct nvme_ns_head, requeue_work);
494 	struct bio *bio, *next;
495 
496 	spin_lock_irq(&head->requeue_lock);
497 	next = bio_list_get(&head->requeue_list);
498 	spin_unlock_irq(&head->requeue_lock);
499 
500 	while ((bio = next) != NULL) {
501 		next = bio->bi_next;
502 		bio->bi_next = NULL;
503 
504 		submit_bio_noacct(bio);
505 	}
506 }
507 
508 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
509 {
510 	bool vwc = false;
511 
512 	mutex_init(&head->lock);
513 	bio_list_init(&head->requeue_list);
514 	spin_lock_init(&head->requeue_lock);
515 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
516 
517 	/*
518 	 * Add a multipath node if the subsystems supports multiple controllers.
519 	 * We also do this for private namespaces as the namespace sharing flag
520 	 * could change after a rescan.
521 	 */
522 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
523 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
524 		return 0;
525 
526 	head->disk = blk_alloc_disk(ctrl->numa_node);
527 	if (!head->disk)
528 		return -ENOMEM;
529 	head->disk->fops = &nvme_ns_head_ops;
530 	head->disk->private_data = head;
531 	sprintf(head->disk->disk_name, "nvme%dn%d",
532 			ctrl->subsys->instance, head->instance);
533 
534 	blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
535 	blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
536 	blk_queue_flag_set(QUEUE_FLAG_IO_STAT, head->disk->queue);
537 	/*
538 	 * This assumes all controllers that refer to a namespace either
539 	 * support poll queues or not.  That is not a strict guarantee,
540 	 * but if the assumption is wrong the effect is only suboptimal
541 	 * performance but not correctness problem.
542 	 */
543 	if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL &&
544 	    ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues)
545 		blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue);
546 
547 	/* set to a default value of 512 until the disk is validated */
548 	blk_queue_logical_block_size(head->disk->queue, 512);
549 	blk_set_stacking_limits(&head->disk->queue->limits);
550 	blk_queue_dma_alignment(head->disk->queue, 3);
551 
552 	/* we need to propagate up the VMC settings */
553 	if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
554 		vwc = true;
555 	blk_queue_write_cache(head->disk->queue, vwc, vwc);
556 	return 0;
557 }
558 
559 static void nvme_mpath_set_live(struct nvme_ns *ns)
560 {
561 	struct nvme_ns_head *head = ns->head;
562 	int rc;
563 
564 	if (!head->disk)
565 		return;
566 
567 	/*
568 	 * test_and_set_bit() is used because it is protecting against two nvme
569 	 * paths simultaneously calling device_add_disk() on the same namespace
570 	 * head.
571 	 */
572 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
573 		rc = device_add_disk(&head->subsys->dev, head->disk,
574 				     nvme_ns_id_attr_groups);
575 		if (rc) {
576 			clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
577 			return;
578 		}
579 		nvme_add_ns_head_cdev(head);
580 	}
581 
582 	mutex_lock(&head->lock);
583 	if (nvme_path_is_optimized(ns)) {
584 		int node, srcu_idx;
585 
586 		srcu_idx = srcu_read_lock(&head->srcu);
587 		for_each_node(node)
588 			__nvme_find_path(head, node);
589 		srcu_read_unlock(&head->srcu, srcu_idx);
590 	}
591 	mutex_unlock(&head->lock);
592 
593 	synchronize_srcu(&head->srcu);
594 	kblockd_schedule_work(&head->requeue_work);
595 }
596 
597 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
598 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
599 			void *))
600 {
601 	void *base = ctrl->ana_log_buf;
602 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
603 	int error, i;
604 
605 	lockdep_assert_held(&ctrl->ana_lock);
606 
607 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
608 		struct nvme_ana_group_desc *desc = base + offset;
609 		u32 nr_nsids;
610 		size_t nsid_buf_size;
611 
612 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
613 			return -EINVAL;
614 
615 		nr_nsids = le32_to_cpu(desc->nnsids);
616 		nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
617 
618 		if (WARN_ON_ONCE(desc->grpid == 0))
619 			return -EINVAL;
620 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
621 			return -EINVAL;
622 		if (WARN_ON_ONCE(desc->state == 0))
623 			return -EINVAL;
624 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
625 			return -EINVAL;
626 
627 		offset += sizeof(*desc);
628 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
629 			return -EINVAL;
630 
631 		error = cb(ctrl, desc, data);
632 		if (error)
633 			return error;
634 
635 		offset += nsid_buf_size;
636 	}
637 
638 	return 0;
639 }
640 
641 static inline bool nvme_state_is_live(enum nvme_ana_state state)
642 {
643 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
644 }
645 
646 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
647 		struct nvme_ns *ns)
648 {
649 	ns->ana_grpid = le32_to_cpu(desc->grpid);
650 	ns->ana_state = desc->state;
651 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
652 	/*
653 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
654 	 * and in turn to this path device.  However we cannot accept this I/O
655 	 * if the controller is not live.  This may deadlock if called from
656 	 * nvme_mpath_init_identify() and the ctrl will never complete
657 	 * initialization, preventing I/O from completing.  For this case we
658 	 * will reprocess the ANA log page in nvme_mpath_update() once the
659 	 * controller is ready.
660 	 */
661 	if (nvme_state_is_live(ns->ana_state) &&
662 	    ns->ctrl->state == NVME_CTRL_LIVE)
663 		nvme_mpath_set_live(ns);
664 }
665 
666 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
667 		struct nvme_ana_group_desc *desc, void *data)
668 {
669 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
670 	unsigned *nr_change_groups = data;
671 	struct nvme_ns *ns;
672 
673 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
674 			le32_to_cpu(desc->grpid),
675 			nvme_ana_state_names[desc->state]);
676 
677 	if (desc->state == NVME_ANA_CHANGE)
678 		(*nr_change_groups)++;
679 
680 	if (!nr_nsids)
681 		return 0;
682 
683 	down_read(&ctrl->namespaces_rwsem);
684 	list_for_each_entry(ns, &ctrl->namespaces, list) {
685 		unsigned nsid;
686 again:
687 		nsid = le32_to_cpu(desc->nsids[n]);
688 		if (ns->head->ns_id < nsid)
689 			continue;
690 		if (ns->head->ns_id == nsid)
691 			nvme_update_ns_ana_state(desc, ns);
692 		if (++n == nr_nsids)
693 			break;
694 		if (ns->head->ns_id > nsid)
695 			goto again;
696 	}
697 	up_read(&ctrl->namespaces_rwsem);
698 	return 0;
699 }
700 
701 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
702 {
703 	u32 nr_change_groups = 0;
704 	int error;
705 
706 	mutex_lock(&ctrl->ana_lock);
707 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
708 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
709 	if (error) {
710 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
711 		goto out_unlock;
712 	}
713 
714 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
715 			nvme_update_ana_state);
716 	if (error)
717 		goto out_unlock;
718 
719 	/*
720 	 * In theory we should have an ANATT timer per group as they might enter
721 	 * the change state at different times.  But that is a lot of overhead
722 	 * just to protect against a target that keeps entering new changes
723 	 * states while never finishing previous ones.  But we'll still
724 	 * eventually time out once all groups are in change state, so this
725 	 * isn't a big deal.
726 	 *
727 	 * We also double the ANATT value to provide some slack for transports
728 	 * or AEN processing overhead.
729 	 */
730 	if (nr_change_groups)
731 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
732 	else
733 		del_timer_sync(&ctrl->anatt_timer);
734 out_unlock:
735 	mutex_unlock(&ctrl->ana_lock);
736 	return error;
737 }
738 
739 static void nvme_ana_work(struct work_struct *work)
740 {
741 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
742 
743 	if (ctrl->state != NVME_CTRL_LIVE)
744 		return;
745 
746 	nvme_read_ana_log(ctrl);
747 }
748 
749 void nvme_mpath_update(struct nvme_ctrl *ctrl)
750 {
751 	u32 nr_change_groups = 0;
752 
753 	if (!ctrl->ana_log_buf)
754 		return;
755 
756 	mutex_lock(&ctrl->ana_lock);
757 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
758 	mutex_unlock(&ctrl->ana_lock);
759 }
760 
761 static void nvme_anatt_timeout(struct timer_list *t)
762 {
763 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
764 
765 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
766 	nvme_reset_ctrl(ctrl);
767 }
768 
769 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
770 {
771 	if (!nvme_ctrl_use_ana(ctrl))
772 		return;
773 	del_timer_sync(&ctrl->anatt_timer);
774 	cancel_work_sync(&ctrl->ana_work);
775 }
776 
777 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
778 	struct device_attribute subsys_attr_##_name =	\
779 		__ATTR(_name, _mode, _show, _store)
780 
781 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
782 		struct device_attribute *attr, char *buf)
783 {
784 	struct nvme_subsystem *subsys =
785 		container_of(dev, struct nvme_subsystem, dev);
786 
787 	return sysfs_emit(buf, "%s\n",
788 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
789 }
790 
791 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
792 		struct device_attribute *attr, const char *buf, size_t count)
793 {
794 	struct nvme_subsystem *subsys =
795 		container_of(dev, struct nvme_subsystem, dev);
796 	int i;
797 
798 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
799 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
800 			WRITE_ONCE(subsys->iopolicy, i);
801 			return count;
802 		}
803 	}
804 
805 	return -EINVAL;
806 }
807 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
808 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
809 
810 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
811 		char *buf)
812 {
813 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
814 }
815 DEVICE_ATTR_RO(ana_grpid);
816 
817 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
818 		char *buf)
819 {
820 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
821 
822 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
823 }
824 DEVICE_ATTR_RO(ana_state);
825 
826 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
827 		struct nvme_ana_group_desc *desc, void *data)
828 {
829 	struct nvme_ana_group_desc *dst = data;
830 
831 	if (desc->grpid != dst->grpid)
832 		return 0;
833 
834 	*dst = *desc;
835 	return -ENXIO; /* just break out of the loop */
836 }
837 
838 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
839 {
840 	if (nvme_ctrl_use_ana(ns->ctrl)) {
841 		struct nvme_ana_group_desc desc = {
842 			.grpid = anagrpid,
843 			.state = 0,
844 		};
845 
846 		mutex_lock(&ns->ctrl->ana_lock);
847 		ns->ana_grpid = le32_to_cpu(anagrpid);
848 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
849 		mutex_unlock(&ns->ctrl->ana_lock);
850 		if (desc.state) {
851 			/* found the group desc: update */
852 			nvme_update_ns_ana_state(&desc, ns);
853 		} else {
854 			/* group desc not found: trigger a re-read */
855 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
856 			queue_work(nvme_wq, &ns->ctrl->ana_work);
857 		}
858 	} else {
859 		ns->ana_state = NVME_ANA_OPTIMIZED;
860 		nvme_mpath_set_live(ns);
861 	}
862 
863 	if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
864 		blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
865 				   ns->head->disk->queue);
866 #ifdef CONFIG_BLK_DEV_ZONED
867 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
868 		ns->head->disk->nr_zones = ns->disk->nr_zones;
869 #endif
870 }
871 
872 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
873 {
874 	if (!head->disk)
875 		return;
876 	kblockd_schedule_work(&head->requeue_work);
877 	if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
878 		nvme_cdev_del(&head->cdev, &head->cdev_device);
879 		del_gendisk(head->disk);
880 	}
881 }
882 
883 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
884 {
885 	if (!head->disk)
886 		return;
887 	/* make sure all pending bios are cleaned up */
888 	kblockd_schedule_work(&head->requeue_work);
889 	flush_work(&head->requeue_work);
890 	put_disk(head->disk);
891 }
892 
893 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
894 {
895 	mutex_init(&ctrl->ana_lock);
896 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
897 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
898 }
899 
900 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
901 {
902 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
903 	size_t ana_log_size;
904 	int error = 0;
905 
906 	/* check if multipath is enabled and we have the capability */
907 	if (!multipath || !ctrl->subsys ||
908 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
909 		return 0;
910 
911 	if (!ctrl->max_namespaces ||
912 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
913 		dev_err(ctrl->device,
914 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
915 		return -EINVAL;
916 	}
917 
918 	ctrl->anacap = id->anacap;
919 	ctrl->anatt = id->anatt;
920 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
921 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
922 
923 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
924 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
925 		ctrl->max_namespaces * sizeof(__le32);
926 	if (ana_log_size > max_transfer_size) {
927 		dev_err(ctrl->device,
928 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
929 			ana_log_size, max_transfer_size);
930 		dev_err(ctrl->device, "disabling ANA support.\n");
931 		goto out_uninit;
932 	}
933 	if (ana_log_size > ctrl->ana_log_size) {
934 		nvme_mpath_stop(ctrl);
935 		nvme_mpath_uninit(ctrl);
936 		ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
937 		if (!ctrl->ana_log_buf)
938 			return -ENOMEM;
939 	}
940 	ctrl->ana_log_size = ana_log_size;
941 	error = nvme_read_ana_log(ctrl);
942 	if (error)
943 		goto out_uninit;
944 	return 0;
945 
946 out_uninit:
947 	nvme_mpath_uninit(ctrl);
948 	return error;
949 }
950 
951 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
952 {
953 	kvfree(ctrl->ana_log_buf);
954 	ctrl->ana_log_buf = NULL;
955 	ctrl->ana_log_size = 0;
956 }
957