xref: /linux/drivers/nvme/host/multipath.c (revision eb01fe7abbe2d0b38824d2a93fdb4cc3eaf2ccc1)
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 		 * The alternate request queue that we may end up submitting
111 		 * the bio to may be frozen temporarily, in this case REQ_NOWAIT
112 		 * will fail the I/O immediately with EAGAIN to the issuer.
113 		 * We are not in the issuer context which cannot block. Clear
114 		 * the flag to avoid spurious EAGAIN I/O failures.
115 		 */
116 		bio->bi_opf &= ~REQ_NOWAIT;
117 	}
118 	blk_steal_bios(&ns->head->requeue_list, req);
119 	spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
120 
121 	blk_mq_end_request(req, 0);
122 	kblockd_schedule_work(&ns->head->requeue_work);
123 }
124 
125 void nvme_mpath_start_request(struct request *rq)
126 {
127 	struct nvme_ns *ns = rq->q->queuedata;
128 	struct gendisk *disk = ns->head->disk;
129 
130 	if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq))
131 		return;
132 
133 	nvme_req(rq)->flags |= NVME_MPATH_IO_STATS;
134 	nvme_req(rq)->start_time = bdev_start_io_acct(disk->part0, req_op(rq),
135 						      jiffies);
136 }
137 EXPORT_SYMBOL_GPL(nvme_mpath_start_request);
138 
139 void nvme_mpath_end_request(struct request *rq)
140 {
141 	struct nvme_ns *ns = rq->q->queuedata;
142 
143 	if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
144 		return;
145 	bdev_end_io_acct(ns->head->disk->part0, req_op(rq),
146 			 blk_rq_bytes(rq) >> SECTOR_SHIFT,
147 			 nvme_req(rq)->start_time);
148 }
149 
150 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
151 {
152 	struct nvme_ns *ns;
153 
154 	down_read(&ctrl->namespaces_rwsem);
155 	list_for_each_entry(ns, &ctrl->namespaces, list) {
156 		if (!ns->head->disk)
157 			continue;
158 		kblockd_schedule_work(&ns->head->requeue_work);
159 		if (nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
160 			disk_uevent(ns->head->disk, KOBJ_CHANGE);
161 	}
162 	up_read(&ctrl->namespaces_rwsem);
163 }
164 
165 static const char *nvme_ana_state_names[] = {
166 	[0]				= "invalid state",
167 	[NVME_ANA_OPTIMIZED]		= "optimized",
168 	[NVME_ANA_NONOPTIMIZED]		= "non-optimized",
169 	[NVME_ANA_INACCESSIBLE]		= "inaccessible",
170 	[NVME_ANA_PERSISTENT_LOSS]	= "persistent-loss",
171 	[NVME_ANA_CHANGE]		= "change",
172 };
173 
174 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
175 {
176 	struct nvme_ns_head *head = ns->head;
177 	bool changed = false;
178 	int node;
179 
180 	if (!head)
181 		goto out;
182 
183 	for_each_node(node) {
184 		if (ns == rcu_access_pointer(head->current_path[node])) {
185 			rcu_assign_pointer(head->current_path[node], NULL);
186 			changed = true;
187 		}
188 	}
189 out:
190 	return changed;
191 }
192 
193 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
194 {
195 	struct nvme_ns *ns;
196 
197 	down_read(&ctrl->namespaces_rwsem);
198 	list_for_each_entry(ns, &ctrl->namespaces, list) {
199 		nvme_mpath_clear_current_path(ns);
200 		kblockd_schedule_work(&ns->head->requeue_work);
201 	}
202 	up_read(&ctrl->namespaces_rwsem);
203 }
204 
205 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
206 {
207 	struct nvme_ns_head *head = ns->head;
208 	sector_t capacity = get_capacity(head->disk);
209 	int node;
210 	int srcu_idx;
211 
212 	srcu_idx = srcu_read_lock(&head->srcu);
213 	list_for_each_entry_rcu(ns, &head->list, siblings) {
214 		if (capacity != get_capacity(ns->disk))
215 			clear_bit(NVME_NS_READY, &ns->flags);
216 	}
217 	srcu_read_unlock(&head->srcu, srcu_idx);
218 
219 	for_each_node(node)
220 		rcu_assign_pointer(head->current_path[node], NULL);
221 	kblockd_schedule_work(&head->requeue_work);
222 }
223 
224 static bool nvme_path_is_disabled(struct nvme_ns *ns)
225 {
226 	enum nvme_ctrl_state state = nvme_ctrl_state(ns->ctrl);
227 
228 	/*
229 	 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
230 	 * still be able to complete assuming that the controller is connected.
231 	 * Otherwise it will fail immediately and return to the requeue list.
232 	 */
233 	if (state != NVME_CTRL_LIVE && state != NVME_CTRL_DELETING)
234 		return true;
235 	if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
236 	    !test_bit(NVME_NS_READY, &ns->flags))
237 		return true;
238 	return false;
239 }
240 
241 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
242 {
243 	int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
244 	struct nvme_ns *found = NULL, *fallback = NULL, *ns;
245 
246 	list_for_each_entry_rcu(ns, &head->list, siblings) {
247 		if (nvme_path_is_disabled(ns))
248 			continue;
249 
250 		if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
251 			distance = node_distance(node, ns->ctrl->numa_node);
252 		else
253 			distance = LOCAL_DISTANCE;
254 
255 		switch (ns->ana_state) {
256 		case NVME_ANA_OPTIMIZED:
257 			if (distance < found_distance) {
258 				found_distance = distance;
259 				found = ns;
260 			}
261 			break;
262 		case NVME_ANA_NONOPTIMIZED:
263 			if (distance < fallback_distance) {
264 				fallback_distance = distance;
265 				fallback = ns;
266 			}
267 			break;
268 		default:
269 			break;
270 		}
271 	}
272 
273 	if (!found)
274 		found = fallback;
275 	if (found)
276 		rcu_assign_pointer(head->current_path[node], found);
277 	return found;
278 }
279 
280 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
281 		struct nvme_ns *ns)
282 {
283 	ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
284 			siblings);
285 	if (ns)
286 		return ns;
287 	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
288 }
289 
290 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
291 		int node, struct nvme_ns *old)
292 {
293 	struct nvme_ns *ns, *found = NULL;
294 
295 	if (list_is_singular(&head->list)) {
296 		if (nvme_path_is_disabled(old))
297 			return NULL;
298 		return old;
299 	}
300 
301 	for (ns = nvme_next_ns(head, old);
302 	     ns && ns != old;
303 	     ns = nvme_next_ns(head, ns)) {
304 		if (nvme_path_is_disabled(ns))
305 			continue;
306 
307 		if (ns->ana_state == NVME_ANA_OPTIMIZED) {
308 			found = ns;
309 			goto out;
310 		}
311 		if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
312 			found = ns;
313 	}
314 
315 	/*
316 	 * The loop above skips the current path for round-robin semantics.
317 	 * Fall back to the current path if either:
318 	 *  - no other optimized path found and current is optimized,
319 	 *  - no other usable path found and current is usable.
320 	 */
321 	if (!nvme_path_is_disabled(old) &&
322 	    (old->ana_state == NVME_ANA_OPTIMIZED ||
323 	     (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
324 		return old;
325 
326 	if (!found)
327 		return NULL;
328 out:
329 	rcu_assign_pointer(head->current_path[node], found);
330 	return found;
331 }
332 
333 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
334 {
335 	return nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE &&
336 		ns->ana_state == NVME_ANA_OPTIMIZED;
337 }
338 
339 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
340 {
341 	int node = numa_node_id();
342 	struct nvme_ns *ns;
343 
344 	ns = srcu_dereference(head->current_path[node], &head->srcu);
345 	if (unlikely(!ns))
346 		return __nvme_find_path(head, node);
347 
348 	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
349 		return nvme_round_robin_path(head, node, ns);
350 	if (unlikely(!nvme_path_is_optimized(ns)))
351 		return __nvme_find_path(head, node);
352 	return ns;
353 }
354 
355 static bool nvme_available_path(struct nvme_ns_head *head)
356 {
357 	struct nvme_ns *ns;
358 
359 	list_for_each_entry_rcu(ns, &head->list, siblings) {
360 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
361 			continue;
362 		switch (nvme_ctrl_state(ns->ctrl)) {
363 		case NVME_CTRL_LIVE:
364 		case NVME_CTRL_RESETTING:
365 		case NVME_CTRL_CONNECTING:
366 			/* fallthru */
367 			return true;
368 		default:
369 			break;
370 		}
371 	}
372 	return false;
373 }
374 
375 static void nvme_ns_head_submit_bio(struct bio *bio)
376 {
377 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
378 	struct device *dev = disk_to_dev(head->disk);
379 	struct nvme_ns *ns;
380 	int srcu_idx;
381 
382 	/*
383 	 * The namespace might be going away and the bio might be moved to a
384 	 * different queue via blk_steal_bios(), so we need to use the bio_split
385 	 * pool from the original queue to allocate the bvecs from.
386 	 */
387 	bio = bio_split_to_limits(bio);
388 	if (!bio)
389 		return;
390 
391 	srcu_idx = srcu_read_lock(&head->srcu);
392 	ns = nvme_find_path(head);
393 	if (likely(ns)) {
394 		bio_set_dev(bio, ns->disk->part0);
395 		bio->bi_opf |= REQ_NVME_MPATH;
396 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
397 				      bio->bi_iter.bi_sector);
398 		submit_bio_noacct(bio);
399 	} else if (nvme_available_path(head)) {
400 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
401 
402 		spin_lock_irq(&head->requeue_lock);
403 		bio_list_add(&head->requeue_list, bio);
404 		spin_unlock_irq(&head->requeue_lock);
405 	} else {
406 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
407 
408 		bio_io_error(bio);
409 	}
410 
411 	srcu_read_unlock(&head->srcu, srcu_idx);
412 }
413 
414 static int nvme_ns_head_open(struct gendisk *disk, blk_mode_t mode)
415 {
416 	if (!nvme_tryget_ns_head(disk->private_data))
417 		return -ENXIO;
418 	return 0;
419 }
420 
421 static void nvme_ns_head_release(struct gendisk *disk)
422 {
423 	nvme_put_ns_head(disk->private_data);
424 }
425 
426 #ifdef CONFIG_BLK_DEV_ZONED
427 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
428 		unsigned int nr_zones, report_zones_cb cb, void *data)
429 {
430 	struct nvme_ns_head *head = disk->private_data;
431 	struct nvme_ns *ns;
432 	int srcu_idx, ret = -EWOULDBLOCK;
433 
434 	srcu_idx = srcu_read_lock(&head->srcu);
435 	ns = nvme_find_path(head);
436 	if (ns)
437 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
438 	srcu_read_unlock(&head->srcu, srcu_idx);
439 	return ret;
440 }
441 #else
442 #define nvme_ns_head_report_zones	NULL
443 #endif /* CONFIG_BLK_DEV_ZONED */
444 
445 const struct block_device_operations nvme_ns_head_ops = {
446 	.owner		= THIS_MODULE,
447 	.submit_bio	= nvme_ns_head_submit_bio,
448 	.open		= nvme_ns_head_open,
449 	.release	= nvme_ns_head_release,
450 	.ioctl		= nvme_ns_head_ioctl,
451 	.compat_ioctl	= blkdev_compat_ptr_ioctl,
452 	.getgeo		= nvme_getgeo,
453 	.report_zones	= nvme_ns_head_report_zones,
454 	.pr_ops		= &nvme_pr_ops,
455 };
456 
457 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
458 {
459 	return container_of(cdev, struct nvme_ns_head, cdev);
460 }
461 
462 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
463 {
464 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
465 		return -ENXIO;
466 	return 0;
467 }
468 
469 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
470 {
471 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
472 	return 0;
473 }
474 
475 static const struct file_operations nvme_ns_head_chr_fops = {
476 	.owner		= THIS_MODULE,
477 	.open		= nvme_ns_head_chr_open,
478 	.release	= nvme_ns_head_chr_release,
479 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
480 	.compat_ioctl	= compat_ptr_ioctl,
481 	.uring_cmd	= nvme_ns_head_chr_uring_cmd,
482 	.uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
483 };
484 
485 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
486 {
487 	int ret;
488 
489 	head->cdev_device.parent = &head->subsys->dev;
490 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
491 			   head->subsys->instance, head->instance);
492 	if (ret)
493 		return ret;
494 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
495 			    &nvme_ns_head_chr_fops, THIS_MODULE);
496 	return ret;
497 }
498 
499 static void nvme_requeue_work(struct work_struct *work)
500 {
501 	struct nvme_ns_head *head =
502 		container_of(work, struct nvme_ns_head, requeue_work);
503 	struct bio *bio, *next;
504 
505 	spin_lock_irq(&head->requeue_lock);
506 	next = bio_list_get(&head->requeue_list);
507 	spin_unlock_irq(&head->requeue_lock);
508 
509 	while ((bio = next) != NULL) {
510 		next = bio->bi_next;
511 		bio->bi_next = NULL;
512 
513 		submit_bio_noacct(bio);
514 	}
515 }
516 
517 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
518 {
519 	struct queue_limits lim;
520 	bool vwc = false;
521 
522 	mutex_init(&head->lock);
523 	bio_list_init(&head->requeue_list);
524 	spin_lock_init(&head->requeue_lock);
525 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
526 
527 	/*
528 	 * Add a multipath node if the subsystems supports multiple controllers.
529 	 * We also do this for private namespaces as the namespace sharing flag
530 	 * could change after a rescan.
531 	 */
532 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
533 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
534 		return 0;
535 
536 	blk_set_stacking_limits(&lim);
537 	lim.dma_alignment = 3;
538 	if (head->ids.csi != NVME_CSI_ZNS)
539 		lim.max_zone_append_sectors = 0;
540 
541 	head->disk = blk_alloc_disk(&lim, ctrl->numa_node);
542 	if (IS_ERR(head->disk))
543 		return PTR_ERR(head->disk);
544 	head->disk->fops = &nvme_ns_head_ops;
545 	head->disk->private_data = head;
546 	sprintf(head->disk->disk_name, "nvme%dn%d",
547 			ctrl->subsys->instance, head->instance);
548 
549 	blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
550 	blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
551 	blk_queue_flag_set(QUEUE_FLAG_IO_STAT, head->disk->queue);
552 	/*
553 	 * This assumes all controllers that refer to a namespace either
554 	 * support poll queues or not.  That is not a strict guarantee,
555 	 * but if the assumption is wrong the effect is only suboptimal
556 	 * performance but not correctness problem.
557 	 */
558 	if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL &&
559 	    ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues)
560 		blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue);
561 
562 	/* we need to propagate up the VMC settings */
563 	if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
564 		vwc = true;
565 	blk_queue_write_cache(head->disk->queue, vwc, vwc);
566 	return 0;
567 }
568 
569 static void nvme_mpath_set_live(struct nvme_ns *ns)
570 {
571 	struct nvme_ns_head *head = ns->head;
572 	int rc;
573 
574 	if (!head->disk)
575 		return;
576 
577 	/*
578 	 * test_and_set_bit() is used because it is protecting against two nvme
579 	 * paths simultaneously calling device_add_disk() on the same namespace
580 	 * head.
581 	 */
582 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
583 		rc = device_add_disk(&head->subsys->dev, head->disk,
584 				     nvme_ns_attr_groups);
585 		if (rc) {
586 			clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
587 			return;
588 		}
589 		nvme_add_ns_head_cdev(head);
590 	}
591 
592 	mutex_lock(&head->lock);
593 	if (nvme_path_is_optimized(ns)) {
594 		int node, srcu_idx;
595 
596 		srcu_idx = srcu_read_lock(&head->srcu);
597 		for_each_node(node)
598 			__nvme_find_path(head, node);
599 		srcu_read_unlock(&head->srcu, srcu_idx);
600 	}
601 	mutex_unlock(&head->lock);
602 
603 	synchronize_srcu(&head->srcu);
604 	kblockd_schedule_work(&head->requeue_work);
605 }
606 
607 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
608 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
609 			void *))
610 {
611 	void *base = ctrl->ana_log_buf;
612 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
613 	int error, i;
614 
615 	lockdep_assert_held(&ctrl->ana_lock);
616 
617 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
618 		struct nvme_ana_group_desc *desc = base + offset;
619 		u32 nr_nsids;
620 		size_t nsid_buf_size;
621 
622 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
623 			return -EINVAL;
624 
625 		nr_nsids = le32_to_cpu(desc->nnsids);
626 		nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
627 
628 		if (WARN_ON_ONCE(desc->grpid == 0))
629 			return -EINVAL;
630 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
631 			return -EINVAL;
632 		if (WARN_ON_ONCE(desc->state == 0))
633 			return -EINVAL;
634 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
635 			return -EINVAL;
636 
637 		offset += sizeof(*desc);
638 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
639 			return -EINVAL;
640 
641 		error = cb(ctrl, desc, data);
642 		if (error)
643 			return error;
644 
645 		offset += nsid_buf_size;
646 	}
647 
648 	return 0;
649 }
650 
651 static inline bool nvme_state_is_live(enum nvme_ana_state state)
652 {
653 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
654 }
655 
656 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
657 		struct nvme_ns *ns)
658 {
659 	ns->ana_grpid = le32_to_cpu(desc->grpid);
660 	ns->ana_state = desc->state;
661 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
662 	/*
663 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
664 	 * and in turn to this path device.  However we cannot accept this I/O
665 	 * if the controller is not live.  This may deadlock if called from
666 	 * nvme_mpath_init_identify() and the ctrl will never complete
667 	 * initialization, preventing I/O from completing.  For this case we
668 	 * will reprocess the ANA log page in nvme_mpath_update() once the
669 	 * controller is ready.
670 	 */
671 	if (nvme_state_is_live(ns->ana_state) &&
672 	    nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
673 		nvme_mpath_set_live(ns);
674 }
675 
676 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
677 		struct nvme_ana_group_desc *desc, void *data)
678 {
679 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
680 	unsigned *nr_change_groups = data;
681 	struct nvme_ns *ns;
682 
683 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
684 			le32_to_cpu(desc->grpid),
685 			nvme_ana_state_names[desc->state]);
686 
687 	if (desc->state == NVME_ANA_CHANGE)
688 		(*nr_change_groups)++;
689 
690 	if (!nr_nsids)
691 		return 0;
692 
693 	down_read(&ctrl->namespaces_rwsem);
694 	list_for_each_entry(ns, &ctrl->namespaces, list) {
695 		unsigned nsid;
696 again:
697 		nsid = le32_to_cpu(desc->nsids[n]);
698 		if (ns->head->ns_id < nsid)
699 			continue;
700 		if (ns->head->ns_id == nsid)
701 			nvme_update_ns_ana_state(desc, ns);
702 		if (++n == nr_nsids)
703 			break;
704 		if (ns->head->ns_id > nsid)
705 			goto again;
706 	}
707 	up_read(&ctrl->namespaces_rwsem);
708 	return 0;
709 }
710 
711 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
712 {
713 	u32 nr_change_groups = 0;
714 	int error;
715 
716 	mutex_lock(&ctrl->ana_lock);
717 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
718 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
719 	if (error) {
720 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
721 		goto out_unlock;
722 	}
723 
724 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
725 			nvme_update_ana_state);
726 	if (error)
727 		goto out_unlock;
728 
729 	/*
730 	 * In theory we should have an ANATT timer per group as they might enter
731 	 * the change state at different times.  But that is a lot of overhead
732 	 * just to protect against a target that keeps entering new changes
733 	 * states while never finishing previous ones.  But we'll still
734 	 * eventually time out once all groups are in change state, so this
735 	 * isn't a big deal.
736 	 *
737 	 * We also double the ANATT value to provide some slack for transports
738 	 * or AEN processing overhead.
739 	 */
740 	if (nr_change_groups)
741 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
742 	else
743 		del_timer_sync(&ctrl->anatt_timer);
744 out_unlock:
745 	mutex_unlock(&ctrl->ana_lock);
746 	return error;
747 }
748 
749 static void nvme_ana_work(struct work_struct *work)
750 {
751 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
752 
753 	if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
754 		return;
755 
756 	nvme_read_ana_log(ctrl);
757 }
758 
759 void nvme_mpath_update(struct nvme_ctrl *ctrl)
760 {
761 	u32 nr_change_groups = 0;
762 
763 	if (!ctrl->ana_log_buf)
764 		return;
765 
766 	mutex_lock(&ctrl->ana_lock);
767 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
768 	mutex_unlock(&ctrl->ana_lock);
769 }
770 
771 static void nvme_anatt_timeout(struct timer_list *t)
772 {
773 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
774 
775 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
776 	nvme_reset_ctrl(ctrl);
777 }
778 
779 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
780 {
781 	if (!nvme_ctrl_use_ana(ctrl))
782 		return;
783 	del_timer_sync(&ctrl->anatt_timer);
784 	cancel_work_sync(&ctrl->ana_work);
785 }
786 
787 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
788 	struct device_attribute subsys_attr_##_name =	\
789 		__ATTR(_name, _mode, _show, _store)
790 
791 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
792 		struct device_attribute *attr, char *buf)
793 {
794 	struct nvme_subsystem *subsys =
795 		container_of(dev, struct nvme_subsystem, dev);
796 
797 	return sysfs_emit(buf, "%s\n",
798 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
799 }
800 
801 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
802 		struct device_attribute *attr, const char *buf, size_t count)
803 {
804 	struct nvme_subsystem *subsys =
805 		container_of(dev, struct nvme_subsystem, dev);
806 	int i;
807 
808 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
809 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
810 			WRITE_ONCE(subsys->iopolicy, i);
811 			return count;
812 		}
813 	}
814 
815 	return -EINVAL;
816 }
817 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
818 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
819 
820 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
821 		char *buf)
822 {
823 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
824 }
825 DEVICE_ATTR_RO(ana_grpid);
826 
827 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
828 		char *buf)
829 {
830 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
831 
832 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
833 }
834 DEVICE_ATTR_RO(ana_state);
835 
836 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
837 		struct nvme_ana_group_desc *desc, void *data)
838 {
839 	struct nvme_ana_group_desc *dst = data;
840 
841 	if (desc->grpid != dst->grpid)
842 		return 0;
843 
844 	*dst = *desc;
845 	return -ENXIO; /* just break out of the loop */
846 }
847 
848 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
849 {
850 	if (nvme_ctrl_use_ana(ns->ctrl)) {
851 		struct nvme_ana_group_desc desc = {
852 			.grpid = anagrpid,
853 			.state = 0,
854 		};
855 
856 		mutex_lock(&ns->ctrl->ana_lock);
857 		ns->ana_grpid = le32_to_cpu(anagrpid);
858 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
859 		mutex_unlock(&ns->ctrl->ana_lock);
860 		if (desc.state) {
861 			/* found the group desc: update */
862 			nvme_update_ns_ana_state(&desc, ns);
863 		} else {
864 			/* group desc not found: trigger a re-read */
865 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
866 			queue_work(nvme_wq, &ns->ctrl->ana_work);
867 		}
868 	} else {
869 		ns->ana_state = NVME_ANA_OPTIMIZED;
870 		nvme_mpath_set_live(ns);
871 	}
872 
873 	if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
874 		blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
875 				   ns->head->disk->queue);
876 #ifdef CONFIG_BLK_DEV_ZONED
877 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
878 		ns->head->disk->nr_zones = ns->disk->nr_zones;
879 #endif
880 }
881 
882 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
883 {
884 	if (!head->disk)
885 		return;
886 	kblockd_schedule_work(&head->requeue_work);
887 	if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
888 		nvme_cdev_del(&head->cdev, &head->cdev_device);
889 		del_gendisk(head->disk);
890 	}
891 }
892 
893 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
894 {
895 	if (!head->disk)
896 		return;
897 	/* make sure all pending bios are cleaned up */
898 	kblockd_schedule_work(&head->requeue_work);
899 	flush_work(&head->requeue_work);
900 	put_disk(head->disk);
901 }
902 
903 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
904 {
905 	mutex_init(&ctrl->ana_lock);
906 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
907 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
908 }
909 
910 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
911 {
912 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
913 	size_t ana_log_size;
914 	int error = 0;
915 
916 	/* check if multipath is enabled and we have the capability */
917 	if (!multipath || !ctrl->subsys ||
918 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
919 		return 0;
920 
921 	if (!ctrl->max_namespaces ||
922 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
923 		dev_err(ctrl->device,
924 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
925 		return -EINVAL;
926 	}
927 
928 	ctrl->anacap = id->anacap;
929 	ctrl->anatt = id->anatt;
930 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
931 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
932 
933 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
934 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
935 		ctrl->max_namespaces * sizeof(__le32);
936 	if (ana_log_size > max_transfer_size) {
937 		dev_err(ctrl->device,
938 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
939 			ana_log_size, max_transfer_size);
940 		dev_err(ctrl->device, "disabling ANA support.\n");
941 		goto out_uninit;
942 	}
943 	if (ana_log_size > ctrl->ana_log_size) {
944 		nvme_mpath_stop(ctrl);
945 		nvme_mpath_uninit(ctrl);
946 		ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
947 		if (!ctrl->ana_log_buf)
948 			return -ENOMEM;
949 	}
950 	ctrl->ana_log_size = ana_log_size;
951 	error = nvme_read_ana_log(ctrl);
952 	if (error)
953 		goto out_uninit;
954 	return 0;
955 
956 out_uninit:
957 	nvme_mpath_uninit(ctrl);
958 	return error;
959 }
960 
961 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
962 {
963 	kvfree(ctrl->ana_log_buf);
964 	ctrl->ana_log_buf = NULL;
965 	ctrl->ana_log_size = 0;
966 }
967