xref: /linux/drivers/nvme/host/multipath.c (revision 5ddb88f22eb97218d9295e69c39e0ff7cc64e09c)
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 	nvme_req(req)->status = 0;
122 	nvme_end_req(req);
123 	kblockd_schedule_work(&ns->head->requeue_work);
124 }
125 
126 void nvme_mpath_start_request(struct request *rq)
127 {
128 	struct nvme_ns *ns = rq->q->queuedata;
129 	struct gendisk *disk = ns->head->disk;
130 
131 	if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq))
132 		return;
133 
134 	nvme_req(rq)->flags |= NVME_MPATH_IO_STATS;
135 	nvme_req(rq)->start_time = bdev_start_io_acct(disk->part0, req_op(rq),
136 						      jiffies);
137 }
138 EXPORT_SYMBOL_GPL(nvme_mpath_start_request);
139 
140 void nvme_mpath_end_request(struct request *rq)
141 {
142 	struct nvme_ns *ns = rq->q->queuedata;
143 
144 	if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
145 		return;
146 	bdev_end_io_acct(ns->head->disk->part0, req_op(rq),
147 			 blk_rq_bytes(rq) >> SECTOR_SHIFT,
148 			 nvme_req(rq)->start_time);
149 }
150 
151 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
152 {
153 	struct nvme_ns *ns;
154 	int srcu_idx;
155 
156 	srcu_idx = srcu_read_lock(&ctrl->srcu);
157 	list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
158 		if (!ns->head->disk)
159 			continue;
160 		kblockd_schedule_work(&ns->head->requeue_work);
161 		if (nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
162 			disk_uevent(ns->head->disk, KOBJ_CHANGE);
163 	}
164 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
165 }
166 
167 static const char *nvme_ana_state_names[] = {
168 	[0]				= "invalid state",
169 	[NVME_ANA_OPTIMIZED]		= "optimized",
170 	[NVME_ANA_NONOPTIMIZED]		= "non-optimized",
171 	[NVME_ANA_INACCESSIBLE]		= "inaccessible",
172 	[NVME_ANA_PERSISTENT_LOSS]	= "persistent-loss",
173 	[NVME_ANA_CHANGE]		= "change",
174 };
175 
176 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
177 {
178 	struct nvme_ns_head *head = ns->head;
179 	bool changed = false;
180 	int node;
181 
182 	if (!head)
183 		goto out;
184 
185 	for_each_node(node) {
186 		if (ns == rcu_access_pointer(head->current_path[node])) {
187 			rcu_assign_pointer(head->current_path[node], NULL);
188 			changed = true;
189 		}
190 	}
191 out:
192 	return changed;
193 }
194 
195 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
196 {
197 	struct nvme_ns *ns;
198 	int srcu_idx;
199 
200 	srcu_idx = srcu_read_lock(&ctrl->srcu);
201 	list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
202 		nvme_mpath_clear_current_path(ns);
203 		kblockd_schedule_work(&ns->head->requeue_work);
204 	}
205 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
206 }
207 
208 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
209 {
210 	struct nvme_ns_head *head = ns->head;
211 	sector_t capacity = get_capacity(head->disk);
212 	int node;
213 	int srcu_idx;
214 
215 	srcu_idx = srcu_read_lock(&head->srcu);
216 	list_for_each_entry_rcu(ns, &head->list, siblings) {
217 		if (capacity != get_capacity(ns->disk))
218 			clear_bit(NVME_NS_READY, &ns->flags);
219 	}
220 	srcu_read_unlock(&head->srcu, srcu_idx);
221 
222 	for_each_node(node)
223 		rcu_assign_pointer(head->current_path[node], NULL);
224 	kblockd_schedule_work(&head->requeue_work);
225 }
226 
227 static bool nvme_path_is_disabled(struct nvme_ns *ns)
228 {
229 	enum nvme_ctrl_state state = nvme_ctrl_state(ns->ctrl);
230 
231 	/*
232 	 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
233 	 * still be able to complete assuming that the controller is connected.
234 	 * Otherwise it will fail immediately and return to the requeue list.
235 	 */
236 	if (state != NVME_CTRL_LIVE && state != NVME_CTRL_DELETING)
237 		return true;
238 	if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
239 	    !test_bit(NVME_NS_READY, &ns->flags))
240 		return true;
241 	return false;
242 }
243 
244 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
245 {
246 	int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
247 	struct nvme_ns *found = NULL, *fallback = NULL, *ns;
248 
249 	list_for_each_entry_rcu(ns, &head->list, siblings) {
250 		if (nvme_path_is_disabled(ns))
251 			continue;
252 
253 		if (ns->ctrl->numa_node != NUMA_NO_NODE &&
254 		    READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
255 			distance = node_distance(node, ns->ctrl->numa_node);
256 		else
257 			distance = LOCAL_DISTANCE;
258 
259 		switch (ns->ana_state) {
260 		case NVME_ANA_OPTIMIZED:
261 			if (distance < found_distance) {
262 				found_distance = distance;
263 				found = ns;
264 			}
265 			break;
266 		case NVME_ANA_NONOPTIMIZED:
267 			if (distance < fallback_distance) {
268 				fallback_distance = distance;
269 				fallback = ns;
270 			}
271 			break;
272 		default:
273 			break;
274 		}
275 	}
276 
277 	if (!found)
278 		found = fallback;
279 	if (found)
280 		rcu_assign_pointer(head->current_path[node], found);
281 	return found;
282 }
283 
284 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
285 		struct nvme_ns *ns)
286 {
287 	ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
288 			siblings);
289 	if (ns)
290 		return ns;
291 	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
292 }
293 
294 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
295 		int node, struct nvme_ns *old)
296 {
297 	struct nvme_ns *ns, *found = NULL;
298 
299 	if (list_is_singular(&head->list)) {
300 		if (nvme_path_is_disabled(old))
301 			return NULL;
302 		return old;
303 	}
304 
305 	for (ns = nvme_next_ns(head, old);
306 	     ns && ns != old;
307 	     ns = nvme_next_ns(head, ns)) {
308 		if (nvme_path_is_disabled(ns))
309 			continue;
310 
311 		if (ns->ana_state == NVME_ANA_OPTIMIZED) {
312 			found = ns;
313 			goto out;
314 		}
315 		if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
316 			found = ns;
317 	}
318 
319 	/*
320 	 * The loop above skips the current path for round-robin semantics.
321 	 * Fall back to the current path if either:
322 	 *  - no other optimized path found and current is optimized,
323 	 *  - no other usable path found and current is usable.
324 	 */
325 	if (!nvme_path_is_disabled(old) &&
326 	    (old->ana_state == NVME_ANA_OPTIMIZED ||
327 	     (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
328 		return old;
329 
330 	if (!found)
331 		return NULL;
332 out:
333 	rcu_assign_pointer(head->current_path[node], found);
334 	return found;
335 }
336 
337 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
338 {
339 	return nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE &&
340 		ns->ana_state == NVME_ANA_OPTIMIZED;
341 }
342 
343 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
344 {
345 	int node = numa_node_id();
346 	struct nvme_ns *ns;
347 
348 	ns = srcu_dereference(head->current_path[node], &head->srcu);
349 	if (unlikely(!ns))
350 		return __nvme_find_path(head, node);
351 
352 	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
353 		return nvme_round_robin_path(head, node, ns);
354 	if (unlikely(!nvme_path_is_optimized(ns)))
355 		return __nvme_find_path(head, node);
356 	return ns;
357 }
358 
359 static bool nvme_available_path(struct nvme_ns_head *head)
360 {
361 	struct nvme_ns *ns;
362 
363 	list_for_each_entry_rcu(ns, &head->list, siblings) {
364 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
365 			continue;
366 		switch (nvme_ctrl_state(ns->ctrl)) {
367 		case NVME_CTRL_LIVE:
368 		case NVME_CTRL_RESETTING:
369 		case NVME_CTRL_CONNECTING:
370 			/* fallthru */
371 			return true;
372 		default:
373 			break;
374 		}
375 	}
376 	return false;
377 }
378 
379 static void nvme_ns_head_submit_bio(struct bio *bio)
380 {
381 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
382 	struct device *dev = disk_to_dev(head->disk);
383 	struct nvme_ns *ns;
384 	int srcu_idx;
385 
386 	/*
387 	 * The namespace might be going away and the bio might be moved to a
388 	 * different queue via blk_steal_bios(), so we need to use the bio_split
389 	 * pool from the original queue to allocate the bvecs from.
390 	 */
391 	bio = bio_split_to_limits(bio);
392 	if (!bio)
393 		return;
394 
395 	srcu_idx = srcu_read_lock(&head->srcu);
396 	ns = nvme_find_path(head);
397 	if (likely(ns)) {
398 		bio_set_dev(bio, ns->disk->part0);
399 		bio->bi_opf |= REQ_NVME_MPATH;
400 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
401 				      bio->bi_iter.bi_sector);
402 		submit_bio_noacct(bio);
403 	} else if (nvme_available_path(head)) {
404 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
405 
406 		spin_lock_irq(&head->requeue_lock);
407 		bio_list_add(&head->requeue_list, bio);
408 		spin_unlock_irq(&head->requeue_lock);
409 	} else {
410 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
411 
412 		bio_io_error(bio);
413 	}
414 
415 	srcu_read_unlock(&head->srcu, srcu_idx);
416 }
417 
418 static int nvme_ns_head_open(struct gendisk *disk, blk_mode_t mode)
419 {
420 	if (!nvme_tryget_ns_head(disk->private_data))
421 		return -ENXIO;
422 	return 0;
423 }
424 
425 static void nvme_ns_head_release(struct gendisk *disk)
426 {
427 	nvme_put_ns_head(disk->private_data);
428 }
429 
430 #ifdef CONFIG_BLK_DEV_ZONED
431 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
432 		unsigned int nr_zones, report_zones_cb cb, void *data)
433 {
434 	struct nvme_ns_head *head = disk->private_data;
435 	struct nvme_ns *ns;
436 	int srcu_idx, ret = -EWOULDBLOCK;
437 
438 	srcu_idx = srcu_read_lock(&head->srcu);
439 	ns = nvme_find_path(head);
440 	if (ns)
441 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
442 	srcu_read_unlock(&head->srcu, srcu_idx);
443 	return ret;
444 }
445 #else
446 #define nvme_ns_head_report_zones	NULL
447 #endif /* CONFIG_BLK_DEV_ZONED */
448 
449 const struct block_device_operations nvme_ns_head_ops = {
450 	.owner		= THIS_MODULE,
451 	.submit_bio	= nvme_ns_head_submit_bio,
452 	.open		= nvme_ns_head_open,
453 	.release	= nvme_ns_head_release,
454 	.ioctl		= nvme_ns_head_ioctl,
455 	.compat_ioctl	= blkdev_compat_ptr_ioctl,
456 	.getgeo		= nvme_getgeo,
457 	.report_zones	= nvme_ns_head_report_zones,
458 	.pr_ops		= &nvme_pr_ops,
459 };
460 
461 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
462 {
463 	return container_of(cdev, struct nvme_ns_head, cdev);
464 }
465 
466 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
467 {
468 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
469 		return -ENXIO;
470 	return 0;
471 }
472 
473 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
474 {
475 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
476 	return 0;
477 }
478 
479 static const struct file_operations nvme_ns_head_chr_fops = {
480 	.owner		= THIS_MODULE,
481 	.open		= nvme_ns_head_chr_open,
482 	.release	= nvme_ns_head_chr_release,
483 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
484 	.compat_ioctl	= compat_ptr_ioctl,
485 	.uring_cmd	= nvme_ns_head_chr_uring_cmd,
486 	.uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
487 };
488 
489 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
490 {
491 	int ret;
492 
493 	head->cdev_device.parent = &head->subsys->dev;
494 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
495 			   head->subsys->instance, head->instance);
496 	if (ret)
497 		return ret;
498 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
499 			    &nvme_ns_head_chr_fops, THIS_MODULE);
500 	return ret;
501 }
502 
503 static void nvme_requeue_work(struct work_struct *work)
504 {
505 	struct nvme_ns_head *head =
506 		container_of(work, struct nvme_ns_head, requeue_work);
507 	struct bio *bio, *next;
508 
509 	spin_lock_irq(&head->requeue_lock);
510 	next = bio_list_get(&head->requeue_list);
511 	spin_unlock_irq(&head->requeue_lock);
512 
513 	while ((bio = next) != NULL) {
514 		next = bio->bi_next;
515 		bio->bi_next = NULL;
516 
517 		submit_bio_noacct(bio);
518 	}
519 }
520 
521 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
522 {
523 	struct queue_limits lim;
524 
525 	mutex_init(&head->lock);
526 	bio_list_init(&head->requeue_list);
527 	spin_lock_init(&head->requeue_lock);
528 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
529 
530 	/*
531 	 * Add a multipath node if the subsystems supports multiple controllers.
532 	 * We also do this for private namespaces as the namespace sharing flag
533 	 * could change after a rescan.
534 	 */
535 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
536 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
537 		return 0;
538 
539 	blk_set_stacking_limits(&lim);
540 	lim.dma_alignment = 3;
541 	lim.features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT | BLK_FEAT_POLL;
542 	if (head->ids.csi != NVME_CSI_ZNS)
543 		lim.max_zone_append_sectors = 0;
544 
545 	head->disk = blk_alloc_disk(&lim, ctrl->numa_node);
546 	if (IS_ERR(head->disk))
547 		return PTR_ERR(head->disk);
548 	head->disk->fops = &nvme_ns_head_ops;
549 	head->disk->private_data = head;
550 	sprintf(head->disk->disk_name, "nvme%dn%d",
551 			ctrl->subsys->instance, head->instance);
552 	return 0;
553 }
554 
555 static void nvme_mpath_set_live(struct nvme_ns *ns)
556 {
557 	struct nvme_ns_head *head = ns->head;
558 	int rc;
559 
560 	if (!head->disk)
561 		return;
562 
563 	/*
564 	 * test_and_set_bit() is used because it is protecting against two nvme
565 	 * paths simultaneously calling device_add_disk() on the same namespace
566 	 * head.
567 	 */
568 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
569 		rc = device_add_disk(&head->subsys->dev, head->disk,
570 				     nvme_ns_attr_groups);
571 		if (rc) {
572 			clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
573 			return;
574 		}
575 		nvme_add_ns_head_cdev(head);
576 	}
577 
578 	mutex_lock(&head->lock);
579 	if (nvme_path_is_optimized(ns)) {
580 		int node, srcu_idx;
581 
582 		srcu_idx = srcu_read_lock(&head->srcu);
583 		for_each_online_node(node)
584 			__nvme_find_path(head, node);
585 		srcu_read_unlock(&head->srcu, srcu_idx);
586 	}
587 	mutex_unlock(&head->lock);
588 
589 	synchronize_srcu(&head->srcu);
590 	kblockd_schedule_work(&head->requeue_work);
591 }
592 
593 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
594 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
595 			void *))
596 {
597 	void *base = ctrl->ana_log_buf;
598 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
599 	int error, i;
600 
601 	lockdep_assert_held(&ctrl->ana_lock);
602 
603 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
604 		struct nvme_ana_group_desc *desc = base + offset;
605 		u32 nr_nsids;
606 		size_t nsid_buf_size;
607 
608 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
609 			return -EINVAL;
610 
611 		nr_nsids = le32_to_cpu(desc->nnsids);
612 		nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
613 
614 		if (WARN_ON_ONCE(desc->grpid == 0))
615 			return -EINVAL;
616 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
617 			return -EINVAL;
618 		if (WARN_ON_ONCE(desc->state == 0))
619 			return -EINVAL;
620 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
621 			return -EINVAL;
622 
623 		offset += sizeof(*desc);
624 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
625 			return -EINVAL;
626 
627 		error = cb(ctrl, desc, data);
628 		if (error)
629 			return error;
630 
631 		offset += nsid_buf_size;
632 	}
633 
634 	return 0;
635 }
636 
637 static inline bool nvme_state_is_live(enum nvme_ana_state state)
638 {
639 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
640 }
641 
642 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
643 		struct nvme_ns *ns)
644 {
645 	ns->ana_grpid = le32_to_cpu(desc->grpid);
646 	ns->ana_state = desc->state;
647 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
648 	/*
649 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
650 	 * and in turn to this path device.  However we cannot accept this I/O
651 	 * if the controller is not live.  This may deadlock if called from
652 	 * nvme_mpath_init_identify() and the ctrl will never complete
653 	 * initialization, preventing I/O from completing.  For this case we
654 	 * will reprocess the ANA log page in nvme_mpath_update() once the
655 	 * controller is ready.
656 	 */
657 	if (nvme_state_is_live(ns->ana_state) &&
658 	    nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
659 		nvme_mpath_set_live(ns);
660 }
661 
662 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
663 		struct nvme_ana_group_desc *desc, void *data)
664 {
665 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
666 	unsigned *nr_change_groups = data;
667 	struct nvme_ns *ns;
668 	int srcu_idx;
669 
670 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
671 			le32_to_cpu(desc->grpid),
672 			nvme_ana_state_names[desc->state]);
673 
674 	if (desc->state == NVME_ANA_CHANGE)
675 		(*nr_change_groups)++;
676 
677 	if (!nr_nsids)
678 		return 0;
679 
680 	srcu_idx = srcu_read_lock(&ctrl->srcu);
681 	list_for_each_entry_rcu(ns, &ctrl->namespaces, list) {
682 		unsigned nsid;
683 again:
684 		nsid = le32_to_cpu(desc->nsids[n]);
685 		if (ns->head->ns_id < nsid)
686 			continue;
687 		if (ns->head->ns_id == nsid)
688 			nvme_update_ns_ana_state(desc, ns);
689 		if (++n == nr_nsids)
690 			break;
691 		if (ns->head->ns_id > nsid)
692 			goto again;
693 	}
694 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
695 	return 0;
696 }
697 
698 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
699 {
700 	u32 nr_change_groups = 0;
701 	int error;
702 
703 	mutex_lock(&ctrl->ana_lock);
704 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
705 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
706 	if (error) {
707 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
708 		goto out_unlock;
709 	}
710 
711 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
712 			nvme_update_ana_state);
713 	if (error)
714 		goto out_unlock;
715 
716 	/*
717 	 * In theory we should have an ANATT timer per group as they might enter
718 	 * the change state at different times.  But that is a lot of overhead
719 	 * just to protect against a target that keeps entering new changes
720 	 * states while never finishing previous ones.  But we'll still
721 	 * eventually time out once all groups are in change state, so this
722 	 * isn't a big deal.
723 	 *
724 	 * We also double the ANATT value to provide some slack for transports
725 	 * or AEN processing overhead.
726 	 */
727 	if (nr_change_groups)
728 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
729 	else
730 		del_timer_sync(&ctrl->anatt_timer);
731 out_unlock:
732 	mutex_unlock(&ctrl->ana_lock);
733 	return error;
734 }
735 
736 static void nvme_ana_work(struct work_struct *work)
737 {
738 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
739 
740 	if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
741 		return;
742 
743 	nvme_read_ana_log(ctrl);
744 }
745 
746 void nvme_mpath_update(struct nvme_ctrl *ctrl)
747 {
748 	u32 nr_change_groups = 0;
749 
750 	if (!ctrl->ana_log_buf)
751 		return;
752 
753 	mutex_lock(&ctrl->ana_lock);
754 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
755 	mutex_unlock(&ctrl->ana_lock);
756 }
757 
758 static void nvme_anatt_timeout(struct timer_list *t)
759 {
760 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
761 
762 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
763 	nvme_reset_ctrl(ctrl);
764 }
765 
766 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
767 {
768 	if (!nvme_ctrl_use_ana(ctrl))
769 		return;
770 	del_timer_sync(&ctrl->anatt_timer);
771 	cancel_work_sync(&ctrl->ana_work);
772 }
773 
774 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
775 	struct device_attribute subsys_attr_##_name =	\
776 		__ATTR(_name, _mode, _show, _store)
777 
778 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
779 		struct device_attribute *attr, char *buf)
780 {
781 	struct nvme_subsystem *subsys =
782 		container_of(dev, struct nvme_subsystem, dev);
783 
784 	return sysfs_emit(buf, "%s\n",
785 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
786 }
787 
788 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
789 		struct device_attribute *attr, const char *buf, size_t count)
790 {
791 	struct nvme_subsystem *subsys =
792 		container_of(dev, struct nvme_subsystem, dev);
793 	int i;
794 
795 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
796 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
797 			WRITE_ONCE(subsys->iopolicy, i);
798 			return count;
799 		}
800 	}
801 
802 	return -EINVAL;
803 }
804 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
805 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
806 
807 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
808 		char *buf)
809 {
810 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
811 }
812 DEVICE_ATTR_RO(ana_grpid);
813 
814 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
815 		char *buf)
816 {
817 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
818 
819 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
820 }
821 DEVICE_ATTR_RO(ana_state);
822 
823 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
824 		struct nvme_ana_group_desc *desc, void *data)
825 {
826 	struct nvme_ana_group_desc *dst = data;
827 
828 	if (desc->grpid != dst->grpid)
829 		return 0;
830 
831 	*dst = *desc;
832 	return -ENXIO; /* just break out of the loop */
833 }
834 
835 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
836 {
837 	if (nvme_ctrl_use_ana(ns->ctrl)) {
838 		struct nvme_ana_group_desc desc = {
839 			.grpid = anagrpid,
840 			.state = 0,
841 		};
842 
843 		mutex_lock(&ns->ctrl->ana_lock);
844 		ns->ana_grpid = le32_to_cpu(anagrpid);
845 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
846 		mutex_unlock(&ns->ctrl->ana_lock);
847 		if (desc.state) {
848 			/* found the group desc: update */
849 			nvme_update_ns_ana_state(&desc, ns);
850 		} else {
851 			/* group desc not found: trigger a re-read */
852 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
853 			queue_work(nvme_wq, &ns->ctrl->ana_work);
854 		}
855 	} else {
856 		ns->ana_state = NVME_ANA_OPTIMIZED;
857 		nvme_mpath_set_live(ns);
858 	}
859 
860 #ifdef CONFIG_BLK_DEV_ZONED
861 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
862 		ns->head->disk->nr_zones = ns->disk->nr_zones;
863 #endif
864 }
865 
866 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
867 {
868 	if (!head->disk)
869 		return;
870 	kblockd_schedule_work(&head->requeue_work);
871 	if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
872 		nvme_cdev_del(&head->cdev, &head->cdev_device);
873 		del_gendisk(head->disk);
874 	}
875 }
876 
877 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
878 {
879 	if (!head->disk)
880 		return;
881 	/* make sure all pending bios are cleaned up */
882 	kblockd_schedule_work(&head->requeue_work);
883 	flush_work(&head->requeue_work);
884 	put_disk(head->disk);
885 }
886 
887 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
888 {
889 	mutex_init(&ctrl->ana_lock);
890 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
891 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
892 }
893 
894 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
895 {
896 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
897 	size_t ana_log_size;
898 	int error = 0;
899 
900 	/* check if multipath is enabled and we have the capability */
901 	if (!multipath || !ctrl->subsys ||
902 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
903 		return 0;
904 
905 	if (!ctrl->max_namespaces ||
906 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
907 		dev_err(ctrl->device,
908 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
909 		return -EINVAL;
910 	}
911 
912 	ctrl->anacap = id->anacap;
913 	ctrl->anatt = id->anatt;
914 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
915 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
916 
917 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
918 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
919 		ctrl->max_namespaces * sizeof(__le32);
920 	if (ana_log_size > max_transfer_size) {
921 		dev_err(ctrl->device,
922 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
923 			ana_log_size, max_transfer_size);
924 		dev_err(ctrl->device, "disabling ANA support.\n");
925 		goto out_uninit;
926 	}
927 	if (ana_log_size > ctrl->ana_log_size) {
928 		nvme_mpath_stop(ctrl);
929 		nvme_mpath_uninit(ctrl);
930 		ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
931 		if (!ctrl->ana_log_buf)
932 			return -ENOMEM;
933 	}
934 	ctrl->ana_log_size = ana_log_size;
935 	error = nvme_read_ana_log(ctrl);
936 	if (error)
937 		goto out_uninit;
938 	return 0;
939 
940 out_uninit:
941 	nvme_mpath_uninit(ctrl);
942 	return error;
943 }
944 
945 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
946 {
947 	kvfree(ctrl->ana_log_buf);
948 	ctrl->ana_log_buf = NULL;
949 	ctrl->ana_log_size = 0;
950 }
951