xref: /linux/drivers/nvme/host/multipath.c (revision 90d32e92011eaae8e70a9169b4e7acf4ca8f9d3a)
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 (ns->ctrl->numa_node != NUMA_NO_NODE &&
251 		    READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
252 			distance = node_distance(node, ns->ctrl->numa_node);
253 		else
254 			distance = LOCAL_DISTANCE;
255 
256 		switch (ns->ana_state) {
257 		case NVME_ANA_OPTIMIZED:
258 			if (distance < found_distance) {
259 				found_distance = distance;
260 				found = ns;
261 			}
262 			break;
263 		case NVME_ANA_NONOPTIMIZED:
264 			if (distance < fallback_distance) {
265 				fallback_distance = distance;
266 				fallback = ns;
267 			}
268 			break;
269 		default:
270 			break;
271 		}
272 	}
273 
274 	if (!found)
275 		found = fallback;
276 	if (found)
277 		rcu_assign_pointer(head->current_path[node], found);
278 	return found;
279 }
280 
281 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
282 		struct nvme_ns *ns)
283 {
284 	ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
285 			siblings);
286 	if (ns)
287 		return ns;
288 	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
289 }
290 
291 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
292 		int node, struct nvme_ns *old)
293 {
294 	struct nvme_ns *ns, *found = NULL;
295 
296 	if (list_is_singular(&head->list)) {
297 		if (nvme_path_is_disabled(old))
298 			return NULL;
299 		return old;
300 	}
301 
302 	for (ns = nvme_next_ns(head, old);
303 	     ns && ns != old;
304 	     ns = nvme_next_ns(head, ns)) {
305 		if (nvme_path_is_disabled(ns))
306 			continue;
307 
308 		if (ns->ana_state == NVME_ANA_OPTIMIZED) {
309 			found = ns;
310 			goto out;
311 		}
312 		if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
313 			found = ns;
314 	}
315 
316 	/*
317 	 * The loop above skips the current path for round-robin semantics.
318 	 * Fall back to the current path if either:
319 	 *  - no other optimized path found and current is optimized,
320 	 *  - no other usable path found and current is usable.
321 	 */
322 	if (!nvme_path_is_disabled(old) &&
323 	    (old->ana_state == NVME_ANA_OPTIMIZED ||
324 	     (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
325 		return old;
326 
327 	if (!found)
328 		return NULL;
329 out:
330 	rcu_assign_pointer(head->current_path[node], found);
331 	return found;
332 }
333 
334 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
335 {
336 	return nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE &&
337 		ns->ana_state == NVME_ANA_OPTIMIZED;
338 }
339 
340 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
341 {
342 	int node = numa_node_id();
343 	struct nvme_ns *ns;
344 
345 	ns = srcu_dereference(head->current_path[node], &head->srcu);
346 	if (unlikely(!ns))
347 		return __nvme_find_path(head, node);
348 
349 	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
350 		return nvme_round_robin_path(head, node, ns);
351 	if (unlikely(!nvme_path_is_optimized(ns)))
352 		return __nvme_find_path(head, node);
353 	return ns;
354 }
355 
356 static bool nvme_available_path(struct nvme_ns_head *head)
357 {
358 	struct nvme_ns *ns;
359 
360 	list_for_each_entry_rcu(ns, &head->list, siblings) {
361 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
362 			continue;
363 		switch (nvme_ctrl_state(ns->ctrl)) {
364 		case NVME_CTRL_LIVE:
365 		case NVME_CTRL_RESETTING:
366 		case NVME_CTRL_CONNECTING:
367 			/* fallthru */
368 			return true;
369 		default:
370 			break;
371 		}
372 	}
373 	return false;
374 }
375 
376 static void nvme_ns_head_submit_bio(struct bio *bio)
377 {
378 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
379 	struct device *dev = disk_to_dev(head->disk);
380 	struct nvme_ns *ns;
381 	int srcu_idx;
382 
383 	/*
384 	 * The namespace might be going away and the bio might be moved to a
385 	 * different queue via blk_steal_bios(), so we need to use the bio_split
386 	 * pool from the original queue to allocate the bvecs from.
387 	 */
388 	bio = bio_split_to_limits(bio);
389 	if (!bio)
390 		return;
391 
392 	srcu_idx = srcu_read_lock(&head->srcu);
393 	ns = nvme_find_path(head);
394 	if (likely(ns)) {
395 		bio_set_dev(bio, ns->disk->part0);
396 		bio->bi_opf |= REQ_NVME_MPATH;
397 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
398 				      bio->bi_iter.bi_sector);
399 		submit_bio_noacct(bio);
400 	} else if (nvme_available_path(head)) {
401 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
402 
403 		spin_lock_irq(&head->requeue_lock);
404 		bio_list_add(&head->requeue_list, bio);
405 		spin_unlock_irq(&head->requeue_lock);
406 	} else {
407 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
408 
409 		bio_io_error(bio);
410 	}
411 
412 	srcu_read_unlock(&head->srcu, srcu_idx);
413 }
414 
415 static int nvme_ns_head_open(struct gendisk *disk, blk_mode_t mode)
416 {
417 	if (!nvme_tryget_ns_head(disk->private_data))
418 		return -ENXIO;
419 	return 0;
420 }
421 
422 static void nvme_ns_head_release(struct gendisk *disk)
423 {
424 	nvme_put_ns_head(disk->private_data);
425 }
426 
427 #ifdef CONFIG_BLK_DEV_ZONED
428 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
429 		unsigned int nr_zones, report_zones_cb cb, void *data)
430 {
431 	struct nvme_ns_head *head = disk->private_data;
432 	struct nvme_ns *ns;
433 	int srcu_idx, ret = -EWOULDBLOCK;
434 
435 	srcu_idx = srcu_read_lock(&head->srcu);
436 	ns = nvme_find_path(head);
437 	if (ns)
438 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
439 	srcu_read_unlock(&head->srcu, srcu_idx);
440 	return ret;
441 }
442 #else
443 #define nvme_ns_head_report_zones	NULL
444 #endif /* CONFIG_BLK_DEV_ZONED */
445 
446 const struct block_device_operations nvme_ns_head_ops = {
447 	.owner		= THIS_MODULE,
448 	.submit_bio	= nvme_ns_head_submit_bio,
449 	.open		= nvme_ns_head_open,
450 	.release	= nvme_ns_head_release,
451 	.ioctl		= nvme_ns_head_ioctl,
452 	.compat_ioctl	= blkdev_compat_ptr_ioctl,
453 	.getgeo		= nvme_getgeo,
454 	.report_zones	= nvme_ns_head_report_zones,
455 	.pr_ops		= &nvme_pr_ops,
456 };
457 
458 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
459 {
460 	return container_of(cdev, struct nvme_ns_head, cdev);
461 }
462 
463 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
464 {
465 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
466 		return -ENXIO;
467 	return 0;
468 }
469 
470 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
471 {
472 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
473 	return 0;
474 }
475 
476 static const struct file_operations nvme_ns_head_chr_fops = {
477 	.owner		= THIS_MODULE,
478 	.open		= nvme_ns_head_chr_open,
479 	.release	= nvme_ns_head_chr_release,
480 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
481 	.compat_ioctl	= compat_ptr_ioctl,
482 	.uring_cmd	= nvme_ns_head_chr_uring_cmd,
483 	.uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
484 };
485 
486 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
487 {
488 	int ret;
489 
490 	head->cdev_device.parent = &head->subsys->dev;
491 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
492 			   head->subsys->instance, head->instance);
493 	if (ret)
494 		return ret;
495 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
496 			    &nvme_ns_head_chr_fops, THIS_MODULE);
497 	return ret;
498 }
499 
500 static void nvme_requeue_work(struct work_struct *work)
501 {
502 	struct nvme_ns_head *head =
503 		container_of(work, struct nvme_ns_head, requeue_work);
504 	struct bio *bio, *next;
505 
506 	spin_lock_irq(&head->requeue_lock);
507 	next = bio_list_get(&head->requeue_list);
508 	spin_unlock_irq(&head->requeue_lock);
509 
510 	while ((bio = next) != NULL) {
511 		next = bio->bi_next;
512 		bio->bi_next = NULL;
513 
514 		submit_bio_noacct(bio);
515 	}
516 }
517 
518 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
519 {
520 	struct queue_limits lim;
521 	bool vwc = false;
522 
523 	mutex_init(&head->lock);
524 	bio_list_init(&head->requeue_list);
525 	spin_lock_init(&head->requeue_lock);
526 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
527 
528 	/*
529 	 * Add a multipath node if the subsystems supports multiple controllers.
530 	 * We also do this for private namespaces as the namespace sharing flag
531 	 * could change after a rescan.
532 	 */
533 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
534 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
535 		return 0;
536 
537 	blk_set_stacking_limits(&lim);
538 	lim.dma_alignment = 3;
539 	if (head->ids.csi != NVME_CSI_ZNS)
540 		lim.max_zone_append_sectors = 0;
541 
542 	head->disk = blk_alloc_disk(&lim, ctrl->numa_node);
543 	if (IS_ERR(head->disk))
544 		return PTR_ERR(head->disk);
545 	head->disk->fops = &nvme_ns_head_ops;
546 	head->disk->private_data = head;
547 	sprintf(head->disk->disk_name, "nvme%dn%d",
548 			ctrl->subsys->instance, head->instance);
549 
550 	blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
551 	blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
552 	blk_queue_flag_set(QUEUE_FLAG_IO_STAT, head->disk->queue);
553 	/*
554 	 * This assumes all controllers that refer to a namespace either
555 	 * support poll queues or not.  That is not a strict guarantee,
556 	 * but if the assumption is wrong the effect is only suboptimal
557 	 * performance but not correctness problem.
558 	 */
559 	if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL &&
560 	    ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues)
561 		blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue);
562 
563 	/* we need to propagate up the VMC settings */
564 	if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
565 		vwc = true;
566 	blk_queue_write_cache(head->disk->queue, vwc, vwc);
567 	return 0;
568 }
569 
570 static void nvme_mpath_set_live(struct nvme_ns *ns)
571 {
572 	struct nvme_ns_head *head = ns->head;
573 	int rc;
574 
575 	if (!head->disk)
576 		return;
577 
578 	/*
579 	 * test_and_set_bit() is used because it is protecting against two nvme
580 	 * paths simultaneously calling device_add_disk() on the same namespace
581 	 * head.
582 	 */
583 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
584 		rc = device_add_disk(&head->subsys->dev, head->disk,
585 				     nvme_ns_attr_groups);
586 		if (rc) {
587 			clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
588 			return;
589 		}
590 		nvme_add_ns_head_cdev(head);
591 	}
592 
593 	mutex_lock(&head->lock);
594 	if (nvme_path_is_optimized(ns)) {
595 		int node, srcu_idx;
596 
597 		srcu_idx = srcu_read_lock(&head->srcu);
598 		for_each_node(node)
599 			__nvme_find_path(head, node);
600 		srcu_read_unlock(&head->srcu, srcu_idx);
601 	}
602 	mutex_unlock(&head->lock);
603 
604 	synchronize_srcu(&head->srcu);
605 	kblockd_schedule_work(&head->requeue_work);
606 }
607 
608 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
609 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
610 			void *))
611 {
612 	void *base = ctrl->ana_log_buf;
613 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
614 	int error, i;
615 
616 	lockdep_assert_held(&ctrl->ana_lock);
617 
618 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
619 		struct nvme_ana_group_desc *desc = base + offset;
620 		u32 nr_nsids;
621 		size_t nsid_buf_size;
622 
623 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
624 			return -EINVAL;
625 
626 		nr_nsids = le32_to_cpu(desc->nnsids);
627 		nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
628 
629 		if (WARN_ON_ONCE(desc->grpid == 0))
630 			return -EINVAL;
631 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
632 			return -EINVAL;
633 		if (WARN_ON_ONCE(desc->state == 0))
634 			return -EINVAL;
635 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
636 			return -EINVAL;
637 
638 		offset += sizeof(*desc);
639 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
640 			return -EINVAL;
641 
642 		error = cb(ctrl, desc, data);
643 		if (error)
644 			return error;
645 
646 		offset += nsid_buf_size;
647 	}
648 
649 	return 0;
650 }
651 
652 static inline bool nvme_state_is_live(enum nvme_ana_state state)
653 {
654 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
655 }
656 
657 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
658 		struct nvme_ns *ns)
659 {
660 	ns->ana_grpid = le32_to_cpu(desc->grpid);
661 	ns->ana_state = desc->state;
662 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
663 	/*
664 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
665 	 * and in turn to this path device.  However we cannot accept this I/O
666 	 * if the controller is not live.  This may deadlock if called from
667 	 * nvme_mpath_init_identify() and the ctrl will never complete
668 	 * initialization, preventing I/O from completing.  For this case we
669 	 * will reprocess the ANA log page in nvme_mpath_update() once the
670 	 * controller is ready.
671 	 */
672 	if (nvme_state_is_live(ns->ana_state) &&
673 	    nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE)
674 		nvme_mpath_set_live(ns);
675 }
676 
677 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
678 		struct nvme_ana_group_desc *desc, void *data)
679 {
680 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
681 	unsigned *nr_change_groups = data;
682 	struct nvme_ns *ns;
683 
684 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
685 			le32_to_cpu(desc->grpid),
686 			nvme_ana_state_names[desc->state]);
687 
688 	if (desc->state == NVME_ANA_CHANGE)
689 		(*nr_change_groups)++;
690 
691 	if (!nr_nsids)
692 		return 0;
693 
694 	down_read(&ctrl->namespaces_rwsem);
695 	list_for_each_entry(ns, &ctrl->namespaces, list) {
696 		unsigned nsid;
697 again:
698 		nsid = le32_to_cpu(desc->nsids[n]);
699 		if (ns->head->ns_id < nsid)
700 			continue;
701 		if (ns->head->ns_id == nsid)
702 			nvme_update_ns_ana_state(desc, ns);
703 		if (++n == nr_nsids)
704 			break;
705 		if (ns->head->ns_id > nsid)
706 			goto again;
707 	}
708 	up_read(&ctrl->namespaces_rwsem);
709 	return 0;
710 }
711 
712 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
713 {
714 	u32 nr_change_groups = 0;
715 	int error;
716 
717 	mutex_lock(&ctrl->ana_lock);
718 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
719 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
720 	if (error) {
721 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
722 		goto out_unlock;
723 	}
724 
725 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
726 			nvme_update_ana_state);
727 	if (error)
728 		goto out_unlock;
729 
730 	/*
731 	 * In theory we should have an ANATT timer per group as they might enter
732 	 * the change state at different times.  But that is a lot of overhead
733 	 * just to protect against a target that keeps entering new changes
734 	 * states while never finishing previous ones.  But we'll still
735 	 * eventually time out once all groups are in change state, so this
736 	 * isn't a big deal.
737 	 *
738 	 * We also double the ANATT value to provide some slack for transports
739 	 * or AEN processing overhead.
740 	 */
741 	if (nr_change_groups)
742 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
743 	else
744 		del_timer_sync(&ctrl->anatt_timer);
745 out_unlock:
746 	mutex_unlock(&ctrl->ana_lock);
747 	return error;
748 }
749 
750 static void nvme_ana_work(struct work_struct *work)
751 {
752 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
753 
754 	if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
755 		return;
756 
757 	nvme_read_ana_log(ctrl);
758 }
759 
760 void nvme_mpath_update(struct nvme_ctrl *ctrl)
761 {
762 	u32 nr_change_groups = 0;
763 
764 	if (!ctrl->ana_log_buf)
765 		return;
766 
767 	mutex_lock(&ctrl->ana_lock);
768 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
769 	mutex_unlock(&ctrl->ana_lock);
770 }
771 
772 static void nvme_anatt_timeout(struct timer_list *t)
773 {
774 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
775 
776 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
777 	nvme_reset_ctrl(ctrl);
778 }
779 
780 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
781 {
782 	if (!nvme_ctrl_use_ana(ctrl))
783 		return;
784 	del_timer_sync(&ctrl->anatt_timer);
785 	cancel_work_sync(&ctrl->ana_work);
786 }
787 
788 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
789 	struct device_attribute subsys_attr_##_name =	\
790 		__ATTR(_name, _mode, _show, _store)
791 
792 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
793 		struct device_attribute *attr, char *buf)
794 {
795 	struct nvme_subsystem *subsys =
796 		container_of(dev, struct nvme_subsystem, dev);
797 
798 	return sysfs_emit(buf, "%s\n",
799 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
800 }
801 
802 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
803 		struct device_attribute *attr, const char *buf, size_t count)
804 {
805 	struct nvme_subsystem *subsys =
806 		container_of(dev, struct nvme_subsystem, dev);
807 	int i;
808 
809 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
810 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
811 			WRITE_ONCE(subsys->iopolicy, i);
812 			return count;
813 		}
814 	}
815 
816 	return -EINVAL;
817 }
818 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
819 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
820 
821 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
822 		char *buf)
823 {
824 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
825 }
826 DEVICE_ATTR_RO(ana_grpid);
827 
828 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
829 		char *buf)
830 {
831 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
832 
833 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
834 }
835 DEVICE_ATTR_RO(ana_state);
836 
837 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
838 		struct nvme_ana_group_desc *desc, void *data)
839 {
840 	struct nvme_ana_group_desc *dst = data;
841 
842 	if (desc->grpid != dst->grpid)
843 		return 0;
844 
845 	*dst = *desc;
846 	return -ENXIO; /* just break out of the loop */
847 }
848 
849 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
850 {
851 	if (nvme_ctrl_use_ana(ns->ctrl)) {
852 		struct nvme_ana_group_desc desc = {
853 			.grpid = anagrpid,
854 			.state = 0,
855 		};
856 
857 		mutex_lock(&ns->ctrl->ana_lock);
858 		ns->ana_grpid = le32_to_cpu(anagrpid);
859 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
860 		mutex_unlock(&ns->ctrl->ana_lock);
861 		if (desc.state) {
862 			/* found the group desc: update */
863 			nvme_update_ns_ana_state(&desc, ns);
864 		} else {
865 			/* group desc not found: trigger a re-read */
866 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
867 			queue_work(nvme_wq, &ns->ctrl->ana_work);
868 		}
869 	} else {
870 		ns->ana_state = NVME_ANA_OPTIMIZED;
871 		nvme_mpath_set_live(ns);
872 	}
873 
874 	if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
875 		blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
876 				   ns->head->disk->queue);
877 #ifdef CONFIG_BLK_DEV_ZONED
878 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
879 		ns->head->disk->nr_zones = ns->disk->nr_zones;
880 #endif
881 }
882 
883 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
884 {
885 	if (!head->disk)
886 		return;
887 	kblockd_schedule_work(&head->requeue_work);
888 	if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
889 		nvme_cdev_del(&head->cdev, &head->cdev_device);
890 		del_gendisk(head->disk);
891 	}
892 }
893 
894 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
895 {
896 	if (!head->disk)
897 		return;
898 	/* make sure all pending bios are cleaned up */
899 	kblockd_schedule_work(&head->requeue_work);
900 	flush_work(&head->requeue_work);
901 	put_disk(head->disk);
902 }
903 
904 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
905 {
906 	mutex_init(&ctrl->ana_lock);
907 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
908 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
909 }
910 
911 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
912 {
913 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
914 	size_t ana_log_size;
915 	int error = 0;
916 
917 	/* check if multipath is enabled and we have the capability */
918 	if (!multipath || !ctrl->subsys ||
919 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
920 		return 0;
921 
922 	if (!ctrl->max_namespaces ||
923 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
924 		dev_err(ctrl->device,
925 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
926 		return -EINVAL;
927 	}
928 
929 	ctrl->anacap = id->anacap;
930 	ctrl->anatt = id->anatt;
931 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
932 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
933 
934 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
935 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
936 		ctrl->max_namespaces * sizeof(__le32);
937 	if (ana_log_size > max_transfer_size) {
938 		dev_err(ctrl->device,
939 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
940 			ana_log_size, max_transfer_size);
941 		dev_err(ctrl->device, "disabling ANA support.\n");
942 		goto out_uninit;
943 	}
944 	if (ana_log_size > ctrl->ana_log_size) {
945 		nvme_mpath_stop(ctrl);
946 		nvme_mpath_uninit(ctrl);
947 		ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
948 		if (!ctrl->ana_log_buf)
949 			return -ENOMEM;
950 	}
951 	ctrl->ana_log_size = ana_log_size;
952 	error = nvme_read_ana_log(ctrl);
953 	if (error)
954 		goto out_uninit;
955 	return 0;
956 
957 out_uninit:
958 	nvme_mpath_uninit(ctrl);
959 	return error;
960 }
961 
962 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
963 {
964 	kvfree(ctrl->ana_log_buf);
965 	ctrl->ana_log_buf = NULL;
966 	ctrl->ana_log_size = 0;
967 }
968