xref: /openbmc/linux/drivers/nvme/host/multipath.c (revision 7671f967)
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_kick_requeue_lists(struct nvme_ctrl *ctrl)
118 {
119 	struct nvme_ns *ns;
120 
121 	down_read(&ctrl->namespaces_rwsem);
122 	list_for_each_entry(ns, &ctrl->namespaces, list) {
123 		if (!ns->head->disk)
124 			continue;
125 		kblockd_schedule_work(&ns->head->requeue_work);
126 		if (ctrl->state == NVME_CTRL_LIVE)
127 			disk_uevent(ns->head->disk, KOBJ_CHANGE);
128 	}
129 	up_read(&ctrl->namespaces_rwsem);
130 }
131 
132 static const char *nvme_ana_state_names[] = {
133 	[0]				= "invalid state",
134 	[NVME_ANA_OPTIMIZED]		= "optimized",
135 	[NVME_ANA_NONOPTIMIZED]		= "non-optimized",
136 	[NVME_ANA_INACCESSIBLE]		= "inaccessible",
137 	[NVME_ANA_PERSISTENT_LOSS]	= "persistent-loss",
138 	[NVME_ANA_CHANGE]		= "change",
139 };
140 
141 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
142 {
143 	struct nvme_ns_head *head = ns->head;
144 	bool changed = false;
145 	int node;
146 
147 	if (!head)
148 		goto out;
149 
150 	for_each_node(node) {
151 		if (ns == rcu_access_pointer(head->current_path[node])) {
152 			rcu_assign_pointer(head->current_path[node], NULL);
153 			changed = true;
154 		}
155 	}
156 out:
157 	return changed;
158 }
159 
160 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
161 {
162 	struct nvme_ns *ns;
163 
164 	down_read(&ctrl->namespaces_rwsem);
165 	list_for_each_entry(ns, &ctrl->namespaces, list) {
166 		nvme_mpath_clear_current_path(ns);
167 		kblockd_schedule_work(&ns->head->requeue_work);
168 	}
169 	up_read(&ctrl->namespaces_rwsem);
170 }
171 
172 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
173 {
174 	struct nvme_ns_head *head = ns->head;
175 	sector_t capacity = get_capacity(head->disk);
176 	int node;
177 
178 	list_for_each_entry_rcu(ns, &head->list, siblings) {
179 		if (capacity != get_capacity(ns->disk))
180 			clear_bit(NVME_NS_READY, &ns->flags);
181 	}
182 
183 	for_each_node(node)
184 		rcu_assign_pointer(head->current_path[node], NULL);
185 }
186 
187 static bool nvme_path_is_disabled(struct nvme_ns *ns)
188 {
189 	/*
190 	 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
191 	 * still be able to complete assuming that the controller is connected.
192 	 * Otherwise it will fail immediately and return to the requeue list.
193 	 */
194 	if (ns->ctrl->state != NVME_CTRL_LIVE &&
195 	    ns->ctrl->state != NVME_CTRL_DELETING)
196 		return true;
197 	if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
198 	    !test_bit(NVME_NS_READY, &ns->flags))
199 		return true;
200 	return false;
201 }
202 
203 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
204 {
205 	int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
206 	struct nvme_ns *found = NULL, *fallback = NULL, *ns;
207 
208 	list_for_each_entry_rcu(ns, &head->list, siblings) {
209 		if (nvme_path_is_disabled(ns))
210 			continue;
211 
212 		if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
213 			distance = node_distance(node, ns->ctrl->numa_node);
214 		else
215 			distance = LOCAL_DISTANCE;
216 
217 		switch (ns->ana_state) {
218 		case NVME_ANA_OPTIMIZED:
219 			if (distance < found_distance) {
220 				found_distance = distance;
221 				found = ns;
222 			}
223 			break;
224 		case NVME_ANA_NONOPTIMIZED:
225 			if (distance < fallback_distance) {
226 				fallback_distance = distance;
227 				fallback = ns;
228 			}
229 			break;
230 		default:
231 			break;
232 		}
233 	}
234 
235 	if (!found)
236 		found = fallback;
237 	if (found)
238 		rcu_assign_pointer(head->current_path[node], found);
239 	return found;
240 }
241 
242 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
243 		struct nvme_ns *ns)
244 {
245 	ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
246 			siblings);
247 	if (ns)
248 		return ns;
249 	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
250 }
251 
252 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
253 		int node, struct nvme_ns *old)
254 {
255 	struct nvme_ns *ns, *found = NULL;
256 
257 	if (list_is_singular(&head->list)) {
258 		if (nvme_path_is_disabled(old))
259 			return NULL;
260 		return old;
261 	}
262 
263 	for (ns = nvme_next_ns(head, old);
264 	     ns && ns != old;
265 	     ns = nvme_next_ns(head, ns)) {
266 		if (nvme_path_is_disabled(ns))
267 			continue;
268 
269 		if (ns->ana_state == NVME_ANA_OPTIMIZED) {
270 			found = ns;
271 			goto out;
272 		}
273 		if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
274 			found = ns;
275 	}
276 
277 	/*
278 	 * The loop above skips the current path for round-robin semantics.
279 	 * Fall back to the current path if either:
280 	 *  - no other optimized path found and current is optimized,
281 	 *  - no other usable path found and current is usable.
282 	 */
283 	if (!nvme_path_is_disabled(old) &&
284 	    (old->ana_state == NVME_ANA_OPTIMIZED ||
285 	     (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
286 		return old;
287 
288 	if (!found)
289 		return NULL;
290 out:
291 	rcu_assign_pointer(head->current_path[node], found);
292 	return found;
293 }
294 
295 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
296 {
297 	return ns->ctrl->state == NVME_CTRL_LIVE &&
298 		ns->ana_state == NVME_ANA_OPTIMIZED;
299 }
300 
301 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
302 {
303 	int node = numa_node_id();
304 	struct nvme_ns *ns;
305 
306 	ns = srcu_dereference(head->current_path[node], &head->srcu);
307 	if (unlikely(!ns))
308 		return __nvme_find_path(head, node);
309 
310 	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
311 		return nvme_round_robin_path(head, node, ns);
312 	if (unlikely(!nvme_path_is_optimized(ns)))
313 		return __nvme_find_path(head, node);
314 	return ns;
315 }
316 
317 static bool nvme_available_path(struct nvme_ns_head *head)
318 {
319 	struct nvme_ns *ns;
320 
321 	list_for_each_entry_rcu(ns, &head->list, siblings) {
322 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
323 			continue;
324 		switch (ns->ctrl->state) {
325 		case NVME_CTRL_LIVE:
326 		case NVME_CTRL_RESETTING:
327 		case NVME_CTRL_CONNECTING:
328 			/* fallthru */
329 			return true;
330 		default:
331 			break;
332 		}
333 	}
334 	return false;
335 }
336 
337 static void nvme_ns_head_submit_bio(struct bio *bio)
338 {
339 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
340 	struct device *dev = disk_to_dev(head->disk);
341 	struct nvme_ns *ns;
342 	int srcu_idx;
343 
344 	/*
345 	 * The namespace might be going away and the bio might be moved to a
346 	 * different queue via blk_steal_bios(), so we need to use the bio_split
347 	 * pool from the original queue to allocate the bvecs from.
348 	 */
349 	blk_queue_split(&bio);
350 
351 	srcu_idx = srcu_read_lock(&head->srcu);
352 	ns = nvme_find_path(head);
353 	if (likely(ns)) {
354 		bio_set_dev(bio, ns->disk->part0);
355 		bio->bi_opf |= REQ_NVME_MPATH;
356 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
357 				      bio->bi_iter.bi_sector);
358 		submit_bio_noacct(bio);
359 	} else if (nvme_available_path(head)) {
360 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
361 
362 		spin_lock_irq(&head->requeue_lock);
363 		bio_list_add(&head->requeue_list, bio);
364 		spin_unlock_irq(&head->requeue_lock);
365 	} else {
366 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
367 
368 		bio_io_error(bio);
369 	}
370 
371 	srcu_read_unlock(&head->srcu, srcu_idx);
372 }
373 
374 static int nvme_ns_head_open(struct block_device *bdev, fmode_t mode)
375 {
376 	if (!nvme_tryget_ns_head(bdev->bd_disk->private_data))
377 		return -ENXIO;
378 	return 0;
379 }
380 
381 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode)
382 {
383 	nvme_put_ns_head(disk->private_data);
384 }
385 
386 #ifdef CONFIG_BLK_DEV_ZONED
387 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
388 		unsigned int nr_zones, report_zones_cb cb, void *data)
389 {
390 	struct nvme_ns_head *head = disk->private_data;
391 	struct nvme_ns *ns;
392 	int srcu_idx, ret = -EWOULDBLOCK;
393 
394 	srcu_idx = srcu_read_lock(&head->srcu);
395 	ns = nvme_find_path(head);
396 	if (ns)
397 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
398 	srcu_read_unlock(&head->srcu, srcu_idx);
399 	return ret;
400 }
401 #else
402 #define nvme_ns_head_report_zones	NULL
403 #endif /* CONFIG_BLK_DEV_ZONED */
404 
405 const struct block_device_operations nvme_ns_head_ops = {
406 	.owner		= THIS_MODULE,
407 	.submit_bio	= nvme_ns_head_submit_bio,
408 	.open		= nvme_ns_head_open,
409 	.release	= nvme_ns_head_release,
410 	.ioctl		= nvme_ns_head_ioctl,
411 	.getgeo		= nvme_getgeo,
412 	.report_zones	= nvme_ns_head_report_zones,
413 	.pr_ops		= &nvme_pr_ops,
414 };
415 
416 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
417 {
418 	return container_of(cdev, struct nvme_ns_head, cdev);
419 }
420 
421 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
422 {
423 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
424 		return -ENXIO;
425 	return 0;
426 }
427 
428 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
429 {
430 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
431 	return 0;
432 }
433 
434 static const struct file_operations nvme_ns_head_chr_fops = {
435 	.owner		= THIS_MODULE,
436 	.open		= nvme_ns_head_chr_open,
437 	.release	= nvme_ns_head_chr_release,
438 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
439 	.compat_ioctl	= compat_ptr_ioctl,
440 };
441 
442 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
443 {
444 	int ret;
445 
446 	head->cdev_device.parent = &head->subsys->dev;
447 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
448 			   head->subsys->instance, head->instance);
449 	if (ret)
450 		return ret;
451 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
452 			    &nvme_ns_head_chr_fops, THIS_MODULE);
453 	return ret;
454 }
455 
456 static void nvme_requeue_work(struct work_struct *work)
457 {
458 	struct nvme_ns_head *head =
459 		container_of(work, struct nvme_ns_head, requeue_work);
460 	struct bio *bio, *next;
461 
462 	spin_lock_irq(&head->requeue_lock);
463 	next = bio_list_get(&head->requeue_list);
464 	spin_unlock_irq(&head->requeue_lock);
465 
466 	while ((bio = next) != NULL) {
467 		next = bio->bi_next;
468 		bio->bi_next = NULL;
469 
470 		submit_bio_noacct(bio);
471 	}
472 }
473 
474 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
475 {
476 	bool vwc = false;
477 
478 	mutex_init(&head->lock);
479 	bio_list_init(&head->requeue_list);
480 	spin_lock_init(&head->requeue_lock);
481 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
482 
483 	/*
484 	 * Add a multipath node if the subsystems supports multiple controllers.
485 	 * We also do this for private namespaces as the namespace sharing flag
486 	 * could change after a rescan.
487 	 */
488 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
489 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
490 		return 0;
491 
492 	head->disk = blk_alloc_disk(ctrl->numa_node);
493 	if (!head->disk)
494 		return -ENOMEM;
495 	head->disk->fops = &nvme_ns_head_ops;
496 	head->disk->private_data = head;
497 	sprintf(head->disk->disk_name, "nvme%dn%d",
498 			ctrl->subsys->instance, head->instance);
499 
500 	blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
501 	blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
502 	/*
503 	 * This assumes all controllers that refer to a namespace either
504 	 * support poll queues or not.  That is not a strict guarantee,
505 	 * but if the assumption is wrong the effect is only suboptimal
506 	 * performance but not correctness problem.
507 	 */
508 	if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL &&
509 	    ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues)
510 		blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue);
511 
512 	/* set to a default value of 512 until the disk is validated */
513 	blk_queue_logical_block_size(head->disk->queue, 512);
514 	blk_set_stacking_limits(&head->disk->queue->limits);
515 
516 	/* we need to propagate up the VMC settings */
517 	if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
518 		vwc = true;
519 	blk_queue_write_cache(head->disk->queue, vwc, vwc);
520 	return 0;
521 }
522 
523 static void nvme_mpath_set_live(struct nvme_ns *ns)
524 {
525 	struct nvme_ns_head *head = ns->head;
526 	int rc;
527 
528 	if (!head->disk)
529 		return;
530 
531 	/*
532 	 * test_and_set_bit() is used because it is protecting against two nvme
533 	 * paths simultaneously calling device_add_disk() on the same namespace
534 	 * head.
535 	 */
536 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
537 		rc = device_add_disk(&head->subsys->dev, head->disk,
538 				     nvme_ns_id_attr_groups);
539 		if (rc) {
540 			clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags);
541 			return;
542 		}
543 		nvme_add_ns_head_cdev(head);
544 	}
545 
546 	mutex_lock(&head->lock);
547 	if (nvme_path_is_optimized(ns)) {
548 		int node, srcu_idx;
549 
550 		srcu_idx = srcu_read_lock(&head->srcu);
551 		for_each_node(node)
552 			__nvme_find_path(head, node);
553 		srcu_read_unlock(&head->srcu, srcu_idx);
554 	}
555 	mutex_unlock(&head->lock);
556 
557 	synchronize_srcu(&head->srcu);
558 	kblockd_schedule_work(&head->requeue_work);
559 }
560 
561 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
562 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
563 			void *))
564 {
565 	void *base = ctrl->ana_log_buf;
566 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
567 	int error, i;
568 
569 	lockdep_assert_held(&ctrl->ana_lock);
570 
571 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
572 		struct nvme_ana_group_desc *desc = base + offset;
573 		u32 nr_nsids;
574 		size_t nsid_buf_size;
575 
576 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
577 			return -EINVAL;
578 
579 		nr_nsids = le32_to_cpu(desc->nnsids);
580 		nsid_buf_size = flex_array_size(desc, nsids, nr_nsids);
581 
582 		if (WARN_ON_ONCE(desc->grpid == 0))
583 			return -EINVAL;
584 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
585 			return -EINVAL;
586 		if (WARN_ON_ONCE(desc->state == 0))
587 			return -EINVAL;
588 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
589 			return -EINVAL;
590 
591 		offset += sizeof(*desc);
592 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
593 			return -EINVAL;
594 
595 		error = cb(ctrl, desc, data);
596 		if (error)
597 			return error;
598 
599 		offset += nsid_buf_size;
600 	}
601 
602 	return 0;
603 }
604 
605 static inline bool nvme_state_is_live(enum nvme_ana_state state)
606 {
607 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
608 }
609 
610 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
611 		struct nvme_ns *ns)
612 {
613 	ns->ana_grpid = le32_to_cpu(desc->grpid);
614 	ns->ana_state = desc->state;
615 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
616 	/*
617 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
618 	 * and in turn to this path device.  However we cannot accept this I/O
619 	 * if the controller is not live.  This may deadlock if called from
620 	 * nvme_mpath_init_identify() and the ctrl will never complete
621 	 * initialization, preventing I/O from completing.  For this case we
622 	 * will reprocess the ANA log page in nvme_mpath_update() once the
623 	 * controller is ready.
624 	 */
625 	if (nvme_state_is_live(ns->ana_state) &&
626 	    ns->ctrl->state == NVME_CTRL_LIVE)
627 		nvme_mpath_set_live(ns);
628 }
629 
630 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
631 		struct nvme_ana_group_desc *desc, void *data)
632 {
633 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
634 	unsigned *nr_change_groups = data;
635 	struct nvme_ns *ns;
636 
637 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
638 			le32_to_cpu(desc->grpid),
639 			nvme_ana_state_names[desc->state]);
640 
641 	if (desc->state == NVME_ANA_CHANGE)
642 		(*nr_change_groups)++;
643 
644 	if (!nr_nsids)
645 		return 0;
646 
647 	down_read(&ctrl->namespaces_rwsem);
648 	list_for_each_entry(ns, &ctrl->namespaces, list) {
649 		unsigned nsid;
650 again:
651 		nsid = le32_to_cpu(desc->nsids[n]);
652 		if (ns->head->ns_id < nsid)
653 			continue;
654 		if (ns->head->ns_id == nsid)
655 			nvme_update_ns_ana_state(desc, ns);
656 		if (++n == nr_nsids)
657 			break;
658 		if (ns->head->ns_id > nsid)
659 			goto again;
660 	}
661 	up_read(&ctrl->namespaces_rwsem);
662 	return 0;
663 }
664 
665 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
666 {
667 	u32 nr_change_groups = 0;
668 	int error;
669 
670 	mutex_lock(&ctrl->ana_lock);
671 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
672 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
673 	if (error) {
674 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
675 		goto out_unlock;
676 	}
677 
678 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
679 			nvme_update_ana_state);
680 	if (error)
681 		goto out_unlock;
682 
683 	/*
684 	 * In theory we should have an ANATT timer per group as they might enter
685 	 * the change state at different times.  But that is a lot of overhead
686 	 * just to protect against a target that keeps entering new changes
687 	 * states while never finishing previous ones.  But we'll still
688 	 * eventually time out once all groups are in change state, so this
689 	 * isn't a big deal.
690 	 *
691 	 * We also double the ANATT value to provide some slack for transports
692 	 * or AEN processing overhead.
693 	 */
694 	if (nr_change_groups)
695 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
696 	else
697 		del_timer_sync(&ctrl->anatt_timer);
698 out_unlock:
699 	mutex_unlock(&ctrl->ana_lock);
700 	return error;
701 }
702 
703 static void nvme_ana_work(struct work_struct *work)
704 {
705 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
706 
707 	if (ctrl->state != NVME_CTRL_LIVE)
708 		return;
709 
710 	nvme_read_ana_log(ctrl);
711 }
712 
713 void nvme_mpath_update(struct nvme_ctrl *ctrl)
714 {
715 	u32 nr_change_groups = 0;
716 
717 	if (!ctrl->ana_log_buf)
718 		return;
719 
720 	mutex_lock(&ctrl->ana_lock);
721 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
722 	mutex_unlock(&ctrl->ana_lock);
723 }
724 
725 static void nvme_anatt_timeout(struct timer_list *t)
726 {
727 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
728 
729 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
730 	nvme_reset_ctrl(ctrl);
731 }
732 
733 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
734 {
735 	if (!nvme_ctrl_use_ana(ctrl))
736 		return;
737 	del_timer_sync(&ctrl->anatt_timer);
738 	cancel_work_sync(&ctrl->ana_work);
739 }
740 
741 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
742 	struct device_attribute subsys_attr_##_name =	\
743 		__ATTR(_name, _mode, _show, _store)
744 
745 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
746 		struct device_attribute *attr, char *buf)
747 {
748 	struct nvme_subsystem *subsys =
749 		container_of(dev, struct nvme_subsystem, dev);
750 
751 	return sysfs_emit(buf, "%s\n",
752 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
753 }
754 
755 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
756 		struct device_attribute *attr, const char *buf, size_t count)
757 {
758 	struct nvme_subsystem *subsys =
759 		container_of(dev, struct nvme_subsystem, dev);
760 	int i;
761 
762 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
763 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
764 			WRITE_ONCE(subsys->iopolicy, i);
765 			return count;
766 		}
767 	}
768 
769 	return -EINVAL;
770 }
771 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
772 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
773 
774 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
775 		char *buf)
776 {
777 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
778 }
779 DEVICE_ATTR_RO(ana_grpid);
780 
781 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
782 		char *buf)
783 {
784 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
785 
786 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
787 }
788 DEVICE_ATTR_RO(ana_state);
789 
790 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
791 		struct nvme_ana_group_desc *desc, void *data)
792 {
793 	struct nvme_ana_group_desc *dst = data;
794 
795 	if (desc->grpid != dst->grpid)
796 		return 0;
797 
798 	*dst = *desc;
799 	return -ENXIO; /* just break out of the loop */
800 }
801 
802 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
803 {
804 	if (nvme_ctrl_use_ana(ns->ctrl)) {
805 		struct nvme_ana_group_desc desc = {
806 			.grpid = id->anagrpid,
807 			.state = 0,
808 		};
809 
810 		mutex_lock(&ns->ctrl->ana_lock);
811 		ns->ana_grpid = le32_to_cpu(id->anagrpid);
812 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
813 		mutex_unlock(&ns->ctrl->ana_lock);
814 		if (desc.state) {
815 			/* found the group desc: update */
816 			nvme_update_ns_ana_state(&desc, ns);
817 		} else {
818 			/* group desc not found: trigger a re-read */
819 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
820 			queue_work(nvme_wq, &ns->ctrl->ana_work);
821 		}
822 	} else {
823 		ns->ana_state = NVME_ANA_OPTIMIZED;
824 		nvme_mpath_set_live(ns);
825 	}
826 
827 	if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
828 		blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
829 				   ns->head->disk->queue);
830 #ifdef CONFIG_BLK_DEV_ZONED
831 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
832 		ns->head->disk->queue->nr_zones = ns->queue->nr_zones;
833 #endif
834 }
835 
836 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
837 {
838 	if (!head->disk)
839 		return;
840 	kblockd_schedule_work(&head->requeue_work);
841 	if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
842 		nvme_cdev_del(&head->cdev, &head->cdev_device);
843 		del_gendisk(head->disk);
844 	}
845 }
846 
847 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
848 {
849 	if (!head->disk)
850 		return;
851 	blk_mark_disk_dead(head->disk);
852 	/* make sure all pending bios are cleaned up */
853 	kblockd_schedule_work(&head->requeue_work);
854 	flush_work(&head->requeue_work);
855 	blk_cleanup_disk(head->disk);
856 }
857 
858 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
859 {
860 	mutex_init(&ctrl->ana_lock);
861 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
862 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
863 }
864 
865 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
866 {
867 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
868 	size_t ana_log_size;
869 	int error = 0;
870 
871 	/* check if multipath is enabled and we have the capability */
872 	if (!multipath || !ctrl->subsys ||
873 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
874 		return 0;
875 
876 	if (!ctrl->max_namespaces ||
877 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
878 		dev_err(ctrl->device,
879 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
880 		return -EINVAL;
881 	}
882 
883 	ctrl->anacap = id->anacap;
884 	ctrl->anatt = id->anatt;
885 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
886 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
887 
888 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
889 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
890 		ctrl->max_namespaces * sizeof(__le32);
891 	if (ana_log_size > max_transfer_size) {
892 		dev_err(ctrl->device,
893 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
894 			ana_log_size, max_transfer_size);
895 		dev_err(ctrl->device, "disabling ANA support.\n");
896 		goto out_uninit;
897 	}
898 	if (ana_log_size > ctrl->ana_log_size) {
899 		nvme_mpath_stop(ctrl);
900 		nvme_mpath_uninit(ctrl);
901 		ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL);
902 		if (!ctrl->ana_log_buf)
903 			return -ENOMEM;
904 	}
905 	ctrl->ana_log_size = ana_log_size;
906 	error = nvme_read_ana_log(ctrl);
907 	if (error)
908 		goto out_uninit;
909 	return 0;
910 
911 out_uninit:
912 	nvme_mpath_uninit(ctrl);
913 	return error;
914 }
915 
916 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
917 {
918 	kvfree(ctrl->ana_log_buf);
919 	ctrl->ana_log_buf = NULL;
920 	ctrl->ana_log_size = 0;
921 }
922