• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <trace/events/block.h>
9 #include "nvme.h"
10 
11 static bool multipath = true;
12 module_param(multipath, bool, 0444);
13 MODULE_PARM_DESC(multipath,
14 	"turn on native support for multiple controllers per subsystem");
15 
nvme_mpath_unfreeze(struct nvme_subsystem * subsys)16 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
17 {
18 	struct nvme_ns_head *h;
19 
20 	lockdep_assert_held(&subsys->lock);
21 	list_for_each_entry(h, &subsys->nsheads, entry)
22 		if (h->disk)
23 			blk_mq_unfreeze_queue(h->disk->queue);
24 }
25 
nvme_mpath_wait_freeze(struct nvme_subsystem * subsys)26 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
27 {
28 	struct nvme_ns_head *h;
29 
30 	lockdep_assert_held(&subsys->lock);
31 	list_for_each_entry(h, &subsys->nsheads, entry)
32 		if (h->disk)
33 			blk_mq_freeze_queue_wait(h->disk->queue);
34 }
35 
nvme_mpath_start_freeze(struct nvme_subsystem * subsys)36 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
37 {
38 	struct nvme_ns_head *h;
39 
40 	lockdep_assert_held(&subsys->lock);
41 	list_for_each_entry(h, &subsys->nsheads, entry)
42 		if (h->disk)
43 			blk_freeze_queue_start(h->disk->queue);
44 }
45 
46 /*
47  * If multipathing is enabled we need to always use the subsystem instance
48  * number for numbering our devices to avoid conflicts between subsystems that
49  * have multiple controllers and thus use the multipath-aware subsystem node
50  * and those that have a single controller and use the controller node
51  * directly.
52  */
nvme_mpath_set_disk_name(struct nvme_ns * ns,char * disk_name,int * flags)53 bool nvme_mpath_set_disk_name(struct nvme_ns *ns, char *disk_name, int *flags)
54 {
55 	if (!multipath)
56 		return false;
57 	if (!ns->head->disk) {
58 		sprintf(disk_name, "nvme%dn%d", ns->ctrl->subsys->instance,
59 			ns->head->instance);
60 		return true;
61 	}
62 	sprintf(disk_name, "nvme%dc%dn%d", ns->ctrl->subsys->instance,
63 		ns->ctrl->instance, ns->head->instance);
64 	*flags = GENHD_FL_HIDDEN;
65 	return true;
66 }
67 
nvme_failover_req(struct request * req)68 void nvme_failover_req(struct request *req)
69 {
70 	struct nvme_ns *ns = req->q->queuedata;
71 	u16 status = nvme_req(req)->status & 0x7ff;
72 	unsigned long flags;
73 	struct bio *bio;
74 
75 	nvme_mpath_clear_current_path(ns);
76 
77 	/*
78 	 * If we got back an ANA error, we know the controller is alive but not
79 	 * ready to serve this namespace.  Kick of a re-read of the ANA
80 	 * information page, and just try any other available path for now.
81 	 */
82 	if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) {
83 		set_bit(NVME_NS_ANA_PENDING, &ns->flags);
84 		queue_work(nvme_wq, &ns->ctrl->ana_work);
85 	}
86 
87 	spin_lock_irqsave(&ns->head->requeue_lock, flags);
88 	for (bio = req->bio; bio; bio = bio->bi_next)
89 		bio_set_dev(bio, ns->head->disk->part0);
90 	blk_steal_bios(&ns->head->requeue_list, req);
91 	spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
92 
93 	blk_mq_end_request(req, 0);
94 	kblockd_schedule_work(&ns->head->requeue_work);
95 }
96 
nvme_kick_requeue_lists(struct nvme_ctrl * ctrl)97 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
98 {
99 	struct nvme_ns *ns;
100 
101 	down_read(&ctrl->namespaces_rwsem);
102 	list_for_each_entry(ns, &ctrl->namespaces, list) {
103 		if (ns->head->disk)
104 			kblockd_schedule_work(&ns->head->requeue_work);
105 	}
106 	up_read(&ctrl->namespaces_rwsem);
107 }
108 
109 static const char *nvme_ana_state_names[] = {
110 	[0]				= "invalid state",
111 	[NVME_ANA_OPTIMIZED]		= "optimized",
112 	[NVME_ANA_NONOPTIMIZED]		= "non-optimized",
113 	[NVME_ANA_INACCESSIBLE]		= "inaccessible",
114 	[NVME_ANA_PERSISTENT_LOSS]	= "persistent-loss",
115 	[NVME_ANA_CHANGE]		= "change",
116 };
117 
nvme_mpath_clear_current_path(struct nvme_ns * ns)118 bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
119 {
120 	struct nvme_ns_head *head = ns->head;
121 	bool changed = false;
122 	int node;
123 
124 	if (!head)
125 		goto out;
126 
127 	for_each_node(node) {
128 		if (ns == rcu_access_pointer(head->current_path[node])) {
129 			rcu_assign_pointer(head->current_path[node], NULL);
130 			changed = true;
131 		}
132 	}
133 out:
134 	return changed;
135 }
136 
nvme_mpath_clear_ctrl_paths(struct nvme_ctrl * ctrl)137 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
138 {
139 	struct nvme_ns *ns;
140 
141 	down_read(&ctrl->namespaces_rwsem);
142 	list_for_each_entry(ns, &ctrl->namespaces, list) {
143 		nvme_mpath_clear_current_path(ns);
144 		kblockd_schedule_work(&ns->head->requeue_work);
145 	}
146 	up_read(&ctrl->namespaces_rwsem);
147 }
148 
nvme_mpath_revalidate_paths(struct nvme_ns * ns)149 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
150 {
151 	struct nvme_ns_head *head = ns->head;
152 	sector_t capacity = get_capacity(head->disk);
153 	int node;
154 	int srcu_idx;
155 
156 	srcu_idx = srcu_read_lock(&head->srcu);
157 	list_for_each_entry_rcu(ns, &head->list, siblings) {
158 		if (capacity != get_capacity(ns->disk))
159 			clear_bit(NVME_NS_READY, &ns->flags);
160 	}
161 	srcu_read_unlock(&head->srcu, srcu_idx);
162 
163 	for_each_node(node)
164 		rcu_assign_pointer(head->current_path[node], NULL);
165 	kblockd_schedule_work(&head->requeue_work);
166 }
167 
nvme_path_is_disabled(struct nvme_ns * ns)168 static bool nvme_path_is_disabled(struct nvme_ns *ns)
169 {
170 	/*
171 	 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should
172 	 * still be able to complete assuming that the controller is connected.
173 	 * Otherwise it will fail immediately and return to the requeue list.
174 	 */
175 	if (ns->ctrl->state != NVME_CTRL_LIVE &&
176 	    ns->ctrl->state != NVME_CTRL_DELETING)
177 		return true;
178 	if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) ||
179 	    !test_bit(NVME_NS_READY, &ns->flags))
180 		return true;
181 	return false;
182 }
183 
__nvme_find_path(struct nvme_ns_head * head,int node)184 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node)
185 {
186 	int found_distance = INT_MAX, fallback_distance = INT_MAX, distance;
187 	struct nvme_ns *found = NULL, *fallback = NULL, *ns;
188 
189 	list_for_each_entry_rcu(ns, &head->list, siblings) {
190 		if (nvme_path_is_disabled(ns))
191 			continue;
192 
193 		if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA)
194 			distance = node_distance(node, ns->ctrl->numa_node);
195 		else
196 			distance = LOCAL_DISTANCE;
197 
198 		switch (ns->ana_state) {
199 		case NVME_ANA_OPTIMIZED:
200 			if (distance < found_distance) {
201 				found_distance = distance;
202 				found = ns;
203 			}
204 			break;
205 		case NVME_ANA_NONOPTIMIZED:
206 			if (distance < fallback_distance) {
207 				fallback_distance = distance;
208 				fallback = ns;
209 			}
210 			break;
211 		default:
212 			break;
213 		}
214 	}
215 
216 	if (!found)
217 		found = fallback;
218 	if (found)
219 		rcu_assign_pointer(head->current_path[node], found);
220 	return found;
221 }
222 
nvme_next_ns(struct nvme_ns_head * head,struct nvme_ns * ns)223 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
224 		struct nvme_ns *ns)
225 {
226 	ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns,
227 			siblings);
228 	if (ns)
229 		return ns;
230 	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
231 }
232 
nvme_round_robin_path(struct nvme_ns_head * head,int node,struct nvme_ns * old)233 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
234 		int node, struct nvme_ns *old)
235 {
236 	struct nvme_ns *ns, *found = NULL;
237 
238 	if (list_is_singular(&head->list)) {
239 		if (nvme_path_is_disabled(old))
240 			return NULL;
241 		return old;
242 	}
243 
244 	for (ns = nvme_next_ns(head, old);
245 	     ns && ns != old;
246 	     ns = nvme_next_ns(head, ns)) {
247 		if (nvme_path_is_disabled(ns))
248 			continue;
249 
250 		if (ns->ana_state == NVME_ANA_OPTIMIZED) {
251 			found = ns;
252 			goto out;
253 		}
254 		if (ns->ana_state == NVME_ANA_NONOPTIMIZED)
255 			found = ns;
256 	}
257 
258 	/*
259 	 * The loop above skips the current path for round-robin semantics.
260 	 * Fall back to the current path if either:
261 	 *  - no other optimized path found and current is optimized,
262 	 *  - no other usable path found and current is usable.
263 	 */
264 	if (!nvme_path_is_disabled(old) &&
265 	    (old->ana_state == NVME_ANA_OPTIMIZED ||
266 	     (!found && old->ana_state == NVME_ANA_NONOPTIMIZED)))
267 		return old;
268 
269 	if (!found)
270 		return NULL;
271 out:
272 	rcu_assign_pointer(head->current_path[node], found);
273 	return found;
274 }
275 
nvme_path_is_optimized(struct nvme_ns * ns)276 static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
277 {
278 	return ns->ctrl->state == NVME_CTRL_LIVE &&
279 		ns->ana_state == NVME_ANA_OPTIMIZED;
280 }
281 
nvme_find_path(struct nvme_ns_head * head)282 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
283 {
284 	int node = numa_node_id();
285 	struct nvme_ns *ns;
286 
287 	ns = srcu_dereference(head->current_path[node], &head->srcu);
288 	if (unlikely(!ns))
289 		return __nvme_find_path(head, node);
290 
291 	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
292 		return nvme_round_robin_path(head, node, ns);
293 	if (unlikely(!nvme_path_is_optimized(ns)))
294 		return __nvme_find_path(head, node);
295 	return ns;
296 }
297 
nvme_available_path(struct nvme_ns_head * head)298 static bool nvme_available_path(struct nvme_ns_head *head)
299 {
300 	struct nvme_ns *ns;
301 
302 	list_for_each_entry_rcu(ns, &head->list, siblings) {
303 		if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags))
304 			continue;
305 		switch (ns->ctrl->state) {
306 		case NVME_CTRL_LIVE:
307 		case NVME_CTRL_RESETTING:
308 		case NVME_CTRL_CONNECTING:
309 			/* fallthru */
310 			return true;
311 		default:
312 			break;
313 		}
314 	}
315 	return false;
316 }
317 
nvme_ns_head_submit_bio(struct bio * bio)318 static blk_qc_t nvme_ns_head_submit_bio(struct bio *bio)
319 {
320 	struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data;
321 	struct device *dev = disk_to_dev(head->disk);
322 	struct nvme_ns *ns;
323 	blk_qc_t ret = BLK_QC_T_NONE;
324 	int srcu_idx;
325 
326 	/*
327 	 * The namespace might be going away and the bio might be moved to a
328 	 * different queue via blk_steal_bios(), so we need to use the bio_split
329 	 * pool from the original queue to allocate the bvecs from.
330 	 */
331 	blk_queue_split(&bio);
332 	if (!bio)
333 		return BLK_QC_T_NONE;
334 
335 	srcu_idx = srcu_read_lock(&head->srcu);
336 	ns = nvme_find_path(head);
337 	if (likely(ns)) {
338 		bio_set_dev(bio, ns->disk->part0);
339 		bio->bi_opf |= REQ_NVME_MPATH;
340 		trace_block_bio_remap(bio, disk_devt(ns->head->disk),
341 				      bio->bi_iter.bi_sector);
342 		ret = submit_bio_noacct(bio);
343 	} else if (nvme_available_path(head)) {
344 		dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n");
345 
346 		spin_lock_irq(&head->requeue_lock);
347 		bio_list_add(&head->requeue_list, bio);
348 		spin_unlock_irq(&head->requeue_lock);
349 	} else {
350 		dev_warn_ratelimited(dev, "no available path - failing I/O\n");
351 
352 		bio->bi_status = BLK_STS_IOERR;
353 		bio_endio(bio);
354 	}
355 
356 	srcu_read_unlock(&head->srcu, srcu_idx);
357 	return ret;
358 }
359 
nvme_ns_head_open(struct block_device * bdev,fmode_t mode)360 static int nvme_ns_head_open(struct block_device *bdev, fmode_t mode)
361 {
362 	if (!nvme_tryget_ns_head(bdev->bd_disk->private_data))
363 		return -ENXIO;
364 	return 0;
365 }
366 
nvme_ns_head_release(struct gendisk * disk,fmode_t mode)367 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode)
368 {
369 	nvme_put_ns_head(disk->private_data);
370 }
371 
372 #ifdef CONFIG_BLK_DEV_ZONED
nvme_ns_head_report_zones(struct gendisk * disk,sector_t sector,unsigned int nr_zones,report_zones_cb cb,void * data)373 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector,
374 		unsigned int nr_zones, report_zones_cb cb, void *data)
375 {
376 	struct nvme_ns_head *head = disk->private_data;
377 	struct nvme_ns *ns;
378 	int srcu_idx, ret = -EWOULDBLOCK;
379 
380 	srcu_idx = srcu_read_lock(&head->srcu);
381 	ns = nvme_find_path(head);
382 	if (ns)
383 		ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data);
384 	srcu_read_unlock(&head->srcu, srcu_idx);
385 	return ret;
386 }
387 #else
388 #define nvme_ns_head_report_zones	NULL
389 #endif /* CONFIG_BLK_DEV_ZONED */
390 
391 const struct block_device_operations nvme_ns_head_ops = {
392 	.owner		= THIS_MODULE,
393 	.submit_bio	= nvme_ns_head_submit_bio,
394 	.open		= nvme_ns_head_open,
395 	.release	= nvme_ns_head_release,
396 	.ioctl		= nvme_ns_head_ioctl,
397 	.compat_ioctl	= blkdev_compat_ptr_ioctl,
398 	.getgeo		= nvme_getgeo,
399 	.report_zones	= nvme_ns_head_report_zones,
400 	.pr_ops		= &nvme_pr_ops,
401 };
402 
cdev_to_ns_head(struct cdev * cdev)403 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev)
404 {
405 	return container_of(cdev, struct nvme_ns_head, cdev);
406 }
407 
nvme_ns_head_chr_open(struct inode * inode,struct file * file)408 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file)
409 {
410 	if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev)))
411 		return -ENXIO;
412 	return 0;
413 }
414 
nvme_ns_head_chr_release(struct inode * inode,struct file * file)415 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file)
416 {
417 	nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev));
418 	return 0;
419 }
420 
421 static const struct file_operations nvme_ns_head_chr_fops = {
422 	.owner		= THIS_MODULE,
423 	.open		= nvme_ns_head_chr_open,
424 	.release	= nvme_ns_head_chr_release,
425 	.unlocked_ioctl	= nvme_ns_head_chr_ioctl,
426 	.compat_ioctl	= compat_ptr_ioctl,
427 };
428 
nvme_add_ns_head_cdev(struct nvme_ns_head * head)429 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head)
430 {
431 	int ret;
432 
433 	head->cdev_device.parent = &head->subsys->dev;
434 	ret = dev_set_name(&head->cdev_device, "ng%dn%d",
435 			   head->subsys->instance, head->instance);
436 	if (ret)
437 		return ret;
438 	ret = nvme_cdev_add(&head->cdev, &head->cdev_device,
439 			    &nvme_ns_head_chr_fops, THIS_MODULE);
440 	return ret;
441 }
442 
nvme_requeue_work(struct work_struct * work)443 static void nvme_requeue_work(struct work_struct *work)
444 {
445 	struct nvme_ns_head *head =
446 		container_of(work, struct nvme_ns_head, requeue_work);
447 	struct bio *bio, *next;
448 
449 	spin_lock_irq(&head->requeue_lock);
450 	next = bio_list_get(&head->requeue_list);
451 	spin_unlock_irq(&head->requeue_lock);
452 
453 	while ((bio = next) != NULL) {
454 		next = bio->bi_next;
455 		bio->bi_next = NULL;
456 
457 		submit_bio_noacct(bio);
458 	}
459 }
460 
nvme_mpath_alloc_disk(struct nvme_ctrl * ctrl,struct nvme_ns_head * head)461 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
462 {
463 	bool vwc = false;
464 
465 	mutex_init(&head->lock);
466 	bio_list_init(&head->requeue_list);
467 	spin_lock_init(&head->requeue_lock);
468 	INIT_WORK(&head->requeue_work, nvme_requeue_work);
469 
470 	/*
471 	 * Add a multipath node if the subsystems supports multiple controllers.
472 	 * We also do this for private namespaces as the namespace sharing flag
473 	 * could change after a rescan.
474 	 */
475 	if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
476 	    !nvme_is_unique_nsid(ctrl, head) || !multipath)
477 		return 0;
478 
479 	head->disk = blk_alloc_disk(ctrl->numa_node);
480 	if (!head->disk)
481 		return -ENOMEM;
482 	head->disk->fops = &nvme_ns_head_ops;
483 	head->disk->private_data = head;
484 	sprintf(head->disk->disk_name, "nvme%dn%d",
485 			ctrl->subsys->instance, head->instance);
486 
487 	blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue);
488 	blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue);
489 
490 	/* set to a default value of 512 until the disk is validated */
491 	blk_queue_logical_block_size(head->disk->queue, 512);
492 	blk_set_stacking_limits(&head->disk->queue->limits);
493 
494 	/* we need to propagate up the VMC settings */
495 	if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
496 		vwc = true;
497 	blk_queue_write_cache(head->disk->queue, vwc, vwc);
498 	return 0;
499 }
500 
nvme_mpath_set_live(struct nvme_ns * ns)501 static void nvme_mpath_set_live(struct nvme_ns *ns)
502 {
503 	struct nvme_ns_head *head = ns->head;
504 
505 	if (!head->disk)
506 		return;
507 
508 	if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
509 		device_add_disk(&head->subsys->dev, head->disk,
510 				nvme_ns_id_attr_groups);
511 		nvme_add_ns_head_cdev(head);
512 	}
513 
514 	mutex_lock(&head->lock);
515 	if (nvme_path_is_optimized(ns)) {
516 		int node, srcu_idx;
517 
518 		srcu_idx = srcu_read_lock(&head->srcu);
519 		for_each_node(node)
520 			__nvme_find_path(head, node);
521 		srcu_read_unlock(&head->srcu, srcu_idx);
522 	}
523 	mutex_unlock(&head->lock);
524 
525 	synchronize_srcu(&head->srcu);
526 	kblockd_schedule_work(&head->requeue_work);
527 }
528 
nvme_parse_ana_log(struct nvme_ctrl * ctrl,void * data,int (* cb)(struct nvme_ctrl * ctrl,struct nvme_ana_group_desc *,void *))529 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data,
530 		int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *,
531 			void *))
532 {
533 	void *base = ctrl->ana_log_buf;
534 	size_t offset = sizeof(struct nvme_ana_rsp_hdr);
535 	int error, i;
536 
537 	lockdep_assert_held(&ctrl->ana_lock);
538 
539 	for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) {
540 		struct nvme_ana_group_desc *desc = base + offset;
541 		u32 nr_nsids;
542 		size_t nsid_buf_size;
543 
544 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc)))
545 			return -EINVAL;
546 
547 		nr_nsids = le32_to_cpu(desc->nnsids);
548 		nsid_buf_size = nr_nsids * sizeof(__le32);
549 
550 		if (WARN_ON_ONCE(desc->grpid == 0))
551 			return -EINVAL;
552 		if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax))
553 			return -EINVAL;
554 		if (WARN_ON_ONCE(desc->state == 0))
555 			return -EINVAL;
556 		if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE))
557 			return -EINVAL;
558 
559 		offset += sizeof(*desc);
560 		if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size))
561 			return -EINVAL;
562 
563 		error = cb(ctrl, desc, data);
564 		if (error)
565 			return error;
566 
567 		offset += nsid_buf_size;
568 	}
569 
570 	return 0;
571 }
572 
nvme_state_is_live(enum nvme_ana_state state)573 static inline bool nvme_state_is_live(enum nvme_ana_state state)
574 {
575 	return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED;
576 }
577 
nvme_update_ns_ana_state(struct nvme_ana_group_desc * desc,struct nvme_ns * ns)578 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc,
579 		struct nvme_ns *ns)
580 {
581 	ns->ana_grpid = le32_to_cpu(desc->grpid);
582 	ns->ana_state = desc->state;
583 	clear_bit(NVME_NS_ANA_PENDING, &ns->flags);
584 	/*
585 	 * nvme_mpath_set_live() will trigger I/O to the multipath path device
586 	 * and in turn to this path device.  However we cannot accept this I/O
587 	 * if the controller is not live.  This may deadlock if called from
588 	 * nvme_mpath_init_identify() and the ctrl will never complete
589 	 * initialization, preventing I/O from completing.  For this case we
590 	 * will reprocess the ANA log page in nvme_mpath_update() once the
591 	 * controller is ready.
592 	 */
593 	if (nvme_state_is_live(ns->ana_state) &&
594 	    ns->ctrl->state == NVME_CTRL_LIVE)
595 		nvme_mpath_set_live(ns);
596 }
597 
nvme_update_ana_state(struct nvme_ctrl * ctrl,struct nvme_ana_group_desc * desc,void * data)598 static int nvme_update_ana_state(struct nvme_ctrl *ctrl,
599 		struct nvme_ana_group_desc *desc, void *data)
600 {
601 	u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0;
602 	unsigned *nr_change_groups = data;
603 	struct nvme_ns *ns;
604 
605 	dev_dbg(ctrl->device, "ANA group %d: %s.\n",
606 			le32_to_cpu(desc->grpid),
607 			nvme_ana_state_names[desc->state]);
608 
609 	if (desc->state == NVME_ANA_CHANGE)
610 		(*nr_change_groups)++;
611 
612 	if (!nr_nsids)
613 		return 0;
614 
615 	down_read(&ctrl->namespaces_rwsem);
616 	list_for_each_entry(ns, &ctrl->namespaces, list) {
617 		unsigned nsid;
618 again:
619 		nsid = le32_to_cpu(desc->nsids[n]);
620 		if (ns->head->ns_id < nsid)
621 			continue;
622 		if (ns->head->ns_id == nsid)
623 			nvme_update_ns_ana_state(desc, ns);
624 		if (++n == nr_nsids)
625 			break;
626 		if (ns->head->ns_id > nsid)
627 			goto again;
628 	}
629 	up_read(&ctrl->namespaces_rwsem);
630 	return 0;
631 }
632 
nvme_read_ana_log(struct nvme_ctrl * ctrl)633 static int nvme_read_ana_log(struct nvme_ctrl *ctrl)
634 {
635 	u32 nr_change_groups = 0;
636 	int error;
637 
638 	mutex_lock(&ctrl->ana_lock);
639 	error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM,
640 			ctrl->ana_log_buf, ctrl->ana_log_size, 0);
641 	if (error) {
642 		dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error);
643 		goto out_unlock;
644 	}
645 
646 	error = nvme_parse_ana_log(ctrl, &nr_change_groups,
647 			nvme_update_ana_state);
648 	if (error)
649 		goto out_unlock;
650 
651 	/*
652 	 * In theory we should have an ANATT timer per group as they might enter
653 	 * the change state at different times.  But that is a lot of overhead
654 	 * just to protect against a target that keeps entering new changes
655 	 * states while never finishing previous ones.  But we'll still
656 	 * eventually time out once all groups are in change state, so this
657 	 * isn't a big deal.
658 	 *
659 	 * We also double the ANATT value to provide some slack for transports
660 	 * or AEN processing overhead.
661 	 */
662 	if (nr_change_groups)
663 		mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies);
664 	else
665 		del_timer_sync(&ctrl->anatt_timer);
666 out_unlock:
667 	mutex_unlock(&ctrl->ana_lock);
668 	return error;
669 }
670 
nvme_ana_work(struct work_struct * work)671 static void nvme_ana_work(struct work_struct *work)
672 {
673 	struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work);
674 
675 	if (ctrl->state != NVME_CTRL_LIVE)
676 		return;
677 
678 	nvme_read_ana_log(ctrl);
679 }
680 
nvme_mpath_update(struct nvme_ctrl * ctrl)681 void nvme_mpath_update(struct nvme_ctrl *ctrl)
682 {
683 	u32 nr_change_groups = 0;
684 
685 	if (!ctrl->ana_log_buf)
686 		return;
687 
688 	mutex_lock(&ctrl->ana_lock);
689 	nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state);
690 	mutex_unlock(&ctrl->ana_lock);
691 }
692 
nvme_anatt_timeout(struct timer_list * t)693 static void nvme_anatt_timeout(struct timer_list *t)
694 {
695 	struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer);
696 
697 	dev_info(ctrl->device, "ANATT timeout, resetting controller.\n");
698 	nvme_reset_ctrl(ctrl);
699 }
700 
nvme_mpath_stop(struct nvme_ctrl * ctrl)701 void nvme_mpath_stop(struct nvme_ctrl *ctrl)
702 {
703 	if (!nvme_ctrl_use_ana(ctrl))
704 		return;
705 	del_timer_sync(&ctrl->anatt_timer);
706 	cancel_work_sync(&ctrl->ana_work);
707 }
708 
709 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store)  \
710 	struct device_attribute subsys_attr_##_name =	\
711 		__ATTR(_name, _mode, _show, _store)
712 
713 static const char *nvme_iopolicy_names[] = {
714 	[NVME_IOPOLICY_NUMA]	= "numa",
715 	[NVME_IOPOLICY_RR]	= "round-robin",
716 };
717 
nvme_subsys_iopolicy_show(struct device * dev,struct device_attribute * attr,char * buf)718 static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
719 		struct device_attribute *attr, char *buf)
720 {
721 	struct nvme_subsystem *subsys =
722 		container_of(dev, struct nvme_subsystem, dev);
723 
724 	return sysfs_emit(buf, "%s\n",
725 			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
726 }
727 
nvme_subsys_iopolicy_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)728 static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
729 		struct device_attribute *attr, const char *buf, size_t count)
730 {
731 	struct nvme_subsystem *subsys =
732 		container_of(dev, struct nvme_subsystem, dev);
733 	int i;
734 
735 	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
736 		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
737 			WRITE_ONCE(subsys->iopolicy, i);
738 			return count;
739 		}
740 	}
741 
742 	return -EINVAL;
743 }
744 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR,
745 		      nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store);
746 
ana_grpid_show(struct device * dev,struct device_attribute * attr,char * buf)747 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr,
748 		char *buf)
749 {
750 	return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid);
751 }
752 DEVICE_ATTR_RO(ana_grpid);
753 
ana_state_show(struct device * dev,struct device_attribute * attr,char * buf)754 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr,
755 		char *buf)
756 {
757 	struct nvme_ns *ns = nvme_get_ns_from_dev(dev);
758 
759 	return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]);
760 }
761 DEVICE_ATTR_RO(ana_state);
762 
nvme_lookup_ana_group_desc(struct nvme_ctrl * ctrl,struct nvme_ana_group_desc * desc,void * data)763 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl,
764 		struct nvme_ana_group_desc *desc, void *data)
765 {
766 	struct nvme_ana_group_desc *dst = data;
767 
768 	if (desc->grpid != dst->grpid)
769 		return 0;
770 
771 	*dst = *desc;
772 	return -ENXIO; /* just break out of the loop */
773 }
774 
nvme_mpath_add_disk(struct nvme_ns * ns,struct nvme_id_ns * id)775 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id)
776 {
777 	if (nvme_ctrl_use_ana(ns->ctrl)) {
778 		struct nvme_ana_group_desc desc = {
779 			.grpid = id->anagrpid,
780 			.state = 0,
781 		};
782 
783 		mutex_lock(&ns->ctrl->ana_lock);
784 		ns->ana_grpid = le32_to_cpu(id->anagrpid);
785 		nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc);
786 		mutex_unlock(&ns->ctrl->ana_lock);
787 		if (desc.state) {
788 			/* found the group desc: update */
789 			nvme_update_ns_ana_state(&desc, ns);
790 		} else {
791 			/* group desc not found: trigger a re-read */
792 			set_bit(NVME_NS_ANA_PENDING, &ns->flags);
793 			queue_work(nvme_wq, &ns->ctrl->ana_work);
794 		}
795 	} else {
796 		ns->ana_state = NVME_ANA_OPTIMIZED;
797 		nvme_mpath_set_live(ns);
798 	}
799 
800 	if (blk_queue_stable_writes(ns->queue) && ns->head->disk)
801 		blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES,
802 				   ns->head->disk->queue);
803 #ifdef CONFIG_BLK_DEV_ZONED
804 	if (blk_queue_is_zoned(ns->queue) && ns->head->disk)
805 		ns->head->disk->queue->nr_zones = ns->queue->nr_zones;
806 #endif
807 }
808 
nvme_mpath_shutdown_disk(struct nvme_ns_head * head)809 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
810 {
811 	if (!head->disk)
812 		return;
813 	kblockd_schedule_work(&head->requeue_work);
814 	if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) {
815 		nvme_cdev_del(&head->cdev, &head->cdev_device);
816 		del_gendisk(head->disk);
817 	}
818 }
819 
nvme_mpath_remove_disk(struct nvme_ns_head * head)820 void nvme_mpath_remove_disk(struct nvme_ns_head *head)
821 {
822 	if (!head->disk)
823 		return;
824 	/* make sure all pending bios are cleaned up */
825 	kblockd_schedule_work(&head->requeue_work);
826 	flush_work(&head->requeue_work);
827 	blk_cleanup_disk(head->disk);
828 }
829 
nvme_mpath_init_ctrl(struct nvme_ctrl * ctrl)830 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
831 {
832 	mutex_init(&ctrl->ana_lock);
833 	timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0);
834 	INIT_WORK(&ctrl->ana_work, nvme_ana_work);
835 }
836 
nvme_mpath_init_identify(struct nvme_ctrl * ctrl,struct nvme_id_ctrl * id)837 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
838 {
839 	size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT;
840 	size_t ana_log_size;
841 	int error = 0;
842 
843 	/* check if multipath is enabled and we have the capability */
844 	if (!multipath || !ctrl->subsys ||
845 	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
846 		return 0;
847 
848 	if (!ctrl->max_namespaces ||
849 	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
850 		dev_err(ctrl->device,
851 			"Invalid MNAN value %u\n", ctrl->max_namespaces);
852 		return -EINVAL;
853 	}
854 
855 	ctrl->anacap = id->anacap;
856 	ctrl->anatt = id->anatt;
857 	ctrl->nanagrpid = le32_to_cpu(id->nanagrpid);
858 	ctrl->anagrpmax = le32_to_cpu(id->anagrpmax);
859 
860 	ana_log_size = sizeof(struct nvme_ana_rsp_hdr) +
861 		ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) +
862 		ctrl->max_namespaces * sizeof(__le32);
863 	if (ana_log_size > max_transfer_size) {
864 		dev_err(ctrl->device,
865 			"ANA log page size (%zd) larger than MDTS (%zd).\n",
866 			ana_log_size, max_transfer_size);
867 		dev_err(ctrl->device, "disabling ANA support.\n");
868 		goto out_uninit;
869 	}
870 	if (ana_log_size > ctrl->ana_log_size) {
871 		nvme_mpath_stop(ctrl);
872 		kfree(ctrl->ana_log_buf);
873 		ctrl->ana_log_buf = kmalloc(ana_log_size, GFP_KERNEL);
874 		if (!ctrl->ana_log_buf)
875 			return -ENOMEM;
876 	}
877 	ctrl->ana_log_size = ana_log_size;
878 	error = nvme_read_ana_log(ctrl);
879 	if (error)
880 		goto out_uninit;
881 	return 0;
882 
883 out_uninit:
884 	nvme_mpath_uninit(ctrl);
885 	return error;
886 }
887 
nvme_mpath_uninit(struct nvme_ctrl * ctrl)888 void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
889 {
890 	kfree(ctrl->ana_log_buf);
891 	ctrl->ana_log_buf = NULL;
892 }
893