xref: /linux/drivers/md/dm-mpath.c (revision be54f8c558027a218423134dd9b8c7c46d92204a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2003 Sistina Software Limited.
4  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5  *
6  * This file is released under the GPL.
7  */
8 
9 #include <linux/device-mapper.h>
10 
11 #include "dm-rq.h"
12 #include "dm-bio-record.h"
13 #include "dm-path-selector.h"
14 #include "dm-uevent.h"
15 
16 #include <linux/blkdev.h>
17 #include <linux/ctype.h>
18 #include <linux/init.h>
19 #include <linux/mempool.h>
20 #include <linux/module.h>
21 #include <linux/pagemap.h>
22 #include <linux/slab.h>
23 #include <linux/time.h>
24 #include <linux/timer.h>
25 #include <linux/workqueue.h>
26 #include <linux/delay.h>
27 #include <scsi/scsi_dh.h>
28 #include <linux/atomic.h>
29 #include <linux/blk-mq.h>
30 
31 static struct workqueue_struct *dm_mpath_wq;
32 
33 #define DM_MSG_PREFIX "multipath"
34 #define DM_PG_INIT_DELAY_MSECS 2000
35 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned int) -1)
36 #define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
37 
38 static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
39 
40 /* Path properties */
41 struct pgpath {
42 	struct list_head list;
43 
44 	struct priority_group *pg;	/* Owning PG */
45 	unsigned int fail_count;		/* Cumulative failure count */
46 
47 	struct dm_path path;
48 	struct delayed_work activate_path;
49 
50 	bool is_active:1;		/* Path status */
51 };
52 
53 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
54 
55 /*
56  * Paths are grouped into Priority Groups and numbered from 1 upwards.
57  * Each has a path selector which controls which path gets used.
58  */
59 struct priority_group {
60 	struct list_head list;
61 
62 	struct multipath *m;		/* Owning multipath instance */
63 	struct path_selector ps;
64 
65 	unsigned int pg_num;		/* Reference number */
66 	unsigned int nr_pgpaths;		/* Number of paths in PG */
67 	struct list_head pgpaths;
68 
69 	bool bypassed:1;		/* Temporarily bypass this PG? */
70 };
71 
72 /* Multipath context */
73 struct multipath {
74 	unsigned long flags;		/* Multipath state flags */
75 
76 	spinlock_t lock;
77 	enum dm_queue_mode queue_mode;
78 
79 	struct pgpath *current_pgpath;
80 	struct priority_group *current_pg;
81 	struct priority_group *next_pg;	/* Switch to this PG if set */
82 	struct priority_group *last_probed_pg;
83 
84 	atomic_t nr_valid_paths;	/* Total number of usable paths */
85 	unsigned int nr_priority_groups;
86 	struct list_head priority_groups;
87 
88 	const char *hw_handler_name;
89 	char *hw_handler_params;
90 	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */
91 	wait_queue_head_t probe_wait;   /* Wait for probing paths */
92 	unsigned int pg_init_retries;	/* Number of times to retry pg_init */
93 	unsigned int pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
94 	atomic_t pg_init_in_progress;	/* Only one pg_init allowed at once */
95 	atomic_t pg_init_count;		/* Number of times pg_init called */
96 
97 	struct mutex work_mutex;
98 	struct work_struct trigger_event;
99 	struct dm_target *ti;
100 
101 	struct work_struct process_queued_bios;
102 	struct bio_list queued_bios;
103 
104 	struct timer_list nopath_timer;	/* Timeout for queue_if_no_path */
105 	bool is_suspending;
106 };
107 
108 /*
109  * Context information attached to each io we process.
110  */
111 struct dm_mpath_io {
112 	struct pgpath *pgpath;
113 	size_t nr_bytes;
114 	u64 start_time_ns;
115 };
116 
117 typedef int (*action_fn) (struct pgpath *pgpath);
118 
119 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
120 static void trigger_event(struct work_struct *work);
121 static void activate_or_offline_path(struct pgpath *pgpath);
122 static void activate_path_work(struct work_struct *work);
123 static void process_queued_bios(struct work_struct *work);
124 static void queue_if_no_path_timeout_work(struct timer_list *t);
125 
126 /*
127  *-----------------------------------------------
128  * Multipath state flags.
129  *-----------------------------------------------
130  */
131 #define MPATHF_QUEUE_IO 0			/* Must we queue all I/O? */
132 #define MPATHF_QUEUE_IF_NO_PATH 1		/* Queue I/O if last path fails? */
133 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2		/* Saved state during suspension */
134 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3	/* If there's already a hw_handler present, don't change it. */
135 #define MPATHF_PG_INIT_DISABLED 4		/* pg_init is not currently allowed */
136 #define MPATHF_PG_INIT_REQUIRED 5		/* pg_init needs calling? */
137 #define MPATHF_PG_INIT_DELAY_RETRY 6		/* Delay pg_init retry? */
138 #define MPATHF_DELAY_PG_SWITCH 7		/* Delay switching pg if it still has paths */
139 #define MPATHF_NEED_PG_SWITCH 8			/* Need to switch pgs after the delay has ended */
140 
mpath_double_check_test_bit(int MPATHF_bit,struct multipath * m)141 static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
142 {
143 	bool r = test_bit(MPATHF_bit, &m->flags);
144 
145 	if (r) {
146 		unsigned long flags;
147 
148 		spin_lock_irqsave(&m->lock, flags);
149 		r = test_bit(MPATHF_bit, &m->flags);
150 		spin_unlock_irqrestore(&m->lock, flags);
151 	}
152 
153 	return r;
154 }
155 
156 /*
157  *-----------------------------------------------
158  * Allocation routines
159  *-----------------------------------------------
160  */
alloc_pgpath(void)161 static struct pgpath *alloc_pgpath(void)
162 {
163 	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
164 
165 	if (!pgpath)
166 		return NULL;
167 
168 	pgpath->is_active = true;
169 
170 	return pgpath;
171 }
172 
free_pgpath(struct pgpath * pgpath)173 static void free_pgpath(struct pgpath *pgpath)
174 {
175 	kfree(pgpath);
176 }
177 
alloc_priority_group(void)178 static struct priority_group *alloc_priority_group(void)
179 {
180 	struct priority_group *pg;
181 
182 	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
183 
184 	if (pg)
185 		INIT_LIST_HEAD(&pg->pgpaths);
186 
187 	return pg;
188 }
189 
free_pgpaths(struct list_head * pgpaths,struct dm_target * ti)190 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
191 {
192 	struct pgpath *pgpath, *tmp;
193 
194 	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
195 		list_del(&pgpath->list);
196 		dm_put_device(ti, pgpath->path.dev);
197 		free_pgpath(pgpath);
198 	}
199 }
200 
free_priority_group(struct priority_group * pg,struct dm_target * ti)201 static void free_priority_group(struct priority_group *pg,
202 				struct dm_target *ti)
203 {
204 	struct path_selector *ps = &pg->ps;
205 
206 	if (ps->type) {
207 		ps->type->destroy(ps);
208 		dm_put_path_selector(ps->type);
209 	}
210 
211 	free_pgpaths(&pg->pgpaths, ti);
212 	kfree(pg);
213 }
214 
alloc_multipath(struct dm_target * ti)215 static struct multipath *alloc_multipath(struct dm_target *ti)
216 {
217 	struct multipath *m;
218 
219 	m = kzalloc(sizeof(*m), GFP_KERNEL);
220 	if (m) {
221 		INIT_LIST_HEAD(&m->priority_groups);
222 		spin_lock_init(&m->lock);
223 		atomic_set(&m->nr_valid_paths, 0);
224 		INIT_WORK(&m->trigger_event, trigger_event);
225 		mutex_init(&m->work_mutex);
226 
227 		m->queue_mode = DM_TYPE_NONE;
228 
229 		m->ti = ti;
230 		ti->private = m;
231 
232 		timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
233 	}
234 
235 	return m;
236 }
237 
alloc_multipath_stage2(struct dm_target * ti,struct multipath * m)238 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
239 {
240 	if (m->queue_mode == DM_TYPE_NONE) {
241 		m->queue_mode = DM_TYPE_REQUEST_BASED;
242 	} else if (m->queue_mode == DM_TYPE_BIO_BASED) {
243 		INIT_WORK(&m->process_queued_bios, process_queued_bios);
244 		/*
245 		 * bio-based doesn't support any direct scsi_dh management;
246 		 * it just discovers if a scsi_dh is attached.
247 		 */
248 		set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
249 	}
250 
251 	dm_table_set_type(ti->table, m->queue_mode);
252 
253 	/*
254 	 * Init fields that are only used when a scsi_dh is attached
255 	 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
256 	 */
257 	set_bit(MPATHF_QUEUE_IO, &m->flags);
258 	atomic_set(&m->pg_init_in_progress, 0);
259 	atomic_set(&m->pg_init_count, 0);
260 	m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
261 	init_waitqueue_head(&m->pg_init_wait);
262 	init_waitqueue_head(&m->probe_wait);
263 
264 	return 0;
265 }
266 
free_multipath(struct multipath * m)267 static void free_multipath(struct multipath *m)
268 {
269 	struct priority_group *pg, *tmp;
270 
271 	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
272 		list_del(&pg->list);
273 		free_priority_group(pg, m->ti);
274 	}
275 
276 	kfree(m->hw_handler_name);
277 	kfree(m->hw_handler_params);
278 	mutex_destroy(&m->work_mutex);
279 	kfree(m);
280 }
281 
get_mpio(union map_info * info)282 static struct dm_mpath_io *get_mpio(union map_info *info)
283 {
284 	return info->ptr;
285 }
286 
multipath_per_bio_data_size(void)287 static size_t multipath_per_bio_data_size(void)
288 {
289 	return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
290 }
291 
get_mpio_from_bio(struct bio * bio)292 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
293 {
294 	return dm_per_bio_data(bio, multipath_per_bio_data_size());
295 }
296 
get_bio_details_from_mpio(struct dm_mpath_io * mpio)297 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
298 {
299 	/* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
300 	void *bio_details = mpio + 1;
301 	return bio_details;
302 }
303 
multipath_init_per_bio_data(struct bio * bio,struct dm_mpath_io ** mpio_p)304 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
305 {
306 	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
307 	struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
308 
309 	mpio->nr_bytes = bio->bi_iter.bi_size;
310 	mpio->pgpath = NULL;
311 	mpio->start_time_ns = 0;
312 	*mpio_p = mpio;
313 
314 	dm_bio_record(bio_details, bio);
315 }
316 
317 /*
318  *-----------------------------------------------
319  * Path selection
320  *-----------------------------------------------
321  */
__pg_init_all_paths(struct multipath * m)322 static int __pg_init_all_paths(struct multipath *m)
323 {
324 	struct pgpath *pgpath;
325 	unsigned long pg_init_delay = 0;
326 
327 	lockdep_assert_held(&m->lock);
328 
329 	if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
330 		return 0;
331 
332 	atomic_inc(&m->pg_init_count);
333 	clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
334 
335 	/* Check here to reset pg_init_required */
336 	if (!m->current_pg)
337 		return 0;
338 
339 	if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
340 		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
341 						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
342 	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
343 		/* Skip failed paths */
344 		if (!pgpath->is_active)
345 			continue;
346 		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
347 				       pg_init_delay))
348 			atomic_inc(&m->pg_init_in_progress);
349 	}
350 	return atomic_read(&m->pg_init_in_progress);
351 }
352 
pg_init_all_paths(struct multipath * m)353 static int pg_init_all_paths(struct multipath *m)
354 {
355 	int ret;
356 	unsigned long flags;
357 
358 	spin_lock_irqsave(&m->lock, flags);
359 	ret = __pg_init_all_paths(m);
360 	spin_unlock_irqrestore(&m->lock, flags);
361 
362 	return ret;
363 }
364 
__switch_pg(struct multipath * m,struct priority_group * pg)365 static void __switch_pg(struct multipath *m, struct priority_group *pg)
366 {
367 	lockdep_assert_held(&m->lock);
368 
369 	m->current_pg = pg;
370 
371 	/* Must we initialise the PG first, and queue I/O till it's ready? */
372 	if (m->hw_handler_name) {
373 		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
374 		set_bit(MPATHF_QUEUE_IO, &m->flags);
375 	} else {
376 		clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
377 		clear_bit(MPATHF_QUEUE_IO, &m->flags);
378 	}
379 
380 	atomic_set(&m->pg_init_count, 0);
381 }
382 
choose_path_in_pg(struct multipath * m,struct priority_group * pg,size_t nr_bytes)383 static struct pgpath *choose_path_in_pg(struct multipath *m,
384 					struct priority_group *pg,
385 					size_t nr_bytes)
386 {
387 	unsigned long flags;
388 	struct dm_path *path;
389 	struct pgpath *pgpath;
390 
391 	path = pg->ps.type->select_path(&pg->ps, nr_bytes);
392 	if (!path)
393 		return ERR_PTR(-ENXIO);
394 
395 	pgpath = path_to_pgpath(path);
396 
397 	if (unlikely(READ_ONCE(m->current_pg) != pg)) {
398 		/* Only update current_pgpath if pg changed */
399 		spin_lock_irqsave(&m->lock, flags);
400 		m->current_pgpath = pgpath;
401 		__switch_pg(m, pg);
402 		spin_unlock_irqrestore(&m->lock, flags);
403 	}
404 
405 	return pgpath;
406 }
407 
choose_pgpath(struct multipath * m,size_t nr_bytes)408 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
409 {
410 	unsigned long flags;
411 	struct priority_group *pg;
412 	struct pgpath *pgpath;
413 	unsigned int bypassed = 1;
414 
415 	if (!atomic_read(&m->nr_valid_paths)) {
416 		spin_lock_irqsave(&m->lock, flags);
417 		clear_bit(MPATHF_QUEUE_IO, &m->flags);
418 		spin_unlock_irqrestore(&m->lock, flags);
419 		goto failed;
420 	}
421 
422 	/* Don't change PG until it has no remaining paths */
423 	pg = READ_ONCE(m->current_pg);
424 	if (pg) {
425 		pgpath = choose_path_in_pg(m, pg, nr_bytes);
426 		if (!IS_ERR_OR_NULL(pgpath))
427 			return pgpath;
428 	}
429 
430 	/* Were we instructed to switch PG? */
431 	if (READ_ONCE(m->next_pg)) {
432 		spin_lock_irqsave(&m->lock, flags);
433 		pg = m->next_pg;
434 		if (!pg) {
435 			spin_unlock_irqrestore(&m->lock, flags);
436 			goto check_all_pgs;
437 		}
438 		m->next_pg = NULL;
439 		spin_unlock_irqrestore(&m->lock, flags);
440 		pgpath = choose_path_in_pg(m, pg, nr_bytes);
441 		if (!IS_ERR_OR_NULL(pgpath))
442 			return pgpath;
443 	}
444 check_all_pgs:
445 	/*
446 	 * Loop through priority groups until we find a valid path.
447 	 * First time we skip PGs marked 'bypassed'.
448 	 * Second time we only try the ones we skipped, but set
449 	 * pg_init_delay_retry so we do not hammer controllers.
450 	 */
451 	do {
452 		list_for_each_entry(pg, &m->priority_groups, list) {
453 			if (pg->bypassed == !!bypassed)
454 				continue;
455 			pgpath = choose_path_in_pg(m, pg, nr_bytes);
456 			if (!IS_ERR_OR_NULL(pgpath)) {
457 				if (!bypassed) {
458 					spin_lock_irqsave(&m->lock, flags);
459 					set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
460 					spin_unlock_irqrestore(&m->lock, flags);
461 				}
462 				return pgpath;
463 			}
464 		}
465 	} while (bypassed--);
466 
467 failed:
468 	spin_lock_irqsave(&m->lock, flags);
469 	m->current_pgpath = NULL;
470 	m->current_pg = NULL;
471 	spin_unlock_irqrestore(&m->lock, flags);
472 
473 	return NULL;
474 }
475 
476 /*
477  * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
478  * report the function name and line number of the function from which
479  * it has been invoked.
480  */
481 #define dm_report_EIO(m)						\
482 	DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
483 		      dm_table_device_name((m)->ti->table),		\
484 		      test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),	\
485 		      test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
486 		      dm_noflush_suspending((m)->ti))
487 
488 /*
489  * Check whether bios must be queued in the device-mapper core rather
490  * than here in the target.
491  */
__must_push_back(struct multipath * m)492 static bool __must_push_back(struct multipath *m)
493 {
494 	return dm_noflush_suspending(m->ti);
495 }
496 
must_push_back_rq(struct multipath * m)497 static bool must_push_back_rq(struct multipath *m)
498 {
499 	unsigned long flags;
500 	bool ret;
501 
502 	spin_lock_irqsave(&m->lock, flags);
503 	ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
504 	spin_unlock_irqrestore(&m->lock, flags);
505 
506 	return ret;
507 }
508 
509 /*
510  * Map cloned requests (request-based multipath)
511  */
multipath_clone_and_map(struct dm_target * ti,struct request * rq,union map_info * map_context,struct request ** __clone)512 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
513 				   union map_info *map_context,
514 				   struct request **__clone)
515 {
516 	struct multipath *m = ti->private;
517 	size_t nr_bytes = blk_rq_bytes(rq);
518 	struct pgpath *pgpath;
519 	struct block_device *bdev;
520 	struct dm_mpath_io *mpio = get_mpio(map_context);
521 	struct request_queue *q;
522 	struct request *clone;
523 
524 	/* Do we need to select a new pgpath? */
525 	pgpath = READ_ONCE(m->current_pgpath);
526 	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
527 		pgpath = choose_pgpath(m, nr_bytes);
528 
529 	if (!pgpath) {
530 		if (must_push_back_rq(m))
531 			return DM_MAPIO_DELAY_REQUEUE;
532 		dm_report_EIO(m);	/* Failed */
533 		return DM_MAPIO_KILL;
534 	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
535 		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
536 		pg_init_all_paths(m);
537 		return DM_MAPIO_DELAY_REQUEUE;
538 	}
539 
540 	mpio->pgpath = pgpath;
541 	mpio->nr_bytes = nr_bytes;
542 
543 	bdev = pgpath->path.dev->bdev;
544 	q = bdev_get_queue(bdev);
545 	clone = blk_mq_alloc_request(q, rq->cmd_flags | REQ_NOMERGE,
546 			BLK_MQ_REQ_NOWAIT);
547 	if (IS_ERR(clone)) {
548 		/* EBUSY, ENODEV or EWOULDBLOCK: requeue */
549 		if (blk_queue_dying(q)) {
550 			atomic_inc(&m->pg_init_in_progress);
551 			activate_or_offline_path(pgpath);
552 			return DM_MAPIO_DELAY_REQUEUE;
553 		}
554 
555 		/*
556 		 * blk-mq's SCHED_RESTART can cover this requeue, so we
557 		 * needn't deal with it by DELAY_REQUEUE. More importantly,
558 		 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
559 		 * get the queue busy feedback (via BLK_STS_RESOURCE),
560 		 * otherwise I/O merging can suffer.
561 		 */
562 		return DM_MAPIO_REQUEUE;
563 	}
564 	clone->bio = clone->biotail = NULL;
565 	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
566 	*__clone = clone;
567 
568 	if (pgpath->pg->ps.type->start_io)
569 		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
570 					      &pgpath->path,
571 					      nr_bytes);
572 	return DM_MAPIO_REMAPPED;
573 }
574 
multipath_release_clone(struct request * clone,union map_info * map_context)575 static void multipath_release_clone(struct request *clone,
576 				    union map_info *map_context)
577 {
578 	if (unlikely(map_context)) {
579 		/*
580 		 * non-NULL map_context means caller is still map
581 		 * method; must undo multipath_clone_and_map()
582 		 */
583 		struct dm_mpath_io *mpio = get_mpio(map_context);
584 		struct pgpath *pgpath = mpio->pgpath;
585 
586 		if (pgpath && pgpath->pg->ps.type->end_io)
587 			pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
588 						    &pgpath->path,
589 						    mpio->nr_bytes,
590 						    clone->io_start_time_ns);
591 	}
592 
593 	blk_mq_free_request(clone);
594 }
595 
596 /*
597  * Map cloned bios (bio-based multipath)
598  */
599 
__multipath_queue_bio(struct multipath * m,struct bio * bio)600 static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
601 {
602 	/* Queue for the daemon to resubmit */
603 	bio_list_add(&m->queued_bios, bio);
604 	if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
605 		queue_work(kmultipathd, &m->process_queued_bios);
606 }
607 
multipath_queue_bio(struct multipath * m,struct bio * bio)608 static void multipath_queue_bio(struct multipath *m, struct bio *bio)
609 {
610 	unsigned long flags;
611 
612 	spin_lock_irqsave(&m->lock, flags);
613 	__multipath_queue_bio(m, bio);
614 	spin_unlock_irqrestore(&m->lock, flags);
615 }
616 
__map_bio(struct multipath * m,struct bio * bio)617 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
618 {
619 	struct pgpath *pgpath;
620 
621 	/* Do we need to select a new pgpath? */
622 	pgpath = READ_ONCE(m->current_pgpath);
623 	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
624 		pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
625 
626 	if (!pgpath) {
627 		spin_lock_irq(&m->lock);
628 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
629 			__multipath_queue_bio(m, bio);
630 			pgpath = ERR_PTR(-EAGAIN);
631 		}
632 		spin_unlock_irq(&m->lock);
633 
634 	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
635 		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
636 		multipath_queue_bio(m, bio);
637 		pg_init_all_paths(m);
638 		return ERR_PTR(-EAGAIN);
639 	}
640 
641 	return pgpath;
642 }
643 
__multipath_map_bio(struct multipath * m,struct bio * bio,struct dm_mpath_io * mpio)644 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
645 			       struct dm_mpath_io *mpio)
646 {
647 	struct pgpath *pgpath = __map_bio(m, bio);
648 
649 	if (IS_ERR(pgpath))
650 		return DM_MAPIO_SUBMITTED;
651 
652 	if (!pgpath) {
653 		if (__must_push_back(m))
654 			return DM_MAPIO_REQUEUE;
655 		dm_report_EIO(m);
656 		return DM_MAPIO_KILL;
657 	}
658 
659 	mpio->pgpath = pgpath;
660 
661 	if (dm_ps_use_hr_timer(pgpath->pg->ps.type))
662 		mpio->start_time_ns = ktime_get_ns();
663 
664 	bio->bi_status = 0;
665 	bio_set_dev(bio, pgpath->path.dev->bdev);
666 	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
667 
668 	if (pgpath->pg->ps.type->start_io)
669 		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
670 					      &pgpath->path,
671 					      mpio->nr_bytes);
672 	return DM_MAPIO_REMAPPED;
673 }
674 
multipath_map_bio(struct dm_target * ti,struct bio * bio)675 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
676 {
677 	struct multipath *m = ti->private;
678 	struct dm_mpath_io *mpio = NULL;
679 
680 	multipath_init_per_bio_data(bio, &mpio);
681 	return __multipath_map_bio(m, bio, mpio);
682 }
683 
process_queued_io_list(struct multipath * m)684 static void process_queued_io_list(struct multipath *m)
685 {
686 	if (m->queue_mode == DM_TYPE_REQUEST_BASED)
687 		dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
688 	else if (m->queue_mode == DM_TYPE_BIO_BASED)
689 		queue_work(kmultipathd, &m->process_queued_bios);
690 }
691 
process_queued_bios(struct work_struct * work)692 static void process_queued_bios(struct work_struct *work)
693 {
694 	int r;
695 	struct bio *bio;
696 	struct bio_list bios;
697 	struct blk_plug plug;
698 	struct multipath *m =
699 		container_of(work, struct multipath, process_queued_bios);
700 
701 	bio_list_init(&bios);
702 
703 	spin_lock_irq(&m->lock);
704 
705 	if (bio_list_empty(&m->queued_bios)) {
706 		spin_unlock_irq(&m->lock);
707 		return;
708 	}
709 
710 	bio_list_merge_init(&bios, &m->queued_bios);
711 
712 	spin_unlock_irq(&m->lock);
713 
714 	blk_start_plug(&plug);
715 	while ((bio = bio_list_pop(&bios))) {
716 		struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
717 
718 		dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
719 		r = __multipath_map_bio(m, bio, mpio);
720 		switch (r) {
721 		case DM_MAPIO_KILL:
722 			bio->bi_status = BLK_STS_IOERR;
723 			bio_endio(bio);
724 			break;
725 		case DM_MAPIO_REQUEUE:
726 			bio->bi_status = BLK_STS_DM_REQUEUE;
727 			bio_endio(bio);
728 			break;
729 		case DM_MAPIO_REMAPPED:
730 			submit_bio_noacct(bio);
731 			break;
732 		case DM_MAPIO_SUBMITTED:
733 			break;
734 		default:
735 			WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
736 		}
737 	}
738 	blk_finish_plug(&plug);
739 }
740 
741 /*
742  * If we run out of usable paths, should we queue I/O or error it?
743  */
queue_if_no_path(struct multipath * m,bool f_queue_if_no_path,bool save_old_value,const char * caller)744 static int queue_if_no_path(struct multipath *m, bool f_queue_if_no_path,
745 			    bool save_old_value, const char *caller)
746 {
747 	unsigned long flags;
748 	bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
749 	const char *dm_dev_name = dm_table_device_name(m->ti->table);
750 
751 	DMDEBUG("%s: %s caller=%s f_queue_if_no_path=%d save_old_value=%d",
752 		dm_dev_name, __func__, caller, f_queue_if_no_path, save_old_value);
753 
754 	spin_lock_irqsave(&m->lock, flags);
755 
756 	queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
757 	saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
758 
759 	if (save_old_value) {
760 		if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
761 			DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
762 			      dm_dev_name);
763 		} else
764 			assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
765 	} else if (!f_queue_if_no_path && saved_queue_if_no_path_bit) {
766 		/* due to "fail_if_no_path" message, need to honor it. */
767 		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
768 	}
769 	assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, f_queue_if_no_path);
770 
771 	DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
772 		dm_dev_name, __func__,
773 		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
774 		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
775 		dm_noflush_suspending(m->ti));
776 
777 	spin_unlock_irqrestore(&m->lock, flags);
778 
779 	if (!f_queue_if_no_path) {
780 		dm_table_run_md_queue_async(m->ti->table);
781 		process_queued_io_list(m);
782 	}
783 
784 	return 0;
785 }
786 
787 /*
788  * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
789  * process any queued I/O.
790  */
queue_if_no_path_timeout_work(struct timer_list * t)791 static void queue_if_no_path_timeout_work(struct timer_list *t)
792 {
793 	struct multipath *m = timer_container_of(m, t, nopath_timer);
794 
795 	DMWARN("queue_if_no_path timeout on %s, failing queued IO",
796 	       dm_table_device_name(m->ti->table));
797 	queue_if_no_path(m, false, false, __func__);
798 }
799 
800 /*
801  * Enable the queue_if_no_path timeout if necessary.
802  * Called with m->lock held.
803  */
enable_nopath_timeout(struct multipath * m)804 static void enable_nopath_timeout(struct multipath *m)
805 {
806 	unsigned long queue_if_no_path_timeout =
807 		READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
808 
809 	lockdep_assert_held(&m->lock);
810 
811 	if (queue_if_no_path_timeout > 0 &&
812 	    atomic_read(&m->nr_valid_paths) == 0 &&
813 	    test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
814 		mod_timer(&m->nopath_timer,
815 			  jiffies + queue_if_no_path_timeout);
816 	}
817 }
818 
disable_nopath_timeout(struct multipath * m)819 static void disable_nopath_timeout(struct multipath *m)
820 {
821 	timer_delete_sync(&m->nopath_timer);
822 }
823 
824 /*
825  * An event is triggered whenever a path is taken out of use.
826  * Includes path failure and PG bypass.
827  */
trigger_event(struct work_struct * work)828 static void trigger_event(struct work_struct *work)
829 {
830 	struct multipath *m =
831 		container_of(work, struct multipath, trigger_event);
832 
833 	dm_table_event(m->ti->table);
834 }
835 
836 /*
837  *---------------------------------------------------------------
838  * Constructor/argument parsing:
839  * <#multipath feature args> [<arg>]*
840  * <#hw_handler args> [hw_handler [<arg>]*]
841  * <#priority groups>
842  * <initial priority group>
843  *     [<selector> <#selector args> [<arg>]*
844  *      <#paths> <#per-path selector args>
845  *         [<path> [<arg>]* ]+ ]+
846  *---------------------------------------------------------------
847  */
parse_path_selector(struct dm_arg_set * as,struct priority_group * pg,struct dm_target * ti)848 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
849 			       struct dm_target *ti)
850 {
851 	int r;
852 	struct path_selector_type *pst;
853 	unsigned int ps_argc;
854 
855 	static const struct dm_arg _args[] = {
856 		{0, 1024, "invalid number of path selector args"},
857 	};
858 
859 	pst = dm_get_path_selector(dm_shift_arg(as));
860 	if (!pst) {
861 		ti->error = "unknown path selector type";
862 		return -EINVAL;
863 	}
864 
865 	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
866 	if (r) {
867 		dm_put_path_selector(pst);
868 		return -EINVAL;
869 	}
870 
871 	r = pst->create(&pg->ps, ps_argc, as->argv);
872 	if (r) {
873 		dm_put_path_selector(pst);
874 		ti->error = "path selector constructor failed";
875 		return r;
876 	}
877 
878 	pg->ps.type = pst;
879 	dm_consume_args(as, ps_argc);
880 
881 	return 0;
882 }
883 
setup_scsi_dh(struct block_device * bdev,struct multipath * m,const char ** attached_handler_name,char ** error)884 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
885 			 const char **attached_handler_name, char **error)
886 {
887 	struct request_queue *q = bdev_get_queue(bdev);
888 	int r;
889 
890 	if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
891 retain:
892 		if (*attached_handler_name) {
893 			/*
894 			 * Clear any hw_handler_params associated with a
895 			 * handler that isn't already attached.
896 			 */
897 			if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
898 				kfree(m->hw_handler_params);
899 				m->hw_handler_params = NULL;
900 			}
901 
902 			/*
903 			 * Reset hw_handler_name to match the attached handler
904 			 *
905 			 * NB. This modifies the table line to show the actual
906 			 * handler instead of the original table passed in.
907 			 */
908 			kfree(m->hw_handler_name);
909 			m->hw_handler_name = *attached_handler_name;
910 			*attached_handler_name = NULL;
911 		}
912 	}
913 
914 	if (m->hw_handler_name) {
915 		r = scsi_dh_attach(q, m->hw_handler_name);
916 		if (r == -EBUSY) {
917 			DMINFO("retaining handler on device %pg", bdev);
918 			goto retain;
919 		}
920 		if (r < 0) {
921 			*error = "error attaching hardware handler";
922 			return r;
923 		}
924 
925 		if (m->hw_handler_params) {
926 			r = scsi_dh_set_params(q, m->hw_handler_params);
927 			if (r < 0) {
928 				*error = "unable to set hardware handler parameters";
929 				return r;
930 			}
931 		}
932 	}
933 
934 	return 0;
935 }
936 
parse_path(struct dm_arg_set * as,struct path_selector * ps,struct dm_target * ti)937 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
938 				 struct dm_target *ti)
939 {
940 	int r;
941 	struct pgpath *p;
942 	struct multipath *m = ti->private;
943 	struct request_queue *q;
944 	const char *attached_handler_name = NULL;
945 
946 	/* we need at least a path arg */
947 	if (as->argc < 1) {
948 		ti->error = "no device given";
949 		return ERR_PTR(-EINVAL);
950 	}
951 
952 	p = alloc_pgpath();
953 	if (!p)
954 		return ERR_PTR(-ENOMEM);
955 
956 	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
957 			  &p->path.dev);
958 	if (r) {
959 		ti->error = "error getting device";
960 		goto bad;
961 	}
962 
963 	q = bdev_get_queue(p->path.dev->bdev);
964 	attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
965 	if (attached_handler_name || m->hw_handler_name) {
966 		INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
967 		r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
968 		kfree(attached_handler_name);
969 		if (r) {
970 			dm_put_device(ti, p->path.dev);
971 			goto bad;
972 		}
973 	}
974 
975 	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
976 	if (r) {
977 		dm_put_device(ti, p->path.dev);
978 		goto bad;
979 	}
980 
981 	return p;
982  bad:
983 	free_pgpath(p);
984 	return ERR_PTR(r);
985 }
986 
parse_priority_group(struct dm_arg_set * as,struct multipath * m)987 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
988 						   struct multipath *m)
989 {
990 	static const struct dm_arg _args[] = {
991 		{1, 1024, "invalid number of paths"},
992 		{0, 1024, "invalid number of selector args"}
993 	};
994 
995 	int r;
996 	unsigned int i, nr_selector_args, nr_args;
997 	struct priority_group *pg;
998 	struct dm_target *ti = m->ti;
999 
1000 	if (as->argc < 2) {
1001 		as->argc = 0;
1002 		ti->error = "not enough priority group arguments";
1003 		return ERR_PTR(-EINVAL);
1004 	}
1005 
1006 	pg = alloc_priority_group();
1007 	if (!pg) {
1008 		ti->error = "couldn't allocate priority group";
1009 		return ERR_PTR(-ENOMEM);
1010 	}
1011 	pg->m = m;
1012 
1013 	r = parse_path_selector(as, pg, ti);
1014 	if (r)
1015 		goto bad;
1016 
1017 	/*
1018 	 * read the paths
1019 	 */
1020 	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1021 	if (r)
1022 		goto bad;
1023 
1024 	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1025 	if (r)
1026 		goto bad;
1027 
1028 	nr_args = 1 + nr_selector_args;
1029 	for (i = 0; i < pg->nr_pgpaths; i++) {
1030 		struct pgpath *pgpath;
1031 		struct dm_arg_set path_args;
1032 
1033 		if (as->argc < nr_args) {
1034 			ti->error = "not enough path parameters";
1035 			r = -EINVAL;
1036 			goto bad;
1037 		}
1038 
1039 		path_args.argc = nr_args;
1040 		path_args.argv = as->argv;
1041 
1042 		pgpath = parse_path(&path_args, &pg->ps, ti);
1043 		if (IS_ERR(pgpath)) {
1044 			r = PTR_ERR(pgpath);
1045 			goto bad;
1046 		}
1047 
1048 		pgpath->pg = pg;
1049 		list_add_tail(&pgpath->list, &pg->pgpaths);
1050 		dm_consume_args(as, nr_args);
1051 	}
1052 
1053 	return pg;
1054 
1055  bad:
1056 	free_priority_group(pg, ti);
1057 	return ERR_PTR(r);
1058 }
1059 
parse_hw_handler(struct dm_arg_set * as,struct multipath * m)1060 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1061 {
1062 	unsigned int hw_argc;
1063 	int ret;
1064 	struct dm_target *ti = m->ti;
1065 
1066 	static const struct dm_arg _args[] = {
1067 		{0, 1024, "invalid number of hardware handler args"},
1068 	};
1069 
1070 	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1071 		return -EINVAL;
1072 
1073 	if (!hw_argc)
1074 		return 0;
1075 
1076 	if (m->queue_mode == DM_TYPE_BIO_BASED) {
1077 		dm_consume_args(as, hw_argc);
1078 		DMERR("bio-based multipath doesn't allow hardware handler args");
1079 		return 0;
1080 	}
1081 
1082 	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1083 	if (!m->hw_handler_name)
1084 		return -EINVAL;
1085 
1086 	if (hw_argc > 1) {
1087 		char *p;
1088 		int i, j, len = 4;
1089 
1090 		for (i = 0; i <= hw_argc - 2; i++)
1091 			len += strlen(as->argv[i]) + 1;
1092 		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1093 		if (!p) {
1094 			ti->error = "memory allocation failed";
1095 			ret = -ENOMEM;
1096 			goto fail;
1097 		}
1098 		j = sprintf(p, "%d", hw_argc - 1);
1099 		for (i = 0, p += j + 1; i <= hw_argc - 2; i++, p += j + 1)
1100 			j = sprintf(p, "%s", as->argv[i]);
1101 	}
1102 	dm_consume_args(as, hw_argc - 1);
1103 
1104 	return 0;
1105 fail:
1106 	kfree(m->hw_handler_name);
1107 	m->hw_handler_name = NULL;
1108 	return ret;
1109 }
1110 
parse_features(struct dm_arg_set * as,struct multipath * m)1111 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1112 {
1113 	int r;
1114 	unsigned int argc;
1115 	struct dm_target *ti = m->ti;
1116 	const char *arg_name;
1117 
1118 	static const struct dm_arg _args[] = {
1119 		{0, 8, "invalid number of feature args"},
1120 		{1, 50, "pg_init_retries must be between 1 and 50"},
1121 		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1122 	};
1123 
1124 	r = dm_read_arg_group(_args, as, &argc, &ti->error);
1125 	if (r)
1126 		return -EINVAL;
1127 
1128 	if (!argc)
1129 		return 0;
1130 
1131 	do {
1132 		arg_name = dm_shift_arg(as);
1133 		argc--;
1134 
1135 		if (!strcasecmp(arg_name, "queue_if_no_path")) {
1136 			r = queue_if_no_path(m, true, false, __func__);
1137 			continue;
1138 		}
1139 
1140 		if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1141 			set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1142 			continue;
1143 		}
1144 
1145 		if (!strcasecmp(arg_name, "pg_init_retries") &&
1146 		    (argc >= 1)) {
1147 			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1148 			argc--;
1149 			continue;
1150 		}
1151 
1152 		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1153 		    (argc >= 1)) {
1154 			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1155 			argc--;
1156 			continue;
1157 		}
1158 
1159 		if (!strcasecmp(arg_name, "queue_mode") &&
1160 		    (argc >= 1)) {
1161 			const char *queue_mode_name = dm_shift_arg(as);
1162 
1163 			if (!strcasecmp(queue_mode_name, "bio"))
1164 				m->queue_mode = DM_TYPE_BIO_BASED;
1165 			else if (!strcasecmp(queue_mode_name, "rq") ||
1166 				 !strcasecmp(queue_mode_name, "mq"))
1167 				m->queue_mode = DM_TYPE_REQUEST_BASED;
1168 			else {
1169 				ti->error = "Unknown 'queue_mode' requested";
1170 				r = -EINVAL;
1171 			}
1172 			argc--;
1173 			continue;
1174 		}
1175 
1176 		ti->error = "Unrecognised multipath feature request";
1177 		r = -EINVAL;
1178 	} while (argc && !r);
1179 
1180 	return r;
1181 }
1182 
multipath_ctr(struct dm_target * ti,unsigned int argc,char ** argv)1183 static int multipath_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1184 {
1185 	/* target arguments */
1186 	static const struct dm_arg _args[] = {
1187 		{0, 1024, "invalid number of priority groups"},
1188 		{0, 1024, "invalid initial priority group number"},
1189 	};
1190 
1191 	int r;
1192 	struct multipath *m;
1193 	struct dm_arg_set as;
1194 	unsigned int pg_count = 0;
1195 	unsigned int next_pg_num;
1196 
1197 	as.argc = argc;
1198 	as.argv = argv;
1199 
1200 	m = alloc_multipath(ti);
1201 	if (!m) {
1202 		ti->error = "can't allocate multipath";
1203 		return -EINVAL;
1204 	}
1205 
1206 	r = parse_features(&as, m);
1207 	if (r)
1208 		goto bad;
1209 
1210 	r = alloc_multipath_stage2(ti, m);
1211 	if (r)
1212 		goto bad;
1213 
1214 	r = parse_hw_handler(&as, m);
1215 	if (r)
1216 		goto bad;
1217 
1218 	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1219 	if (r)
1220 		goto bad;
1221 
1222 	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1223 	if (r)
1224 		goto bad;
1225 
1226 	if ((!m->nr_priority_groups && next_pg_num) ||
1227 	    (m->nr_priority_groups && !next_pg_num)) {
1228 		ti->error = "invalid initial priority group";
1229 		r = -EINVAL;
1230 		goto bad;
1231 	}
1232 
1233 	/* parse the priority groups */
1234 	while (as.argc) {
1235 		struct priority_group *pg;
1236 		unsigned int nr_valid_paths = atomic_read(&m->nr_valid_paths);
1237 
1238 		pg = parse_priority_group(&as, m);
1239 		if (IS_ERR(pg)) {
1240 			r = PTR_ERR(pg);
1241 			goto bad;
1242 		}
1243 
1244 		nr_valid_paths += pg->nr_pgpaths;
1245 		atomic_set(&m->nr_valid_paths, nr_valid_paths);
1246 
1247 		list_add_tail(&pg->list, &m->priority_groups);
1248 		pg_count++;
1249 		pg->pg_num = pg_count;
1250 		if (!--next_pg_num)
1251 			m->next_pg = pg;
1252 	}
1253 
1254 	if (pg_count != m->nr_priority_groups) {
1255 		ti->error = "priority group count mismatch";
1256 		r = -EINVAL;
1257 		goto bad;
1258 	}
1259 
1260 	spin_lock_irq(&m->lock);
1261 	enable_nopath_timeout(m);
1262 	spin_unlock_irq(&m->lock);
1263 
1264 	ti->num_flush_bios = 1;
1265 	ti->num_discard_bios = 1;
1266 	ti->num_write_zeroes_bios = 1;
1267 	if (m->queue_mode == DM_TYPE_BIO_BASED)
1268 		ti->per_io_data_size = multipath_per_bio_data_size();
1269 	else
1270 		ti->per_io_data_size = sizeof(struct dm_mpath_io);
1271 
1272 	return 0;
1273 
1274  bad:
1275 	free_multipath(m);
1276 	return r;
1277 }
1278 
multipath_wait_for_pg_init_completion(struct multipath * m)1279 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1280 {
1281 	DEFINE_WAIT(wait);
1282 
1283 	while (1) {
1284 		prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1285 
1286 		if (!atomic_read(&m->pg_init_in_progress))
1287 			break;
1288 
1289 		io_schedule();
1290 	}
1291 	finish_wait(&m->pg_init_wait, &wait);
1292 }
1293 
flush_multipath_work(struct multipath * m)1294 static void flush_multipath_work(struct multipath *m)
1295 {
1296 	if (m->hw_handler_name) {
1297 		if (!atomic_read(&m->pg_init_in_progress))
1298 			goto skip;
1299 
1300 		spin_lock_irq(&m->lock);
1301 		if (atomic_read(&m->pg_init_in_progress) &&
1302 		    !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1303 			spin_unlock_irq(&m->lock);
1304 
1305 			flush_workqueue(kmpath_handlerd);
1306 			multipath_wait_for_pg_init_completion(m);
1307 
1308 			spin_lock_irq(&m->lock);
1309 			clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1310 		}
1311 		spin_unlock_irq(&m->lock);
1312 	}
1313 skip:
1314 	if (m->queue_mode == DM_TYPE_BIO_BASED)
1315 		flush_work(&m->process_queued_bios);
1316 	flush_work(&m->trigger_event);
1317 }
1318 
multipath_dtr(struct dm_target * ti)1319 static void multipath_dtr(struct dm_target *ti)
1320 {
1321 	struct multipath *m = ti->private;
1322 
1323 	disable_nopath_timeout(m);
1324 	flush_multipath_work(m);
1325 	free_multipath(m);
1326 }
1327 
1328 /*
1329  * Take a path out of use.
1330  */
fail_path(struct pgpath * pgpath)1331 static int fail_path(struct pgpath *pgpath)
1332 {
1333 	unsigned long flags;
1334 	struct multipath *m = pgpath->pg->m;
1335 
1336 	spin_lock_irqsave(&m->lock, flags);
1337 
1338 	if (!pgpath->is_active)
1339 		goto out;
1340 
1341 	DMWARN("%s: Failing path %s.",
1342 	       dm_table_device_name(m->ti->table),
1343 	       pgpath->path.dev->name);
1344 
1345 	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1346 	pgpath->is_active = false;
1347 	pgpath->fail_count++;
1348 
1349 	atomic_dec(&m->nr_valid_paths);
1350 
1351 	if (pgpath == m->current_pgpath)
1352 		m->current_pgpath = NULL;
1353 
1354 	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1355 		       pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1356 
1357 	queue_work(dm_mpath_wq, &m->trigger_event);
1358 
1359 	enable_nopath_timeout(m);
1360 
1361 out:
1362 	spin_unlock_irqrestore(&m->lock, flags);
1363 
1364 	return 0;
1365 }
1366 
1367 /*
1368  * Reinstate a previously-failed path
1369  */
reinstate_path(struct pgpath * pgpath)1370 static int reinstate_path(struct pgpath *pgpath)
1371 {
1372 	int r = 0, run_queue = 0;
1373 	struct multipath *m = pgpath->pg->m;
1374 	unsigned int nr_valid_paths;
1375 
1376 	spin_lock_irq(&m->lock);
1377 
1378 	if (pgpath->is_active)
1379 		goto out;
1380 
1381 	DMWARN("%s: Reinstating path %s.",
1382 	       dm_table_device_name(m->ti->table),
1383 	       pgpath->path.dev->name);
1384 
1385 	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1386 	if (r)
1387 		goto out;
1388 
1389 	pgpath->is_active = true;
1390 
1391 	nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1392 	if (nr_valid_paths == 1) {
1393 		m->current_pgpath = NULL;
1394 		run_queue = 1;
1395 	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1396 		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1397 			atomic_inc(&m->pg_init_in_progress);
1398 	}
1399 
1400 	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1401 		       pgpath->path.dev->name, nr_valid_paths);
1402 
1403 	schedule_work(&m->trigger_event);
1404 
1405 out:
1406 	spin_unlock_irq(&m->lock);
1407 	if (run_queue) {
1408 		dm_table_run_md_queue_async(m->ti->table);
1409 		process_queued_io_list(m);
1410 	}
1411 
1412 	if (pgpath->is_active)
1413 		disable_nopath_timeout(m);
1414 
1415 	return r;
1416 }
1417 
1418 /*
1419  * Fail or reinstate all paths that match the provided struct dm_dev.
1420  */
action_dev(struct multipath * m,dev_t dev,action_fn action)1421 static int action_dev(struct multipath *m, dev_t dev, action_fn action)
1422 {
1423 	int r = -EINVAL;
1424 	struct pgpath *pgpath;
1425 	struct priority_group *pg;
1426 
1427 	list_for_each_entry(pg, &m->priority_groups, list) {
1428 		list_for_each_entry(pgpath, &pg->pgpaths, list) {
1429 			if (pgpath->path.dev->bdev->bd_dev == dev)
1430 				r = action(pgpath);
1431 		}
1432 	}
1433 
1434 	return r;
1435 }
1436 
1437 /*
1438  * Temporarily try to avoid having to use the specified PG
1439  */
bypass_pg(struct multipath * m,struct priority_group * pg,bool bypassed,bool can_be_delayed)1440 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1441 		      bool bypassed, bool can_be_delayed)
1442 {
1443 	unsigned long flags;
1444 
1445 	spin_lock_irqsave(&m->lock, flags);
1446 
1447 	pg->bypassed = bypassed;
1448 	if (can_be_delayed && test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags))
1449 		set_bit(MPATHF_NEED_PG_SWITCH, &m->flags);
1450 	else {
1451 		m->current_pgpath = NULL;
1452 		m->current_pg = NULL;
1453 	}
1454 
1455 	spin_unlock_irqrestore(&m->lock, flags);
1456 
1457 	schedule_work(&m->trigger_event);
1458 }
1459 
1460 /*
1461  * Switch to using the specified PG from the next I/O that gets mapped
1462  */
switch_pg_num(struct multipath * m,const char * pgstr)1463 static int switch_pg_num(struct multipath *m, const char *pgstr)
1464 {
1465 	struct priority_group *pg;
1466 	unsigned int pgnum;
1467 	char dummy;
1468 
1469 	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1470 	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1471 		DMWARN("invalid PG number supplied to %s", __func__);
1472 		return -EINVAL;
1473 	}
1474 
1475 	spin_lock_irq(&m->lock);
1476 	list_for_each_entry(pg, &m->priority_groups, list) {
1477 		pg->bypassed = false;
1478 		if (--pgnum)
1479 			continue;
1480 
1481 		if (test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags))
1482 			set_bit(MPATHF_NEED_PG_SWITCH, &m->flags);
1483 		else {
1484 			m->current_pgpath = NULL;
1485 			m->current_pg = NULL;
1486 		}
1487 		m->next_pg = pg;
1488 	}
1489 	spin_unlock_irq(&m->lock);
1490 
1491 	schedule_work(&m->trigger_event);
1492 	return 0;
1493 }
1494 
1495 /*
1496  * Set/clear bypassed status of a PG.
1497  * PGs are numbered upwards from 1 in the order they were declared.
1498  */
bypass_pg_num(struct multipath * m,const char * pgstr,bool bypassed)1499 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1500 {
1501 	struct priority_group *pg;
1502 	unsigned int pgnum;
1503 	char dummy;
1504 
1505 	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1506 	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1507 		DMWARN("invalid PG number supplied to bypass_pg");
1508 		return -EINVAL;
1509 	}
1510 
1511 	list_for_each_entry(pg, &m->priority_groups, list) {
1512 		if (!--pgnum)
1513 			break;
1514 	}
1515 
1516 	bypass_pg(m, pg, bypassed, true);
1517 	return 0;
1518 }
1519 
1520 /*
1521  * Should we retry pg_init immediately?
1522  */
pg_init_limit_reached(struct multipath * m,struct pgpath * pgpath)1523 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1524 {
1525 	unsigned long flags;
1526 	bool limit_reached = false;
1527 
1528 	spin_lock_irqsave(&m->lock, flags);
1529 
1530 	if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1531 	    !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1532 		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1533 	else
1534 		limit_reached = true;
1535 
1536 	spin_unlock_irqrestore(&m->lock, flags);
1537 
1538 	return limit_reached;
1539 }
1540 
pg_init_done(void * data,int errors)1541 static void pg_init_done(void *data, int errors)
1542 {
1543 	struct pgpath *pgpath = data;
1544 	struct priority_group *pg = pgpath->pg;
1545 	struct multipath *m = pg->m;
1546 	unsigned long flags;
1547 	bool delay_retry = false;
1548 
1549 	/* device or driver problems */
1550 	switch (errors) {
1551 	case SCSI_DH_OK:
1552 		break;
1553 	case SCSI_DH_NOSYS:
1554 		if (!m->hw_handler_name) {
1555 			errors = 0;
1556 			break;
1557 		}
1558 		DMERR("Could not failover the device: Handler scsi_dh_%s "
1559 		      "Error %d.", m->hw_handler_name, errors);
1560 		/*
1561 		 * Fail path for now, so we do not ping pong
1562 		 */
1563 		fail_path(pgpath);
1564 		break;
1565 	case SCSI_DH_DEV_TEMP_BUSY:
1566 		/*
1567 		 * Probably doing something like FW upgrade on the
1568 		 * controller so try the other pg.
1569 		 */
1570 		bypass_pg(m, pg, true, false);
1571 		break;
1572 	case SCSI_DH_RETRY:
1573 		/* Wait before retrying. */
1574 		delay_retry = true;
1575 		fallthrough;
1576 	case SCSI_DH_IMM_RETRY:
1577 	case SCSI_DH_RES_TEMP_UNAVAIL:
1578 		if (pg_init_limit_reached(m, pgpath))
1579 			fail_path(pgpath);
1580 		errors = 0;
1581 		break;
1582 	case SCSI_DH_DEV_OFFLINED:
1583 	default:
1584 		/*
1585 		 * We probably do not want to fail the path for a device
1586 		 * error, but this is what the old dm did. In future
1587 		 * patches we can do more advanced handling.
1588 		 */
1589 		fail_path(pgpath);
1590 	}
1591 
1592 	spin_lock_irqsave(&m->lock, flags);
1593 	if (errors) {
1594 		if (pgpath == m->current_pgpath) {
1595 			DMERR("Could not failover device. Error %d.", errors);
1596 			m->current_pgpath = NULL;
1597 			m->current_pg = NULL;
1598 		}
1599 	} else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1600 		pg->bypassed = false;
1601 
1602 	if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1603 		/* Activations of other paths are still on going */
1604 		goto out;
1605 
1606 	if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1607 		if (delay_retry)
1608 			set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1609 		else
1610 			clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1611 
1612 		if (__pg_init_all_paths(m))
1613 			goto out;
1614 	}
1615 	clear_bit(MPATHF_QUEUE_IO, &m->flags);
1616 
1617 	process_queued_io_list(m);
1618 
1619 	/*
1620 	 * Wake up any thread waiting to suspend.
1621 	 */
1622 	wake_up(&m->pg_init_wait);
1623 
1624 out:
1625 	spin_unlock_irqrestore(&m->lock, flags);
1626 }
1627 
activate_or_offline_path(struct pgpath * pgpath)1628 static void activate_or_offline_path(struct pgpath *pgpath)
1629 {
1630 	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1631 
1632 	if (pgpath->is_active && !blk_queue_dying(q))
1633 		scsi_dh_activate(q, pg_init_done, pgpath);
1634 	else
1635 		pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1636 }
1637 
activate_path_work(struct work_struct * work)1638 static void activate_path_work(struct work_struct *work)
1639 {
1640 	struct pgpath *pgpath =
1641 		container_of(work, struct pgpath, activate_path.work);
1642 
1643 	activate_or_offline_path(pgpath);
1644 }
1645 
multipath_end_io(struct dm_target * ti,struct request * clone,blk_status_t error,union map_info * map_context)1646 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1647 			    blk_status_t error, union map_info *map_context)
1648 {
1649 	struct dm_mpath_io *mpio = get_mpio(map_context);
1650 	struct pgpath *pgpath = mpio->pgpath;
1651 	int r = DM_ENDIO_DONE;
1652 
1653 	/*
1654 	 * We don't queue any clone request inside the multipath target
1655 	 * during end I/O handling, since those clone requests don't have
1656 	 * bio clones.  If we queue them inside the multipath target,
1657 	 * we need to make bio clones, that requires memory allocation.
1658 	 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1659 	 *  don't have bio clones.)
1660 	 * Instead of queueing the clone request here, we queue the original
1661 	 * request into dm core, which will remake a clone request and
1662 	 * clone bios for it and resubmit it later.
1663 	 */
1664 	if (error && blk_path_error(error)) {
1665 		struct multipath *m = ti->private;
1666 
1667 		if (error == BLK_STS_RESOURCE)
1668 			r = DM_ENDIO_DELAY_REQUEUE;
1669 		else
1670 			r = DM_ENDIO_REQUEUE;
1671 
1672 		if (pgpath)
1673 			fail_path(pgpath);
1674 
1675 		if (!atomic_read(&m->nr_valid_paths) &&
1676 		    !must_push_back_rq(m)) {
1677 			if (error == BLK_STS_IOERR)
1678 				dm_report_EIO(m);
1679 			/* complete with the original error */
1680 			r = DM_ENDIO_DONE;
1681 		}
1682 	}
1683 
1684 	if (pgpath) {
1685 		struct path_selector *ps = &pgpath->pg->ps;
1686 
1687 		if (ps->type->end_io)
1688 			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1689 					 clone->io_start_time_ns);
1690 	}
1691 
1692 	return r;
1693 }
1694 
multipath_end_io_bio(struct dm_target * ti,struct bio * clone,blk_status_t * error)1695 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1696 				blk_status_t *error)
1697 {
1698 	struct multipath *m = ti->private;
1699 	struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1700 	struct pgpath *pgpath = mpio->pgpath;
1701 	unsigned long flags;
1702 	int r = DM_ENDIO_DONE;
1703 
1704 	if (!*error || !blk_path_error(*error))
1705 		goto done;
1706 
1707 	if (pgpath)
1708 		fail_path(pgpath);
1709 
1710 	if (!atomic_read(&m->nr_valid_paths)) {
1711 		spin_lock_irqsave(&m->lock, flags);
1712 		if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1713 			if (__must_push_back(m)) {
1714 				r = DM_ENDIO_REQUEUE;
1715 			} else {
1716 				dm_report_EIO(m);
1717 				*error = BLK_STS_IOERR;
1718 			}
1719 			spin_unlock_irqrestore(&m->lock, flags);
1720 			goto done;
1721 		}
1722 		spin_unlock_irqrestore(&m->lock, flags);
1723 	}
1724 
1725 	multipath_queue_bio(m, clone);
1726 	r = DM_ENDIO_INCOMPLETE;
1727 done:
1728 	if (pgpath) {
1729 		struct path_selector *ps = &pgpath->pg->ps;
1730 
1731 		if (ps->type->end_io)
1732 			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1733 					 (mpio->start_time_ns ?:
1734 					  dm_start_time_ns_from_clone(clone)));
1735 	}
1736 
1737 	return r;
1738 }
1739 
1740 /*
1741  * Suspend with flush can't complete until all the I/O is processed
1742  * so if the last path fails we must error any remaining I/O.
1743  * - Note that if the freeze_bdev fails while suspending, the
1744  *   queue_if_no_path state is lost - userspace should reset it.
1745  * Otherwise, during noflush suspend, queue_if_no_path will not change.
1746  */
multipath_presuspend(struct dm_target * ti)1747 static void multipath_presuspend(struct dm_target *ti)
1748 {
1749 	struct multipath *m = ti->private;
1750 
1751 	spin_lock_irq(&m->lock);
1752 	m->is_suspending = true;
1753 	spin_unlock_irq(&m->lock);
1754 	/* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1755 	if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1756 		queue_if_no_path(m, false, true, __func__);
1757 }
1758 
multipath_postsuspend(struct dm_target * ti)1759 static void multipath_postsuspend(struct dm_target *ti)
1760 {
1761 	struct multipath *m = ti->private;
1762 
1763 	mutex_lock(&m->work_mutex);
1764 	flush_multipath_work(m);
1765 	mutex_unlock(&m->work_mutex);
1766 }
1767 
1768 /*
1769  * Restore the queue_if_no_path setting.
1770  */
multipath_resume(struct dm_target * ti)1771 static void multipath_resume(struct dm_target *ti)
1772 {
1773 	struct multipath *m = ti->private;
1774 
1775 	spin_lock_irq(&m->lock);
1776 	m->is_suspending = false;
1777 	if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1778 		set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1779 		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1780 	}
1781 
1782 	DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1783 		dm_table_device_name(m->ti->table), __func__,
1784 		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1785 		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1786 
1787 	spin_unlock_irq(&m->lock);
1788 }
1789 
1790 /*
1791  * Info output has the following format:
1792  * num_multipath_feature_args [multipath_feature_args]*
1793  * num_handler_status_args [handler_status_args]*
1794  * num_groups init_group_number
1795  *            [A|D|E num_ps_status_args [ps_status_args]*
1796  *             num_paths num_selector_args
1797  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1798  *
1799  * Table output has the following format (identical to the constructor string):
1800  * num_feature_args [features_args]*
1801  * num_handler_args hw_handler [hw_handler_args]*
1802  * num_groups init_group_number
1803  *     [priority selector-name num_ps_args [ps_args]*
1804  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1805  */
multipath_status(struct dm_target * ti,status_type_t type,unsigned int status_flags,char * result,unsigned int maxlen)1806 static void multipath_status(struct dm_target *ti, status_type_t type,
1807 			     unsigned int status_flags, char *result, unsigned int maxlen)
1808 {
1809 	int sz = 0, pg_counter, pgpath_counter;
1810 	struct multipath *m = ti->private;
1811 	struct priority_group *pg;
1812 	struct pgpath *p;
1813 	unsigned int pg_num;
1814 	char state;
1815 
1816 	spin_lock_irq(&m->lock);
1817 
1818 	/* Features */
1819 	if (type == STATUSTYPE_INFO)
1820 		DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1821 		       atomic_read(&m->pg_init_count));
1822 	else {
1823 		DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1824 			      (m->pg_init_retries > 0) * 2 +
1825 			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1826 			      test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1827 			      (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1828 
1829 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1830 			DMEMIT("queue_if_no_path ");
1831 		if (m->pg_init_retries)
1832 			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1833 		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1834 			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1835 		if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1836 			DMEMIT("retain_attached_hw_handler ");
1837 		if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1838 			switch (m->queue_mode) {
1839 			case DM_TYPE_BIO_BASED:
1840 				DMEMIT("queue_mode bio ");
1841 				break;
1842 			default:
1843 				WARN_ON_ONCE(true);
1844 				break;
1845 			}
1846 		}
1847 	}
1848 
1849 	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1850 		DMEMIT("0 ");
1851 	else
1852 		DMEMIT("1 %s ", m->hw_handler_name);
1853 
1854 	DMEMIT("%u ", m->nr_priority_groups);
1855 
1856 	if (m->current_pg)
1857 		pg_num = m->current_pg->pg_num;
1858 	else if (m->next_pg)
1859 		pg_num = m->next_pg->pg_num;
1860 	else
1861 		pg_num = (m->nr_priority_groups ? 1 : 0);
1862 
1863 	DMEMIT("%u ", pg_num);
1864 
1865 	switch (type) {
1866 	case STATUSTYPE_INFO:
1867 		list_for_each_entry(pg, &m->priority_groups, list) {
1868 			if (pg->bypassed)
1869 				state = 'D';	/* Disabled */
1870 			else if (pg == m->current_pg)
1871 				state = 'A';	/* Currently Active */
1872 			else
1873 				state = 'E';	/* Enabled */
1874 
1875 			DMEMIT("%c ", state);
1876 
1877 			if (pg->ps.type->status)
1878 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1879 							  result + sz,
1880 							  maxlen - sz);
1881 			else
1882 				DMEMIT("0 ");
1883 
1884 			DMEMIT("%u %u ", pg->nr_pgpaths,
1885 			       pg->ps.type->info_args);
1886 
1887 			list_for_each_entry(p, &pg->pgpaths, list) {
1888 				DMEMIT("%s %s %u ", p->path.dev->name,
1889 				       p->is_active ? "A" : "F",
1890 				       p->fail_count);
1891 				if (pg->ps.type->status)
1892 					sz += pg->ps.type->status(&pg->ps,
1893 					      &p->path, type, result + sz,
1894 					      maxlen - sz);
1895 			}
1896 		}
1897 		break;
1898 
1899 	case STATUSTYPE_TABLE:
1900 		list_for_each_entry(pg, &m->priority_groups, list) {
1901 			DMEMIT("%s ", pg->ps.type->name);
1902 
1903 			if (pg->ps.type->status)
1904 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1905 							  result + sz,
1906 							  maxlen - sz);
1907 			else
1908 				DMEMIT("0 ");
1909 
1910 			DMEMIT("%u %u ", pg->nr_pgpaths,
1911 			       pg->ps.type->table_args);
1912 
1913 			list_for_each_entry(p, &pg->pgpaths, list) {
1914 				DMEMIT("%s ", p->path.dev->name);
1915 				if (pg->ps.type->status)
1916 					sz += pg->ps.type->status(&pg->ps,
1917 					      &p->path, type, result + sz,
1918 					      maxlen - sz);
1919 			}
1920 		}
1921 		break;
1922 
1923 	case STATUSTYPE_IMA:
1924 		sz = 0; /*reset the result pointer*/
1925 
1926 		DMEMIT_TARGET_NAME_VERSION(ti->type);
1927 		DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1928 
1929 		pg_counter = 0;
1930 		list_for_each_entry(pg, &m->priority_groups, list) {
1931 			if (pg->bypassed)
1932 				state = 'D';	/* Disabled */
1933 			else if (pg == m->current_pg)
1934 				state = 'A';	/* Currently Active */
1935 			else
1936 				state = 'E';	/* Enabled */
1937 			DMEMIT(",pg_state_%d=%c", pg_counter, state);
1938 			DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1939 			DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1940 
1941 			pgpath_counter = 0;
1942 			list_for_each_entry(p, &pg->pgpaths, list) {
1943 				DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1944 				       pg_counter, pgpath_counter, p->path.dev->name,
1945 				       pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1946 				       pg_counter, pgpath_counter, p->fail_count);
1947 				if (pg->ps.type->status) {
1948 					DMEMIT(",path_selector_status_%d_%d=",
1949 					       pg_counter, pgpath_counter);
1950 					sz += pg->ps.type->status(&pg->ps, &p->path,
1951 								  type, result + sz,
1952 								  maxlen - sz);
1953 				}
1954 				pgpath_counter++;
1955 			}
1956 			pg_counter++;
1957 		}
1958 		DMEMIT(";");
1959 		break;
1960 	}
1961 
1962 	spin_unlock_irq(&m->lock);
1963 }
1964 
multipath_message(struct dm_target * ti,unsigned int argc,char ** argv,char * result,unsigned int maxlen)1965 static int multipath_message(struct dm_target *ti, unsigned int argc, char **argv,
1966 			     char *result, unsigned int maxlen)
1967 {
1968 	int r = -EINVAL;
1969 	dev_t dev;
1970 	struct multipath *m = ti->private;
1971 	action_fn action;
1972 
1973 	mutex_lock(&m->work_mutex);
1974 
1975 	if (dm_suspended(ti)) {
1976 		r = -EBUSY;
1977 		goto out;
1978 	}
1979 
1980 	if (argc == 1) {
1981 		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1982 			r = queue_if_no_path(m, true, false, __func__);
1983 			spin_lock_irq(&m->lock);
1984 			enable_nopath_timeout(m);
1985 			spin_unlock_irq(&m->lock);
1986 			goto out;
1987 		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1988 			r = queue_if_no_path(m, false, false, __func__);
1989 			disable_nopath_timeout(m);
1990 			goto out;
1991 		}
1992 	}
1993 
1994 	if (argc != 2) {
1995 		DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1996 		goto out;
1997 	}
1998 
1999 	if (!strcasecmp(argv[0], "disable_group")) {
2000 		r = bypass_pg_num(m, argv[1], true);
2001 		goto out;
2002 	} else if (!strcasecmp(argv[0], "enable_group")) {
2003 		r = bypass_pg_num(m, argv[1], false);
2004 		goto out;
2005 	} else if (!strcasecmp(argv[0], "switch_group")) {
2006 		r = switch_pg_num(m, argv[1]);
2007 		goto out;
2008 	} else if (!strcasecmp(argv[0], "reinstate_path"))
2009 		action = reinstate_path;
2010 	else if (!strcasecmp(argv[0], "fail_path"))
2011 		action = fail_path;
2012 	else {
2013 		DMWARN("Unrecognised multipath message received: %s", argv[0]);
2014 		goto out;
2015 	}
2016 
2017 	r = dm_devt_from_path(argv[1], &dev);
2018 	if (r) {
2019 		DMWARN("message: error getting device %s",
2020 		       argv[1]);
2021 		goto out;
2022 	}
2023 
2024 	r = action_dev(m, dev, action);
2025 
2026 out:
2027 	mutex_unlock(&m->work_mutex);
2028 	return r;
2029 }
2030 
2031 /*
2032  * Perform a minimal read from the given path to find out whether the
2033  * path still works.  If a path error occurs, fail it.
2034  */
probe_path(struct pgpath * pgpath)2035 static int probe_path(struct pgpath *pgpath)
2036 {
2037 	struct block_device *bdev = pgpath->path.dev->bdev;
2038 	unsigned int read_size = bdev_logical_block_size(bdev);
2039 	struct page *page;
2040 	struct bio *bio;
2041 	blk_status_t status;
2042 	int r = 0;
2043 
2044 	if (WARN_ON_ONCE(read_size > PAGE_SIZE))
2045 		return -EINVAL;
2046 
2047 	page = alloc_page(GFP_KERNEL);
2048 	if (!page)
2049 		return -ENOMEM;
2050 
2051 	/* Perform a minimal read: Sector 0, length read_size */
2052 	bio = bio_alloc(bdev, 1, REQ_OP_READ, GFP_KERNEL);
2053 	if (!bio) {
2054 		r = -ENOMEM;
2055 		goto out;
2056 	}
2057 
2058 	bio->bi_iter.bi_sector = 0;
2059 	__bio_add_page(bio, page, read_size, 0);
2060 	submit_bio_wait(bio);
2061 	status = bio->bi_status;
2062 	bio_put(bio);
2063 
2064 	if (status && blk_path_error(status))
2065 		fail_path(pgpath);
2066 
2067 out:
2068 	__free_page(page);
2069 	return r;
2070 }
2071 
2072 /*
2073  * Probe all active paths in current_pg to find out whether they still work.
2074  * Fail all paths that do not work.
2075  *
2076  * Return -ENOTCONN if no valid path is left (even outside of current_pg). We
2077  * cannot probe paths in other pgs without switching current_pg, so if valid
2078  * paths are only in different pgs, they may or may not work. Additionally
2079  * we should not probe paths in a pathgroup that is in the process of
2080  * Initializing. Userspace can submit a request and we'll switch and wait
2081  * for the pathgroup to be initialized. If the request fails, it may need to
2082  * probe again.
2083  */
probe_active_paths(struct multipath * m)2084 static int probe_active_paths(struct multipath *m)
2085 {
2086 	struct pgpath *pgpath;
2087 	struct priority_group *pg = NULL;
2088 	int r = 0;
2089 
2090 	spin_lock_irq(&m->lock);
2091 	if (test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags)) {
2092 		wait_event_lock_irq(m->probe_wait,
2093 				    !test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags),
2094 				    m->lock);
2095 		/*
2096 		 * if we waited because a probe was already in progress,
2097 		 * and it probed the current active pathgroup, don't
2098 		 * reprobe. Just return the number of valid paths
2099 		 */
2100 		if (m->current_pg == m->last_probed_pg)
2101 			goto skip_probe;
2102 	}
2103 	if (!m->current_pg || m->is_suspending ||
2104 	    test_bit(MPATHF_QUEUE_IO, &m->flags))
2105 		goto skip_probe;
2106 	set_bit(MPATHF_DELAY_PG_SWITCH, &m->flags);
2107 	pg = m->last_probed_pg = m->current_pg;
2108 	spin_unlock_irq(&m->lock);
2109 
2110 	list_for_each_entry(pgpath, &pg->pgpaths, list) {
2111 		if (pg != READ_ONCE(m->current_pg) ||
2112 		    READ_ONCE(m->is_suspending))
2113 			goto out;
2114 		if (!pgpath->is_active)
2115 			continue;
2116 
2117 		r = probe_path(pgpath);
2118 		if (r < 0)
2119 			goto out;
2120 	}
2121 
2122 out:
2123 	spin_lock_irq(&m->lock);
2124 	clear_bit(MPATHF_DELAY_PG_SWITCH, &m->flags);
2125 	if (test_and_clear_bit(MPATHF_NEED_PG_SWITCH, &m->flags)) {
2126 		m->current_pgpath = NULL;
2127 		m->current_pg = NULL;
2128 	}
2129 skip_probe:
2130 	if (r == 0 && !atomic_read(&m->nr_valid_paths))
2131 		r = -ENOTCONN;
2132 	spin_unlock_irq(&m->lock);
2133 	if (pg)
2134 		wake_up(&m->probe_wait);
2135 	return r;
2136 }
2137 
multipath_prepare_ioctl(struct dm_target * ti,struct block_device ** bdev,unsigned int cmd,unsigned long arg,bool * forward)2138 static int multipath_prepare_ioctl(struct dm_target *ti,
2139 				   struct block_device **bdev,
2140 				   unsigned int cmd, unsigned long arg,
2141 				   bool *forward)
2142 {
2143 	struct multipath *m = ti->private;
2144 	struct pgpath *pgpath;
2145 	int r;
2146 
2147 	if (_IOC_TYPE(cmd) == DM_IOCTL) {
2148 		*forward = false;
2149 		switch (cmd) {
2150 		case DM_MPATH_PROBE_PATHS:
2151 			return probe_active_paths(m);
2152 		default:
2153 			return -ENOTTY;
2154 		}
2155 	}
2156 
2157 	pgpath = READ_ONCE(m->current_pgpath);
2158 	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2159 		pgpath = choose_pgpath(m, 0);
2160 
2161 	if (pgpath) {
2162 		if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2163 			*bdev = pgpath->path.dev->bdev;
2164 			r = 0;
2165 		} else {
2166 			/* pg_init has not started or completed */
2167 			r = -ENOTCONN;
2168 		}
2169 	} else {
2170 		/* No path is available */
2171 		r = -EIO;
2172 		spin_lock_irq(&m->lock);
2173 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2174 			r = -ENOTCONN;
2175 		spin_unlock_irq(&m->lock);
2176 	}
2177 
2178 	if (r == -ENOTCONN) {
2179 		if (!READ_ONCE(m->current_pg)) {
2180 			/* Path status changed, redo selection */
2181 			(void) choose_pgpath(m, 0);
2182 		}
2183 		spin_lock_irq(&m->lock);
2184 		if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2185 			(void) __pg_init_all_paths(m);
2186 		spin_unlock_irq(&m->lock);
2187 		dm_table_run_md_queue_async(m->ti->table);
2188 		process_queued_io_list(m);
2189 	}
2190 
2191 	/*
2192 	 * Only pass ioctls through if the device sizes match exactly.
2193 	 */
2194 	if (!r && ti->len != bdev_nr_sectors((*bdev)))
2195 		return 1;
2196 	return r;
2197 }
2198 
multipath_iterate_devices(struct dm_target * ti,iterate_devices_callout_fn fn,void * data)2199 static int multipath_iterate_devices(struct dm_target *ti,
2200 				     iterate_devices_callout_fn fn, void *data)
2201 {
2202 	struct multipath *m = ti->private;
2203 	struct priority_group *pg;
2204 	struct pgpath *p;
2205 	int ret = 0;
2206 
2207 	list_for_each_entry(pg, &m->priority_groups, list) {
2208 		list_for_each_entry(p, &pg->pgpaths, list) {
2209 			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2210 			if (ret)
2211 				goto out;
2212 		}
2213 	}
2214 
2215 out:
2216 	return ret;
2217 }
2218 
pgpath_busy(struct pgpath * pgpath)2219 static int pgpath_busy(struct pgpath *pgpath)
2220 {
2221 	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2222 
2223 	return blk_lld_busy(q);
2224 }
2225 
2226 /*
2227  * We return "busy", only when we can map I/Os but underlying devices
2228  * are busy (so even if we map I/Os now, the I/Os will wait on
2229  * the underlying queue).
2230  * In other words, if we want to kill I/Os or queue them inside us
2231  * due to map unavailability, we don't return "busy".  Otherwise,
2232  * dm core won't give us the I/Os and we can't do what we want.
2233  */
multipath_busy(struct dm_target * ti)2234 static int multipath_busy(struct dm_target *ti)
2235 {
2236 	bool busy = false, has_active = false;
2237 	struct multipath *m = ti->private;
2238 	struct priority_group *pg, *next_pg;
2239 	struct pgpath *pgpath;
2240 
2241 	/* pg_init in progress */
2242 	if (atomic_read(&m->pg_init_in_progress))
2243 		return true;
2244 
2245 	/* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2246 	if (!atomic_read(&m->nr_valid_paths)) {
2247 		unsigned long flags;
2248 
2249 		spin_lock_irqsave(&m->lock, flags);
2250 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2251 			spin_unlock_irqrestore(&m->lock, flags);
2252 			return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2253 		}
2254 		spin_unlock_irqrestore(&m->lock, flags);
2255 	}
2256 
2257 	/* Guess which priority_group will be used at next mapping time */
2258 	pg = READ_ONCE(m->current_pg);
2259 	next_pg = READ_ONCE(m->next_pg);
2260 	if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2261 		pg = next_pg;
2262 
2263 	if (!pg) {
2264 		/*
2265 		 * We don't know which pg will be used at next mapping time.
2266 		 * We don't call choose_pgpath() here to avoid to trigger
2267 		 * pg_init just by busy checking.
2268 		 * So we don't know whether underlying devices we will be using
2269 		 * at next mapping time are busy or not. Just try mapping.
2270 		 */
2271 		return busy;
2272 	}
2273 
2274 	/*
2275 	 * If there is one non-busy active path at least, the path selector
2276 	 * will be able to select it. So we consider such a pg as not busy.
2277 	 */
2278 	busy = true;
2279 	list_for_each_entry(pgpath, &pg->pgpaths, list) {
2280 		if (pgpath->is_active) {
2281 			has_active = true;
2282 			if (!pgpath_busy(pgpath)) {
2283 				busy = false;
2284 				break;
2285 			}
2286 		}
2287 	}
2288 
2289 	if (!has_active) {
2290 		/*
2291 		 * No active path in this pg, so this pg won't be used and
2292 		 * the current_pg will be changed at next mapping time.
2293 		 * We need to try mapping to determine it.
2294 		 */
2295 		busy = false;
2296 	}
2297 
2298 	return busy;
2299 }
2300 
2301 /*
2302  *---------------------------------------------------------------
2303  * Module setup
2304  *---------------------------------------------------------------
2305  */
2306 static struct target_type multipath_target = {
2307 	.name = "multipath",
2308 	.version = {1, 15, 0},
2309 	.features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2310 		    DM_TARGET_PASSES_INTEGRITY,
2311 	.module = THIS_MODULE,
2312 	.ctr = multipath_ctr,
2313 	.dtr = multipath_dtr,
2314 	.clone_and_map_rq = multipath_clone_and_map,
2315 	.release_clone_rq = multipath_release_clone,
2316 	.rq_end_io = multipath_end_io,
2317 	.map = multipath_map_bio,
2318 	.end_io = multipath_end_io_bio,
2319 	.presuspend = multipath_presuspend,
2320 	.postsuspend = multipath_postsuspend,
2321 	.resume = multipath_resume,
2322 	.status = multipath_status,
2323 	.message = multipath_message,
2324 	.prepare_ioctl = multipath_prepare_ioctl,
2325 	.iterate_devices = multipath_iterate_devices,
2326 	.busy = multipath_busy,
2327 };
2328 
dm_multipath_init(void)2329 static int __init dm_multipath_init(void)
2330 {
2331 	int r = -ENOMEM;
2332 
2333 	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2334 	if (!kmultipathd) {
2335 		DMERR("failed to create workqueue kmpathd");
2336 		goto bad_alloc_kmultipathd;
2337 	}
2338 
2339 	/*
2340 	 * A separate workqueue is used to handle the device handlers
2341 	 * to avoid overloading existing workqueue. Overloading the
2342 	 * old workqueue would also create a bottleneck in the
2343 	 * path of the storage hardware device activation.
2344 	 */
2345 	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2346 						  WQ_MEM_RECLAIM);
2347 	if (!kmpath_handlerd) {
2348 		DMERR("failed to create workqueue kmpath_handlerd");
2349 		goto bad_alloc_kmpath_handlerd;
2350 	}
2351 
2352 	dm_mpath_wq = alloc_workqueue("dm_mpath_wq", 0, 0);
2353 	if (!dm_mpath_wq) {
2354 		DMERR("failed to create workqueue dm_mpath_wq");
2355 		goto bad_alloc_dm_mpath_wq;
2356 	}
2357 
2358 	r = dm_register_target(&multipath_target);
2359 	if (r < 0)
2360 		goto bad_register_target;
2361 
2362 	return 0;
2363 
2364 bad_register_target:
2365 	destroy_workqueue(dm_mpath_wq);
2366 bad_alloc_dm_mpath_wq:
2367 	destroy_workqueue(kmpath_handlerd);
2368 bad_alloc_kmpath_handlerd:
2369 	destroy_workqueue(kmultipathd);
2370 bad_alloc_kmultipathd:
2371 	return r;
2372 }
2373 
dm_multipath_exit(void)2374 static void __exit dm_multipath_exit(void)
2375 {
2376 	destroy_workqueue(dm_mpath_wq);
2377 	destroy_workqueue(kmpath_handlerd);
2378 	destroy_workqueue(kmultipathd);
2379 
2380 	dm_unregister_target(&multipath_target);
2381 }
2382 
2383 module_init(dm_multipath_init);
2384 module_exit(dm_multipath_exit);
2385 
2386 module_param_named(queue_if_no_path_timeout_secs, queue_if_no_path_timeout_secs, ulong, 0644);
2387 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2388 
2389 MODULE_DESCRIPTION(DM_NAME " multipath target");
2390 MODULE_AUTHOR("Sistina Software <dm-devel@lists.linux.dev>");
2391 MODULE_LICENSE("GPL");
2392