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