Loading...
1/*
2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2005-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-bio-record.h"
9
10#include <linux/init.h>
11#include <linux/mempool.h>
12#include <linux/module.h>
13#include <linux/pagemap.h>
14#include <linux/slab.h>
15#include <linux/workqueue.h>
16#include <linux/device-mapper.h>
17#include <linux/dm-io.h>
18#include <linux/dm-dirty-log.h>
19#include <linux/dm-kcopyd.h>
20#include <linux/dm-region-hash.h>
21
22#define DM_MSG_PREFIX "raid1"
23
24#define MAX_RECOVERY 1 /* Maximum number of regions recovered in parallel. */
25
26#define MAX_NR_MIRRORS (DM_KCOPYD_MAX_REGIONS + 1)
27
28#define DM_RAID1_HANDLE_ERRORS 0x01
29#define DM_RAID1_KEEP_LOG 0x02
30#define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
31#define keep_log(p) ((p)->features & DM_RAID1_KEEP_LOG)
32
33static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
34
35/*-----------------------------------------------------------------
36 * Mirror set structures.
37 *---------------------------------------------------------------*/
38enum dm_raid1_error {
39 DM_RAID1_WRITE_ERROR,
40 DM_RAID1_FLUSH_ERROR,
41 DM_RAID1_SYNC_ERROR,
42 DM_RAID1_READ_ERROR
43};
44
45struct mirror {
46 struct mirror_set *ms;
47 atomic_t error_count;
48 unsigned long error_type;
49 struct dm_dev *dev;
50 sector_t offset;
51};
52
53struct mirror_set {
54 struct dm_target *ti;
55 struct list_head list;
56
57 uint64_t features;
58
59 spinlock_t lock; /* protects the lists */
60 struct bio_list reads;
61 struct bio_list writes;
62 struct bio_list failures;
63 struct bio_list holds; /* bios are waiting until suspend */
64
65 struct dm_region_hash *rh;
66 struct dm_kcopyd_client *kcopyd_client;
67 struct dm_io_client *io_client;
68
69 /* recovery */
70 region_t nr_regions;
71 int in_sync;
72 int log_failure;
73 int leg_failure;
74 atomic_t suspend;
75
76 atomic_t default_mirror; /* Default mirror */
77
78 struct workqueue_struct *kmirrord_wq;
79 struct work_struct kmirrord_work;
80 struct timer_list timer;
81 unsigned long timer_pending;
82
83 struct work_struct trigger_event;
84
85 unsigned nr_mirrors;
86 struct mirror mirror[];
87};
88
89DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(raid1_resync_throttle,
90 "A percentage of time allocated for raid resynchronization");
91
92static void wakeup_mirrord(void *context)
93{
94 struct mirror_set *ms = context;
95
96 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
97}
98
99static void delayed_wake_fn(struct timer_list *t)
100{
101 struct mirror_set *ms = from_timer(ms, t, timer);
102
103 clear_bit(0, &ms->timer_pending);
104 wakeup_mirrord(ms);
105}
106
107static void delayed_wake(struct mirror_set *ms)
108{
109 if (test_and_set_bit(0, &ms->timer_pending))
110 return;
111
112 ms->timer.expires = jiffies + HZ / 5;
113 add_timer(&ms->timer);
114}
115
116static void wakeup_all_recovery_waiters(void *context)
117{
118 wake_up_all(&_kmirrord_recovery_stopped);
119}
120
121static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
122{
123 unsigned long flags;
124 int should_wake = 0;
125 struct bio_list *bl;
126
127 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
128 spin_lock_irqsave(&ms->lock, flags);
129 should_wake = !(bl->head);
130 bio_list_add(bl, bio);
131 spin_unlock_irqrestore(&ms->lock, flags);
132
133 if (should_wake)
134 wakeup_mirrord(ms);
135}
136
137static void dispatch_bios(void *context, struct bio_list *bio_list)
138{
139 struct mirror_set *ms = context;
140 struct bio *bio;
141
142 while ((bio = bio_list_pop(bio_list)))
143 queue_bio(ms, bio, WRITE);
144}
145
146struct dm_raid1_bio_record {
147 struct mirror *m;
148 /* if details->bi_disk == NULL, details were not saved */
149 struct dm_bio_details details;
150 region_t write_region;
151};
152
153/*
154 * Every mirror should look like this one.
155 */
156#define DEFAULT_MIRROR 0
157
158/*
159 * This is yucky. We squirrel the mirror struct away inside
160 * bi_next for read/write buffers. This is safe since the bh
161 * doesn't get submitted to the lower levels of block layer.
162 */
163static struct mirror *bio_get_m(struct bio *bio)
164{
165 return (struct mirror *) bio->bi_next;
166}
167
168static void bio_set_m(struct bio *bio, struct mirror *m)
169{
170 bio->bi_next = (struct bio *) m;
171}
172
173static struct mirror *get_default_mirror(struct mirror_set *ms)
174{
175 return &ms->mirror[atomic_read(&ms->default_mirror)];
176}
177
178static void set_default_mirror(struct mirror *m)
179{
180 struct mirror_set *ms = m->ms;
181 struct mirror *m0 = &(ms->mirror[0]);
182
183 atomic_set(&ms->default_mirror, m - m0);
184}
185
186static struct mirror *get_valid_mirror(struct mirror_set *ms)
187{
188 struct mirror *m;
189
190 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
191 if (!atomic_read(&m->error_count))
192 return m;
193
194 return NULL;
195}
196
197/* fail_mirror
198 * @m: mirror device to fail
199 * @error_type: one of the enum's, DM_RAID1_*_ERROR
200 *
201 * If errors are being handled, record the type of
202 * error encountered for this device. If this type
203 * of error has already been recorded, we can return;
204 * otherwise, we must signal userspace by triggering
205 * an event. Additionally, if the device is the
206 * primary device, we must choose a new primary, but
207 * only if the mirror is in-sync.
208 *
209 * This function must not block.
210 */
211static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
212{
213 struct mirror_set *ms = m->ms;
214 struct mirror *new;
215
216 ms->leg_failure = 1;
217
218 /*
219 * error_count is used for nothing more than a
220 * simple way to tell if a device has encountered
221 * errors.
222 */
223 atomic_inc(&m->error_count);
224
225 if (test_and_set_bit(error_type, &m->error_type))
226 return;
227
228 if (!errors_handled(ms))
229 return;
230
231 if (m != get_default_mirror(ms))
232 goto out;
233
234 if (!ms->in_sync && !keep_log(ms)) {
235 /*
236 * Better to issue requests to same failing device
237 * than to risk returning corrupt data.
238 */
239 DMERR("Primary mirror (%s) failed while out-of-sync: "
240 "Reads may fail.", m->dev->name);
241 goto out;
242 }
243
244 new = get_valid_mirror(ms);
245 if (new)
246 set_default_mirror(new);
247 else
248 DMWARN("All sides of mirror have failed.");
249
250out:
251 schedule_work(&ms->trigger_event);
252}
253
254static int mirror_flush(struct dm_target *ti)
255{
256 struct mirror_set *ms = ti->private;
257 unsigned long error_bits;
258
259 unsigned int i;
260 struct dm_io_region io[MAX_NR_MIRRORS];
261 struct mirror *m;
262 struct dm_io_request io_req = {
263 .bi_op = REQ_OP_WRITE,
264 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
265 .mem.type = DM_IO_KMEM,
266 .mem.ptr.addr = NULL,
267 .client = ms->io_client,
268 };
269
270 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
271 io[i].bdev = m->dev->bdev;
272 io[i].sector = 0;
273 io[i].count = 0;
274 }
275
276 error_bits = -1;
277 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
278 if (unlikely(error_bits != 0)) {
279 for (i = 0; i < ms->nr_mirrors; i++)
280 if (test_bit(i, &error_bits))
281 fail_mirror(ms->mirror + i,
282 DM_RAID1_FLUSH_ERROR);
283 return -EIO;
284 }
285
286 return 0;
287}
288
289/*-----------------------------------------------------------------
290 * Recovery.
291 *
292 * When a mirror is first activated we may find that some regions
293 * are in the no-sync state. We have to recover these by
294 * recopying from the default mirror to all the others.
295 *---------------------------------------------------------------*/
296static void recovery_complete(int read_err, unsigned long write_err,
297 void *context)
298{
299 struct dm_region *reg = context;
300 struct mirror_set *ms = dm_rh_region_context(reg);
301 int m, bit = 0;
302
303 if (read_err) {
304 /* Read error means the failure of default mirror. */
305 DMERR_LIMIT("Unable to read primary mirror during recovery");
306 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
307 }
308
309 if (write_err) {
310 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
311 write_err);
312 /*
313 * Bits correspond to devices (excluding default mirror).
314 * The default mirror cannot change during recovery.
315 */
316 for (m = 0; m < ms->nr_mirrors; m++) {
317 if (&ms->mirror[m] == get_default_mirror(ms))
318 continue;
319 if (test_bit(bit, &write_err))
320 fail_mirror(ms->mirror + m,
321 DM_RAID1_SYNC_ERROR);
322 bit++;
323 }
324 }
325
326 dm_rh_recovery_end(reg, !(read_err || write_err));
327}
328
329static void recover(struct mirror_set *ms, struct dm_region *reg)
330{
331 unsigned i;
332 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
333 struct mirror *m;
334 unsigned long flags = 0;
335 region_t key = dm_rh_get_region_key(reg);
336 sector_t region_size = dm_rh_get_region_size(ms->rh);
337
338 /* fill in the source */
339 m = get_default_mirror(ms);
340 from.bdev = m->dev->bdev;
341 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
342 if (key == (ms->nr_regions - 1)) {
343 /*
344 * The final region may be smaller than
345 * region_size.
346 */
347 from.count = ms->ti->len & (region_size - 1);
348 if (!from.count)
349 from.count = region_size;
350 } else
351 from.count = region_size;
352
353 /* fill in the destinations */
354 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
355 if (&ms->mirror[i] == get_default_mirror(ms))
356 continue;
357
358 m = ms->mirror + i;
359 dest->bdev = m->dev->bdev;
360 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
361 dest->count = from.count;
362 dest++;
363 }
364
365 /* hand to kcopyd */
366 if (!errors_handled(ms))
367 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
368
369 dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
370 flags, recovery_complete, reg);
371}
372
373static void reset_ms_flags(struct mirror_set *ms)
374{
375 unsigned int m;
376
377 ms->leg_failure = 0;
378 for (m = 0; m < ms->nr_mirrors; m++) {
379 atomic_set(&(ms->mirror[m].error_count), 0);
380 ms->mirror[m].error_type = 0;
381 }
382}
383
384static void do_recovery(struct mirror_set *ms)
385{
386 struct dm_region *reg;
387 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
388
389 /*
390 * Start quiescing some regions.
391 */
392 dm_rh_recovery_prepare(ms->rh);
393
394 /*
395 * Copy any already quiesced regions.
396 */
397 while ((reg = dm_rh_recovery_start(ms->rh)))
398 recover(ms, reg);
399
400 /*
401 * Update the in sync flag.
402 */
403 if (!ms->in_sync &&
404 (log->type->get_sync_count(log) == ms->nr_regions)) {
405 /* the sync is complete */
406 dm_table_event(ms->ti->table);
407 ms->in_sync = 1;
408 reset_ms_flags(ms);
409 }
410}
411
412/*-----------------------------------------------------------------
413 * Reads
414 *---------------------------------------------------------------*/
415static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
416{
417 struct mirror *m = get_default_mirror(ms);
418
419 do {
420 if (likely(!atomic_read(&m->error_count)))
421 return m;
422
423 if (m-- == ms->mirror)
424 m += ms->nr_mirrors;
425 } while (m != get_default_mirror(ms));
426
427 return NULL;
428}
429
430static int default_ok(struct mirror *m)
431{
432 struct mirror *default_mirror = get_default_mirror(m->ms);
433
434 return !atomic_read(&default_mirror->error_count);
435}
436
437static int mirror_available(struct mirror_set *ms, struct bio *bio)
438{
439 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
440 region_t region = dm_rh_bio_to_region(ms->rh, bio);
441
442 if (log->type->in_sync(log, region, 0))
443 return choose_mirror(ms, bio->bi_iter.bi_sector) ? 1 : 0;
444
445 return 0;
446}
447
448/*
449 * remap a buffer to a particular mirror.
450 */
451static sector_t map_sector(struct mirror *m, struct bio *bio)
452{
453 if (unlikely(!bio->bi_iter.bi_size))
454 return 0;
455 return m->offset + dm_target_offset(m->ms->ti, bio->bi_iter.bi_sector);
456}
457
458static void map_bio(struct mirror *m, struct bio *bio)
459{
460 bio_set_dev(bio, m->dev->bdev);
461 bio->bi_iter.bi_sector = map_sector(m, bio);
462}
463
464static void map_region(struct dm_io_region *io, struct mirror *m,
465 struct bio *bio)
466{
467 io->bdev = m->dev->bdev;
468 io->sector = map_sector(m, bio);
469 io->count = bio_sectors(bio);
470}
471
472static void hold_bio(struct mirror_set *ms, struct bio *bio)
473{
474 /*
475 * Lock is required to avoid race condition during suspend
476 * process.
477 */
478 spin_lock_irq(&ms->lock);
479
480 if (atomic_read(&ms->suspend)) {
481 spin_unlock_irq(&ms->lock);
482
483 /*
484 * If device is suspended, complete the bio.
485 */
486 if (dm_noflush_suspending(ms->ti))
487 bio->bi_status = BLK_STS_DM_REQUEUE;
488 else
489 bio->bi_status = BLK_STS_IOERR;
490
491 bio_endio(bio);
492 return;
493 }
494
495 /*
496 * Hold bio until the suspend is complete.
497 */
498 bio_list_add(&ms->holds, bio);
499 spin_unlock_irq(&ms->lock);
500}
501
502/*-----------------------------------------------------------------
503 * Reads
504 *---------------------------------------------------------------*/
505static void read_callback(unsigned long error, void *context)
506{
507 struct bio *bio = context;
508 struct mirror *m;
509
510 m = bio_get_m(bio);
511 bio_set_m(bio, NULL);
512
513 if (likely(!error)) {
514 bio_endio(bio);
515 return;
516 }
517
518 fail_mirror(m, DM_RAID1_READ_ERROR);
519
520 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
521 DMWARN_LIMIT("Read failure on mirror device %s. "
522 "Trying alternative device.",
523 m->dev->name);
524 queue_bio(m->ms, bio, bio_data_dir(bio));
525 return;
526 }
527
528 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
529 m->dev->name);
530 bio_io_error(bio);
531}
532
533/* Asynchronous read. */
534static void read_async_bio(struct mirror *m, struct bio *bio)
535{
536 struct dm_io_region io;
537 struct dm_io_request io_req = {
538 .bi_op = REQ_OP_READ,
539 .bi_op_flags = 0,
540 .mem.type = DM_IO_BIO,
541 .mem.ptr.bio = bio,
542 .notify.fn = read_callback,
543 .notify.context = bio,
544 .client = m->ms->io_client,
545 };
546
547 map_region(&io, m, bio);
548 bio_set_m(bio, m);
549 BUG_ON(dm_io(&io_req, 1, &io, NULL));
550}
551
552static inline int region_in_sync(struct mirror_set *ms, region_t region,
553 int may_block)
554{
555 int state = dm_rh_get_state(ms->rh, region, may_block);
556 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
557}
558
559static void do_reads(struct mirror_set *ms, struct bio_list *reads)
560{
561 region_t region;
562 struct bio *bio;
563 struct mirror *m;
564
565 while ((bio = bio_list_pop(reads))) {
566 region = dm_rh_bio_to_region(ms->rh, bio);
567 m = get_default_mirror(ms);
568
569 /*
570 * We can only read balance if the region is in sync.
571 */
572 if (likely(region_in_sync(ms, region, 1)))
573 m = choose_mirror(ms, bio->bi_iter.bi_sector);
574 else if (m && atomic_read(&m->error_count))
575 m = NULL;
576
577 if (likely(m))
578 read_async_bio(m, bio);
579 else
580 bio_io_error(bio);
581 }
582}
583
584/*-----------------------------------------------------------------
585 * Writes.
586 *
587 * We do different things with the write io depending on the
588 * state of the region that it's in:
589 *
590 * SYNC: increment pending, use kcopyd to write to *all* mirrors
591 * RECOVERING: delay the io until recovery completes
592 * NOSYNC: increment pending, just write to the default mirror
593 *---------------------------------------------------------------*/
594
595
596static void write_callback(unsigned long error, void *context)
597{
598 unsigned i;
599 struct bio *bio = (struct bio *) context;
600 struct mirror_set *ms;
601 int should_wake = 0;
602 unsigned long flags;
603
604 ms = bio_get_m(bio)->ms;
605 bio_set_m(bio, NULL);
606
607 /*
608 * NOTE: We don't decrement the pending count here,
609 * instead it is done by the targets endio function.
610 * This way we handle both writes to SYNC and NOSYNC
611 * regions with the same code.
612 */
613 if (likely(!error)) {
614 bio_endio(bio);
615 return;
616 }
617
618 /*
619 * If the bio is discard, return an error, but do not
620 * degrade the array.
621 */
622 if (bio_op(bio) == REQ_OP_DISCARD) {
623 bio->bi_status = BLK_STS_NOTSUPP;
624 bio_endio(bio);
625 return;
626 }
627
628 for (i = 0; i < ms->nr_mirrors; i++)
629 if (test_bit(i, &error))
630 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
631
632 /*
633 * Need to raise event. Since raising
634 * events can block, we need to do it in
635 * the main thread.
636 */
637 spin_lock_irqsave(&ms->lock, flags);
638 if (!ms->failures.head)
639 should_wake = 1;
640 bio_list_add(&ms->failures, bio);
641 spin_unlock_irqrestore(&ms->lock, flags);
642 if (should_wake)
643 wakeup_mirrord(ms);
644}
645
646static void do_write(struct mirror_set *ms, struct bio *bio)
647{
648 unsigned int i;
649 struct dm_io_region io[MAX_NR_MIRRORS], *dest = io;
650 struct mirror *m;
651 struct dm_io_request io_req = {
652 .bi_op = REQ_OP_WRITE,
653 .bi_op_flags = bio->bi_opf & (REQ_FUA | REQ_PREFLUSH),
654 .mem.type = DM_IO_BIO,
655 .mem.ptr.bio = bio,
656 .notify.fn = write_callback,
657 .notify.context = bio,
658 .client = ms->io_client,
659 };
660
661 if (bio_op(bio) == REQ_OP_DISCARD) {
662 io_req.bi_op = REQ_OP_DISCARD;
663 io_req.mem.type = DM_IO_KMEM;
664 io_req.mem.ptr.addr = NULL;
665 }
666
667 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
668 map_region(dest++, m, bio);
669
670 /*
671 * Use default mirror because we only need it to retrieve the reference
672 * to the mirror set in write_callback().
673 */
674 bio_set_m(bio, get_default_mirror(ms));
675
676 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
677}
678
679static void do_writes(struct mirror_set *ms, struct bio_list *writes)
680{
681 int state;
682 struct bio *bio;
683 struct bio_list sync, nosync, recover, *this_list = NULL;
684 struct bio_list requeue;
685 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
686 region_t region;
687
688 if (!writes->head)
689 return;
690
691 /*
692 * Classify each write.
693 */
694 bio_list_init(&sync);
695 bio_list_init(&nosync);
696 bio_list_init(&recover);
697 bio_list_init(&requeue);
698
699 while ((bio = bio_list_pop(writes))) {
700 if ((bio->bi_opf & REQ_PREFLUSH) ||
701 (bio_op(bio) == REQ_OP_DISCARD)) {
702 bio_list_add(&sync, bio);
703 continue;
704 }
705
706 region = dm_rh_bio_to_region(ms->rh, bio);
707
708 if (log->type->is_remote_recovering &&
709 log->type->is_remote_recovering(log, region)) {
710 bio_list_add(&requeue, bio);
711 continue;
712 }
713
714 state = dm_rh_get_state(ms->rh, region, 1);
715 switch (state) {
716 case DM_RH_CLEAN:
717 case DM_RH_DIRTY:
718 this_list = &sync;
719 break;
720
721 case DM_RH_NOSYNC:
722 this_list = &nosync;
723 break;
724
725 case DM_RH_RECOVERING:
726 this_list = &recover;
727 break;
728 }
729
730 bio_list_add(this_list, bio);
731 }
732
733 /*
734 * Add bios that are delayed due to remote recovery
735 * back on to the write queue
736 */
737 if (unlikely(requeue.head)) {
738 spin_lock_irq(&ms->lock);
739 bio_list_merge(&ms->writes, &requeue);
740 spin_unlock_irq(&ms->lock);
741 delayed_wake(ms);
742 }
743
744 /*
745 * Increment the pending counts for any regions that will
746 * be written to (writes to recover regions are going to
747 * be delayed).
748 */
749 dm_rh_inc_pending(ms->rh, &sync);
750 dm_rh_inc_pending(ms->rh, &nosync);
751
752 /*
753 * If the flush fails on a previous call and succeeds here,
754 * we must not reset the log_failure variable. We need
755 * userspace interaction to do that.
756 */
757 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
758
759 /*
760 * Dispatch io.
761 */
762 if (unlikely(ms->log_failure) && errors_handled(ms)) {
763 spin_lock_irq(&ms->lock);
764 bio_list_merge(&ms->failures, &sync);
765 spin_unlock_irq(&ms->lock);
766 wakeup_mirrord(ms);
767 } else
768 while ((bio = bio_list_pop(&sync)))
769 do_write(ms, bio);
770
771 while ((bio = bio_list_pop(&recover)))
772 dm_rh_delay(ms->rh, bio);
773
774 while ((bio = bio_list_pop(&nosync))) {
775 if (unlikely(ms->leg_failure) && errors_handled(ms) && !keep_log(ms)) {
776 spin_lock_irq(&ms->lock);
777 bio_list_add(&ms->failures, bio);
778 spin_unlock_irq(&ms->lock);
779 wakeup_mirrord(ms);
780 } else {
781 map_bio(get_default_mirror(ms), bio);
782 submit_bio_noacct(bio);
783 }
784 }
785}
786
787static void do_failures(struct mirror_set *ms, struct bio_list *failures)
788{
789 struct bio *bio;
790
791 if (likely(!failures->head))
792 return;
793
794 /*
795 * If the log has failed, unattempted writes are being
796 * put on the holds list. We can't issue those writes
797 * until a log has been marked, so we must store them.
798 *
799 * If a 'noflush' suspend is in progress, we can requeue
800 * the I/O's to the core. This give userspace a chance
801 * to reconfigure the mirror, at which point the core
802 * will reissue the writes. If the 'noflush' flag is
803 * not set, we have no choice but to return errors.
804 *
805 * Some writes on the failures list may have been
806 * submitted before the log failure and represent a
807 * failure to write to one of the devices. It is ok
808 * for us to treat them the same and requeue them
809 * as well.
810 */
811 while ((bio = bio_list_pop(failures))) {
812 if (!ms->log_failure) {
813 ms->in_sync = 0;
814 dm_rh_mark_nosync(ms->rh, bio);
815 }
816
817 /*
818 * If all the legs are dead, fail the I/O.
819 * If the device has failed and keep_log is enabled,
820 * fail the I/O.
821 *
822 * If we have been told to handle errors, and keep_log
823 * isn't enabled, hold the bio and wait for userspace to
824 * deal with the problem.
825 *
826 * Otherwise pretend that the I/O succeeded. (This would
827 * be wrong if the failed leg returned after reboot and
828 * got replicated back to the good legs.)
829 */
830 if (unlikely(!get_valid_mirror(ms) || (keep_log(ms) && ms->log_failure)))
831 bio_io_error(bio);
832 else if (errors_handled(ms) && !keep_log(ms))
833 hold_bio(ms, bio);
834 else
835 bio_endio(bio);
836 }
837}
838
839static void trigger_event(struct work_struct *work)
840{
841 struct mirror_set *ms =
842 container_of(work, struct mirror_set, trigger_event);
843
844 dm_table_event(ms->ti->table);
845}
846
847/*-----------------------------------------------------------------
848 * kmirrord
849 *---------------------------------------------------------------*/
850static void do_mirror(struct work_struct *work)
851{
852 struct mirror_set *ms = container_of(work, struct mirror_set,
853 kmirrord_work);
854 struct bio_list reads, writes, failures;
855 unsigned long flags;
856
857 spin_lock_irqsave(&ms->lock, flags);
858 reads = ms->reads;
859 writes = ms->writes;
860 failures = ms->failures;
861 bio_list_init(&ms->reads);
862 bio_list_init(&ms->writes);
863 bio_list_init(&ms->failures);
864 spin_unlock_irqrestore(&ms->lock, flags);
865
866 dm_rh_update_states(ms->rh, errors_handled(ms));
867 do_recovery(ms);
868 do_reads(ms, &reads);
869 do_writes(ms, &writes);
870 do_failures(ms, &failures);
871}
872
873/*-----------------------------------------------------------------
874 * Target functions
875 *---------------------------------------------------------------*/
876static struct mirror_set *alloc_context(unsigned int nr_mirrors,
877 uint32_t region_size,
878 struct dm_target *ti,
879 struct dm_dirty_log *dl)
880{
881 struct mirror_set *ms =
882 kzalloc(struct_size(ms, mirror, nr_mirrors), GFP_KERNEL);
883
884 if (!ms) {
885 ti->error = "Cannot allocate mirror context";
886 return NULL;
887 }
888
889 spin_lock_init(&ms->lock);
890 bio_list_init(&ms->reads);
891 bio_list_init(&ms->writes);
892 bio_list_init(&ms->failures);
893 bio_list_init(&ms->holds);
894
895 ms->ti = ti;
896 ms->nr_mirrors = nr_mirrors;
897 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
898 ms->in_sync = 0;
899 ms->log_failure = 0;
900 ms->leg_failure = 0;
901 atomic_set(&ms->suspend, 0);
902 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
903
904 ms->io_client = dm_io_client_create();
905 if (IS_ERR(ms->io_client)) {
906 ti->error = "Error creating dm_io client";
907 kfree(ms);
908 return NULL;
909 }
910
911 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
912 wakeup_all_recovery_waiters,
913 ms->ti->begin, MAX_RECOVERY,
914 dl, region_size, ms->nr_regions);
915 if (IS_ERR(ms->rh)) {
916 ti->error = "Error creating dirty region hash";
917 dm_io_client_destroy(ms->io_client);
918 kfree(ms);
919 return NULL;
920 }
921
922 return ms;
923}
924
925static void free_context(struct mirror_set *ms, struct dm_target *ti,
926 unsigned int m)
927{
928 while (m--)
929 dm_put_device(ti, ms->mirror[m].dev);
930
931 dm_io_client_destroy(ms->io_client);
932 dm_region_hash_destroy(ms->rh);
933 kfree(ms);
934}
935
936static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
937 unsigned int mirror, char **argv)
938{
939 unsigned long long offset;
940 char dummy;
941 int ret;
942
943 if (sscanf(argv[1], "%llu%c", &offset, &dummy) != 1 ||
944 offset != (sector_t)offset) {
945 ti->error = "Invalid offset";
946 return -EINVAL;
947 }
948
949 ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
950 &ms->mirror[mirror].dev);
951 if (ret) {
952 ti->error = "Device lookup failure";
953 return ret;
954 }
955
956 ms->mirror[mirror].ms = ms;
957 atomic_set(&(ms->mirror[mirror].error_count), 0);
958 ms->mirror[mirror].error_type = 0;
959 ms->mirror[mirror].offset = offset;
960
961 return 0;
962}
963
964/*
965 * Create dirty log: log_type #log_params <log_params>
966 */
967static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
968 unsigned argc, char **argv,
969 unsigned *args_used)
970{
971 unsigned param_count;
972 struct dm_dirty_log *dl;
973 char dummy;
974
975 if (argc < 2) {
976 ti->error = "Insufficient mirror log arguments";
977 return NULL;
978 }
979
980 if (sscanf(argv[1], "%u%c", ¶m_count, &dummy) != 1) {
981 ti->error = "Invalid mirror log argument count";
982 return NULL;
983 }
984
985 *args_used = 2 + param_count;
986
987 if (argc < *args_used) {
988 ti->error = "Insufficient mirror log arguments";
989 return NULL;
990 }
991
992 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
993 argv + 2);
994 if (!dl) {
995 ti->error = "Error creating mirror dirty log";
996 return NULL;
997 }
998
999 return dl;
1000}
1001
1002static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
1003 unsigned *args_used)
1004{
1005 unsigned num_features;
1006 struct dm_target *ti = ms->ti;
1007 char dummy;
1008 int i;
1009
1010 *args_used = 0;
1011
1012 if (!argc)
1013 return 0;
1014
1015 if (sscanf(argv[0], "%u%c", &num_features, &dummy) != 1) {
1016 ti->error = "Invalid number of features";
1017 return -EINVAL;
1018 }
1019
1020 argc--;
1021 argv++;
1022 (*args_used)++;
1023
1024 if (num_features > argc) {
1025 ti->error = "Not enough arguments to support feature count";
1026 return -EINVAL;
1027 }
1028
1029 for (i = 0; i < num_features; i++) {
1030 if (!strcmp("handle_errors", argv[0]))
1031 ms->features |= DM_RAID1_HANDLE_ERRORS;
1032 else if (!strcmp("keep_log", argv[0]))
1033 ms->features |= DM_RAID1_KEEP_LOG;
1034 else {
1035 ti->error = "Unrecognised feature requested";
1036 return -EINVAL;
1037 }
1038
1039 argc--;
1040 argv++;
1041 (*args_used)++;
1042 }
1043 if (!errors_handled(ms) && keep_log(ms)) {
1044 ti->error = "keep_log feature requires the handle_errors feature";
1045 return -EINVAL;
1046 }
1047
1048 return 0;
1049}
1050
1051/*
1052 * Construct a mirror mapping:
1053 *
1054 * log_type #log_params <log_params>
1055 * #mirrors [mirror_path offset]{2,}
1056 * [#features <features>]
1057 *
1058 * log_type is "core" or "disk"
1059 * #log_params is between 1 and 3
1060 *
1061 * If present, supported features are "handle_errors" and "keep_log".
1062 */
1063static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1064{
1065 int r;
1066 unsigned int nr_mirrors, m, args_used;
1067 struct mirror_set *ms;
1068 struct dm_dirty_log *dl;
1069 char dummy;
1070
1071 dl = create_dirty_log(ti, argc, argv, &args_used);
1072 if (!dl)
1073 return -EINVAL;
1074
1075 argv += args_used;
1076 argc -= args_used;
1077
1078 if (!argc || sscanf(argv[0], "%u%c", &nr_mirrors, &dummy) != 1 ||
1079 nr_mirrors < 2 || nr_mirrors > MAX_NR_MIRRORS) {
1080 ti->error = "Invalid number of mirrors";
1081 dm_dirty_log_destroy(dl);
1082 return -EINVAL;
1083 }
1084
1085 argv++, argc--;
1086
1087 if (argc < nr_mirrors * 2) {
1088 ti->error = "Too few mirror arguments";
1089 dm_dirty_log_destroy(dl);
1090 return -EINVAL;
1091 }
1092
1093 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1094 if (!ms) {
1095 dm_dirty_log_destroy(dl);
1096 return -ENOMEM;
1097 }
1098
1099 /* Get the mirror parameter sets */
1100 for (m = 0; m < nr_mirrors; m++) {
1101 r = get_mirror(ms, ti, m, argv);
1102 if (r) {
1103 free_context(ms, ti, m);
1104 return r;
1105 }
1106 argv += 2;
1107 argc -= 2;
1108 }
1109
1110 ti->private = ms;
1111
1112 r = dm_set_target_max_io_len(ti, dm_rh_get_region_size(ms->rh));
1113 if (r)
1114 goto err_free_context;
1115
1116 ti->num_flush_bios = 1;
1117 ti->num_discard_bios = 1;
1118 ti->per_io_data_size = sizeof(struct dm_raid1_bio_record);
1119
1120 ms->kmirrord_wq = alloc_workqueue("kmirrord", WQ_MEM_RECLAIM, 0);
1121 if (!ms->kmirrord_wq) {
1122 DMERR("couldn't start kmirrord");
1123 r = -ENOMEM;
1124 goto err_free_context;
1125 }
1126 INIT_WORK(&ms->kmirrord_work, do_mirror);
1127 timer_setup(&ms->timer, delayed_wake_fn, 0);
1128 ms->timer_pending = 0;
1129 INIT_WORK(&ms->trigger_event, trigger_event);
1130
1131 r = parse_features(ms, argc, argv, &args_used);
1132 if (r)
1133 goto err_destroy_wq;
1134
1135 argv += args_used;
1136 argc -= args_used;
1137
1138 /*
1139 * Any read-balancing addition depends on the
1140 * DM_RAID1_HANDLE_ERRORS flag being present.
1141 * This is because the decision to balance depends
1142 * on the sync state of a region. If the above
1143 * flag is not present, we ignore errors; and
1144 * the sync state may be inaccurate.
1145 */
1146
1147 if (argc) {
1148 ti->error = "Too many mirror arguments";
1149 r = -EINVAL;
1150 goto err_destroy_wq;
1151 }
1152
1153 ms->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1154 if (IS_ERR(ms->kcopyd_client)) {
1155 r = PTR_ERR(ms->kcopyd_client);
1156 goto err_destroy_wq;
1157 }
1158
1159 wakeup_mirrord(ms);
1160 return 0;
1161
1162err_destroy_wq:
1163 destroy_workqueue(ms->kmirrord_wq);
1164err_free_context:
1165 free_context(ms, ti, ms->nr_mirrors);
1166 return r;
1167}
1168
1169static void mirror_dtr(struct dm_target *ti)
1170{
1171 struct mirror_set *ms = (struct mirror_set *) ti->private;
1172
1173 del_timer_sync(&ms->timer);
1174 flush_workqueue(ms->kmirrord_wq);
1175 flush_work(&ms->trigger_event);
1176 dm_kcopyd_client_destroy(ms->kcopyd_client);
1177 destroy_workqueue(ms->kmirrord_wq);
1178 free_context(ms, ti, ms->nr_mirrors);
1179}
1180
1181/*
1182 * Mirror mapping function
1183 */
1184static int mirror_map(struct dm_target *ti, struct bio *bio)
1185{
1186 int r, rw = bio_data_dir(bio);
1187 struct mirror *m;
1188 struct mirror_set *ms = ti->private;
1189 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1190 struct dm_raid1_bio_record *bio_record =
1191 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1192
1193 bio_record->details.bi_disk = NULL;
1194
1195 if (rw == WRITE) {
1196 /* Save region for mirror_end_io() handler */
1197 bio_record->write_region = dm_rh_bio_to_region(ms->rh, bio);
1198 queue_bio(ms, bio, rw);
1199 return DM_MAPIO_SUBMITTED;
1200 }
1201
1202 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1203 if (r < 0 && r != -EWOULDBLOCK)
1204 return DM_MAPIO_KILL;
1205
1206 /*
1207 * If region is not in-sync queue the bio.
1208 */
1209 if (!r || (r == -EWOULDBLOCK)) {
1210 if (bio->bi_opf & REQ_RAHEAD)
1211 return DM_MAPIO_KILL;
1212
1213 queue_bio(ms, bio, rw);
1214 return DM_MAPIO_SUBMITTED;
1215 }
1216
1217 /*
1218 * The region is in-sync and we can perform reads directly.
1219 * Store enough information so we can retry if it fails.
1220 */
1221 m = choose_mirror(ms, bio->bi_iter.bi_sector);
1222 if (unlikely(!m))
1223 return DM_MAPIO_KILL;
1224
1225 dm_bio_record(&bio_record->details, bio);
1226 bio_record->m = m;
1227
1228 map_bio(m, bio);
1229
1230 return DM_MAPIO_REMAPPED;
1231}
1232
1233static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1234 blk_status_t *error)
1235{
1236 int rw = bio_data_dir(bio);
1237 struct mirror_set *ms = (struct mirror_set *) ti->private;
1238 struct mirror *m = NULL;
1239 struct dm_bio_details *bd = NULL;
1240 struct dm_raid1_bio_record *bio_record =
1241 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1242
1243 /*
1244 * We need to dec pending if this was a write.
1245 */
1246 if (rw == WRITE) {
1247 if (!(bio->bi_opf & REQ_PREFLUSH) &&
1248 bio_op(bio) != REQ_OP_DISCARD)
1249 dm_rh_dec(ms->rh, bio_record->write_region);
1250 return DM_ENDIO_DONE;
1251 }
1252
1253 if (*error == BLK_STS_NOTSUPP)
1254 goto out;
1255
1256 if (bio->bi_opf & REQ_RAHEAD)
1257 goto out;
1258
1259 if (unlikely(*error)) {
1260 if (!bio_record->details.bi_disk) {
1261 /*
1262 * There wasn't enough memory to record necessary
1263 * information for a retry or there was no other
1264 * mirror in-sync.
1265 */
1266 DMERR_LIMIT("Mirror read failed.");
1267 return DM_ENDIO_DONE;
1268 }
1269
1270 m = bio_record->m;
1271
1272 DMERR("Mirror read failed from %s. Trying alternative device.",
1273 m->dev->name);
1274
1275 fail_mirror(m, DM_RAID1_READ_ERROR);
1276
1277 /*
1278 * A failed read is requeued for another attempt using an intact
1279 * mirror.
1280 */
1281 if (default_ok(m) || mirror_available(ms, bio)) {
1282 bd = &bio_record->details;
1283
1284 dm_bio_restore(bd, bio);
1285 bio_record->details.bi_disk = NULL;
1286 bio->bi_status = 0;
1287
1288 queue_bio(ms, bio, rw);
1289 return DM_ENDIO_INCOMPLETE;
1290 }
1291 DMERR("All replicated volumes dead, failing I/O");
1292 }
1293
1294out:
1295 bio_record->details.bi_disk = NULL;
1296
1297 return DM_ENDIO_DONE;
1298}
1299
1300static void mirror_presuspend(struct dm_target *ti)
1301{
1302 struct mirror_set *ms = (struct mirror_set *) ti->private;
1303 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1304
1305 struct bio_list holds;
1306 struct bio *bio;
1307
1308 atomic_set(&ms->suspend, 1);
1309
1310 /*
1311 * Process bios in the hold list to start recovery waiting
1312 * for bios in the hold list. After the process, no bio has
1313 * a chance to be added in the hold list because ms->suspend
1314 * is set.
1315 */
1316 spin_lock_irq(&ms->lock);
1317 holds = ms->holds;
1318 bio_list_init(&ms->holds);
1319 spin_unlock_irq(&ms->lock);
1320
1321 while ((bio = bio_list_pop(&holds)))
1322 hold_bio(ms, bio);
1323
1324 /*
1325 * We must finish up all the work that we've
1326 * generated (i.e. recovery work).
1327 */
1328 dm_rh_stop_recovery(ms->rh);
1329
1330 wait_event(_kmirrord_recovery_stopped,
1331 !dm_rh_recovery_in_flight(ms->rh));
1332
1333 if (log->type->presuspend && log->type->presuspend(log))
1334 /* FIXME: need better error handling */
1335 DMWARN("log presuspend failed");
1336
1337 /*
1338 * Now that recovery is complete/stopped and the
1339 * delayed bios are queued, we need to wait for
1340 * the worker thread to complete. This way,
1341 * we know that all of our I/O has been pushed.
1342 */
1343 flush_workqueue(ms->kmirrord_wq);
1344}
1345
1346static void mirror_postsuspend(struct dm_target *ti)
1347{
1348 struct mirror_set *ms = ti->private;
1349 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1350
1351 if (log->type->postsuspend && log->type->postsuspend(log))
1352 /* FIXME: need better error handling */
1353 DMWARN("log postsuspend failed");
1354}
1355
1356static void mirror_resume(struct dm_target *ti)
1357{
1358 struct mirror_set *ms = ti->private;
1359 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1360
1361 atomic_set(&ms->suspend, 0);
1362 if (log->type->resume && log->type->resume(log))
1363 /* FIXME: need better error handling */
1364 DMWARN("log resume failed");
1365 dm_rh_start_recovery(ms->rh);
1366}
1367
1368/*
1369 * device_status_char
1370 * @m: mirror device/leg we want the status of
1371 *
1372 * We return one character representing the most severe error
1373 * we have encountered.
1374 * A => Alive - No failures
1375 * D => Dead - A write failure occurred leaving mirror out-of-sync
1376 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1377 * R => Read - A read failure occurred, mirror data unaffected
1378 *
1379 * Returns: <char>
1380 */
1381static char device_status_char(struct mirror *m)
1382{
1383 if (!atomic_read(&(m->error_count)))
1384 return 'A';
1385
1386 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1387 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1388 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1389 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1390}
1391
1392
1393static void mirror_status(struct dm_target *ti, status_type_t type,
1394 unsigned status_flags, char *result, unsigned maxlen)
1395{
1396 unsigned int m, sz = 0;
1397 int num_feature_args = 0;
1398 struct mirror_set *ms = (struct mirror_set *) ti->private;
1399 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1400 char buffer[MAX_NR_MIRRORS + 1];
1401
1402 switch (type) {
1403 case STATUSTYPE_INFO:
1404 DMEMIT("%d ", ms->nr_mirrors);
1405 for (m = 0; m < ms->nr_mirrors; m++) {
1406 DMEMIT("%s ", ms->mirror[m].dev->name);
1407 buffer[m] = device_status_char(&(ms->mirror[m]));
1408 }
1409 buffer[m] = '\0';
1410
1411 DMEMIT("%llu/%llu 1 %s ",
1412 (unsigned long long)log->type->get_sync_count(log),
1413 (unsigned long long)ms->nr_regions, buffer);
1414
1415 sz += log->type->status(log, type, result+sz, maxlen-sz);
1416
1417 break;
1418
1419 case STATUSTYPE_TABLE:
1420 sz = log->type->status(log, type, result, maxlen);
1421
1422 DMEMIT("%d", ms->nr_mirrors);
1423 for (m = 0; m < ms->nr_mirrors; m++)
1424 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1425 (unsigned long long)ms->mirror[m].offset);
1426
1427 num_feature_args += !!errors_handled(ms);
1428 num_feature_args += !!keep_log(ms);
1429 if (num_feature_args) {
1430 DMEMIT(" %d", num_feature_args);
1431 if (errors_handled(ms))
1432 DMEMIT(" handle_errors");
1433 if (keep_log(ms))
1434 DMEMIT(" keep_log");
1435 }
1436
1437 break;
1438 }
1439}
1440
1441static int mirror_iterate_devices(struct dm_target *ti,
1442 iterate_devices_callout_fn fn, void *data)
1443{
1444 struct mirror_set *ms = ti->private;
1445 int ret = 0;
1446 unsigned i;
1447
1448 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1449 ret = fn(ti, ms->mirror[i].dev,
1450 ms->mirror[i].offset, ti->len, data);
1451
1452 return ret;
1453}
1454
1455static struct target_type mirror_target = {
1456 .name = "mirror",
1457 .version = {1, 14, 0},
1458 .module = THIS_MODULE,
1459 .ctr = mirror_ctr,
1460 .dtr = mirror_dtr,
1461 .map = mirror_map,
1462 .end_io = mirror_end_io,
1463 .presuspend = mirror_presuspend,
1464 .postsuspend = mirror_postsuspend,
1465 .resume = mirror_resume,
1466 .status = mirror_status,
1467 .iterate_devices = mirror_iterate_devices,
1468};
1469
1470static int __init dm_mirror_init(void)
1471{
1472 int r;
1473
1474 r = dm_register_target(&mirror_target);
1475 if (r < 0) {
1476 DMERR("Failed to register mirror target");
1477 goto bad_target;
1478 }
1479
1480 return 0;
1481
1482bad_target:
1483 return r;
1484}
1485
1486static void __exit dm_mirror_exit(void)
1487{
1488 dm_unregister_target(&mirror_target);
1489}
1490
1491/* Module hooks */
1492module_init(dm_mirror_init);
1493module_exit(dm_mirror_exit);
1494
1495MODULE_DESCRIPTION(DM_NAME " mirror target");
1496MODULE_AUTHOR("Joe Thornber");
1497MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2003 Sistina Software Limited.
4 * Copyright (C) 2005-2008 Red Hat, Inc. All rights reserved.
5 *
6 * This file is released under the GPL.
7 */
8
9#include "dm-bio-record.h"
10
11#include <linux/init.h>
12#include <linux/mempool.h>
13#include <linux/module.h>
14#include <linux/pagemap.h>
15#include <linux/slab.h>
16#include <linux/workqueue.h>
17#include <linux/device-mapper.h>
18#include <linux/dm-io.h>
19#include <linux/dm-dirty-log.h>
20#include <linux/dm-kcopyd.h>
21#include <linux/dm-region-hash.h>
22
23static struct workqueue_struct *dm_raid1_wq;
24
25#define DM_MSG_PREFIX "raid1"
26
27#define MAX_RECOVERY 1 /* Maximum number of regions recovered in parallel. */
28
29#define MAX_NR_MIRRORS (DM_KCOPYD_MAX_REGIONS + 1)
30
31#define DM_RAID1_HANDLE_ERRORS 0x01
32#define DM_RAID1_KEEP_LOG 0x02
33#define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
34#define keep_log(p) ((p)->features & DM_RAID1_KEEP_LOG)
35
36static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
37
38/*
39 *---------------------------------------------------------------
40 * Mirror set structures.
41 *---------------------------------------------------------------
42 */
43enum dm_raid1_error {
44 DM_RAID1_WRITE_ERROR,
45 DM_RAID1_FLUSH_ERROR,
46 DM_RAID1_SYNC_ERROR,
47 DM_RAID1_READ_ERROR
48};
49
50struct mirror {
51 struct mirror_set *ms;
52 atomic_t error_count;
53 unsigned long error_type;
54 struct dm_dev *dev;
55 sector_t offset;
56};
57
58struct mirror_set {
59 struct dm_target *ti;
60 struct list_head list;
61
62 uint64_t features;
63
64 spinlock_t lock; /* protects the lists */
65 struct bio_list reads;
66 struct bio_list writes;
67 struct bio_list failures;
68 struct bio_list holds; /* bios are waiting until suspend */
69
70 struct dm_region_hash *rh;
71 struct dm_kcopyd_client *kcopyd_client;
72 struct dm_io_client *io_client;
73
74 /* recovery */
75 region_t nr_regions;
76 int in_sync;
77 int log_failure;
78 int leg_failure;
79 atomic_t suspend;
80
81 atomic_t default_mirror; /* Default mirror */
82
83 struct workqueue_struct *kmirrord_wq;
84 struct work_struct kmirrord_work;
85 struct timer_list timer;
86 unsigned long timer_pending;
87
88 struct work_struct trigger_event;
89
90 unsigned int nr_mirrors;
91 struct mirror mirror[];
92};
93
94DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(raid1_resync_throttle,
95 "A percentage of time allocated for raid resynchronization");
96
97static void wakeup_mirrord(void *context)
98{
99 struct mirror_set *ms = context;
100
101 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
102}
103
104static void delayed_wake_fn(struct timer_list *t)
105{
106 struct mirror_set *ms = from_timer(ms, t, timer);
107
108 clear_bit(0, &ms->timer_pending);
109 wakeup_mirrord(ms);
110}
111
112static void delayed_wake(struct mirror_set *ms)
113{
114 if (test_and_set_bit(0, &ms->timer_pending))
115 return;
116
117 ms->timer.expires = jiffies + HZ / 5;
118 add_timer(&ms->timer);
119}
120
121static void wakeup_all_recovery_waiters(void *context)
122{
123 wake_up_all(&_kmirrord_recovery_stopped);
124}
125
126static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
127{
128 unsigned long flags;
129 int should_wake = 0;
130 struct bio_list *bl;
131
132 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
133 spin_lock_irqsave(&ms->lock, flags);
134 should_wake = !(bl->head);
135 bio_list_add(bl, bio);
136 spin_unlock_irqrestore(&ms->lock, flags);
137
138 if (should_wake)
139 wakeup_mirrord(ms);
140}
141
142static void dispatch_bios(void *context, struct bio_list *bio_list)
143{
144 struct mirror_set *ms = context;
145 struct bio *bio;
146
147 while ((bio = bio_list_pop(bio_list)))
148 queue_bio(ms, bio, WRITE);
149}
150
151struct dm_raid1_bio_record {
152 struct mirror *m;
153 /* if details->bi_bdev == NULL, details were not saved */
154 struct dm_bio_details details;
155 region_t write_region;
156};
157
158/*
159 * Every mirror should look like this one.
160 */
161#define DEFAULT_MIRROR 0
162
163/*
164 * This is yucky. We squirrel the mirror struct away inside
165 * bi_next for read/write buffers. This is safe since the bh
166 * doesn't get submitted to the lower levels of block layer.
167 */
168static struct mirror *bio_get_m(struct bio *bio)
169{
170 return (struct mirror *) bio->bi_next;
171}
172
173static void bio_set_m(struct bio *bio, struct mirror *m)
174{
175 bio->bi_next = (struct bio *) m;
176}
177
178static struct mirror *get_default_mirror(struct mirror_set *ms)
179{
180 return &ms->mirror[atomic_read(&ms->default_mirror)];
181}
182
183static void set_default_mirror(struct mirror *m)
184{
185 struct mirror_set *ms = m->ms;
186 struct mirror *m0 = &(ms->mirror[0]);
187
188 atomic_set(&ms->default_mirror, m - m0);
189}
190
191static struct mirror *get_valid_mirror(struct mirror_set *ms)
192{
193 struct mirror *m;
194
195 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
196 if (!atomic_read(&m->error_count))
197 return m;
198
199 return NULL;
200}
201
202/* fail_mirror
203 * @m: mirror device to fail
204 * @error_type: one of the enum's, DM_RAID1_*_ERROR
205 *
206 * If errors are being handled, record the type of
207 * error encountered for this device. If this type
208 * of error has already been recorded, we can return;
209 * otherwise, we must signal userspace by triggering
210 * an event. Additionally, if the device is the
211 * primary device, we must choose a new primary, but
212 * only if the mirror is in-sync.
213 *
214 * This function must not block.
215 */
216static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
217{
218 struct mirror_set *ms = m->ms;
219 struct mirror *new;
220
221 ms->leg_failure = 1;
222
223 /*
224 * error_count is used for nothing more than a
225 * simple way to tell if a device has encountered
226 * errors.
227 */
228 atomic_inc(&m->error_count);
229
230 if (test_and_set_bit(error_type, &m->error_type))
231 return;
232
233 if (!errors_handled(ms))
234 return;
235
236 if (m != get_default_mirror(ms))
237 goto out;
238
239 if (!ms->in_sync && !keep_log(ms)) {
240 /*
241 * Better to issue requests to same failing device
242 * than to risk returning corrupt data.
243 */
244 DMERR("Primary mirror (%s) failed while out-of-sync: Reads may fail.",
245 m->dev->name);
246 goto out;
247 }
248
249 new = get_valid_mirror(ms);
250 if (new)
251 set_default_mirror(new);
252 else
253 DMWARN("All sides of mirror have failed.");
254
255out:
256 queue_work(dm_raid1_wq, &ms->trigger_event);
257}
258
259static int mirror_flush(struct dm_target *ti)
260{
261 struct mirror_set *ms = ti->private;
262 unsigned long error_bits;
263
264 unsigned int i;
265 struct dm_io_region io[MAX_NR_MIRRORS];
266 struct mirror *m;
267 struct dm_io_request io_req = {
268 .bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
269 .mem.type = DM_IO_KMEM,
270 .mem.ptr.addr = NULL,
271 .client = ms->io_client,
272 };
273
274 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
275 io[i].bdev = m->dev->bdev;
276 io[i].sector = 0;
277 io[i].count = 0;
278 }
279
280 error_bits = -1;
281 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
282 if (unlikely(error_bits != 0)) {
283 for (i = 0; i < ms->nr_mirrors; i++)
284 if (test_bit(i, &error_bits))
285 fail_mirror(ms->mirror + i,
286 DM_RAID1_FLUSH_ERROR);
287 return -EIO;
288 }
289
290 return 0;
291}
292
293/*
294 *---------------------------------------------------------------
295 * Recovery.
296 *
297 * When a mirror is first activated we may find that some regions
298 * are in the no-sync state. We have to recover these by
299 * recopying from the default mirror to all the others.
300 *---------------------------------------------------------------
301 */
302static void recovery_complete(int read_err, unsigned long write_err,
303 void *context)
304{
305 struct dm_region *reg = context;
306 struct mirror_set *ms = dm_rh_region_context(reg);
307 int m, bit = 0;
308
309 if (read_err) {
310 /* Read error means the failure of default mirror. */
311 DMERR_LIMIT("Unable to read primary mirror during recovery");
312 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
313 }
314
315 if (write_err) {
316 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
317 write_err);
318 /*
319 * Bits correspond to devices (excluding default mirror).
320 * The default mirror cannot change during recovery.
321 */
322 for (m = 0; m < ms->nr_mirrors; m++) {
323 if (&ms->mirror[m] == get_default_mirror(ms))
324 continue;
325 if (test_bit(bit, &write_err))
326 fail_mirror(ms->mirror + m,
327 DM_RAID1_SYNC_ERROR);
328 bit++;
329 }
330 }
331
332 dm_rh_recovery_end(reg, !(read_err || write_err));
333}
334
335static void recover(struct mirror_set *ms, struct dm_region *reg)
336{
337 unsigned int i;
338 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
339 struct mirror *m;
340 unsigned long flags = 0;
341 region_t key = dm_rh_get_region_key(reg);
342 sector_t region_size = dm_rh_get_region_size(ms->rh);
343
344 /* fill in the source */
345 m = get_default_mirror(ms);
346 from.bdev = m->dev->bdev;
347 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
348 if (key == (ms->nr_regions - 1)) {
349 /*
350 * The final region may be smaller than
351 * region_size.
352 */
353 from.count = ms->ti->len & (region_size - 1);
354 if (!from.count)
355 from.count = region_size;
356 } else
357 from.count = region_size;
358
359 /* fill in the destinations */
360 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
361 if (&ms->mirror[i] == get_default_mirror(ms))
362 continue;
363
364 m = ms->mirror + i;
365 dest->bdev = m->dev->bdev;
366 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
367 dest->count = from.count;
368 dest++;
369 }
370
371 /* hand to kcopyd */
372 if (!errors_handled(ms))
373 flags |= BIT(DM_KCOPYD_IGNORE_ERROR);
374
375 dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
376 flags, recovery_complete, reg);
377}
378
379static void reset_ms_flags(struct mirror_set *ms)
380{
381 unsigned int m;
382
383 ms->leg_failure = 0;
384 for (m = 0; m < ms->nr_mirrors; m++) {
385 atomic_set(&(ms->mirror[m].error_count), 0);
386 ms->mirror[m].error_type = 0;
387 }
388}
389
390static void do_recovery(struct mirror_set *ms)
391{
392 struct dm_region *reg;
393 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
394
395 /*
396 * Start quiescing some regions.
397 */
398 dm_rh_recovery_prepare(ms->rh);
399
400 /*
401 * Copy any already quiesced regions.
402 */
403 while ((reg = dm_rh_recovery_start(ms->rh)))
404 recover(ms, reg);
405
406 /*
407 * Update the in sync flag.
408 */
409 if (!ms->in_sync &&
410 (log->type->get_sync_count(log) == ms->nr_regions)) {
411 /* the sync is complete */
412 dm_table_event(ms->ti->table);
413 ms->in_sync = 1;
414 reset_ms_flags(ms);
415 }
416}
417
418/*
419 *---------------------------------------------------------------
420 * Reads
421 *---------------------------------------------------------------
422 */
423static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
424{
425 struct mirror *m = get_default_mirror(ms);
426
427 do {
428 if (likely(!atomic_read(&m->error_count)))
429 return m;
430
431 if (m-- == ms->mirror)
432 m += ms->nr_mirrors;
433 } while (m != get_default_mirror(ms));
434
435 return NULL;
436}
437
438static int default_ok(struct mirror *m)
439{
440 struct mirror *default_mirror = get_default_mirror(m->ms);
441
442 return !atomic_read(&default_mirror->error_count);
443}
444
445static int mirror_available(struct mirror_set *ms, struct bio *bio)
446{
447 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
448 region_t region = dm_rh_bio_to_region(ms->rh, bio);
449
450 if (log->type->in_sync(log, region, 0))
451 return choose_mirror(ms, bio->bi_iter.bi_sector) ? 1 : 0;
452
453 return 0;
454}
455
456/*
457 * remap a buffer to a particular mirror.
458 */
459static sector_t map_sector(struct mirror *m, struct bio *bio)
460{
461 if (unlikely(!bio->bi_iter.bi_size))
462 return 0;
463 return m->offset + dm_target_offset(m->ms->ti, bio->bi_iter.bi_sector);
464}
465
466static void map_bio(struct mirror *m, struct bio *bio)
467{
468 bio_set_dev(bio, m->dev->bdev);
469 bio->bi_iter.bi_sector = map_sector(m, bio);
470}
471
472static void map_region(struct dm_io_region *io, struct mirror *m,
473 struct bio *bio)
474{
475 io->bdev = m->dev->bdev;
476 io->sector = map_sector(m, bio);
477 io->count = bio_sectors(bio);
478}
479
480static void hold_bio(struct mirror_set *ms, struct bio *bio)
481{
482 /*
483 * Lock is required to avoid race condition during suspend
484 * process.
485 */
486 spin_lock_irq(&ms->lock);
487
488 if (atomic_read(&ms->suspend)) {
489 spin_unlock_irq(&ms->lock);
490
491 /*
492 * If device is suspended, complete the bio.
493 */
494 if (dm_noflush_suspending(ms->ti))
495 bio->bi_status = BLK_STS_DM_REQUEUE;
496 else
497 bio->bi_status = BLK_STS_IOERR;
498
499 bio_endio(bio);
500 return;
501 }
502
503 /*
504 * Hold bio until the suspend is complete.
505 */
506 bio_list_add(&ms->holds, bio);
507 spin_unlock_irq(&ms->lock);
508}
509
510/*
511 *---------------------------------------------------------------
512 * Reads
513 *---------------------------------------------------------------
514 */
515static void read_callback(unsigned long error, void *context)
516{
517 struct bio *bio = context;
518 struct mirror *m;
519
520 m = bio_get_m(bio);
521 bio_set_m(bio, NULL);
522
523 if (likely(!error)) {
524 bio_endio(bio);
525 return;
526 }
527
528 fail_mirror(m, DM_RAID1_READ_ERROR);
529
530 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
531 DMWARN_LIMIT("Read failure on mirror device %s. Trying alternative device.",
532 m->dev->name);
533 queue_bio(m->ms, bio, bio_data_dir(bio));
534 return;
535 }
536
537 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
538 m->dev->name);
539 bio_io_error(bio);
540}
541
542/* Asynchronous read. */
543static void read_async_bio(struct mirror *m, struct bio *bio)
544{
545 struct dm_io_region io;
546 struct dm_io_request io_req = {
547 .bi_opf = REQ_OP_READ,
548 .mem.type = DM_IO_BIO,
549 .mem.ptr.bio = bio,
550 .notify.fn = read_callback,
551 .notify.context = bio,
552 .client = m->ms->io_client,
553 };
554
555 map_region(&io, m, bio);
556 bio_set_m(bio, m);
557 BUG_ON(dm_io(&io_req, 1, &io, NULL));
558}
559
560static inline int region_in_sync(struct mirror_set *ms, region_t region,
561 int may_block)
562{
563 int state = dm_rh_get_state(ms->rh, region, may_block);
564 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
565}
566
567static void do_reads(struct mirror_set *ms, struct bio_list *reads)
568{
569 region_t region;
570 struct bio *bio;
571 struct mirror *m;
572
573 while ((bio = bio_list_pop(reads))) {
574 region = dm_rh_bio_to_region(ms->rh, bio);
575 m = get_default_mirror(ms);
576
577 /*
578 * We can only read balance if the region is in sync.
579 */
580 if (likely(region_in_sync(ms, region, 1)))
581 m = choose_mirror(ms, bio->bi_iter.bi_sector);
582 else if (m && atomic_read(&m->error_count))
583 m = NULL;
584
585 if (likely(m))
586 read_async_bio(m, bio);
587 else
588 bio_io_error(bio);
589 }
590}
591
592/*
593 *---------------------------------------------------------------------
594 * Writes.
595 *
596 * We do different things with the write io depending on the
597 * state of the region that it's in:
598 *
599 * SYNC: increment pending, use kcopyd to write to *all* mirrors
600 * RECOVERING: delay the io until recovery completes
601 * NOSYNC: increment pending, just write to the default mirror
602 *---------------------------------------------------------------------
603 */
604static void write_callback(unsigned long error, void *context)
605{
606 unsigned int i;
607 struct bio *bio = context;
608 struct mirror_set *ms;
609 int should_wake = 0;
610 unsigned long flags;
611
612 ms = bio_get_m(bio)->ms;
613 bio_set_m(bio, NULL);
614
615 /*
616 * NOTE: We don't decrement the pending count here,
617 * instead it is done by the targets endio function.
618 * This way we handle both writes to SYNC and NOSYNC
619 * regions with the same code.
620 */
621 if (likely(!error)) {
622 bio_endio(bio);
623 return;
624 }
625
626 /*
627 * If the bio is discard, return an error, but do not
628 * degrade the array.
629 */
630 if (bio_op(bio) == REQ_OP_DISCARD) {
631 bio->bi_status = BLK_STS_NOTSUPP;
632 bio_endio(bio);
633 return;
634 }
635
636 for (i = 0; i < ms->nr_mirrors; i++)
637 if (test_bit(i, &error))
638 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
639
640 /*
641 * Need to raise event. Since raising
642 * events can block, we need to do it in
643 * the main thread.
644 */
645 spin_lock_irqsave(&ms->lock, flags);
646 if (!ms->failures.head)
647 should_wake = 1;
648 bio_list_add(&ms->failures, bio);
649 spin_unlock_irqrestore(&ms->lock, flags);
650 if (should_wake)
651 wakeup_mirrord(ms);
652}
653
654static void do_write(struct mirror_set *ms, struct bio *bio)
655{
656 unsigned int i;
657 struct dm_io_region io[MAX_NR_MIRRORS], *dest = io;
658 struct mirror *m;
659 blk_opf_t op_flags = bio->bi_opf & (REQ_FUA | REQ_PREFLUSH);
660 struct dm_io_request io_req = {
661 .bi_opf = REQ_OP_WRITE | op_flags,
662 .mem.type = DM_IO_BIO,
663 .mem.ptr.bio = bio,
664 .notify.fn = write_callback,
665 .notify.context = bio,
666 .client = ms->io_client,
667 };
668
669 if (bio_op(bio) == REQ_OP_DISCARD) {
670 io_req.bi_opf = REQ_OP_DISCARD | op_flags;
671 io_req.mem.type = DM_IO_KMEM;
672 io_req.mem.ptr.addr = NULL;
673 }
674
675 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
676 map_region(dest++, m, bio);
677
678 /*
679 * Use default mirror because we only need it to retrieve the reference
680 * to the mirror set in write_callback().
681 */
682 bio_set_m(bio, get_default_mirror(ms));
683
684 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
685}
686
687static void do_writes(struct mirror_set *ms, struct bio_list *writes)
688{
689 int state;
690 struct bio *bio;
691 struct bio_list sync, nosync, recover, *this_list = NULL;
692 struct bio_list requeue;
693 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
694 region_t region;
695
696 if (!writes->head)
697 return;
698
699 /*
700 * Classify each write.
701 */
702 bio_list_init(&sync);
703 bio_list_init(&nosync);
704 bio_list_init(&recover);
705 bio_list_init(&requeue);
706
707 while ((bio = bio_list_pop(writes))) {
708 if ((bio->bi_opf & REQ_PREFLUSH) ||
709 (bio_op(bio) == REQ_OP_DISCARD)) {
710 bio_list_add(&sync, bio);
711 continue;
712 }
713
714 region = dm_rh_bio_to_region(ms->rh, bio);
715
716 if (log->type->is_remote_recovering &&
717 log->type->is_remote_recovering(log, region)) {
718 bio_list_add(&requeue, bio);
719 continue;
720 }
721
722 state = dm_rh_get_state(ms->rh, region, 1);
723 switch (state) {
724 case DM_RH_CLEAN:
725 case DM_RH_DIRTY:
726 this_list = &sync;
727 break;
728
729 case DM_RH_NOSYNC:
730 this_list = &nosync;
731 break;
732
733 case DM_RH_RECOVERING:
734 this_list = &recover;
735 break;
736 }
737
738 bio_list_add(this_list, bio);
739 }
740
741 /*
742 * Add bios that are delayed due to remote recovery
743 * back on to the write queue
744 */
745 if (unlikely(requeue.head)) {
746 spin_lock_irq(&ms->lock);
747 bio_list_merge(&ms->writes, &requeue);
748 spin_unlock_irq(&ms->lock);
749 delayed_wake(ms);
750 }
751
752 /*
753 * Increment the pending counts for any regions that will
754 * be written to (writes to recover regions are going to
755 * be delayed).
756 */
757 dm_rh_inc_pending(ms->rh, &sync);
758 dm_rh_inc_pending(ms->rh, &nosync);
759
760 /*
761 * If the flush fails on a previous call and succeeds here,
762 * we must not reset the log_failure variable. We need
763 * userspace interaction to do that.
764 */
765 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
766
767 /*
768 * Dispatch io.
769 */
770 if (unlikely(ms->log_failure) && errors_handled(ms)) {
771 spin_lock_irq(&ms->lock);
772 bio_list_merge(&ms->failures, &sync);
773 spin_unlock_irq(&ms->lock);
774 wakeup_mirrord(ms);
775 } else
776 while ((bio = bio_list_pop(&sync)))
777 do_write(ms, bio);
778
779 while ((bio = bio_list_pop(&recover)))
780 dm_rh_delay(ms->rh, bio);
781
782 while ((bio = bio_list_pop(&nosync))) {
783 if (unlikely(ms->leg_failure) && errors_handled(ms) && !keep_log(ms)) {
784 spin_lock_irq(&ms->lock);
785 bio_list_add(&ms->failures, bio);
786 spin_unlock_irq(&ms->lock);
787 wakeup_mirrord(ms);
788 } else {
789 map_bio(get_default_mirror(ms), bio);
790 submit_bio_noacct(bio);
791 }
792 }
793}
794
795static void do_failures(struct mirror_set *ms, struct bio_list *failures)
796{
797 struct bio *bio;
798
799 if (likely(!failures->head))
800 return;
801
802 /*
803 * If the log has failed, unattempted writes are being
804 * put on the holds list. We can't issue those writes
805 * until a log has been marked, so we must store them.
806 *
807 * If a 'noflush' suspend is in progress, we can requeue
808 * the I/O's to the core. This give userspace a chance
809 * to reconfigure the mirror, at which point the core
810 * will reissue the writes. If the 'noflush' flag is
811 * not set, we have no choice but to return errors.
812 *
813 * Some writes on the failures list may have been
814 * submitted before the log failure and represent a
815 * failure to write to one of the devices. It is ok
816 * for us to treat them the same and requeue them
817 * as well.
818 */
819 while ((bio = bio_list_pop(failures))) {
820 if (!ms->log_failure) {
821 ms->in_sync = 0;
822 dm_rh_mark_nosync(ms->rh, bio);
823 }
824
825 /*
826 * If all the legs are dead, fail the I/O.
827 * If the device has failed and keep_log is enabled,
828 * fail the I/O.
829 *
830 * If we have been told to handle errors, and keep_log
831 * isn't enabled, hold the bio and wait for userspace to
832 * deal with the problem.
833 *
834 * Otherwise pretend that the I/O succeeded. (This would
835 * be wrong if the failed leg returned after reboot and
836 * got replicated back to the good legs.)
837 */
838 if (unlikely(!get_valid_mirror(ms) || (keep_log(ms) && ms->log_failure)))
839 bio_io_error(bio);
840 else if (errors_handled(ms) && !keep_log(ms))
841 hold_bio(ms, bio);
842 else
843 bio_endio(bio);
844 }
845}
846
847static void trigger_event(struct work_struct *work)
848{
849 struct mirror_set *ms =
850 container_of(work, struct mirror_set, trigger_event);
851
852 dm_table_event(ms->ti->table);
853}
854
855/*
856 *---------------------------------------------------------------
857 * kmirrord
858 *---------------------------------------------------------------
859 */
860static void do_mirror(struct work_struct *work)
861{
862 struct mirror_set *ms = container_of(work, struct mirror_set,
863 kmirrord_work);
864 struct bio_list reads, writes, failures;
865 unsigned long flags;
866
867 spin_lock_irqsave(&ms->lock, flags);
868 reads = ms->reads;
869 writes = ms->writes;
870 failures = ms->failures;
871 bio_list_init(&ms->reads);
872 bio_list_init(&ms->writes);
873 bio_list_init(&ms->failures);
874 spin_unlock_irqrestore(&ms->lock, flags);
875
876 dm_rh_update_states(ms->rh, errors_handled(ms));
877 do_recovery(ms);
878 do_reads(ms, &reads);
879 do_writes(ms, &writes);
880 do_failures(ms, &failures);
881}
882
883/*
884 *---------------------------------------------------------------
885 * Target functions
886 *---------------------------------------------------------------
887 */
888static struct mirror_set *alloc_context(unsigned int nr_mirrors,
889 uint32_t region_size,
890 struct dm_target *ti,
891 struct dm_dirty_log *dl)
892{
893 struct mirror_set *ms =
894 kzalloc(struct_size(ms, mirror, nr_mirrors), GFP_KERNEL);
895
896 if (!ms) {
897 ti->error = "Cannot allocate mirror context";
898 return NULL;
899 }
900
901 spin_lock_init(&ms->lock);
902 bio_list_init(&ms->reads);
903 bio_list_init(&ms->writes);
904 bio_list_init(&ms->failures);
905 bio_list_init(&ms->holds);
906
907 ms->ti = ti;
908 ms->nr_mirrors = nr_mirrors;
909 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
910 ms->in_sync = 0;
911 ms->log_failure = 0;
912 ms->leg_failure = 0;
913 atomic_set(&ms->suspend, 0);
914 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
915
916 ms->io_client = dm_io_client_create();
917 if (IS_ERR(ms->io_client)) {
918 ti->error = "Error creating dm_io client";
919 kfree(ms);
920 return NULL;
921 }
922
923 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
924 wakeup_all_recovery_waiters,
925 ms->ti->begin, MAX_RECOVERY,
926 dl, region_size, ms->nr_regions);
927 if (IS_ERR(ms->rh)) {
928 ti->error = "Error creating dirty region hash";
929 dm_io_client_destroy(ms->io_client);
930 kfree(ms);
931 return NULL;
932 }
933
934 return ms;
935}
936
937static void free_context(struct mirror_set *ms, struct dm_target *ti,
938 unsigned int m)
939{
940 while (m--)
941 dm_put_device(ti, ms->mirror[m].dev);
942
943 dm_io_client_destroy(ms->io_client);
944 dm_region_hash_destroy(ms->rh);
945 kfree(ms);
946}
947
948static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
949 unsigned int mirror, char **argv)
950{
951 unsigned long long offset;
952 char dummy;
953 int ret;
954
955 if (sscanf(argv[1], "%llu%c", &offset, &dummy) != 1 ||
956 offset != (sector_t)offset) {
957 ti->error = "Invalid offset";
958 return -EINVAL;
959 }
960
961 ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
962 &ms->mirror[mirror].dev);
963 if (ret) {
964 ti->error = "Device lookup failure";
965 return ret;
966 }
967
968 ms->mirror[mirror].ms = ms;
969 atomic_set(&(ms->mirror[mirror].error_count), 0);
970 ms->mirror[mirror].error_type = 0;
971 ms->mirror[mirror].offset = offset;
972
973 return 0;
974}
975
976/*
977 * Create dirty log: log_type #log_params <log_params>
978 */
979static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
980 unsigned int argc, char **argv,
981 unsigned int *args_used)
982{
983 unsigned int param_count;
984 struct dm_dirty_log *dl;
985 char dummy;
986
987 if (argc < 2) {
988 ti->error = "Insufficient mirror log arguments";
989 return NULL;
990 }
991
992 if (sscanf(argv[1], "%u%c", ¶m_count, &dummy) != 1) {
993 ti->error = "Invalid mirror log argument count";
994 return NULL;
995 }
996
997 *args_used = 2 + param_count;
998
999 if (argc < *args_used) {
1000 ti->error = "Insufficient mirror log arguments";
1001 return NULL;
1002 }
1003
1004 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
1005 argv + 2);
1006 if (!dl) {
1007 ti->error = "Error creating mirror dirty log";
1008 return NULL;
1009 }
1010
1011 return dl;
1012}
1013
1014static int parse_features(struct mirror_set *ms, unsigned int argc, char **argv,
1015 unsigned int *args_used)
1016{
1017 unsigned int num_features;
1018 struct dm_target *ti = ms->ti;
1019 char dummy;
1020 int i;
1021
1022 *args_used = 0;
1023
1024 if (!argc)
1025 return 0;
1026
1027 if (sscanf(argv[0], "%u%c", &num_features, &dummy) != 1) {
1028 ti->error = "Invalid number of features";
1029 return -EINVAL;
1030 }
1031
1032 argc--;
1033 argv++;
1034 (*args_used)++;
1035
1036 if (num_features > argc) {
1037 ti->error = "Not enough arguments to support feature count";
1038 return -EINVAL;
1039 }
1040
1041 for (i = 0; i < num_features; i++) {
1042 if (!strcmp("handle_errors", argv[0]))
1043 ms->features |= DM_RAID1_HANDLE_ERRORS;
1044 else if (!strcmp("keep_log", argv[0]))
1045 ms->features |= DM_RAID1_KEEP_LOG;
1046 else {
1047 ti->error = "Unrecognised feature requested";
1048 return -EINVAL;
1049 }
1050
1051 argc--;
1052 argv++;
1053 (*args_used)++;
1054 }
1055 if (!errors_handled(ms) && keep_log(ms)) {
1056 ti->error = "keep_log feature requires the handle_errors feature";
1057 return -EINVAL;
1058 }
1059
1060 return 0;
1061}
1062
1063/*
1064 * Construct a mirror mapping:
1065 *
1066 * log_type #log_params <log_params>
1067 * #mirrors [mirror_path offset]{2,}
1068 * [#features <features>]
1069 *
1070 * log_type is "core" or "disk"
1071 * #log_params is between 1 and 3
1072 *
1073 * If present, supported features are "handle_errors" and "keep_log".
1074 */
1075static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1076{
1077 int r;
1078 unsigned int nr_mirrors, m, args_used;
1079 struct mirror_set *ms;
1080 struct dm_dirty_log *dl;
1081 char dummy;
1082
1083 dl = create_dirty_log(ti, argc, argv, &args_used);
1084 if (!dl)
1085 return -EINVAL;
1086
1087 argv += args_used;
1088 argc -= args_used;
1089
1090 if (!argc || sscanf(argv[0], "%u%c", &nr_mirrors, &dummy) != 1 ||
1091 nr_mirrors < 2 || nr_mirrors > MAX_NR_MIRRORS) {
1092 ti->error = "Invalid number of mirrors";
1093 dm_dirty_log_destroy(dl);
1094 return -EINVAL;
1095 }
1096
1097 argv++, argc--;
1098
1099 if (argc < nr_mirrors * 2) {
1100 ti->error = "Too few mirror arguments";
1101 dm_dirty_log_destroy(dl);
1102 return -EINVAL;
1103 }
1104
1105 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1106 if (!ms) {
1107 dm_dirty_log_destroy(dl);
1108 return -ENOMEM;
1109 }
1110
1111 /* Get the mirror parameter sets */
1112 for (m = 0; m < nr_mirrors; m++) {
1113 r = get_mirror(ms, ti, m, argv);
1114 if (r) {
1115 free_context(ms, ti, m);
1116 return r;
1117 }
1118 argv += 2;
1119 argc -= 2;
1120 }
1121
1122 ti->private = ms;
1123
1124 r = dm_set_target_max_io_len(ti, dm_rh_get_region_size(ms->rh));
1125 if (r)
1126 goto err_free_context;
1127
1128 ti->num_flush_bios = 1;
1129 ti->num_discard_bios = 1;
1130 ti->per_io_data_size = sizeof(struct dm_raid1_bio_record);
1131
1132 ms->kmirrord_wq = alloc_workqueue("kmirrord", WQ_MEM_RECLAIM, 0);
1133 if (!ms->kmirrord_wq) {
1134 DMERR("couldn't start kmirrord");
1135 r = -ENOMEM;
1136 goto err_free_context;
1137 }
1138 INIT_WORK(&ms->kmirrord_work, do_mirror);
1139 timer_setup(&ms->timer, delayed_wake_fn, 0);
1140 ms->timer_pending = 0;
1141 INIT_WORK(&ms->trigger_event, trigger_event);
1142
1143 r = parse_features(ms, argc, argv, &args_used);
1144 if (r)
1145 goto err_destroy_wq;
1146
1147 argv += args_used;
1148 argc -= args_used;
1149
1150 /*
1151 * Any read-balancing addition depends on the
1152 * DM_RAID1_HANDLE_ERRORS flag being present.
1153 * This is because the decision to balance depends
1154 * on the sync state of a region. If the above
1155 * flag is not present, we ignore errors; and
1156 * the sync state may be inaccurate.
1157 */
1158
1159 if (argc) {
1160 ti->error = "Too many mirror arguments";
1161 r = -EINVAL;
1162 goto err_destroy_wq;
1163 }
1164
1165 ms->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1166 if (IS_ERR(ms->kcopyd_client)) {
1167 r = PTR_ERR(ms->kcopyd_client);
1168 goto err_destroy_wq;
1169 }
1170
1171 wakeup_mirrord(ms);
1172 return 0;
1173
1174err_destroy_wq:
1175 destroy_workqueue(ms->kmirrord_wq);
1176err_free_context:
1177 free_context(ms, ti, ms->nr_mirrors);
1178 return r;
1179}
1180
1181static void mirror_dtr(struct dm_target *ti)
1182{
1183 struct mirror_set *ms = ti->private;
1184
1185 del_timer_sync(&ms->timer);
1186 flush_workqueue(ms->kmirrord_wq);
1187 flush_work(&ms->trigger_event);
1188 dm_kcopyd_client_destroy(ms->kcopyd_client);
1189 destroy_workqueue(ms->kmirrord_wq);
1190 free_context(ms, ti, ms->nr_mirrors);
1191}
1192
1193/*
1194 * Mirror mapping function
1195 */
1196static int mirror_map(struct dm_target *ti, struct bio *bio)
1197{
1198 int r, rw = bio_data_dir(bio);
1199 struct mirror *m;
1200 struct mirror_set *ms = ti->private;
1201 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1202 struct dm_raid1_bio_record *bio_record =
1203 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1204
1205 bio_record->details.bi_bdev = NULL;
1206
1207 if (rw == WRITE) {
1208 /* Save region for mirror_end_io() handler */
1209 bio_record->write_region = dm_rh_bio_to_region(ms->rh, bio);
1210 queue_bio(ms, bio, rw);
1211 return DM_MAPIO_SUBMITTED;
1212 }
1213
1214 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1215 if (r < 0 && r != -EWOULDBLOCK)
1216 return DM_MAPIO_KILL;
1217
1218 /*
1219 * If region is not in-sync queue the bio.
1220 */
1221 if (!r || (r == -EWOULDBLOCK)) {
1222 if (bio->bi_opf & REQ_RAHEAD)
1223 return DM_MAPIO_KILL;
1224
1225 queue_bio(ms, bio, rw);
1226 return DM_MAPIO_SUBMITTED;
1227 }
1228
1229 /*
1230 * The region is in-sync and we can perform reads directly.
1231 * Store enough information so we can retry if it fails.
1232 */
1233 m = choose_mirror(ms, bio->bi_iter.bi_sector);
1234 if (unlikely(!m))
1235 return DM_MAPIO_KILL;
1236
1237 dm_bio_record(&bio_record->details, bio);
1238 bio_record->m = m;
1239
1240 map_bio(m, bio);
1241
1242 return DM_MAPIO_REMAPPED;
1243}
1244
1245static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1246 blk_status_t *error)
1247{
1248 int rw = bio_data_dir(bio);
1249 struct mirror_set *ms = ti->private;
1250 struct mirror *m = NULL;
1251 struct dm_bio_details *bd = NULL;
1252 struct dm_raid1_bio_record *bio_record =
1253 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1254
1255 /*
1256 * We need to dec pending if this was a write.
1257 */
1258 if (rw == WRITE) {
1259 if (!(bio->bi_opf & REQ_PREFLUSH) &&
1260 bio_op(bio) != REQ_OP_DISCARD)
1261 dm_rh_dec(ms->rh, bio_record->write_region);
1262 return DM_ENDIO_DONE;
1263 }
1264
1265 if (*error == BLK_STS_NOTSUPP)
1266 goto out;
1267
1268 if (bio->bi_opf & REQ_RAHEAD)
1269 goto out;
1270
1271 if (unlikely(*error)) {
1272 if (!bio_record->details.bi_bdev) {
1273 /*
1274 * There wasn't enough memory to record necessary
1275 * information for a retry or there was no other
1276 * mirror in-sync.
1277 */
1278 DMERR_LIMIT("Mirror read failed.");
1279 return DM_ENDIO_DONE;
1280 }
1281
1282 m = bio_record->m;
1283
1284 DMERR("Mirror read failed from %s. Trying alternative device.",
1285 m->dev->name);
1286
1287 fail_mirror(m, DM_RAID1_READ_ERROR);
1288
1289 /*
1290 * A failed read is requeued for another attempt using an intact
1291 * mirror.
1292 */
1293 if (default_ok(m) || mirror_available(ms, bio)) {
1294 bd = &bio_record->details;
1295
1296 dm_bio_restore(bd, bio);
1297 bio_record->details.bi_bdev = NULL;
1298 bio->bi_status = 0;
1299
1300 queue_bio(ms, bio, rw);
1301 return DM_ENDIO_INCOMPLETE;
1302 }
1303 DMERR("All replicated volumes dead, failing I/O");
1304 }
1305
1306out:
1307 bio_record->details.bi_bdev = NULL;
1308
1309 return DM_ENDIO_DONE;
1310}
1311
1312static void mirror_presuspend(struct dm_target *ti)
1313{
1314 struct mirror_set *ms = ti->private;
1315 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1316
1317 struct bio_list holds;
1318 struct bio *bio;
1319
1320 atomic_set(&ms->suspend, 1);
1321
1322 /*
1323 * Process bios in the hold list to start recovery waiting
1324 * for bios in the hold list. After the process, no bio has
1325 * a chance to be added in the hold list because ms->suspend
1326 * is set.
1327 */
1328 spin_lock_irq(&ms->lock);
1329 holds = ms->holds;
1330 bio_list_init(&ms->holds);
1331 spin_unlock_irq(&ms->lock);
1332
1333 while ((bio = bio_list_pop(&holds)))
1334 hold_bio(ms, bio);
1335
1336 /*
1337 * We must finish up all the work that we've
1338 * generated (i.e. recovery work).
1339 */
1340 dm_rh_stop_recovery(ms->rh);
1341
1342 wait_event(_kmirrord_recovery_stopped,
1343 !dm_rh_recovery_in_flight(ms->rh));
1344
1345 if (log->type->presuspend && log->type->presuspend(log))
1346 /* FIXME: need better error handling */
1347 DMWARN("log presuspend failed");
1348
1349 /*
1350 * Now that recovery is complete/stopped and the
1351 * delayed bios are queued, we need to wait for
1352 * the worker thread to complete. This way,
1353 * we know that all of our I/O has been pushed.
1354 */
1355 flush_workqueue(ms->kmirrord_wq);
1356}
1357
1358static void mirror_postsuspend(struct dm_target *ti)
1359{
1360 struct mirror_set *ms = ti->private;
1361 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1362
1363 if (log->type->postsuspend && log->type->postsuspend(log))
1364 /* FIXME: need better error handling */
1365 DMWARN("log postsuspend failed");
1366}
1367
1368static void mirror_resume(struct dm_target *ti)
1369{
1370 struct mirror_set *ms = ti->private;
1371 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1372
1373 atomic_set(&ms->suspend, 0);
1374 if (log->type->resume && log->type->resume(log))
1375 /* FIXME: need better error handling */
1376 DMWARN("log resume failed");
1377 dm_rh_start_recovery(ms->rh);
1378}
1379
1380/*
1381 * device_status_char
1382 * @m: mirror device/leg we want the status of
1383 *
1384 * We return one character representing the most severe error
1385 * we have encountered.
1386 * A => Alive - No failures
1387 * D => Dead - A write failure occurred leaving mirror out-of-sync
1388 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1389 * R => Read - A read failure occurred, mirror data unaffected
1390 *
1391 * Returns: <char>
1392 */
1393static char device_status_char(struct mirror *m)
1394{
1395 if (!atomic_read(&(m->error_count)))
1396 return 'A';
1397
1398 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1399 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1400 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1401 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1402}
1403
1404
1405static void mirror_status(struct dm_target *ti, status_type_t type,
1406 unsigned int status_flags, char *result, unsigned int maxlen)
1407{
1408 unsigned int m, sz = 0;
1409 int num_feature_args = 0;
1410 struct mirror_set *ms = ti->private;
1411 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1412 char buffer[MAX_NR_MIRRORS + 1];
1413
1414 switch (type) {
1415 case STATUSTYPE_INFO:
1416 DMEMIT("%d ", ms->nr_mirrors);
1417 for (m = 0; m < ms->nr_mirrors; m++) {
1418 DMEMIT("%s ", ms->mirror[m].dev->name);
1419 buffer[m] = device_status_char(&(ms->mirror[m]));
1420 }
1421 buffer[m] = '\0';
1422
1423 DMEMIT("%llu/%llu 1 %s ",
1424 (unsigned long long)log->type->get_sync_count(log),
1425 (unsigned long long)ms->nr_regions, buffer);
1426
1427 sz += log->type->status(log, type, result+sz, maxlen-sz);
1428
1429 break;
1430
1431 case STATUSTYPE_TABLE:
1432 sz = log->type->status(log, type, result, maxlen);
1433
1434 DMEMIT("%d", ms->nr_mirrors);
1435 for (m = 0; m < ms->nr_mirrors; m++)
1436 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1437 (unsigned long long)ms->mirror[m].offset);
1438
1439 num_feature_args += !!errors_handled(ms);
1440 num_feature_args += !!keep_log(ms);
1441 if (num_feature_args) {
1442 DMEMIT(" %d", num_feature_args);
1443 if (errors_handled(ms))
1444 DMEMIT(" handle_errors");
1445 if (keep_log(ms))
1446 DMEMIT(" keep_log");
1447 }
1448
1449 break;
1450
1451 case STATUSTYPE_IMA:
1452 DMEMIT_TARGET_NAME_VERSION(ti->type);
1453 DMEMIT(",nr_mirrors=%d", ms->nr_mirrors);
1454 for (m = 0; m < ms->nr_mirrors; m++) {
1455 DMEMIT(",mirror_device_%d=%s", m, ms->mirror[m].dev->name);
1456 DMEMIT(",mirror_device_%d_status=%c",
1457 m, device_status_char(&(ms->mirror[m])));
1458 }
1459
1460 DMEMIT(",handle_errors=%c", errors_handled(ms) ? 'y' : 'n');
1461 DMEMIT(",keep_log=%c", keep_log(ms) ? 'y' : 'n');
1462
1463 DMEMIT(",log_type_status=");
1464 sz += log->type->status(log, type, result+sz, maxlen-sz);
1465 DMEMIT(";");
1466 break;
1467 }
1468}
1469
1470static int mirror_iterate_devices(struct dm_target *ti,
1471 iterate_devices_callout_fn fn, void *data)
1472{
1473 struct mirror_set *ms = ti->private;
1474 int ret = 0;
1475 unsigned int i;
1476
1477 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1478 ret = fn(ti, ms->mirror[i].dev,
1479 ms->mirror[i].offset, ti->len, data);
1480
1481 return ret;
1482}
1483
1484static struct target_type mirror_target = {
1485 .name = "mirror",
1486 .version = {1, 14, 0},
1487 .module = THIS_MODULE,
1488 .ctr = mirror_ctr,
1489 .dtr = mirror_dtr,
1490 .map = mirror_map,
1491 .end_io = mirror_end_io,
1492 .presuspend = mirror_presuspend,
1493 .postsuspend = mirror_postsuspend,
1494 .resume = mirror_resume,
1495 .status = mirror_status,
1496 .iterate_devices = mirror_iterate_devices,
1497};
1498
1499static int __init dm_mirror_init(void)
1500{
1501 int r;
1502
1503 dm_raid1_wq = alloc_workqueue("dm_raid1_wq", 0, 0);
1504 if (!dm_raid1_wq) {
1505 DMERR("Failed to alloc workqueue");
1506 return -ENOMEM;
1507 }
1508
1509 r = dm_register_target(&mirror_target);
1510 if (r < 0) {
1511 destroy_workqueue(dm_raid1_wq);
1512 return r;
1513 }
1514
1515 return 0;
1516}
1517
1518static void __exit dm_mirror_exit(void)
1519{
1520 destroy_workqueue(dm_raid1_wq);
1521 dm_unregister_target(&mirror_target);
1522}
1523
1524/* Module hooks */
1525module_init(dm_mirror_init);
1526module_exit(dm_mirror_exit);
1527
1528MODULE_DESCRIPTION(DM_NAME " mirror target");
1529MODULE_AUTHOR("Joe Thornber");
1530MODULE_LICENSE("GPL");