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