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 DM_RAID1_HANDLE_ERRORS 0x01
27#define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
28
29static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
30
31/*-----------------------------------------------------------------
32 * Mirror set structures.
33 *---------------------------------------------------------------*/
34enum dm_raid1_error {
35 DM_RAID1_WRITE_ERROR,
36 DM_RAID1_FLUSH_ERROR,
37 DM_RAID1_SYNC_ERROR,
38 DM_RAID1_READ_ERROR
39};
40
41struct mirror {
42 struct mirror_set *ms;
43 atomic_t error_count;
44 unsigned long error_type;
45 struct dm_dev *dev;
46 sector_t offset;
47};
48
49struct mirror_set {
50 struct dm_target *ti;
51 struct list_head list;
52
53 uint64_t features;
54
55 spinlock_t lock; /* protects the lists */
56 struct bio_list reads;
57 struct bio_list writes;
58 struct bio_list failures;
59 struct bio_list holds; /* bios are waiting until suspend */
60
61 struct dm_region_hash *rh;
62 struct dm_kcopyd_client *kcopyd_client;
63 struct dm_io_client *io_client;
64 mempool_t *read_record_pool;
65
66 /* recovery */
67 region_t nr_regions;
68 int in_sync;
69 int log_failure;
70 int leg_failure;
71 atomic_t suspend;
72
73 atomic_t default_mirror; /* Default mirror */
74
75 struct workqueue_struct *kmirrord_wq;
76 struct work_struct kmirrord_work;
77 struct timer_list timer;
78 unsigned long timer_pending;
79
80 struct work_struct trigger_event;
81
82 unsigned nr_mirrors;
83 struct mirror mirror[0];
84};
85
86static void wakeup_mirrord(void *context)
87{
88 struct mirror_set *ms = context;
89
90 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
91}
92
93static void delayed_wake_fn(unsigned long data)
94{
95 struct mirror_set *ms = (struct mirror_set *) data;
96
97 clear_bit(0, &ms->timer_pending);
98 wakeup_mirrord(ms);
99}
100
101static void delayed_wake(struct mirror_set *ms)
102{
103 if (test_and_set_bit(0, &ms->timer_pending))
104 return;
105
106 ms->timer.expires = jiffies + HZ / 5;
107 ms->timer.data = (unsigned long) ms;
108 ms->timer.function = delayed_wake_fn;
109 add_timer(&ms->timer);
110}
111
112static void wakeup_all_recovery_waiters(void *context)
113{
114 wake_up_all(&_kmirrord_recovery_stopped);
115}
116
117static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
118{
119 unsigned long flags;
120 int should_wake = 0;
121 struct bio_list *bl;
122
123 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
124 spin_lock_irqsave(&ms->lock, flags);
125 should_wake = !(bl->head);
126 bio_list_add(bl, bio);
127 spin_unlock_irqrestore(&ms->lock, flags);
128
129 if (should_wake)
130 wakeup_mirrord(ms);
131}
132
133static void dispatch_bios(void *context, struct bio_list *bio_list)
134{
135 struct mirror_set *ms = context;
136 struct bio *bio;
137
138 while ((bio = bio_list_pop(bio_list)))
139 queue_bio(ms, bio, WRITE);
140}
141
142#define MIN_READ_RECORDS 20
143struct dm_raid1_read_record {
144 struct mirror *m;
145 struct dm_bio_details details;
146};
147
148static struct kmem_cache *_dm_raid1_read_record_cache;
149
150/*
151 * Every mirror should look like this one.
152 */
153#define DEFAULT_MIRROR 0
154
155/*
156 * This is yucky. We squirrel the mirror struct away inside
157 * bi_next for read/write buffers. This is safe since the bh
158 * doesn't get submitted to the lower levels of block layer.
159 */
160static struct mirror *bio_get_m(struct bio *bio)
161{
162 return (struct mirror *) bio->bi_next;
163}
164
165static void bio_set_m(struct bio *bio, struct mirror *m)
166{
167 bio->bi_next = (struct bio *) m;
168}
169
170static struct mirror *get_default_mirror(struct mirror_set *ms)
171{
172 return &ms->mirror[atomic_read(&ms->default_mirror)];
173}
174
175static void set_default_mirror(struct mirror *m)
176{
177 struct mirror_set *ms = m->ms;
178 struct mirror *m0 = &(ms->mirror[0]);
179
180 atomic_set(&ms->default_mirror, m - m0);
181}
182
183static struct mirror *get_valid_mirror(struct mirror_set *ms)
184{
185 struct mirror *m;
186
187 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
188 if (!atomic_read(&m->error_count))
189 return m;
190
191 return NULL;
192}
193
194/* fail_mirror
195 * @m: mirror device to fail
196 * @error_type: one of the enum's, DM_RAID1_*_ERROR
197 *
198 * If errors are being handled, record the type of
199 * error encountered for this device. If this type
200 * of error has already been recorded, we can return;
201 * otherwise, we must signal userspace by triggering
202 * an event. Additionally, if the device is the
203 * primary device, we must choose a new primary, but
204 * only if the mirror is in-sync.
205 *
206 * This function must not block.
207 */
208static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
209{
210 struct mirror_set *ms = m->ms;
211 struct mirror *new;
212
213 ms->leg_failure = 1;
214
215 /*
216 * error_count is used for nothing more than a
217 * simple way to tell if a device has encountered
218 * errors.
219 */
220 atomic_inc(&m->error_count);
221
222 if (test_and_set_bit(error_type, &m->error_type))
223 return;
224
225 if (!errors_handled(ms))
226 return;
227
228 if (m != get_default_mirror(ms))
229 goto out;
230
231 if (!ms->in_sync) {
232 /*
233 * Better to issue requests to same failing device
234 * than to risk returning corrupt data.
235 */
236 DMERR("Primary mirror (%s) failed while out-of-sync: "
237 "Reads may fail.", m->dev->name);
238 goto out;
239 }
240
241 new = get_valid_mirror(ms);
242 if (new)
243 set_default_mirror(new);
244 else
245 DMWARN("All sides of mirror have failed.");
246
247out:
248 schedule_work(&ms->trigger_event);
249}
250
251static int mirror_flush(struct dm_target *ti)
252{
253 struct mirror_set *ms = ti->private;
254 unsigned long error_bits;
255
256 unsigned int i;
257 struct dm_io_region io[ms->nr_mirrors];
258 struct mirror *m;
259 struct dm_io_request io_req = {
260 .bi_rw = WRITE_FLUSH,
261 .mem.type = DM_IO_KMEM,
262 .mem.ptr.addr = NULL,
263 .client = ms->io_client,
264 };
265
266 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
267 io[i].bdev = m->dev->bdev;
268 io[i].sector = 0;
269 io[i].count = 0;
270 }
271
272 error_bits = -1;
273 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
274 if (unlikely(error_bits != 0)) {
275 for (i = 0; i < ms->nr_mirrors; i++)
276 if (test_bit(i, &error_bits))
277 fail_mirror(ms->mirror + i,
278 DM_RAID1_FLUSH_ERROR);
279 return -EIO;
280 }
281
282 return 0;
283}
284
285/*-----------------------------------------------------------------
286 * Recovery.
287 *
288 * When a mirror is first activated we may find that some regions
289 * are in the no-sync state. We have to recover these by
290 * recopying from the default mirror to all the others.
291 *---------------------------------------------------------------*/
292static void recovery_complete(int read_err, unsigned long write_err,
293 void *context)
294{
295 struct dm_region *reg = context;
296 struct mirror_set *ms = dm_rh_region_context(reg);
297 int m, bit = 0;
298
299 if (read_err) {
300 /* Read error means the failure of default mirror. */
301 DMERR_LIMIT("Unable to read primary mirror during recovery");
302 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
303 }
304
305 if (write_err) {
306 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
307 write_err);
308 /*
309 * Bits correspond to devices (excluding default mirror).
310 * The default mirror cannot change during recovery.
311 */
312 for (m = 0; m < ms->nr_mirrors; m++) {
313 if (&ms->mirror[m] == get_default_mirror(ms))
314 continue;
315 if (test_bit(bit, &write_err))
316 fail_mirror(ms->mirror + m,
317 DM_RAID1_SYNC_ERROR);
318 bit++;
319 }
320 }
321
322 dm_rh_recovery_end(reg, !(read_err || write_err));
323}
324
325static int recover(struct mirror_set *ms, struct dm_region *reg)
326{
327 int r;
328 unsigned i;
329 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
330 struct mirror *m;
331 unsigned long flags = 0;
332 region_t key = dm_rh_get_region_key(reg);
333 sector_t region_size = dm_rh_get_region_size(ms->rh);
334
335 /* fill in the source */
336 m = get_default_mirror(ms);
337 from.bdev = m->dev->bdev;
338 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
339 if (key == (ms->nr_regions - 1)) {
340 /*
341 * The final region may be smaller than
342 * region_size.
343 */
344 from.count = ms->ti->len & (region_size - 1);
345 if (!from.count)
346 from.count = region_size;
347 } else
348 from.count = region_size;
349
350 /* fill in the destinations */
351 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
352 if (&ms->mirror[i] == get_default_mirror(ms))
353 continue;
354
355 m = ms->mirror + i;
356 dest->bdev = m->dev->bdev;
357 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
358 dest->count = from.count;
359 dest++;
360 }
361
362 /* hand to kcopyd */
363 if (!errors_handled(ms))
364 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
365
366 r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
367 flags, recovery_complete, reg);
368
369 return r;
370}
371
372static void do_recovery(struct mirror_set *ms)
373{
374 struct dm_region *reg;
375 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
376 int r;
377
378 /*
379 * Start quiescing some regions.
380 */
381 dm_rh_recovery_prepare(ms->rh);
382
383 /*
384 * Copy any already quiesced regions.
385 */
386 while ((reg = dm_rh_recovery_start(ms->rh))) {
387 r = recover(ms, reg);
388 if (r)
389 dm_rh_recovery_end(reg, 0);
390 }
391
392 /*
393 * Update the in sync flag.
394 */
395 if (!ms->in_sync &&
396 (log->type->get_sync_count(log) == ms->nr_regions)) {
397 /* the sync is complete */
398 dm_table_event(ms->ti->table);
399 ms->in_sync = 1;
400 }
401}
402
403/*-----------------------------------------------------------------
404 * Reads
405 *---------------------------------------------------------------*/
406static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
407{
408 struct mirror *m = get_default_mirror(ms);
409
410 do {
411 if (likely(!atomic_read(&m->error_count)))
412 return m;
413
414 if (m-- == ms->mirror)
415 m += ms->nr_mirrors;
416 } while (m != get_default_mirror(ms));
417
418 return NULL;
419}
420
421static int default_ok(struct mirror *m)
422{
423 struct mirror *default_mirror = get_default_mirror(m->ms);
424
425 return !atomic_read(&default_mirror->error_count);
426}
427
428static int mirror_available(struct mirror_set *ms, struct bio *bio)
429{
430 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
431 region_t region = dm_rh_bio_to_region(ms->rh, bio);
432
433 if (log->type->in_sync(log, region, 0))
434 return choose_mirror(ms, bio->bi_sector) ? 1 : 0;
435
436 return 0;
437}
438
439/*
440 * remap a buffer to a particular mirror.
441 */
442static sector_t map_sector(struct mirror *m, struct bio *bio)
443{
444 if (unlikely(!bio->bi_size))
445 return 0;
446 return m->offset + dm_target_offset(m->ms->ti, bio->bi_sector);
447}
448
449static void map_bio(struct mirror *m, struct bio *bio)
450{
451 bio->bi_bdev = m->dev->bdev;
452 bio->bi_sector = map_sector(m, bio);
453}
454
455static void map_region(struct dm_io_region *io, struct mirror *m,
456 struct bio *bio)
457{
458 io->bdev = m->dev->bdev;
459 io->sector = map_sector(m, bio);
460 io->count = bio->bi_size >> 9;
461}
462
463static void hold_bio(struct mirror_set *ms, struct bio *bio)
464{
465 /*
466 * Lock is required to avoid race condition during suspend
467 * process.
468 */
469 spin_lock_irq(&ms->lock);
470
471 if (atomic_read(&ms->suspend)) {
472 spin_unlock_irq(&ms->lock);
473
474 /*
475 * If device is suspended, complete the bio.
476 */
477 if (dm_noflush_suspending(ms->ti))
478 bio_endio(bio, DM_ENDIO_REQUEUE);
479 else
480 bio_endio(bio, -EIO);
481 return;
482 }
483
484 /*
485 * Hold bio until the suspend is complete.
486 */
487 bio_list_add(&ms->holds, bio);
488 spin_unlock_irq(&ms->lock);
489}
490
491/*-----------------------------------------------------------------
492 * Reads
493 *---------------------------------------------------------------*/
494static void read_callback(unsigned long error, void *context)
495{
496 struct bio *bio = context;
497 struct mirror *m;
498
499 m = bio_get_m(bio);
500 bio_set_m(bio, NULL);
501
502 if (likely(!error)) {
503 bio_endio(bio, 0);
504 return;
505 }
506
507 fail_mirror(m, DM_RAID1_READ_ERROR);
508
509 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
510 DMWARN_LIMIT("Read failure on mirror device %s. "
511 "Trying alternative device.",
512 m->dev->name);
513 queue_bio(m->ms, bio, bio_rw(bio));
514 return;
515 }
516
517 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
518 m->dev->name);
519 bio_endio(bio, -EIO);
520}
521
522/* Asynchronous read. */
523static void read_async_bio(struct mirror *m, struct bio *bio)
524{
525 struct dm_io_region io;
526 struct dm_io_request io_req = {
527 .bi_rw = READ,
528 .mem.type = DM_IO_BVEC,
529 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
530 .notify.fn = read_callback,
531 .notify.context = bio,
532 .client = m->ms->io_client,
533 };
534
535 map_region(&io, m, bio);
536 bio_set_m(bio, m);
537 BUG_ON(dm_io(&io_req, 1, &io, NULL));
538}
539
540static inline int region_in_sync(struct mirror_set *ms, region_t region,
541 int may_block)
542{
543 int state = dm_rh_get_state(ms->rh, region, may_block);
544 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
545}
546
547static void do_reads(struct mirror_set *ms, struct bio_list *reads)
548{
549 region_t region;
550 struct bio *bio;
551 struct mirror *m;
552
553 while ((bio = bio_list_pop(reads))) {
554 region = dm_rh_bio_to_region(ms->rh, bio);
555 m = get_default_mirror(ms);
556
557 /*
558 * We can only read balance if the region is in sync.
559 */
560 if (likely(region_in_sync(ms, region, 1)))
561 m = choose_mirror(ms, bio->bi_sector);
562 else if (m && atomic_read(&m->error_count))
563 m = NULL;
564
565 if (likely(m))
566 read_async_bio(m, bio);
567 else
568 bio_endio(bio, -EIO);
569 }
570}
571
572/*-----------------------------------------------------------------
573 * Writes.
574 *
575 * We do different things with the write io depending on the
576 * state of the region that it's in:
577 *
578 * SYNC: increment pending, use kcopyd to write to *all* mirrors
579 * RECOVERING: delay the io until recovery completes
580 * NOSYNC: increment pending, just write to the default mirror
581 *---------------------------------------------------------------*/
582
583
584static void write_callback(unsigned long error, void *context)
585{
586 unsigned i, ret = 0;
587 struct bio *bio = (struct bio *) context;
588 struct mirror_set *ms;
589 int should_wake = 0;
590 unsigned long flags;
591
592 ms = bio_get_m(bio)->ms;
593 bio_set_m(bio, NULL);
594
595 /*
596 * NOTE: We don't decrement the pending count here,
597 * instead it is done by the targets endio function.
598 * This way we handle both writes to SYNC and NOSYNC
599 * regions with the same code.
600 */
601 if (likely(!error)) {
602 bio_endio(bio, ret);
603 return;
604 }
605
606 for (i = 0; i < ms->nr_mirrors; i++)
607 if (test_bit(i, &error))
608 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
609
610 /*
611 * Need to raise event. Since raising
612 * events can block, we need to do it in
613 * the main thread.
614 */
615 spin_lock_irqsave(&ms->lock, flags);
616 if (!ms->failures.head)
617 should_wake = 1;
618 bio_list_add(&ms->failures, bio);
619 spin_unlock_irqrestore(&ms->lock, flags);
620 if (should_wake)
621 wakeup_mirrord(ms);
622}
623
624static void do_write(struct mirror_set *ms, struct bio *bio)
625{
626 unsigned int i;
627 struct dm_io_region io[ms->nr_mirrors], *dest = io;
628 struct mirror *m;
629 struct dm_io_request io_req = {
630 .bi_rw = WRITE | (bio->bi_rw & WRITE_FLUSH_FUA),
631 .mem.type = DM_IO_BVEC,
632 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
633 .notify.fn = write_callback,
634 .notify.context = bio,
635 .client = ms->io_client,
636 };
637
638 if (bio->bi_rw & REQ_DISCARD) {
639 io_req.bi_rw |= REQ_DISCARD;
640 io_req.mem.type = DM_IO_KMEM;
641 io_req.mem.ptr.addr = NULL;
642 }
643
644 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
645 map_region(dest++, m, bio);
646
647 /*
648 * Use default mirror because we only need it to retrieve the reference
649 * to the mirror set in write_callback().
650 */
651 bio_set_m(bio, get_default_mirror(ms));
652
653 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
654}
655
656static void do_writes(struct mirror_set *ms, struct bio_list *writes)
657{
658 int state;
659 struct bio *bio;
660 struct bio_list sync, nosync, recover, *this_list = NULL;
661 struct bio_list requeue;
662 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
663 region_t region;
664
665 if (!writes->head)
666 return;
667
668 /*
669 * Classify each write.
670 */
671 bio_list_init(&sync);
672 bio_list_init(&nosync);
673 bio_list_init(&recover);
674 bio_list_init(&requeue);
675
676 while ((bio = bio_list_pop(writes))) {
677 if ((bio->bi_rw & REQ_FLUSH) ||
678 (bio->bi_rw & REQ_DISCARD)) {
679 bio_list_add(&sync, bio);
680 continue;
681 }
682
683 region = dm_rh_bio_to_region(ms->rh, bio);
684
685 if (log->type->is_remote_recovering &&
686 log->type->is_remote_recovering(log, region)) {
687 bio_list_add(&requeue, bio);
688 continue;
689 }
690
691 state = dm_rh_get_state(ms->rh, region, 1);
692 switch (state) {
693 case DM_RH_CLEAN:
694 case DM_RH_DIRTY:
695 this_list = &sync;
696 break;
697
698 case DM_RH_NOSYNC:
699 this_list = &nosync;
700 break;
701
702 case DM_RH_RECOVERING:
703 this_list = &recover;
704 break;
705 }
706
707 bio_list_add(this_list, bio);
708 }
709
710 /*
711 * Add bios that are delayed due to remote recovery
712 * back on to the write queue
713 */
714 if (unlikely(requeue.head)) {
715 spin_lock_irq(&ms->lock);
716 bio_list_merge(&ms->writes, &requeue);
717 spin_unlock_irq(&ms->lock);
718 delayed_wake(ms);
719 }
720
721 /*
722 * Increment the pending counts for any regions that will
723 * be written to (writes to recover regions are going to
724 * be delayed).
725 */
726 dm_rh_inc_pending(ms->rh, &sync);
727 dm_rh_inc_pending(ms->rh, &nosync);
728
729 /*
730 * If the flush fails on a previous call and succeeds here,
731 * we must not reset the log_failure variable. We need
732 * userspace interaction to do that.
733 */
734 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
735
736 /*
737 * Dispatch io.
738 */
739 if (unlikely(ms->log_failure) && errors_handled(ms)) {
740 spin_lock_irq(&ms->lock);
741 bio_list_merge(&ms->failures, &sync);
742 spin_unlock_irq(&ms->lock);
743 wakeup_mirrord(ms);
744 } else
745 while ((bio = bio_list_pop(&sync)))
746 do_write(ms, bio);
747
748 while ((bio = bio_list_pop(&recover)))
749 dm_rh_delay(ms->rh, bio);
750
751 while ((bio = bio_list_pop(&nosync))) {
752 if (unlikely(ms->leg_failure) && errors_handled(ms)) {
753 spin_lock_irq(&ms->lock);
754 bio_list_add(&ms->failures, bio);
755 spin_unlock_irq(&ms->lock);
756 wakeup_mirrord(ms);
757 } else {
758 map_bio(get_default_mirror(ms), bio);
759 generic_make_request(bio);
760 }
761 }
762}
763
764static void do_failures(struct mirror_set *ms, struct bio_list *failures)
765{
766 struct bio *bio;
767
768 if (likely(!failures->head))
769 return;
770
771 /*
772 * If the log has failed, unattempted writes are being
773 * put on the holds list. We can't issue those writes
774 * until a log has been marked, so we must store them.
775 *
776 * If a 'noflush' suspend is in progress, we can requeue
777 * the I/O's to the core. This give userspace a chance
778 * to reconfigure the mirror, at which point the core
779 * will reissue the writes. If the 'noflush' flag is
780 * not set, we have no choice but to return errors.
781 *
782 * Some writes on the failures list may have been
783 * submitted before the log failure and represent a
784 * failure to write to one of the devices. It is ok
785 * for us to treat them the same and requeue them
786 * as well.
787 */
788 while ((bio = bio_list_pop(failures))) {
789 if (!ms->log_failure) {
790 ms->in_sync = 0;
791 dm_rh_mark_nosync(ms->rh, bio);
792 }
793
794 /*
795 * If all the legs are dead, fail the I/O.
796 * If we have been told to handle errors, hold the bio
797 * and wait for userspace to deal with the problem.
798 * Otherwise pretend that the I/O succeeded. (This would
799 * be wrong if the failed leg returned after reboot and
800 * got replicated back to the good legs.)
801 */
802 if (!get_valid_mirror(ms))
803 bio_endio(bio, -EIO);
804 else if (errors_handled(ms))
805 hold_bio(ms, bio);
806 else
807 bio_endio(bio, 0);
808 }
809}
810
811static void trigger_event(struct work_struct *work)
812{
813 struct mirror_set *ms =
814 container_of(work, struct mirror_set, trigger_event);
815
816 dm_table_event(ms->ti->table);
817}
818
819/*-----------------------------------------------------------------
820 * kmirrord
821 *---------------------------------------------------------------*/
822static void do_mirror(struct work_struct *work)
823{
824 struct mirror_set *ms = container_of(work, struct mirror_set,
825 kmirrord_work);
826 struct bio_list reads, writes, failures;
827 unsigned long flags;
828
829 spin_lock_irqsave(&ms->lock, flags);
830 reads = ms->reads;
831 writes = ms->writes;
832 failures = ms->failures;
833 bio_list_init(&ms->reads);
834 bio_list_init(&ms->writes);
835 bio_list_init(&ms->failures);
836 spin_unlock_irqrestore(&ms->lock, flags);
837
838 dm_rh_update_states(ms->rh, errors_handled(ms));
839 do_recovery(ms);
840 do_reads(ms, &reads);
841 do_writes(ms, &writes);
842 do_failures(ms, &failures);
843}
844
845/*-----------------------------------------------------------------
846 * Target functions
847 *---------------------------------------------------------------*/
848static struct mirror_set *alloc_context(unsigned int nr_mirrors,
849 uint32_t region_size,
850 struct dm_target *ti,
851 struct dm_dirty_log *dl)
852{
853 size_t len;
854 struct mirror_set *ms = NULL;
855
856 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
857
858 ms = kzalloc(len, GFP_KERNEL);
859 if (!ms) {
860 ti->error = "Cannot allocate mirror context";
861 return NULL;
862 }
863
864 spin_lock_init(&ms->lock);
865 bio_list_init(&ms->reads);
866 bio_list_init(&ms->writes);
867 bio_list_init(&ms->failures);
868 bio_list_init(&ms->holds);
869
870 ms->ti = ti;
871 ms->nr_mirrors = nr_mirrors;
872 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
873 ms->in_sync = 0;
874 ms->log_failure = 0;
875 ms->leg_failure = 0;
876 atomic_set(&ms->suspend, 0);
877 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
878
879 ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
880 _dm_raid1_read_record_cache);
881
882 if (!ms->read_record_pool) {
883 ti->error = "Error creating mirror read_record_pool";
884 kfree(ms);
885 return NULL;
886 }
887
888 ms->io_client = dm_io_client_create();
889 if (IS_ERR(ms->io_client)) {
890 ti->error = "Error creating dm_io client";
891 mempool_destroy(ms->read_record_pool);
892 kfree(ms);
893 return NULL;
894 }
895
896 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
897 wakeup_all_recovery_waiters,
898 ms->ti->begin, MAX_RECOVERY,
899 dl, region_size, ms->nr_regions);
900 if (IS_ERR(ms->rh)) {
901 ti->error = "Error creating dirty region hash";
902 dm_io_client_destroy(ms->io_client);
903 mempool_destroy(ms->read_record_pool);
904 kfree(ms);
905 return NULL;
906 }
907
908 return ms;
909}
910
911static void free_context(struct mirror_set *ms, struct dm_target *ti,
912 unsigned int m)
913{
914 while (m--)
915 dm_put_device(ti, ms->mirror[m].dev);
916
917 dm_io_client_destroy(ms->io_client);
918 dm_region_hash_destroy(ms->rh);
919 mempool_destroy(ms->read_record_pool);
920 kfree(ms);
921}
922
923static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
924 unsigned int mirror, char **argv)
925{
926 unsigned long long offset;
927 char dummy;
928
929 if (sscanf(argv[1], "%llu%c", &offset, &dummy) != 1) {
930 ti->error = "Invalid offset";
931 return -EINVAL;
932 }
933
934 if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
935 &ms->mirror[mirror].dev)) {
936 ti->error = "Device lookup failure";
937 return -ENXIO;
938 }
939
940 ms->mirror[mirror].ms = ms;
941 atomic_set(&(ms->mirror[mirror].error_count), 0);
942 ms->mirror[mirror].error_type = 0;
943 ms->mirror[mirror].offset = offset;
944
945 return 0;
946}
947
948/*
949 * Create dirty log: log_type #log_params <log_params>
950 */
951static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
952 unsigned argc, char **argv,
953 unsigned *args_used)
954{
955 unsigned param_count;
956 struct dm_dirty_log *dl;
957 char dummy;
958
959 if (argc < 2) {
960 ti->error = "Insufficient mirror log arguments";
961 return NULL;
962 }
963
964 if (sscanf(argv[1], "%u%c", ¶m_count, &dummy) != 1) {
965 ti->error = "Invalid mirror log argument count";
966 return NULL;
967 }
968
969 *args_used = 2 + param_count;
970
971 if (argc < *args_used) {
972 ti->error = "Insufficient mirror log arguments";
973 return NULL;
974 }
975
976 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
977 argv + 2);
978 if (!dl) {
979 ti->error = "Error creating mirror dirty log";
980 return NULL;
981 }
982
983 return dl;
984}
985
986static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
987 unsigned *args_used)
988{
989 unsigned num_features;
990 struct dm_target *ti = ms->ti;
991 char dummy;
992
993 *args_used = 0;
994
995 if (!argc)
996 return 0;
997
998 if (sscanf(argv[0], "%u%c", &num_features, &dummy) != 1) {
999 ti->error = "Invalid number of features";
1000 return -EINVAL;
1001 }
1002
1003 argc--;
1004 argv++;
1005 (*args_used)++;
1006
1007 if (num_features > argc) {
1008 ti->error = "Not enough arguments to support feature count";
1009 return -EINVAL;
1010 }
1011
1012 if (!strcmp("handle_errors", argv[0]))
1013 ms->features |= DM_RAID1_HANDLE_ERRORS;
1014 else {
1015 ti->error = "Unrecognised feature requested";
1016 return -EINVAL;
1017 }
1018
1019 (*args_used)++;
1020
1021 return 0;
1022}
1023
1024/*
1025 * Construct a mirror mapping:
1026 *
1027 * log_type #log_params <log_params>
1028 * #mirrors [mirror_path offset]{2,}
1029 * [#features <features>]
1030 *
1031 * log_type is "core" or "disk"
1032 * #log_params is between 1 and 3
1033 *
1034 * If present, features must be "handle_errors".
1035 */
1036static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1037{
1038 int r;
1039 unsigned int nr_mirrors, m, args_used;
1040 struct mirror_set *ms;
1041 struct dm_dirty_log *dl;
1042 char dummy;
1043
1044 dl = create_dirty_log(ti, argc, argv, &args_used);
1045 if (!dl)
1046 return -EINVAL;
1047
1048 argv += args_used;
1049 argc -= args_used;
1050
1051 if (!argc || sscanf(argv[0], "%u%c", &nr_mirrors, &dummy) != 1 ||
1052 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
1053 ti->error = "Invalid number of mirrors";
1054 dm_dirty_log_destroy(dl);
1055 return -EINVAL;
1056 }
1057
1058 argv++, argc--;
1059
1060 if (argc < nr_mirrors * 2) {
1061 ti->error = "Too few mirror arguments";
1062 dm_dirty_log_destroy(dl);
1063 return -EINVAL;
1064 }
1065
1066 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1067 if (!ms) {
1068 dm_dirty_log_destroy(dl);
1069 return -ENOMEM;
1070 }
1071
1072 /* Get the mirror parameter sets */
1073 for (m = 0; m < nr_mirrors; m++) {
1074 r = get_mirror(ms, ti, m, argv);
1075 if (r) {
1076 free_context(ms, ti, m);
1077 return r;
1078 }
1079 argv += 2;
1080 argc -= 2;
1081 }
1082
1083 ti->private = ms;
1084 ti->split_io = dm_rh_get_region_size(ms->rh);
1085 ti->num_flush_requests = 1;
1086 ti->num_discard_requests = 1;
1087 ti->discard_zeroes_data_unsupported = 1;
1088
1089 ms->kmirrord_wq = alloc_workqueue("kmirrord",
1090 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
1091 if (!ms->kmirrord_wq) {
1092 DMERR("couldn't start kmirrord");
1093 r = -ENOMEM;
1094 goto err_free_context;
1095 }
1096 INIT_WORK(&ms->kmirrord_work, do_mirror);
1097 init_timer(&ms->timer);
1098 ms->timer_pending = 0;
1099 INIT_WORK(&ms->trigger_event, trigger_event);
1100
1101 r = parse_features(ms, argc, argv, &args_used);
1102 if (r)
1103 goto err_destroy_wq;
1104
1105 argv += args_used;
1106 argc -= args_used;
1107
1108 /*
1109 * Any read-balancing addition depends on the
1110 * DM_RAID1_HANDLE_ERRORS flag being present.
1111 * This is because the decision to balance depends
1112 * on the sync state of a region. If the above
1113 * flag is not present, we ignore errors; and
1114 * the sync state may be inaccurate.
1115 */
1116
1117 if (argc) {
1118 ti->error = "Too many mirror arguments";
1119 r = -EINVAL;
1120 goto err_destroy_wq;
1121 }
1122
1123 ms->kcopyd_client = dm_kcopyd_client_create();
1124 if (IS_ERR(ms->kcopyd_client)) {
1125 r = PTR_ERR(ms->kcopyd_client);
1126 goto err_destroy_wq;
1127 }
1128
1129 wakeup_mirrord(ms);
1130 return 0;
1131
1132err_destroy_wq:
1133 destroy_workqueue(ms->kmirrord_wq);
1134err_free_context:
1135 free_context(ms, ti, ms->nr_mirrors);
1136 return r;
1137}
1138
1139static void mirror_dtr(struct dm_target *ti)
1140{
1141 struct mirror_set *ms = (struct mirror_set *) ti->private;
1142
1143 del_timer_sync(&ms->timer);
1144 flush_workqueue(ms->kmirrord_wq);
1145 flush_work_sync(&ms->trigger_event);
1146 dm_kcopyd_client_destroy(ms->kcopyd_client);
1147 destroy_workqueue(ms->kmirrord_wq);
1148 free_context(ms, ti, ms->nr_mirrors);
1149}
1150
1151/*
1152 * Mirror mapping function
1153 */
1154static int mirror_map(struct dm_target *ti, struct bio *bio,
1155 union map_info *map_context)
1156{
1157 int r, rw = bio_rw(bio);
1158 struct mirror *m;
1159 struct mirror_set *ms = ti->private;
1160 struct dm_raid1_read_record *read_record = NULL;
1161 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1162
1163 if (rw == WRITE) {
1164 /* Save region for mirror_end_io() handler */
1165 map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1166 queue_bio(ms, bio, rw);
1167 return DM_MAPIO_SUBMITTED;
1168 }
1169
1170 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1171 if (r < 0 && r != -EWOULDBLOCK)
1172 return r;
1173
1174 /*
1175 * If region is not in-sync queue the bio.
1176 */
1177 if (!r || (r == -EWOULDBLOCK)) {
1178 if (rw == READA)
1179 return -EWOULDBLOCK;
1180
1181 queue_bio(ms, bio, rw);
1182 return DM_MAPIO_SUBMITTED;
1183 }
1184
1185 /*
1186 * The region is in-sync and we can perform reads directly.
1187 * Store enough information so we can retry if it fails.
1188 */
1189 m = choose_mirror(ms, bio->bi_sector);
1190 if (unlikely(!m))
1191 return -EIO;
1192
1193 read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1194 if (likely(read_record)) {
1195 dm_bio_record(&read_record->details, bio);
1196 map_context->ptr = read_record;
1197 read_record->m = m;
1198 }
1199
1200 map_bio(m, bio);
1201
1202 return DM_MAPIO_REMAPPED;
1203}
1204
1205static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1206 int error, union map_info *map_context)
1207{
1208 int rw = bio_rw(bio);
1209 struct mirror_set *ms = (struct mirror_set *) ti->private;
1210 struct mirror *m = NULL;
1211 struct dm_bio_details *bd = NULL;
1212 struct dm_raid1_read_record *read_record = map_context->ptr;
1213
1214 /*
1215 * We need to dec pending if this was a write.
1216 */
1217 if (rw == WRITE) {
1218 if (!(bio->bi_rw & (REQ_FLUSH | REQ_DISCARD)))
1219 dm_rh_dec(ms->rh, map_context->ll);
1220 return error;
1221 }
1222
1223 if (error == -EOPNOTSUPP)
1224 goto out;
1225
1226 if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
1227 goto out;
1228
1229 if (unlikely(error)) {
1230 if (!read_record) {
1231 /*
1232 * There wasn't enough memory to record necessary
1233 * information for a retry or there was no other
1234 * mirror in-sync.
1235 */
1236 DMERR_LIMIT("Mirror read failed.");
1237 return -EIO;
1238 }
1239
1240 m = read_record->m;
1241
1242 DMERR("Mirror read failed from %s. Trying alternative device.",
1243 m->dev->name);
1244
1245 fail_mirror(m, DM_RAID1_READ_ERROR);
1246
1247 /*
1248 * A failed read is requeued for another attempt using an intact
1249 * mirror.
1250 */
1251 if (default_ok(m) || mirror_available(ms, bio)) {
1252 bd = &read_record->details;
1253
1254 dm_bio_restore(bd, bio);
1255 mempool_free(read_record, ms->read_record_pool);
1256 map_context->ptr = NULL;
1257 queue_bio(ms, bio, rw);
1258 return 1;
1259 }
1260 DMERR("All replicated volumes dead, failing I/O");
1261 }
1262
1263out:
1264 if (read_record) {
1265 mempool_free(read_record, ms->read_record_pool);
1266 map_context->ptr = NULL;
1267 }
1268
1269 return error;
1270}
1271
1272static void mirror_presuspend(struct dm_target *ti)
1273{
1274 struct mirror_set *ms = (struct mirror_set *) ti->private;
1275 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1276
1277 struct bio_list holds;
1278 struct bio *bio;
1279
1280 atomic_set(&ms->suspend, 1);
1281
1282 /*
1283 * Process bios in the hold list to start recovery waiting
1284 * for bios in the hold list. After the process, no bio has
1285 * a chance to be added in the hold list because ms->suspend
1286 * is set.
1287 */
1288 spin_lock_irq(&ms->lock);
1289 holds = ms->holds;
1290 bio_list_init(&ms->holds);
1291 spin_unlock_irq(&ms->lock);
1292
1293 while ((bio = bio_list_pop(&holds)))
1294 hold_bio(ms, bio);
1295
1296 /*
1297 * We must finish up all the work that we've
1298 * generated (i.e. recovery work).
1299 */
1300 dm_rh_stop_recovery(ms->rh);
1301
1302 wait_event(_kmirrord_recovery_stopped,
1303 !dm_rh_recovery_in_flight(ms->rh));
1304
1305 if (log->type->presuspend && log->type->presuspend(log))
1306 /* FIXME: need better error handling */
1307 DMWARN("log presuspend failed");
1308
1309 /*
1310 * Now that recovery is complete/stopped and the
1311 * delayed bios are queued, we need to wait for
1312 * the worker thread to complete. This way,
1313 * we know that all of our I/O has been pushed.
1314 */
1315 flush_workqueue(ms->kmirrord_wq);
1316}
1317
1318static void mirror_postsuspend(struct dm_target *ti)
1319{
1320 struct mirror_set *ms = ti->private;
1321 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1322
1323 if (log->type->postsuspend && log->type->postsuspend(log))
1324 /* FIXME: need better error handling */
1325 DMWARN("log postsuspend failed");
1326}
1327
1328static void mirror_resume(struct dm_target *ti)
1329{
1330 struct mirror_set *ms = ti->private;
1331 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1332
1333 atomic_set(&ms->suspend, 0);
1334 if (log->type->resume && log->type->resume(log))
1335 /* FIXME: need better error handling */
1336 DMWARN("log resume failed");
1337 dm_rh_start_recovery(ms->rh);
1338}
1339
1340/*
1341 * device_status_char
1342 * @m: mirror device/leg we want the status of
1343 *
1344 * We return one character representing the most severe error
1345 * we have encountered.
1346 * A => Alive - No failures
1347 * D => Dead - A write failure occurred leaving mirror out-of-sync
1348 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1349 * R => Read - A read failure occurred, mirror data unaffected
1350 *
1351 * Returns: <char>
1352 */
1353static char device_status_char(struct mirror *m)
1354{
1355 if (!atomic_read(&(m->error_count)))
1356 return 'A';
1357
1358 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1359 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1360 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1361 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1362}
1363
1364
1365static int mirror_status(struct dm_target *ti, status_type_t type,
1366 char *result, unsigned int maxlen)
1367{
1368 unsigned int m, sz = 0;
1369 struct mirror_set *ms = (struct mirror_set *) ti->private;
1370 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1371 char buffer[ms->nr_mirrors + 1];
1372
1373 switch (type) {
1374 case STATUSTYPE_INFO:
1375 DMEMIT("%d ", ms->nr_mirrors);
1376 for (m = 0; m < ms->nr_mirrors; m++) {
1377 DMEMIT("%s ", ms->mirror[m].dev->name);
1378 buffer[m] = device_status_char(&(ms->mirror[m]));
1379 }
1380 buffer[m] = '\0';
1381
1382 DMEMIT("%llu/%llu 1 %s ",
1383 (unsigned long long)log->type->get_sync_count(log),
1384 (unsigned long long)ms->nr_regions, buffer);
1385
1386 sz += log->type->status(log, type, result+sz, maxlen-sz);
1387
1388 break;
1389
1390 case STATUSTYPE_TABLE:
1391 sz = log->type->status(log, type, result, maxlen);
1392
1393 DMEMIT("%d", ms->nr_mirrors);
1394 for (m = 0; m < ms->nr_mirrors; m++)
1395 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1396 (unsigned long long)ms->mirror[m].offset);
1397
1398 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1399 DMEMIT(" 1 handle_errors");
1400 }
1401
1402 return 0;
1403}
1404
1405static int mirror_iterate_devices(struct dm_target *ti,
1406 iterate_devices_callout_fn fn, void *data)
1407{
1408 struct mirror_set *ms = ti->private;
1409 int ret = 0;
1410 unsigned i;
1411
1412 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1413 ret = fn(ti, ms->mirror[i].dev,
1414 ms->mirror[i].offset, ti->len, data);
1415
1416 return ret;
1417}
1418
1419static struct target_type mirror_target = {
1420 .name = "mirror",
1421 .version = {1, 12, 1},
1422 .module = THIS_MODULE,
1423 .ctr = mirror_ctr,
1424 .dtr = mirror_dtr,
1425 .map = mirror_map,
1426 .end_io = mirror_end_io,
1427 .presuspend = mirror_presuspend,
1428 .postsuspend = mirror_postsuspend,
1429 .resume = mirror_resume,
1430 .status = mirror_status,
1431 .iterate_devices = mirror_iterate_devices,
1432};
1433
1434static int __init dm_mirror_init(void)
1435{
1436 int r;
1437
1438 _dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1439 if (!_dm_raid1_read_record_cache) {
1440 DMERR("Can't allocate dm_raid1_read_record cache");
1441 r = -ENOMEM;
1442 goto bad_cache;
1443 }
1444
1445 r = dm_register_target(&mirror_target);
1446 if (r < 0) {
1447 DMERR("Failed to register mirror target");
1448 goto bad_target;
1449 }
1450
1451 return 0;
1452
1453bad_target:
1454 kmem_cache_destroy(_dm_raid1_read_record_cache);
1455bad_cache:
1456 return r;
1457}
1458
1459static void __exit dm_mirror_exit(void)
1460{
1461 dm_unregister_target(&mirror_target);
1462 kmem_cache_destroy(_dm_raid1_read_record_cache);
1463}
1464
1465/* Module hooks */
1466module_init(dm_mirror_init);
1467module_exit(dm_mirror_exit);
1468
1469MODULE_DESCRIPTION(DM_NAME " mirror target");
1470MODULE_AUTHOR("Joe Thornber");
1471MODULE_LICENSE("GPL");