2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2005-2008 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include "dm-bio-record.h"
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>
22 #define DM_MSG_PREFIX "raid1"
24 #define MAX_RECOVERY 1 /* Maximum number of regions recovered in parallel. */
25 #define DM_IO_PAGES 64
26 #define DM_KCOPYD_PAGES 64
28 #define DM_RAID1_HANDLE_ERRORS 0x01
29 #define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
31 static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
33 /*-----------------------------------------------------------------
34 * Mirror set structures.
35 *---------------------------------------------------------------*/
43 struct mirror_set *ms;
45 unsigned long error_type;
52 struct list_head list;
56 spinlock_t lock; /* protects the lists */
57 struct bio_list reads;
58 struct bio_list writes;
59 struct bio_list failures;
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;
72 atomic_t default_mirror; /* Default mirror */
74 struct workqueue_struct *kmirrord_wq;
75 struct work_struct kmirrord_work;
76 struct timer_list timer;
77 unsigned long timer_pending;
79 struct work_struct trigger_event;
82 struct mirror mirror[0];
85 static void wakeup_mirrord(void *context)
87 struct mirror_set *ms = context;
89 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
92 static void delayed_wake_fn(unsigned long data)
94 struct mirror_set *ms = (struct mirror_set *) data;
96 clear_bit(0, &ms->timer_pending);
100 static void delayed_wake(struct mirror_set *ms)
102 if (test_and_set_bit(0, &ms->timer_pending))
105 ms->timer.expires = jiffies + HZ / 5;
106 ms->timer.data = (unsigned long) ms;
107 ms->timer.function = delayed_wake_fn;
108 add_timer(&ms->timer);
111 static void wakeup_all_recovery_waiters(void *context)
113 wake_up_all(&_kmirrord_recovery_stopped);
116 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
122 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
123 spin_lock_irqsave(&ms->lock, flags);
124 should_wake = !(bl->head);
125 bio_list_add(bl, bio);
126 spin_unlock_irqrestore(&ms->lock, flags);
132 static void dispatch_bios(void *context, struct bio_list *bio_list)
134 struct mirror_set *ms = context;
137 while ((bio = bio_list_pop(bio_list)))
138 queue_bio(ms, bio, WRITE);
141 #define MIN_READ_RECORDS 20
142 struct dm_raid1_read_record {
144 struct dm_bio_details details;
147 static struct kmem_cache *_dm_raid1_read_record_cache;
150 * Every mirror should look like this one.
152 #define DEFAULT_MIRROR 0
155 * This is yucky. We squirrel the mirror struct away inside
156 * bi_next for read/write buffers. This is safe since the bh
157 * doesn't get submitted to the lower levels of block layer.
159 static struct mirror *bio_get_m(struct bio *bio)
161 return (struct mirror *) bio->bi_next;
164 static void bio_set_m(struct bio *bio, struct mirror *m)
166 bio->bi_next = (struct bio *) m;
169 static struct mirror *get_default_mirror(struct mirror_set *ms)
171 return &ms->mirror[atomic_read(&ms->default_mirror)];
174 static void set_default_mirror(struct mirror *m)
176 struct mirror_set *ms = m->ms;
177 struct mirror *m0 = &(ms->mirror[0]);
179 atomic_set(&ms->default_mirror, m - m0);
183 * @m: mirror device to fail
184 * @error_type: one of the enum's, DM_RAID1_*_ERROR
186 * If errors are being handled, record the type of
187 * error encountered for this device. If this type
188 * of error has already been recorded, we can return;
189 * otherwise, we must signal userspace by triggering
190 * an event. Additionally, if the device is the
191 * primary device, we must choose a new primary, but
192 * only if the mirror is in-sync.
194 * This function must not block.
196 static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
198 struct mirror_set *ms = m->ms;
202 * error_count is used for nothing more than a
203 * simple way to tell if a device has encountered
206 atomic_inc(&m->error_count);
208 if (test_and_set_bit(error_type, &m->error_type))
211 if (!errors_handled(ms))
214 if (m != get_default_mirror(ms))
219 * Better to issue requests to same failing device
220 * than to risk returning corrupt data.
222 DMERR("Primary mirror (%s) failed while out-of-sync: "
223 "Reads may fail.", m->dev->name);
227 for (new = ms->mirror; new < ms->mirror + ms->nr_mirrors; new++)
228 if (!atomic_read(&new->error_count)) {
229 set_default_mirror(new);
233 if (unlikely(new == ms->mirror + ms->nr_mirrors))
234 DMWARN("All sides of mirror have failed.");
237 schedule_work(&ms->trigger_event);
240 /*-----------------------------------------------------------------
243 * When a mirror is first activated we may find that some regions
244 * are in the no-sync state. We have to recover these by
245 * recopying from the default mirror to all the others.
246 *---------------------------------------------------------------*/
247 static void recovery_complete(int read_err, unsigned long write_err,
250 struct dm_region *reg = context;
251 struct mirror_set *ms = dm_rh_region_context(reg);
255 /* Read error means the failure of default mirror. */
256 DMERR_LIMIT("Unable to read primary mirror during recovery");
257 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
261 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
264 * Bits correspond to devices (excluding default mirror).
265 * The default mirror cannot change during recovery.
267 for (m = 0; m < ms->nr_mirrors; m++) {
268 if (&ms->mirror[m] == get_default_mirror(ms))
270 if (test_bit(bit, &write_err))
271 fail_mirror(ms->mirror + m,
272 DM_RAID1_SYNC_ERROR);
277 dm_rh_recovery_end(reg, !(read_err || write_err));
280 static int recover(struct mirror_set *ms, struct dm_region *reg)
284 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
286 unsigned long flags = 0;
287 region_t key = dm_rh_get_region_key(reg);
288 sector_t region_size = dm_rh_get_region_size(ms->rh);
290 /* fill in the source */
291 m = get_default_mirror(ms);
292 from.bdev = m->dev->bdev;
293 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
294 if (key == (ms->nr_regions - 1)) {
296 * The final region may be smaller than
299 from.count = ms->ti->len & (region_size - 1);
301 from.count = region_size;
303 from.count = region_size;
305 /* fill in the destinations */
306 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
307 if (&ms->mirror[i] == get_default_mirror(ms))
311 dest->bdev = m->dev->bdev;
312 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
313 dest->count = from.count;
318 if (!errors_handled(ms))
319 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
321 r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
322 flags, recovery_complete, reg);
327 static void do_recovery(struct mirror_set *ms)
329 struct dm_region *reg;
330 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
334 * Start quiescing some regions.
336 dm_rh_recovery_prepare(ms->rh);
339 * Copy any already quiesced regions.
341 while ((reg = dm_rh_recovery_start(ms->rh))) {
342 r = recover(ms, reg);
344 dm_rh_recovery_end(reg, 0);
348 * Update the in sync flag.
351 (log->type->get_sync_count(log) == ms->nr_regions)) {
352 /* the sync is complete */
353 dm_table_event(ms->ti->table);
358 /*-----------------------------------------------------------------
360 *---------------------------------------------------------------*/
361 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
363 struct mirror *m = get_default_mirror(ms);
366 if (likely(!atomic_read(&m->error_count)))
369 if (m-- == ms->mirror)
371 } while (m != get_default_mirror(ms));
376 static int default_ok(struct mirror *m)
378 struct mirror *default_mirror = get_default_mirror(m->ms);
380 return !atomic_read(&default_mirror->error_count);
383 static int mirror_available(struct mirror_set *ms, struct bio *bio)
385 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
386 region_t region = dm_rh_bio_to_region(ms->rh, bio);
388 if (log->type->in_sync(log, region, 0))
389 return choose_mirror(ms, bio->bi_sector) ? 1 : 0;
395 * remap a buffer to a particular mirror.
397 static sector_t map_sector(struct mirror *m, struct bio *bio)
399 return m->offset + (bio->bi_sector - m->ms->ti->begin);
402 static void map_bio(struct mirror *m, struct bio *bio)
404 bio->bi_bdev = m->dev->bdev;
405 bio->bi_sector = map_sector(m, bio);
408 static void map_region(struct dm_io_region *io, struct mirror *m,
411 io->bdev = m->dev->bdev;
412 io->sector = map_sector(m, bio);
413 io->count = bio->bi_size >> 9;
416 /*-----------------------------------------------------------------
418 *---------------------------------------------------------------*/
419 static void read_callback(unsigned long error, void *context)
421 struct bio *bio = context;
425 bio_set_m(bio, NULL);
427 if (likely(!error)) {
432 fail_mirror(m, DM_RAID1_READ_ERROR);
434 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
435 DMWARN_LIMIT("Read failure on mirror device %s. "
436 "Trying alternative device.",
438 queue_bio(m->ms, bio, bio_rw(bio));
442 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
444 bio_endio(bio, -EIO);
447 /* Asynchronous read. */
448 static void read_async_bio(struct mirror *m, struct bio *bio)
450 struct dm_io_region io;
451 struct dm_io_request io_req = {
453 .mem.type = DM_IO_BVEC,
454 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
455 .notify.fn = read_callback,
456 .notify.context = bio,
457 .client = m->ms->io_client,
460 map_region(&io, m, bio);
462 BUG_ON(dm_io(&io_req, 1, &io, NULL));
465 static inline int region_in_sync(struct mirror_set *ms, region_t region,
468 int state = dm_rh_get_state(ms->rh, region, may_block);
469 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
472 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
478 while ((bio = bio_list_pop(reads))) {
479 region = dm_rh_bio_to_region(ms->rh, bio);
480 m = get_default_mirror(ms);
483 * We can only read balance if the region is in sync.
485 if (likely(region_in_sync(ms, region, 1)))
486 m = choose_mirror(ms, bio->bi_sector);
487 else if (m && atomic_read(&m->error_count))
491 read_async_bio(m, bio);
493 bio_endio(bio, -EIO);
497 /*-----------------------------------------------------------------
500 * We do different things with the write io depending on the
501 * state of the region that it's in:
503 * SYNC: increment pending, use kcopyd to write to *all* mirrors
504 * RECOVERING: delay the io until recovery completes
505 * NOSYNC: increment pending, just write to the default mirror
506 *---------------------------------------------------------------*/
509 static void write_callback(unsigned long error, void *context)
512 struct bio *bio = (struct bio *) context;
513 struct mirror_set *ms;
518 ms = bio_get_m(bio)->ms;
519 bio_set_m(bio, NULL);
522 * NOTE: We don't decrement the pending count here,
523 * instead it is done by the targets endio function.
524 * This way we handle both writes to SYNC and NOSYNC
525 * regions with the same code.
530 for (i = 0; i < ms->nr_mirrors; i++)
531 if (test_bit(i, &error))
532 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
536 if (unlikely(!uptodate)) {
537 DMERR("All replicated volumes dead, failing I/O");
538 /* None of the writes succeeded, fail the I/O. */
540 } else if (errors_handled(ms)) {
542 * Need to raise event. Since raising
543 * events can block, we need to do it in
546 spin_lock_irqsave(&ms->lock, flags);
547 if (!ms->failures.head)
549 bio_list_add(&ms->failures, bio);
550 spin_unlock_irqrestore(&ms->lock, flags);
559 static void do_write(struct mirror_set *ms, struct bio *bio)
562 struct dm_io_region io[ms->nr_mirrors], *dest = io;
564 struct dm_io_request io_req = {
566 .mem.type = DM_IO_BVEC,
567 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
568 .notify.fn = write_callback,
569 .notify.context = bio,
570 .client = ms->io_client,
573 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
574 map_region(dest++, m, bio);
577 * Use default mirror because we only need it to retrieve the reference
578 * to the mirror set in write_callback().
580 bio_set_m(bio, get_default_mirror(ms));
582 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
585 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
589 struct bio_list sync, nosync, recover, *this_list = NULL;
590 struct bio_list requeue;
591 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
598 * Classify each write.
600 bio_list_init(&sync);
601 bio_list_init(&nosync);
602 bio_list_init(&recover);
603 bio_list_init(&requeue);
605 while ((bio = bio_list_pop(writes))) {
606 region = dm_rh_bio_to_region(ms->rh, bio);
608 if (log->type->is_remote_recovering &&
609 log->type->is_remote_recovering(log, region)) {
610 bio_list_add(&requeue, bio);
614 state = dm_rh_get_state(ms->rh, region, 1);
625 case DM_RH_RECOVERING:
626 this_list = &recover;
630 bio_list_add(this_list, bio);
634 * Add bios that are delayed due to remote recovery
635 * back on to the write queue
637 if (unlikely(requeue.head)) {
638 spin_lock_irq(&ms->lock);
639 bio_list_merge(&ms->writes, &requeue);
640 spin_unlock_irq(&ms->lock);
644 * Increment the pending counts for any regions that will
645 * be written to (writes to recover regions are going to
648 dm_rh_inc_pending(ms->rh, &sync);
649 dm_rh_inc_pending(ms->rh, &nosync);
650 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : 0;
655 if (unlikely(ms->log_failure)) {
656 spin_lock_irq(&ms->lock);
657 bio_list_merge(&ms->failures, &sync);
658 spin_unlock_irq(&ms->lock);
661 while ((bio = bio_list_pop(&sync)))
664 while ((bio = bio_list_pop(&recover)))
665 dm_rh_delay(ms->rh, bio);
667 while ((bio = bio_list_pop(&nosync))) {
668 map_bio(get_default_mirror(ms), bio);
669 generic_make_request(bio);
673 static void do_failures(struct mirror_set *ms, struct bio_list *failures)
680 if (!ms->log_failure) {
681 while ((bio = bio_list_pop(failures))) {
683 dm_rh_mark_nosync(ms->rh, bio, bio->bi_size, 0);
689 * If the log has failed, unattempted writes are being
690 * put on the failures list. We can't issue those writes
691 * until a log has been marked, so we must store them.
693 * If a 'noflush' suspend is in progress, we can requeue
694 * the I/O's to the core. This give userspace a chance
695 * to reconfigure the mirror, at which point the core
696 * will reissue the writes. If the 'noflush' flag is
697 * not set, we have no choice but to return errors.
699 * Some writes on the failures list may have been
700 * submitted before the log failure and represent a
701 * failure to write to one of the devices. It is ok
702 * for us to treat them the same and requeue them
705 if (dm_noflush_suspending(ms->ti)) {
706 while ((bio = bio_list_pop(failures)))
707 bio_endio(bio, DM_ENDIO_REQUEUE);
711 if (atomic_read(&ms->suspend)) {
712 while ((bio = bio_list_pop(failures)))
713 bio_endio(bio, -EIO);
717 spin_lock_irq(&ms->lock);
718 bio_list_merge(&ms->failures, failures);
719 spin_unlock_irq(&ms->lock);
724 static void trigger_event(struct work_struct *work)
726 struct mirror_set *ms =
727 container_of(work, struct mirror_set, trigger_event);
729 dm_table_event(ms->ti->table);
732 /*-----------------------------------------------------------------
734 *---------------------------------------------------------------*/
735 static void do_mirror(struct work_struct *work)
737 struct mirror_set *ms = container_of(work, struct mirror_set,
739 struct bio_list reads, writes, failures;
742 spin_lock_irqsave(&ms->lock, flags);
745 failures = ms->failures;
746 bio_list_init(&ms->reads);
747 bio_list_init(&ms->writes);
748 bio_list_init(&ms->failures);
749 spin_unlock_irqrestore(&ms->lock, flags);
751 dm_rh_update_states(ms->rh, errors_handled(ms));
753 do_reads(ms, &reads);
754 do_writes(ms, &writes);
755 do_failures(ms, &failures);
757 dm_table_unplug_all(ms->ti->table);
760 /*-----------------------------------------------------------------
762 *---------------------------------------------------------------*/
763 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
764 uint32_t region_size,
765 struct dm_target *ti,
766 struct dm_dirty_log *dl)
769 struct mirror_set *ms = NULL;
771 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
773 ms = kzalloc(len, GFP_KERNEL);
775 ti->error = "Cannot allocate mirror context";
779 spin_lock_init(&ms->lock);
782 ms->nr_mirrors = nr_mirrors;
783 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
786 atomic_set(&ms->suspend, 0);
787 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
789 ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
790 _dm_raid1_read_record_cache);
792 if (!ms->read_record_pool) {
793 ti->error = "Error creating mirror read_record_pool";
798 ms->io_client = dm_io_client_create(DM_IO_PAGES);
799 if (IS_ERR(ms->io_client)) {
800 ti->error = "Error creating dm_io client";
801 mempool_destroy(ms->read_record_pool);
806 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
807 wakeup_all_recovery_waiters,
808 ms->ti->begin, MAX_RECOVERY,
809 dl, region_size, ms->nr_regions);
810 if (IS_ERR(ms->rh)) {
811 ti->error = "Error creating dirty region hash";
812 dm_io_client_destroy(ms->io_client);
813 mempool_destroy(ms->read_record_pool);
821 static void free_context(struct mirror_set *ms, struct dm_target *ti,
825 dm_put_device(ti, ms->mirror[m].dev);
827 dm_io_client_destroy(ms->io_client);
828 dm_region_hash_destroy(ms->rh);
829 mempool_destroy(ms->read_record_pool);
833 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
834 unsigned int mirror, char **argv)
836 unsigned long long offset;
838 if (sscanf(argv[1], "%llu", &offset) != 1) {
839 ti->error = "Invalid offset";
843 if (dm_get_device(ti, argv[0], offset, ti->len,
844 dm_table_get_mode(ti->table),
845 &ms->mirror[mirror].dev)) {
846 ti->error = "Device lookup failure";
850 ms->mirror[mirror].ms = ms;
851 atomic_set(&(ms->mirror[mirror].error_count), 0);
852 ms->mirror[mirror].error_type = 0;
853 ms->mirror[mirror].offset = offset;
859 * Create dirty log: log_type #log_params <log_params>
861 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
862 unsigned argc, char **argv,
865 unsigned param_count;
866 struct dm_dirty_log *dl;
869 ti->error = "Insufficient mirror log arguments";
873 if (sscanf(argv[1], "%u", ¶m_count) != 1) {
874 ti->error = "Invalid mirror log argument count";
878 *args_used = 2 + param_count;
880 if (argc < *args_used) {
881 ti->error = "Insufficient mirror log arguments";
885 dl = dm_dirty_log_create(argv[0], ti, param_count, argv + 2);
887 ti->error = "Error creating mirror dirty log";
894 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
897 unsigned num_features;
898 struct dm_target *ti = ms->ti;
905 if (sscanf(argv[0], "%u", &num_features) != 1) {
906 ti->error = "Invalid number of features";
914 if (num_features > argc) {
915 ti->error = "Not enough arguments to support feature count";
919 if (!strcmp("handle_errors", argv[0]))
920 ms->features |= DM_RAID1_HANDLE_ERRORS;
922 ti->error = "Unrecognised feature requested";
932 * Construct a mirror mapping:
934 * log_type #log_params <log_params>
935 * #mirrors [mirror_path offset]{2,}
936 * [#features <features>]
938 * log_type is "core" or "disk"
939 * #log_params is between 1 and 3
941 * If present, features must be "handle_errors".
943 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
946 unsigned int nr_mirrors, m, args_used;
947 struct mirror_set *ms;
948 struct dm_dirty_log *dl;
950 dl = create_dirty_log(ti, argc, argv, &args_used);
957 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
958 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
959 ti->error = "Invalid number of mirrors";
960 dm_dirty_log_destroy(dl);
966 if (argc < nr_mirrors * 2) {
967 ti->error = "Too few mirror arguments";
968 dm_dirty_log_destroy(dl);
972 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
974 dm_dirty_log_destroy(dl);
978 /* Get the mirror parameter sets */
979 for (m = 0; m < nr_mirrors; m++) {
980 r = get_mirror(ms, ti, m, argv);
982 free_context(ms, ti, m);
990 ti->split_io = dm_rh_get_region_size(ms->rh);
992 ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
993 if (!ms->kmirrord_wq) {
994 DMERR("couldn't start kmirrord");
996 goto err_free_context;
998 INIT_WORK(&ms->kmirrord_work, do_mirror);
999 init_timer(&ms->timer);
1000 ms->timer_pending = 0;
1001 INIT_WORK(&ms->trigger_event, trigger_event);
1003 r = parse_features(ms, argc, argv, &args_used);
1005 goto err_destroy_wq;
1011 * Any read-balancing addition depends on the
1012 * DM_RAID1_HANDLE_ERRORS flag being present.
1013 * This is because the decision to balance depends
1014 * on the sync state of a region. If the above
1015 * flag is not present, we ignore errors; and
1016 * the sync state may be inaccurate.
1020 ti->error = "Too many mirror arguments";
1022 goto err_destroy_wq;
1025 r = dm_kcopyd_client_create(DM_KCOPYD_PAGES, &ms->kcopyd_client);
1027 goto err_destroy_wq;
1033 destroy_workqueue(ms->kmirrord_wq);
1035 free_context(ms, ti, ms->nr_mirrors);
1039 static void mirror_dtr(struct dm_target *ti)
1041 struct mirror_set *ms = (struct mirror_set *) ti->private;
1043 del_timer_sync(&ms->timer);
1044 flush_workqueue(ms->kmirrord_wq);
1045 flush_scheduled_work();
1046 dm_kcopyd_client_destroy(ms->kcopyd_client);
1047 destroy_workqueue(ms->kmirrord_wq);
1048 free_context(ms, ti, ms->nr_mirrors);
1052 * Mirror mapping function
1054 static int mirror_map(struct dm_target *ti, struct bio *bio,
1055 union map_info *map_context)
1057 int r, rw = bio_rw(bio);
1059 struct mirror_set *ms = ti->private;
1060 struct dm_raid1_read_record *read_record = NULL;
1061 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1064 /* Save region for mirror_end_io() handler */
1065 map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1066 queue_bio(ms, bio, rw);
1067 return DM_MAPIO_SUBMITTED;
1070 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1071 if (r < 0 && r != -EWOULDBLOCK)
1075 * If region is not in-sync queue the bio.
1077 if (!r || (r == -EWOULDBLOCK)) {
1079 return -EWOULDBLOCK;
1081 queue_bio(ms, bio, rw);
1082 return DM_MAPIO_SUBMITTED;
1086 * The region is in-sync and we can perform reads directly.
1087 * Store enough information so we can retry if it fails.
1089 m = choose_mirror(ms, bio->bi_sector);
1093 read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1094 if (likely(read_record)) {
1095 dm_bio_record(&read_record->details, bio);
1096 map_context->ptr = read_record;
1102 return DM_MAPIO_REMAPPED;
1105 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1106 int error, union map_info *map_context)
1108 int rw = bio_rw(bio);
1109 struct mirror_set *ms = (struct mirror_set *) ti->private;
1110 struct mirror *m = NULL;
1111 struct dm_bio_details *bd = NULL;
1112 struct dm_raid1_read_record *read_record = map_context->ptr;
1115 * We need to dec pending if this was a write.
1118 dm_rh_dec(ms->rh, map_context->ll);
1122 if (error == -EOPNOTSUPP)
1125 if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1128 if (unlikely(error)) {
1131 * There wasn't enough memory to record necessary
1132 * information for a retry or there was no other
1135 DMERR_LIMIT("Mirror read failed.");
1141 DMERR("Mirror read failed from %s. Trying alternative device.",
1144 fail_mirror(m, DM_RAID1_READ_ERROR);
1147 * A failed read is requeued for another attempt using an intact
1150 if (default_ok(m) || mirror_available(ms, bio)) {
1151 bd = &read_record->details;
1153 dm_bio_restore(bd, bio);
1154 mempool_free(read_record, ms->read_record_pool);
1155 map_context->ptr = NULL;
1156 queue_bio(ms, bio, rw);
1159 DMERR("All replicated volumes dead, failing I/O");
1164 mempool_free(read_record, ms->read_record_pool);
1165 map_context->ptr = NULL;
1171 static void mirror_presuspend(struct dm_target *ti)
1173 struct mirror_set *ms = (struct mirror_set *) ti->private;
1174 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1176 atomic_set(&ms->suspend, 1);
1179 * We must finish up all the work that we've
1180 * generated (i.e. recovery work).
1182 dm_rh_stop_recovery(ms->rh);
1184 wait_event(_kmirrord_recovery_stopped,
1185 !dm_rh_recovery_in_flight(ms->rh));
1187 if (log->type->presuspend && log->type->presuspend(log))
1188 /* FIXME: need better error handling */
1189 DMWARN("log presuspend failed");
1192 * Now that recovery is complete/stopped and the
1193 * delayed bios are queued, we need to wait for
1194 * the worker thread to complete. This way,
1195 * we know that all of our I/O has been pushed.
1197 flush_workqueue(ms->kmirrord_wq);
1200 static void mirror_postsuspend(struct dm_target *ti)
1202 struct mirror_set *ms = ti->private;
1203 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1205 if (log->type->postsuspend && log->type->postsuspend(log))
1206 /* FIXME: need better error handling */
1207 DMWARN("log postsuspend failed");
1210 static void mirror_resume(struct dm_target *ti)
1212 struct mirror_set *ms = ti->private;
1213 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1215 atomic_set(&ms->suspend, 0);
1216 if (log->type->resume && log->type->resume(log))
1217 /* FIXME: need better error handling */
1218 DMWARN("log resume failed");
1219 dm_rh_start_recovery(ms->rh);
1223 * device_status_char
1224 * @m: mirror device/leg we want the status of
1226 * We return one character representing the most severe error
1227 * we have encountered.
1228 * A => Alive - No failures
1229 * D => Dead - A write failure occurred leaving mirror out-of-sync
1230 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1231 * R => Read - A read failure occurred, mirror data unaffected
1235 static char device_status_char(struct mirror *m)
1237 if (!atomic_read(&(m->error_count)))
1240 return (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1241 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1242 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1246 static int mirror_status(struct dm_target *ti, status_type_t type,
1247 char *result, unsigned int maxlen)
1249 unsigned int m, sz = 0;
1250 struct mirror_set *ms = (struct mirror_set *) ti->private;
1251 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1252 char buffer[ms->nr_mirrors + 1];
1255 case STATUSTYPE_INFO:
1256 DMEMIT("%d ", ms->nr_mirrors);
1257 for (m = 0; m < ms->nr_mirrors; m++) {
1258 DMEMIT("%s ", ms->mirror[m].dev->name);
1259 buffer[m] = device_status_char(&(ms->mirror[m]));
1263 DMEMIT("%llu/%llu 1 %s ",
1264 (unsigned long long)log->type->get_sync_count(log),
1265 (unsigned long long)ms->nr_regions, buffer);
1267 sz += log->type->status(log, type, result+sz, maxlen-sz);
1271 case STATUSTYPE_TABLE:
1272 sz = log->type->status(log, type, result, maxlen);
1274 DMEMIT("%d", ms->nr_mirrors);
1275 for (m = 0; m < ms->nr_mirrors; m++)
1276 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1277 (unsigned long long)ms->mirror[m].offset);
1279 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1280 DMEMIT(" 1 handle_errors");
1286 static struct target_type mirror_target = {
1288 .version = {1, 0, 20},
1289 .module = THIS_MODULE,
1293 .end_io = mirror_end_io,
1294 .presuspend = mirror_presuspend,
1295 .postsuspend = mirror_postsuspend,
1296 .resume = mirror_resume,
1297 .status = mirror_status,
1300 static int __init dm_mirror_init(void)
1304 _dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1305 if (!_dm_raid1_read_record_cache) {
1306 DMERR("Can't allocate dm_raid1_read_record cache");
1311 r = dm_register_target(&mirror_target);
1313 DMERR("Failed to register mirror target");
1320 kmem_cache_destroy(_dm_raid1_read_record_cache);
1325 static void __exit dm_mirror_exit(void)
1327 dm_unregister_target(&mirror_target);
1328 kmem_cache_destroy(_dm_raid1_read_record_cache);
1332 module_init(dm_mirror_init);
1333 module_exit(dm_mirror_exit);
1335 MODULE_DESCRIPTION(DM_NAME " mirror target");
1336 MODULE_AUTHOR("Joe Thornber");
1337 MODULE_LICENSE("GPL");