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 *---------------------------------------------------------------*/
44 struct mirror_set *ms;
46 unsigned long error_type;
53 struct list_head list;
57 spinlock_t lock; /* protects the lists */
58 struct bio_list reads;
59 struct bio_list writes;
60 struct bio_list failures;
61 struct bio_list holds; /* bios are waiting until suspend */
63 struct dm_region_hash *rh;
64 struct dm_kcopyd_client *kcopyd_client;
65 struct dm_io_client *io_client;
66 mempool_t *read_record_pool;
74 atomic_t default_mirror; /* Default mirror */
76 struct workqueue_struct *kmirrord_wq;
77 struct work_struct kmirrord_work;
78 struct timer_list timer;
79 unsigned long timer_pending;
81 struct work_struct trigger_event;
84 struct mirror mirror[0];
87 static void wakeup_mirrord(void *context)
89 struct mirror_set *ms = context;
91 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
94 static void delayed_wake_fn(unsigned long data)
96 struct mirror_set *ms = (struct mirror_set *) data;
98 clear_bit(0, &ms->timer_pending);
102 static void delayed_wake(struct mirror_set *ms)
104 if (test_and_set_bit(0, &ms->timer_pending))
107 ms->timer.expires = jiffies + HZ / 5;
108 ms->timer.data = (unsigned long) ms;
109 ms->timer.function = delayed_wake_fn;
110 add_timer(&ms->timer);
113 static void wakeup_all_recovery_waiters(void *context)
115 wake_up_all(&_kmirrord_recovery_stopped);
118 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
124 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
125 spin_lock_irqsave(&ms->lock, flags);
126 should_wake = !(bl->head);
127 bio_list_add(bl, bio);
128 spin_unlock_irqrestore(&ms->lock, flags);
134 static void dispatch_bios(void *context, struct bio_list *bio_list)
136 struct mirror_set *ms = context;
139 while ((bio = bio_list_pop(bio_list)))
140 queue_bio(ms, bio, WRITE);
143 #define MIN_READ_RECORDS 20
144 struct dm_raid1_read_record {
146 struct dm_bio_details details;
149 static struct kmem_cache *_dm_raid1_read_record_cache;
152 * Every mirror should look like this one.
154 #define DEFAULT_MIRROR 0
157 * This is yucky. We squirrel the mirror struct away inside
158 * bi_next for read/write buffers. This is safe since the bh
159 * doesn't get submitted to the lower levels of block layer.
161 static struct mirror *bio_get_m(struct bio *bio)
163 return (struct mirror *) bio->bi_next;
166 static void bio_set_m(struct bio *bio, struct mirror *m)
168 bio->bi_next = (struct bio *) m;
171 static struct mirror *get_default_mirror(struct mirror_set *ms)
173 return &ms->mirror[atomic_read(&ms->default_mirror)];
176 static void set_default_mirror(struct mirror *m)
178 struct mirror_set *ms = m->ms;
179 struct mirror *m0 = &(ms->mirror[0]);
181 atomic_set(&ms->default_mirror, m - m0);
185 * @m: mirror device to fail
186 * @error_type: one of the enum's, DM_RAID1_*_ERROR
188 * If errors are being handled, record the type of
189 * error encountered for this device. If this type
190 * of error has already been recorded, we can return;
191 * otherwise, we must signal userspace by triggering
192 * an event. Additionally, if the device is the
193 * primary device, we must choose a new primary, but
194 * only if the mirror is in-sync.
196 * This function must not block.
198 static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
200 struct mirror_set *ms = m->ms;
204 * error_count is used for nothing more than a
205 * simple way to tell if a device has encountered
208 atomic_inc(&m->error_count);
210 if (test_and_set_bit(error_type, &m->error_type))
213 if (!errors_handled(ms))
216 if (m != get_default_mirror(ms))
221 * Better to issue requests to same failing device
222 * than to risk returning corrupt data.
224 DMERR("Primary mirror (%s) failed while out-of-sync: "
225 "Reads may fail.", m->dev->name);
229 for (new = ms->mirror; new < ms->mirror + ms->nr_mirrors; new++)
230 if (!atomic_read(&new->error_count)) {
231 set_default_mirror(new);
235 if (unlikely(new == ms->mirror + ms->nr_mirrors))
236 DMWARN("All sides of mirror have failed.");
239 schedule_work(&ms->trigger_event);
242 static int mirror_flush(struct dm_target *ti)
244 struct mirror_set *ms = ti->private;
245 unsigned long error_bits;
248 struct dm_io_region io[ms->nr_mirrors];
250 struct dm_io_request io_req = {
251 .bi_rw = WRITE_BARRIER,
252 .mem.type = DM_IO_KMEM,
253 .mem.ptr.bvec = NULL,
254 .client = ms->io_client,
257 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
258 io[i].bdev = m->dev->bdev;
264 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
265 if (unlikely(error_bits != 0)) {
266 for (i = 0; i < ms->nr_mirrors; i++)
267 if (test_bit(i, &error_bits))
268 fail_mirror(ms->mirror + i,
269 DM_RAID1_FLUSH_ERROR);
276 /*-----------------------------------------------------------------
279 * When a mirror is first activated we may find that some regions
280 * are in the no-sync state. We have to recover these by
281 * recopying from the default mirror to all the others.
282 *---------------------------------------------------------------*/
283 static void recovery_complete(int read_err, unsigned long write_err,
286 struct dm_region *reg = context;
287 struct mirror_set *ms = dm_rh_region_context(reg);
291 /* Read error means the failure of default mirror. */
292 DMERR_LIMIT("Unable to read primary mirror during recovery");
293 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
297 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
300 * Bits correspond to devices (excluding default mirror).
301 * The default mirror cannot change during recovery.
303 for (m = 0; m < ms->nr_mirrors; m++) {
304 if (&ms->mirror[m] == get_default_mirror(ms))
306 if (test_bit(bit, &write_err))
307 fail_mirror(ms->mirror + m,
308 DM_RAID1_SYNC_ERROR);
313 dm_rh_recovery_end(reg, !(read_err || write_err));
316 static int recover(struct mirror_set *ms, struct dm_region *reg)
320 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
322 unsigned long flags = 0;
323 region_t key = dm_rh_get_region_key(reg);
324 sector_t region_size = dm_rh_get_region_size(ms->rh);
326 /* fill in the source */
327 m = get_default_mirror(ms);
328 from.bdev = m->dev->bdev;
329 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
330 if (key == (ms->nr_regions - 1)) {
332 * The final region may be smaller than
335 from.count = ms->ti->len & (region_size - 1);
337 from.count = region_size;
339 from.count = region_size;
341 /* fill in the destinations */
342 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
343 if (&ms->mirror[i] == get_default_mirror(ms))
347 dest->bdev = m->dev->bdev;
348 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
349 dest->count = from.count;
354 if (!errors_handled(ms))
355 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
357 r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
358 flags, recovery_complete, reg);
363 static void do_recovery(struct mirror_set *ms)
365 struct dm_region *reg;
366 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
370 * Start quiescing some regions.
372 dm_rh_recovery_prepare(ms->rh);
375 * Copy any already quiesced regions.
377 while ((reg = dm_rh_recovery_start(ms->rh))) {
378 r = recover(ms, reg);
380 dm_rh_recovery_end(reg, 0);
384 * Update the in sync flag.
387 (log->type->get_sync_count(log) == ms->nr_regions)) {
388 /* the sync is complete */
389 dm_table_event(ms->ti->table);
394 /*-----------------------------------------------------------------
396 *---------------------------------------------------------------*/
397 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
399 struct mirror *m = get_default_mirror(ms);
402 if (likely(!atomic_read(&m->error_count)))
405 if (m-- == ms->mirror)
407 } while (m != get_default_mirror(ms));
412 static int default_ok(struct mirror *m)
414 struct mirror *default_mirror = get_default_mirror(m->ms);
416 return !atomic_read(&default_mirror->error_count);
419 static int mirror_available(struct mirror_set *ms, struct bio *bio)
421 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
422 region_t region = dm_rh_bio_to_region(ms->rh, bio);
424 if (log->type->in_sync(log, region, 0))
425 return choose_mirror(ms, bio->bi_sector) ? 1 : 0;
431 * remap a buffer to a particular mirror.
433 static sector_t map_sector(struct mirror *m, struct bio *bio)
435 if (unlikely(!bio->bi_size))
437 return m->offset + (bio->bi_sector - m->ms->ti->begin);
440 static void map_bio(struct mirror *m, struct bio *bio)
442 bio->bi_bdev = m->dev->bdev;
443 bio->bi_sector = map_sector(m, bio);
446 static void map_region(struct dm_io_region *io, struct mirror *m,
449 io->bdev = m->dev->bdev;
450 io->sector = map_sector(m, bio);
451 io->count = bio->bi_size >> 9;
454 static void hold_bio(struct mirror_set *ms, struct bio *bio)
457 * If device is suspended, complete the bio.
459 if (atomic_read(&ms->suspend)) {
460 if (dm_noflush_suspending(ms->ti))
461 bio_endio(bio, DM_ENDIO_REQUEUE);
463 bio_endio(bio, -EIO);
468 * Hold bio until the suspend is complete.
470 spin_lock_irq(&ms->lock);
471 bio_list_add(&ms->holds, bio);
472 spin_unlock_irq(&ms->lock);
475 /*-----------------------------------------------------------------
477 *---------------------------------------------------------------*/
478 static void read_callback(unsigned long error, void *context)
480 struct bio *bio = context;
484 bio_set_m(bio, NULL);
486 if (likely(!error)) {
491 fail_mirror(m, DM_RAID1_READ_ERROR);
493 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
494 DMWARN_LIMIT("Read failure on mirror device %s. "
495 "Trying alternative device.",
497 queue_bio(m->ms, bio, bio_rw(bio));
501 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
503 bio_endio(bio, -EIO);
506 /* Asynchronous read. */
507 static void read_async_bio(struct mirror *m, struct bio *bio)
509 struct dm_io_region io;
510 struct dm_io_request io_req = {
512 .mem.type = DM_IO_BVEC,
513 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
514 .notify.fn = read_callback,
515 .notify.context = bio,
516 .client = m->ms->io_client,
519 map_region(&io, m, bio);
521 BUG_ON(dm_io(&io_req, 1, &io, NULL));
524 static inline int region_in_sync(struct mirror_set *ms, region_t region,
527 int state = dm_rh_get_state(ms->rh, region, may_block);
528 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
531 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
537 while ((bio = bio_list_pop(reads))) {
538 region = dm_rh_bio_to_region(ms->rh, bio);
539 m = get_default_mirror(ms);
542 * We can only read balance if the region is in sync.
544 if (likely(region_in_sync(ms, region, 1)))
545 m = choose_mirror(ms, bio->bi_sector);
546 else if (m && atomic_read(&m->error_count))
550 read_async_bio(m, bio);
552 bio_endio(bio, -EIO);
556 /*-----------------------------------------------------------------
559 * We do different things with the write io depending on the
560 * state of the region that it's in:
562 * SYNC: increment pending, use kcopyd to write to *all* mirrors
563 * RECOVERING: delay the io until recovery completes
564 * NOSYNC: increment pending, just write to the default mirror
565 *---------------------------------------------------------------*/
568 static void write_callback(unsigned long error, void *context)
571 struct bio *bio = (struct bio *) context;
572 struct mirror_set *ms;
577 ms = bio_get_m(bio)->ms;
578 bio_set_m(bio, NULL);
581 * NOTE: We don't decrement the pending count here,
582 * instead it is done by the targets endio function.
583 * This way we handle both writes to SYNC and NOSYNC
584 * regions with the same code.
589 for (i = 0; i < ms->nr_mirrors; i++)
590 if (test_bit(i, &error))
591 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
595 if (unlikely(!uptodate)) {
596 DMERR("All replicated volumes dead, failing I/O");
597 /* None of the writes succeeded, fail the I/O. */
599 } else if (errors_handled(ms)) {
601 * Need to raise event. Since raising
602 * events can block, we need to do it in
605 spin_lock_irqsave(&ms->lock, flags);
606 if (!ms->failures.head)
608 bio_list_add(&ms->failures, bio);
609 spin_unlock_irqrestore(&ms->lock, flags);
618 static void do_write(struct mirror_set *ms, struct bio *bio)
621 struct dm_io_region io[ms->nr_mirrors], *dest = io;
623 struct dm_io_request io_req = {
624 .bi_rw = WRITE | (bio->bi_rw & WRITE_BARRIER),
625 .mem.type = DM_IO_BVEC,
626 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
627 .notify.fn = write_callback,
628 .notify.context = bio,
629 .client = ms->io_client,
632 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
633 map_region(dest++, m, bio);
636 * Use default mirror because we only need it to retrieve the reference
637 * to the mirror set in write_callback().
639 bio_set_m(bio, get_default_mirror(ms));
641 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
644 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
648 struct bio_list sync, nosync, recover, *this_list = NULL;
649 struct bio_list requeue;
650 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
657 * Classify each write.
659 bio_list_init(&sync);
660 bio_list_init(&nosync);
661 bio_list_init(&recover);
662 bio_list_init(&requeue);
664 while ((bio = bio_list_pop(writes))) {
665 if (unlikely(bio_empty_barrier(bio))) {
666 bio_list_add(&sync, bio);
670 region = dm_rh_bio_to_region(ms->rh, bio);
672 if (log->type->is_remote_recovering &&
673 log->type->is_remote_recovering(log, region)) {
674 bio_list_add(&requeue, bio);
678 state = dm_rh_get_state(ms->rh, region, 1);
689 case DM_RH_RECOVERING:
690 this_list = &recover;
694 bio_list_add(this_list, bio);
698 * Add bios that are delayed due to remote recovery
699 * back on to the write queue
701 if (unlikely(requeue.head)) {
702 spin_lock_irq(&ms->lock);
703 bio_list_merge(&ms->writes, &requeue);
704 spin_unlock_irq(&ms->lock);
709 * Increment the pending counts for any regions that will
710 * be written to (writes to recover regions are going to
713 dm_rh_inc_pending(ms->rh, &sync);
714 dm_rh_inc_pending(ms->rh, &nosync);
717 * If the flush fails on a previous call and succeeds here,
718 * we must not reset the log_failure variable. We need
719 * userspace interaction to do that.
721 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
726 if (unlikely(ms->log_failure)) {
727 spin_lock_irq(&ms->lock);
728 bio_list_merge(&ms->failures, &sync);
729 spin_unlock_irq(&ms->lock);
732 while ((bio = bio_list_pop(&sync)))
735 while ((bio = bio_list_pop(&recover)))
736 dm_rh_delay(ms->rh, bio);
738 while ((bio = bio_list_pop(&nosync))) {
739 map_bio(get_default_mirror(ms), bio);
740 generic_make_request(bio);
744 static void do_failures(struct mirror_set *ms, struct bio_list *failures)
751 if (!ms->log_failure) {
752 while ((bio = bio_list_pop(failures))) {
754 dm_rh_mark_nosync(ms->rh, bio, bio->bi_size, 0);
760 * If the log has failed, unattempted writes are being
761 * put on the failures list. We can't issue those writes
762 * until a log has been marked, so we must store them.
764 * If a 'noflush' suspend is in progress, we can requeue
765 * the I/O's to the core. This give userspace a chance
766 * to reconfigure the mirror, at which point the core
767 * will reissue the writes. If the 'noflush' flag is
768 * not set, we have no choice but to return errors.
770 * Some writes on the failures list may have been
771 * submitted before the log failure and represent a
772 * failure to write to one of the devices. It is ok
773 * for us to treat them the same and requeue them
776 if (dm_noflush_suspending(ms->ti)) {
777 while ((bio = bio_list_pop(failures)))
778 bio_endio(bio, DM_ENDIO_REQUEUE);
782 if (atomic_read(&ms->suspend)) {
783 while ((bio = bio_list_pop(failures)))
784 bio_endio(bio, -EIO);
788 spin_lock_irq(&ms->lock);
789 bio_list_merge(&ms->failures, failures);
790 spin_unlock_irq(&ms->lock);
795 static void trigger_event(struct work_struct *work)
797 struct mirror_set *ms =
798 container_of(work, struct mirror_set, trigger_event);
800 dm_table_event(ms->ti->table);
803 /*-----------------------------------------------------------------
805 *---------------------------------------------------------------*/
806 static void do_mirror(struct work_struct *work)
808 struct mirror_set *ms = container_of(work, struct mirror_set,
810 struct bio_list reads, writes, failures;
813 spin_lock_irqsave(&ms->lock, flags);
816 failures = ms->failures;
817 bio_list_init(&ms->reads);
818 bio_list_init(&ms->writes);
819 bio_list_init(&ms->failures);
820 spin_unlock_irqrestore(&ms->lock, flags);
822 dm_rh_update_states(ms->rh, errors_handled(ms));
824 do_reads(ms, &reads);
825 do_writes(ms, &writes);
826 do_failures(ms, &failures);
828 dm_table_unplug_all(ms->ti->table);
831 /*-----------------------------------------------------------------
833 *---------------------------------------------------------------*/
834 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
835 uint32_t region_size,
836 struct dm_target *ti,
837 struct dm_dirty_log *dl)
840 struct mirror_set *ms = NULL;
842 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
844 ms = kzalloc(len, GFP_KERNEL);
846 ti->error = "Cannot allocate mirror context";
850 spin_lock_init(&ms->lock);
853 ms->nr_mirrors = nr_mirrors;
854 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
857 atomic_set(&ms->suspend, 0);
858 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
860 ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
861 _dm_raid1_read_record_cache);
863 if (!ms->read_record_pool) {
864 ti->error = "Error creating mirror read_record_pool";
869 ms->io_client = dm_io_client_create(DM_IO_PAGES);
870 if (IS_ERR(ms->io_client)) {
871 ti->error = "Error creating dm_io client";
872 mempool_destroy(ms->read_record_pool);
877 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
878 wakeup_all_recovery_waiters,
879 ms->ti->begin, MAX_RECOVERY,
880 dl, region_size, ms->nr_regions);
881 if (IS_ERR(ms->rh)) {
882 ti->error = "Error creating dirty region hash";
883 dm_io_client_destroy(ms->io_client);
884 mempool_destroy(ms->read_record_pool);
892 static void free_context(struct mirror_set *ms, struct dm_target *ti,
896 dm_put_device(ti, ms->mirror[m].dev);
898 dm_io_client_destroy(ms->io_client);
899 dm_region_hash_destroy(ms->rh);
900 mempool_destroy(ms->read_record_pool);
904 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
905 unsigned int mirror, char **argv)
907 unsigned long long offset;
909 if (sscanf(argv[1], "%llu", &offset) != 1) {
910 ti->error = "Invalid offset";
914 if (dm_get_device(ti, argv[0], offset, ti->len,
915 dm_table_get_mode(ti->table),
916 &ms->mirror[mirror].dev)) {
917 ti->error = "Device lookup failure";
921 ms->mirror[mirror].ms = ms;
922 atomic_set(&(ms->mirror[mirror].error_count), 0);
923 ms->mirror[mirror].error_type = 0;
924 ms->mirror[mirror].offset = offset;
930 * Create dirty log: log_type #log_params <log_params>
932 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
933 unsigned argc, char **argv,
936 unsigned param_count;
937 struct dm_dirty_log *dl;
940 ti->error = "Insufficient mirror log arguments";
944 if (sscanf(argv[1], "%u", ¶m_count) != 1) {
945 ti->error = "Invalid mirror log argument count";
949 *args_used = 2 + param_count;
951 if (argc < *args_used) {
952 ti->error = "Insufficient mirror log arguments";
956 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
959 ti->error = "Error creating mirror dirty log";
966 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
969 unsigned num_features;
970 struct dm_target *ti = ms->ti;
977 if (sscanf(argv[0], "%u", &num_features) != 1) {
978 ti->error = "Invalid number of features";
986 if (num_features > argc) {
987 ti->error = "Not enough arguments to support feature count";
991 if (!strcmp("handle_errors", argv[0]))
992 ms->features |= DM_RAID1_HANDLE_ERRORS;
994 ti->error = "Unrecognised feature requested";
1004 * Construct a mirror mapping:
1006 * log_type #log_params <log_params>
1007 * #mirrors [mirror_path offset]{2,}
1008 * [#features <features>]
1010 * log_type is "core" or "disk"
1011 * #log_params is between 1 and 3
1013 * If present, features must be "handle_errors".
1015 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1018 unsigned int nr_mirrors, m, args_used;
1019 struct mirror_set *ms;
1020 struct dm_dirty_log *dl;
1022 dl = create_dirty_log(ti, argc, argv, &args_used);
1029 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1030 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
1031 ti->error = "Invalid number of mirrors";
1032 dm_dirty_log_destroy(dl);
1038 if (argc < nr_mirrors * 2) {
1039 ti->error = "Too few mirror arguments";
1040 dm_dirty_log_destroy(dl);
1044 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1046 dm_dirty_log_destroy(dl);
1050 /* Get the mirror parameter sets */
1051 for (m = 0; m < nr_mirrors; m++) {
1052 r = get_mirror(ms, ti, m, argv);
1054 free_context(ms, ti, m);
1062 ti->split_io = dm_rh_get_region_size(ms->rh);
1063 ti->num_flush_requests = 1;
1065 ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
1066 if (!ms->kmirrord_wq) {
1067 DMERR("couldn't start kmirrord");
1069 goto err_free_context;
1071 INIT_WORK(&ms->kmirrord_work, do_mirror);
1072 init_timer(&ms->timer);
1073 ms->timer_pending = 0;
1074 INIT_WORK(&ms->trigger_event, trigger_event);
1076 r = parse_features(ms, argc, argv, &args_used);
1078 goto err_destroy_wq;
1084 * Any read-balancing addition depends on the
1085 * DM_RAID1_HANDLE_ERRORS flag being present.
1086 * This is because the decision to balance depends
1087 * on the sync state of a region. If the above
1088 * flag is not present, we ignore errors; and
1089 * the sync state may be inaccurate.
1093 ti->error = "Too many mirror arguments";
1095 goto err_destroy_wq;
1098 r = dm_kcopyd_client_create(DM_KCOPYD_PAGES, &ms->kcopyd_client);
1100 goto err_destroy_wq;
1106 destroy_workqueue(ms->kmirrord_wq);
1108 free_context(ms, ti, ms->nr_mirrors);
1112 static void mirror_dtr(struct dm_target *ti)
1114 struct mirror_set *ms = (struct mirror_set *) ti->private;
1116 del_timer_sync(&ms->timer);
1117 flush_workqueue(ms->kmirrord_wq);
1118 flush_scheduled_work();
1119 dm_kcopyd_client_destroy(ms->kcopyd_client);
1120 destroy_workqueue(ms->kmirrord_wq);
1121 free_context(ms, ti, ms->nr_mirrors);
1125 * Mirror mapping function
1127 static int mirror_map(struct dm_target *ti, struct bio *bio,
1128 union map_info *map_context)
1130 int r, rw = bio_rw(bio);
1132 struct mirror_set *ms = ti->private;
1133 struct dm_raid1_read_record *read_record = NULL;
1134 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1137 /* Save region for mirror_end_io() handler */
1138 map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1139 queue_bio(ms, bio, rw);
1140 return DM_MAPIO_SUBMITTED;
1143 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1144 if (r < 0 && r != -EWOULDBLOCK)
1148 * If region is not in-sync queue the bio.
1150 if (!r || (r == -EWOULDBLOCK)) {
1152 return -EWOULDBLOCK;
1154 queue_bio(ms, bio, rw);
1155 return DM_MAPIO_SUBMITTED;
1159 * The region is in-sync and we can perform reads directly.
1160 * Store enough information so we can retry if it fails.
1162 m = choose_mirror(ms, bio->bi_sector);
1166 read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1167 if (likely(read_record)) {
1168 dm_bio_record(&read_record->details, bio);
1169 map_context->ptr = read_record;
1175 return DM_MAPIO_REMAPPED;
1178 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1179 int error, union map_info *map_context)
1181 int rw = bio_rw(bio);
1182 struct mirror_set *ms = (struct mirror_set *) ti->private;
1183 struct mirror *m = NULL;
1184 struct dm_bio_details *bd = NULL;
1185 struct dm_raid1_read_record *read_record = map_context->ptr;
1188 * We need to dec pending if this was a write.
1191 if (likely(!bio_empty_barrier(bio)))
1192 dm_rh_dec(ms->rh, map_context->ll);
1196 if (error == -EOPNOTSUPP)
1199 if ((error == -EWOULDBLOCK) && bio_rw_flagged(bio, BIO_RW_AHEAD))
1202 if (unlikely(error)) {
1205 * There wasn't enough memory to record necessary
1206 * information for a retry or there was no other
1209 DMERR_LIMIT("Mirror read failed.");
1215 DMERR("Mirror read failed from %s. Trying alternative device.",
1218 fail_mirror(m, DM_RAID1_READ_ERROR);
1221 * A failed read is requeued for another attempt using an intact
1224 if (default_ok(m) || mirror_available(ms, bio)) {
1225 bd = &read_record->details;
1227 dm_bio_restore(bd, bio);
1228 mempool_free(read_record, ms->read_record_pool);
1229 map_context->ptr = NULL;
1230 queue_bio(ms, bio, rw);
1233 DMERR("All replicated volumes dead, failing I/O");
1238 mempool_free(read_record, ms->read_record_pool);
1239 map_context->ptr = NULL;
1245 static void mirror_presuspend(struct dm_target *ti)
1247 struct mirror_set *ms = (struct mirror_set *) ti->private;
1248 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1250 struct bio_list holds;
1253 atomic_set(&ms->suspend, 1);
1256 * We must finish up all the work that we've
1257 * generated (i.e. recovery work).
1259 dm_rh_stop_recovery(ms->rh);
1261 wait_event(_kmirrord_recovery_stopped,
1262 !dm_rh_recovery_in_flight(ms->rh));
1264 if (log->type->presuspend && log->type->presuspend(log))
1265 /* FIXME: need better error handling */
1266 DMWARN("log presuspend failed");
1269 * Now that recovery is complete/stopped and the
1270 * delayed bios are queued, we need to wait for
1271 * the worker thread to complete. This way,
1272 * we know that all of our I/O has been pushed.
1274 flush_workqueue(ms->kmirrord_wq);
1277 * Now set ms->suspend is set and the workqueue flushed, no more
1278 * entries can be added to ms->hold list, so process it.
1280 * Bios can still arrive concurrently with or after this
1281 * presuspend function, but they cannot join the hold list
1282 * because ms->suspend is set.
1284 spin_lock_irq(&ms->lock);
1286 bio_list_init(&ms->holds);
1287 spin_unlock_irq(&ms->lock);
1289 while ((bio = bio_list_pop(&holds)))
1293 static void mirror_postsuspend(struct dm_target *ti)
1295 struct mirror_set *ms = ti->private;
1296 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1298 if (log->type->postsuspend && log->type->postsuspend(log))
1299 /* FIXME: need better error handling */
1300 DMWARN("log postsuspend failed");
1303 static void mirror_resume(struct dm_target *ti)
1305 struct mirror_set *ms = ti->private;
1306 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1308 atomic_set(&ms->suspend, 0);
1309 if (log->type->resume && log->type->resume(log))
1310 /* FIXME: need better error handling */
1311 DMWARN("log resume failed");
1312 dm_rh_start_recovery(ms->rh);
1316 * device_status_char
1317 * @m: mirror device/leg we want the status of
1319 * We return one character representing the most severe error
1320 * we have encountered.
1321 * A => Alive - No failures
1322 * D => Dead - A write failure occurred leaving mirror out-of-sync
1323 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1324 * R => Read - A read failure occurred, mirror data unaffected
1328 static char device_status_char(struct mirror *m)
1330 if (!atomic_read(&(m->error_count)))
1333 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1334 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1335 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1336 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1340 static int mirror_status(struct dm_target *ti, status_type_t type,
1341 char *result, unsigned int maxlen)
1343 unsigned int m, sz = 0;
1344 struct mirror_set *ms = (struct mirror_set *) ti->private;
1345 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1346 char buffer[ms->nr_mirrors + 1];
1349 case STATUSTYPE_INFO:
1350 DMEMIT("%d ", ms->nr_mirrors);
1351 for (m = 0; m < ms->nr_mirrors; m++) {
1352 DMEMIT("%s ", ms->mirror[m].dev->name);
1353 buffer[m] = device_status_char(&(ms->mirror[m]));
1357 DMEMIT("%llu/%llu 1 %s ",
1358 (unsigned long long)log->type->get_sync_count(log),
1359 (unsigned long long)ms->nr_regions, buffer);
1361 sz += log->type->status(log, type, result+sz, maxlen-sz);
1365 case STATUSTYPE_TABLE:
1366 sz = log->type->status(log, type, result, maxlen);
1368 DMEMIT("%d", ms->nr_mirrors);
1369 for (m = 0; m < ms->nr_mirrors; m++)
1370 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1371 (unsigned long long)ms->mirror[m].offset);
1373 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1374 DMEMIT(" 1 handle_errors");
1380 static int mirror_iterate_devices(struct dm_target *ti,
1381 iterate_devices_callout_fn fn, void *data)
1383 struct mirror_set *ms = ti->private;
1387 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1388 ret = fn(ti, ms->mirror[i].dev,
1389 ms->mirror[i].offset, ti->len, data);
1394 static struct target_type mirror_target = {
1396 .version = {1, 12, 0},
1397 .module = THIS_MODULE,
1401 .end_io = mirror_end_io,
1402 .presuspend = mirror_presuspend,
1403 .postsuspend = mirror_postsuspend,
1404 .resume = mirror_resume,
1405 .status = mirror_status,
1406 .iterate_devices = mirror_iterate_devices,
1409 static int __init dm_mirror_init(void)
1413 _dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1414 if (!_dm_raid1_read_record_cache) {
1415 DMERR("Can't allocate dm_raid1_read_record cache");
1420 r = dm_register_target(&mirror_target);
1422 DMERR("Failed to register mirror target");
1429 kmem_cache_destroy(_dm_raid1_read_record_cache);
1434 static void __exit dm_mirror_exit(void)
1436 dm_unregister_target(&mirror_target);
1437 kmem_cache_destroy(_dm_raid1_read_record_cache);
1441 module_init(dm_mirror_init);
1442 module_exit(dm_mirror_exit);
1444 MODULE_DESCRIPTION(DM_NAME " mirror target");
1445 MODULE_AUTHOR("Joe Thornber");
1446 MODULE_LICENSE("GPL");