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. */
26 #define DM_RAID1_HANDLE_ERRORS 0x01
27 #define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
29 static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
31 /*-----------------------------------------------------------------
32 * Mirror set structures.
33 *---------------------------------------------------------------*/
42 struct mirror_set *ms;
44 unsigned long error_type;
51 struct list_head list;
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 */
61 struct dm_region_hash *rh;
62 struct dm_kcopyd_client *kcopyd_client;
63 struct dm_io_client *io_client;
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 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(raid1_resync_throttle,
86 "A percentage of time allocated for raid resynchronization");
88 static void wakeup_mirrord(void *context)
90 struct mirror_set *ms = context;
92 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
95 static void delayed_wake_fn(unsigned long data)
97 struct mirror_set *ms = (struct mirror_set *) data;
99 clear_bit(0, &ms->timer_pending);
103 static void delayed_wake(struct mirror_set *ms)
105 if (test_and_set_bit(0, &ms->timer_pending))
108 ms->timer.expires = jiffies + HZ / 5;
109 ms->timer.data = (unsigned long) ms;
110 ms->timer.function = delayed_wake_fn;
111 add_timer(&ms->timer);
114 static void wakeup_all_recovery_waiters(void *context)
116 wake_up_all(&_kmirrord_recovery_stopped);
119 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
125 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
126 spin_lock_irqsave(&ms->lock, flags);
127 should_wake = !(bl->head);
128 bio_list_add(bl, bio);
129 spin_unlock_irqrestore(&ms->lock, flags);
135 static void dispatch_bios(void *context, struct bio_list *bio_list)
137 struct mirror_set *ms = context;
140 while ((bio = bio_list_pop(bio_list)))
141 queue_bio(ms, bio, WRITE);
144 struct dm_raid1_bio_record {
146 /* if details->bi_bdev == NULL, details were not saved */
147 struct dm_bio_details details;
148 region_t write_region;
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);
184 static struct mirror *get_valid_mirror(struct mirror_set *ms)
188 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
189 if (!atomic_read(&m->error_count))
196 * @m: mirror device to fail
197 * @error_type: one of the enum's, DM_RAID1_*_ERROR
199 * If errors are being handled, record the type of
200 * error encountered for this device. If this type
201 * of error has already been recorded, we can return;
202 * otherwise, we must signal userspace by triggering
203 * an event. Additionally, if the device is the
204 * primary device, we must choose a new primary, but
205 * only if the mirror is in-sync.
207 * This function must not block.
209 static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
211 struct mirror_set *ms = m->ms;
217 * error_count is used for nothing more than a
218 * simple way to tell if a device has encountered
221 atomic_inc(&m->error_count);
223 if (test_and_set_bit(error_type, &m->error_type))
226 if (!errors_handled(ms))
229 if (m != get_default_mirror(ms))
234 * Better to issue requests to same failing device
235 * than to risk returning corrupt data.
237 DMERR("Primary mirror (%s) failed while out-of-sync: "
238 "Reads may fail.", m->dev->name);
242 new = get_valid_mirror(ms);
244 set_default_mirror(new);
246 DMWARN("All sides of mirror have failed.");
249 schedule_work(&ms->trigger_event);
252 static int mirror_flush(struct dm_target *ti)
254 struct mirror_set *ms = ti->private;
255 unsigned long error_bits;
258 struct dm_io_region io[ms->nr_mirrors];
260 struct dm_io_request io_req = {
261 .bi_rw = WRITE_FLUSH,
262 .mem.type = DM_IO_KMEM,
263 .mem.ptr.addr = NULL,
264 .client = ms->io_client,
267 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
268 io[i].bdev = m->dev->bdev;
274 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
275 if (unlikely(error_bits != 0)) {
276 for (i = 0; i < ms->nr_mirrors; i++)
277 if (test_bit(i, &error_bits))
278 fail_mirror(ms->mirror + i,
279 DM_RAID1_FLUSH_ERROR);
286 /*-----------------------------------------------------------------
289 * When a mirror is first activated we may find that some regions
290 * are in the no-sync state. We have to recover these by
291 * recopying from the default mirror to all the others.
292 *---------------------------------------------------------------*/
293 static void recovery_complete(int read_err, unsigned long write_err,
296 struct dm_region *reg = context;
297 struct mirror_set *ms = dm_rh_region_context(reg);
301 /* Read error means the failure of default mirror. */
302 DMERR_LIMIT("Unable to read primary mirror during recovery");
303 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
307 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
310 * Bits correspond to devices (excluding default mirror).
311 * The default mirror cannot change during recovery.
313 for (m = 0; m < ms->nr_mirrors; m++) {
314 if (&ms->mirror[m] == get_default_mirror(ms))
316 if (test_bit(bit, &write_err))
317 fail_mirror(ms->mirror + m,
318 DM_RAID1_SYNC_ERROR);
323 dm_rh_recovery_end(reg, !(read_err || write_err));
326 static int recover(struct mirror_set *ms, struct dm_region *reg)
330 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
332 unsigned long flags = 0;
333 region_t key = dm_rh_get_region_key(reg);
334 sector_t region_size = dm_rh_get_region_size(ms->rh);
336 /* fill in the source */
337 m = get_default_mirror(ms);
338 from.bdev = m->dev->bdev;
339 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
340 if (key == (ms->nr_regions - 1)) {
342 * The final region may be smaller than
345 from.count = ms->ti->len & (region_size - 1);
347 from.count = region_size;
349 from.count = region_size;
351 /* fill in the destinations */
352 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
353 if (&ms->mirror[i] == get_default_mirror(ms))
357 dest->bdev = m->dev->bdev;
358 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
359 dest->count = from.count;
364 if (!errors_handled(ms))
365 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
367 r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
368 flags, recovery_complete, reg);
373 static void do_recovery(struct mirror_set *ms)
375 struct dm_region *reg;
376 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
380 * Start quiescing some regions.
382 dm_rh_recovery_prepare(ms->rh);
385 * Copy any already quiesced regions.
387 while ((reg = dm_rh_recovery_start(ms->rh))) {
388 r = recover(ms, reg);
390 dm_rh_recovery_end(reg, 0);
394 * Update the in sync flag.
397 (log->type->get_sync_count(log) == ms->nr_regions)) {
398 /* the sync is complete */
399 dm_table_event(ms->ti->table);
404 /*-----------------------------------------------------------------
406 *---------------------------------------------------------------*/
407 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
409 struct mirror *m = get_default_mirror(ms);
412 if (likely(!atomic_read(&m->error_count)))
415 if (m-- == ms->mirror)
417 } while (m != get_default_mirror(ms));
422 static int default_ok(struct mirror *m)
424 struct mirror *default_mirror = get_default_mirror(m->ms);
426 return !atomic_read(&default_mirror->error_count);
429 static int mirror_available(struct mirror_set *ms, struct bio *bio)
431 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
432 region_t region = dm_rh_bio_to_region(ms->rh, bio);
434 if (log->type->in_sync(log, region, 0))
435 return choose_mirror(ms, bio->bi_iter.bi_sector) ? 1 : 0;
441 * remap a buffer to a particular mirror.
443 static sector_t map_sector(struct mirror *m, struct bio *bio)
445 if (unlikely(!bio->bi_iter.bi_size))
447 return m->offset + dm_target_offset(m->ms->ti, bio->bi_iter.bi_sector);
450 static void map_bio(struct mirror *m, struct bio *bio)
452 bio->bi_bdev = m->dev->bdev;
453 bio->bi_iter.bi_sector = map_sector(m, bio);
456 static void map_region(struct dm_io_region *io, struct mirror *m,
459 io->bdev = m->dev->bdev;
460 io->sector = map_sector(m, bio);
461 io->count = bio_sectors(bio);
464 static void hold_bio(struct mirror_set *ms, struct bio *bio)
467 * Lock is required to avoid race condition during suspend
470 spin_lock_irq(&ms->lock);
472 if (atomic_read(&ms->suspend)) {
473 spin_unlock_irq(&ms->lock);
476 * If device is suspended, complete the bio.
478 if (dm_noflush_suspending(ms->ti))
479 bio_endio(bio, DM_ENDIO_REQUEUE);
481 bio_endio(bio, -EIO);
486 * Hold bio until the suspend is complete.
488 bio_list_add(&ms->holds, bio);
489 spin_unlock_irq(&ms->lock);
492 /*-----------------------------------------------------------------
494 *---------------------------------------------------------------*/
495 static void read_callback(unsigned long error, void *context)
497 struct bio *bio = context;
501 bio_set_m(bio, NULL);
503 if (likely(!error)) {
508 fail_mirror(m, DM_RAID1_READ_ERROR);
510 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
511 DMWARN_LIMIT("Read failure on mirror device %s. "
512 "Trying alternative device.",
514 queue_bio(m->ms, bio, bio_rw(bio));
518 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
520 bio_endio(bio, -EIO);
523 /* Asynchronous read. */
524 static void read_async_bio(struct mirror *m, struct bio *bio)
526 struct dm_io_region io;
527 struct dm_io_request io_req = {
529 .mem.type = DM_IO_BIO,
531 .notify.fn = read_callback,
532 .notify.context = bio,
533 .client = m->ms->io_client,
536 map_region(&io, m, bio);
538 BUG_ON(dm_io(&io_req, 1, &io, NULL));
541 static inline int region_in_sync(struct mirror_set *ms, region_t region,
544 int state = dm_rh_get_state(ms->rh, region, may_block);
545 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
548 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
554 while ((bio = bio_list_pop(reads))) {
555 region = dm_rh_bio_to_region(ms->rh, bio);
556 m = get_default_mirror(ms);
559 * We can only read balance if the region is in sync.
561 if (likely(region_in_sync(ms, region, 1)))
562 m = choose_mirror(ms, bio->bi_iter.bi_sector);
563 else if (m && atomic_read(&m->error_count))
567 read_async_bio(m, bio);
569 bio_endio(bio, -EIO);
573 /*-----------------------------------------------------------------
576 * We do different things with the write io depending on the
577 * state of the region that it's in:
579 * SYNC: increment pending, use kcopyd to write to *all* mirrors
580 * RECOVERING: delay the io until recovery completes
581 * NOSYNC: increment pending, just write to the default mirror
582 *---------------------------------------------------------------*/
585 static void write_callback(unsigned long error, void *context)
588 struct bio *bio = (struct bio *) context;
589 struct mirror_set *ms;
593 ms = bio_get_m(bio)->ms;
594 bio_set_m(bio, NULL);
597 * NOTE: We don't decrement the pending count here,
598 * instead it is done by the targets endio function.
599 * This way we handle both writes to SYNC and NOSYNC
600 * regions with the same code.
602 if (likely(!error)) {
608 * If the bio is discard, return an error, but do not
611 if (bio->bi_rw & REQ_DISCARD) {
612 bio_endio(bio, -EOPNOTSUPP);
616 for (i = 0; i < ms->nr_mirrors; i++)
617 if (test_bit(i, &error))
618 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
621 * Need to raise event. Since raising
622 * events can block, we need to do it in
625 spin_lock_irqsave(&ms->lock, flags);
626 if (!ms->failures.head)
628 bio_list_add(&ms->failures, bio);
629 spin_unlock_irqrestore(&ms->lock, flags);
634 static void do_write(struct mirror_set *ms, struct bio *bio)
637 struct dm_io_region io[ms->nr_mirrors], *dest = io;
639 struct dm_io_request io_req = {
640 .bi_rw = WRITE | (bio->bi_rw & WRITE_FLUSH_FUA),
641 .mem.type = DM_IO_BIO,
643 .notify.fn = write_callback,
644 .notify.context = bio,
645 .client = ms->io_client,
648 if (bio->bi_rw & REQ_DISCARD) {
649 io_req.bi_rw |= REQ_DISCARD;
650 io_req.mem.type = DM_IO_KMEM;
651 io_req.mem.ptr.addr = NULL;
654 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
655 map_region(dest++, m, bio);
658 * Use default mirror because we only need it to retrieve the reference
659 * to the mirror set in write_callback().
661 bio_set_m(bio, get_default_mirror(ms));
663 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
666 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
670 struct bio_list sync, nosync, recover, *this_list = NULL;
671 struct bio_list requeue;
672 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
679 * Classify each write.
681 bio_list_init(&sync);
682 bio_list_init(&nosync);
683 bio_list_init(&recover);
684 bio_list_init(&requeue);
686 while ((bio = bio_list_pop(writes))) {
687 if ((bio->bi_rw & REQ_FLUSH) ||
688 (bio->bi_rw & REQ_DISCARD)) {
689 bio_list_add(&sync, bio);
693 region = dm_rh_bio_to_region(ms->rh, bio);
695 if (log->type->is_remote_recovering &&
696 log->type->is_remote_recovering(log, region)) {
697 bio_list_add(&requeue, bio);
701 state = dm_rh_get_state(ms->rh, region, 1);
712 case DM_RH_RECOVERING:
713 this_list = &recover;
717 bio_list_add(this_list, bio);
721 * Add bios that are delayed due to remote recovery
722 * back on to the write queue
724 if (unlikely(requeue.head)) {
725 spin_lock_irq(&ms->lock);
726 bio_list_merge(&ms->writes, &requeue);
727 spin_unlock_irq(&ms->lock);
732 * Increment the pending counts for any regions that will
733 * be written to (writes to recover regions are going to
736 dm_rh_inc_pending(ms->rh, &sync);
737 dm_rh_inc_pending(ms->rh, &nosync);
740 * If the flush fails on a previous call and succeeds here,
741 * we must not reset the log_failure variable. We need
742 * userspace interaction to do that.
744 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
749 if (unlikely(ms->log_failure) && errors_handled(ms)) {
750 spin_lock_irq(&ms->lock);
751 bio_list_merge(&ms->failures, &sync);
752 spin_unlock_irq(&ms->lock);
755 while ((bio = bio_list_pop(&sync)))
758 while ((bio = bio_list_pop(&recover)))
759 dm_rh_delay(ms->rh, bio);
761 while ((bio = bio_list_pop(&nosync))) {
762 if (unlikely(ms->leg_failure) && errors_handled(ms)) {
763 spin_lock_irq(&ms->lock);
764 bio_list_add(&ms->failures, bio);
765 spin_unlock_irq(&ms->lock);
768 map_bio(get_default_mirror(ms), bio);
769 generic_make_request(bio);
774 static void do_failures(struct mirror_set *ms, struct bio_list *failures)
778 if (likely(!failures->head))
782 * If the log has failed, unattempted writes are being
783 * put on the holds list. We can't issue those writes
784 * until a log has been marked, so we must store them.
786 * If a 'noflush' suspend is in progress, we can requeue
787 * the I/O's to the core. This give userspace a chance
788 * to reconfigure the mirror, at which point the core
789 * will reissue the writes. If the 'noflush' flag is
790 * not set, we have no choice but to return errors.
792 * Some writes on the failures list may have been
793 * submitted before the log failure and represent a
794 * failure to write to one of the devices. It is ok
795 * for us to treat them the same and requeue them
798 while ((bio = bio_list_pop(failures))) {
799 if (!ms->log_failure) {
801 dm_rh_mark_nosync(ms->rh, bio);
805 * If all the legs are dead, fail the I/O.
806 * If we have been told to handle errors, hold the bio
807 * and wait for userspace to deal with the problem.
808 * Otherwise pretend that the I/O succeeded. (This would
809 * be wrong if the failed leg returned after reboot and
810 * got replicated back to the good legs.)
812 if (!get_valid_mirror(ms))
813 bio_endio(bio, -EIO);
814 else if (errors_handled(ms))
821 static void trigger_event(struct work_struct *work)
823 struct mirror_set *ms =
824 container_of(work, struct mirror_set, trigger_event);
826 dm_table_event(ms->ti->table);
829 /*-----------------------------------------------------------------
831 *---------------------------------------------------------------*/
832 static void do_mirror(struct work_struct *work)
834 struct mirror_set *ms = container_of(work, struct mirror_set,
836 struct bio_list reads, writes, failures;
839 spin_lock_irqsave(&ms->lock, flags);
842 failures = ms->failures;
843 bio_list_init(&ms->reads);
844 bio_list_init(&ms->writes);
845 bio_list_init(&ms->failures);
846 spin_unlock_irqrestore(&ms->lock, flags);
848 dm_rh_update_states(ms->rh, errors_handled(ms));
850 do_reads(ms, &reads);
851 do_writes(ms, &writes);
852 do_failures(ms, &failures);
855 /*-----------------------------------------------------------------
857 *---------------------------------------------------------------*/
858 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
859 uint32_t region_size,
860 struct dm_target *ti,
861 struct dm_dirty_log *dl)
864 struct mirror_set *ms = NULL;
866 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
868 ms = kzalloc(len, GFP_KERNEL);
870 ti->error = "Cannot allocate mirror context";
874 spin_lock_init(&ms->lock);
875 bio_list_init(&ms->reads);
876 bio_list_init(&ms->writes);
877 bio_list_init(&ms->failures);
878 bio_list_init(&ms->holds);
881 ms->nr_mirrors = nr_mirrors;
882 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
886 atomic_set(&ms->suspend, 0);
887 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
889 ms->io_client = dm_io_client_create();
890 if (IS_ERR(ms->io_client)) {
891 ti->error = "Error creating dm_io client";
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);
910 static void free_context(struct mirror_set *ms, struct dm_target *ti,
914 dm_put_device(ti, ms->mirror[m].dev);
916 dm_io_client_destroy(ms->io_client);
917 dm_region_hash_destroy(ms->rh);
921 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
922 unsigned int mirror, char **argv)
924 unsigned long long offset;
927 if (sscanf(argv[1], "%llu%c", &offset, &dummy) != 1) {
928 ti->error = "Invalid offset";
932 if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
933 &ms->mirror[mirror].dev)) {
934 ti->error = "Device lookup failure";
938 ms->mirror[mirror].ms = ms;
939 atomic_set(&(ms->mirror[mirror].error_count), 0);
940 ms->mirror[mirror].error_type = 0;
941 ms->mirror[mirror].offset = offset;
947 * Create dirty log: log_type #log_params <log_params>
949 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
950 unsigned argc, char **argv,
953 unsigned param_count;
954 struct dm_dirty_log *dl;
958 ti->error = "Insufficient mirror log arguments";
962 if (sscanf(argv[1], "%u%c", ¶m_count, &dummy) != 1) {
963 ti->error = "Invalid mirror log argument count";
967 *args_used = 2 + param_count;
969 if (argc < *args_used) {
970 ti->error = "Insufficient mirror log arguments";
974 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
977 ti->error = "Error creating mirror dirty log";
984 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
987 unsigned num_features;
988 struct dm_target *ti = ms->ti;
996 if (sscanf(argv[0], "%u%c", &num_features, &dummy) != 1) {
997 ti->error = "Invalid number of features";
1005 if (num_features > argc) {
1006 ti->error = "Not enough arguments to support feature count";
1010 if (!strcmp("handle_errors", argv[0]))
1011 ms->features |= DM_RAID1_HANDLE_ERRORS;
1013 ti->error = "Unrecognised feature requested";
1023 * Construct a mirror mapping:
1025 * log_type #log_params <log_params>
1026 * #mirrors [mirror_path offset]{2,}
1027 * [#features <features>]
1029 * log_type is "core" or "disk"
1030 * #log_params is between 1 and 3
1032 * If present, features must be "handle_errors".
1034 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1037 unsigned int nr_mirrors, m, args_used;
1038 struct mirror_set *ms;
1039 struct dm_dirty_log *dl;
1042 dl = create_dirty_log(ti, argc, argv, &args_used);
1049 if (!argc || sscanf(argv[0], "%u%c", &nr_mirrors, &dummy) != 1 ||
1050 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
1051 ti->error = "Invalid number of mirrors";
1052 dm_dirty_log_destroy(dl);
1058 if (argc < nr_mirrors * 2) {
1059 ti->error = "Too few mirror arguments";
1060 dm_dirty_log_destroy(dl);
1064 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1066 dm_dirty_log_destroy(dl);
1070 /* Get the mirror parameter sets */
1071 for (m = 0; m < nr_mirrors; m++) {
1072 r = get_mirror(ms, ti, m, argv);
1074 free_context(ms, ti, m);
1083 r = dm_set_target_max_io_len(ti, dm_rh_get_region_size(ms->rh));
1085 goto err_free_context;
1087 ti->num_flush_bios = 1;
1088 ti->num_discard_bios = 1;
1089 ti->per_bio_data_size = sizeof(struct dm_raid1_bio_record);
1090 ti->discard_zeroes_data_unsupported = true;
1092 ms->kmirrord_wq = alloc_workqueue("kmirrord", WQ_MEM_RECLAIM, 0);
1093 if (!ms->kmirrord_wq) {
1094 DMERR("couldn't start kmirrord");
1096 goto err_free_context;
1098 INIT_WORK(&ms->kmirrord_work, do_mirror);
1099 init_timer(&ms->timer);
1100 ms->timer_pending = 0;
1101 INIT_WORK(&ms->trigger_event, trigger_event);
1103 r = parse_features(ms, argc, argv, &args_used);
1105 goto err_destroy_wq;
1111 * Any read-balancing addition depends on the
1112 * DM_RAID1_HANDLE_ERRORS flag being present.
1113 * This is because the decision to balance depends
1114 * on the sync state of a region. If the above
1115 * flag is not present, we ignore errors; and
1116 * the sync state may be inaccurate.
1120 ti->error = "Too many mirror arguments";
1122 goto err_destroy_wq;
1125 ms->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1126 if (IS_ERR(ms->kcopyd_client)) {
1127 r = PTR_ERR(ms->kcopyd_client);
1128 goto err_destroy_wq;
1135 destroy_workqueue(ms->kmirrord_wq);
1137 free_context(ms, ti, ms->nr_mirrors);
1141 static void mirror_dtr(struct dm_target *ti)
1143 struct mirror_set *ms = (struct mirror_set *) ti->private;
1145 del_timer_sync(&ms->timer);
1146 flush_workqueue(ms->kmirrord_wq);
1147 flush_work(&ms->trigger_event);
1148 dm_kcopyd_client_destroy(ms->kcopyd_client);
1149 destroy_workqueue(ms->kmirrord_wq);
1150 free_context(ms, ti, ms->nr_mirrors);
1154 * Mirror mapping function
1156 static int mirror_map(struct dm_target *ti, struct bio *bio)
1158 int r, rw = bio_rw(bio);
1160 struct mirror_set *ms = ti->private;
1161 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1162 struct dm_raid1_bio_record *bio_record =
1163 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1165 bio_record->details.bi_bdev = NULL;
1168 /* Save region for mirror_end_io() handler */
1169 bio_record->write_region = dm_rh_bio_to_region(ms->rh, bio);
1170 queue_bio(ms, bio, rw);
1171 return DM_MAPIO_SUBMITTED;
1174 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1175 if (r < 0 && r != -EWOULDBLOCK)
1179 * If region is not in-sync queue the bio.
1181 if (!r || (r == -EWOULDBLOCK)) {
1183 return -EWOULDBLOCK;
1185 queue_bio(ms, bio, rw);
1186 return DM_MAPIO_SUBMITTED;
1190 * The region is in-sync and we can perform reads directly.
1191 * Store enough information so we can retry if it fails.
1193 m = choose_mirror(ms, bio->bi_iter.bi_sector);
1197 dm_bio_record(&bio_record->details, bio);
1202 return DM_MAPIO_REMAPPED;
1205 static int mirror_end_io(struct dm_target *ti, struct bio *bio, int error)
1207 int rw = bio_rw(bio);
1208 struct mirror_set *ms = (struct mirror_set *) ti->private;
1209 struct mirror *m = NULL;
1210 struct dm_bio_details *bd = NULL;
1211 struct dm_raid1_bio_record *bio_record =
1212 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1215 * We need to dec pending if this was a write.
1218 if (!(bio->bi_rw & (REQ_FLUSH | REQ_DISCARD)))
1219 dm_rh_dec(ms->rh, bio_record->write_region);
1223 if (error == -EOPNOTSUPP)
1226 if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
1229 if (unlikely(error)) {
1230 if (!bio_record->details.bi_bdev) {
1232 * There wasn't enough memory to record necessary
1233 * information for a retry or there was no other
1236 DMERR_LIMIT("Mirror read failed.");
1242 DMERR("Mirror read failed from %s. Trying alternative device.",
1245 fail_mirror(m, DM_RAID1_READ_ERROR);
1248 * A failed read is requeued for another attempt using an intact
1251 if (default_ok(m) || mirror_available(ms, bio)) {
1252 bd = &bio_record->details;
1254 dm_bio_restore(bd, bio);
1255 bio_record->details.bi_bdev = NULL;
1257 atomic_inc(&bio->bi_remaining);
1259 queue_bio(ms, bio, rw);
1260 return DM_ENDIO_INCOMPLETE;
1262 DMERR("All replicated volumes dead, failing I/O");
1266 bio_record->details.bi_bdev = NULL;
1271 static void mirror_presuspend(struct dm_target *ti)
1273 struct mirror_set *ms = (struct mirror_set *) ti->private;
1274 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1276 struct bio_list holds;
1279 atomic_set(&ms->suspend, 1);
1282 * Process bios in the hold list to start recovery waiting
1283 * for bios in the hold list. After the process, no bio has
1284 * a chance to be added in the hold list because ms->suspend
1287 spin_lock_irq(&ms->lock);
1289 bio_list_init(&ms->holds);
1290 spin_unlock_irq(&ms->lock);
1292 while ((bio = bio_list_pop(&holds)))
1296 * We must finish up all the work that we've
1297 * generated (i.e. recovery work).
1299 dm_rh_stop_recovery(ms->rh);
1301 wait_event(_kmirrord_recovery_stopped,
1302 !dm_rh_recovery_in_flight(ms->rh));
1304 if (log->type->presuspend && log->type->presuspend(log))
1305 /* FIXME: need better error handling */
1306 DMWARN("log presuspend failed");
1309 * Now that recovery is complete/stopped and the
1310 * delayed bios are queued, we need to wait for
1311 * the worker thread to complete. This way,
1312 * we know that all of our I/O has been pushed.
1314 flush_workqueue(ms->kmirrord_wq);
1317 static void mirror_postsuspend(struct dm_target *ti)
1319 struct mirror_set *ms = ti->private;
1320 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1322 if (log->type->postsuspend && log->type->postsuspend(log))
1323 /* FIXME: need better error handling */
1324 DMWARN("log postsuspend failed");
1327 static void mirror_resume(struct dm_target *ti)
1329 struct mirror_set *ms = ti->private;
1330 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1332 atomic_set(&ms->suspend, 0);
1333 if (log->type->resume && log->type->resume(log))
1334 /* FIXME: need better error handling */
1335 DMWARN("log resume failed");
1336 dm_rh_start_recovery(ms->rh);
1340 * device_status_char
1341 * @m: mirror device/leg we want the status of
1343 * We return one character representing the most severe error
1344 * we have encountered.
1345 * A => Alive - No failures
1346 * D => Dead - A write failure occurred leaving mirror out-of-sync
1347 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1348 * R => Read - A read failure occurred, mirror data unaffected
1352 static char device_status_char(struct mirror *m)
1354 if (!atomic_read(&(m->error_count)))
1357 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1358 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1359 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1360 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1364 static void mirror_status(struct dm_target *ti, status_type_t type,
1365 unsigned status_flags, char *result, unsigned maxlen)
1367 unsigned int m, sz = 0;
1368 struct mirror_set *ms = (struct mirror_set *) ti->private;
1369 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1370 char buffer[ms->nr_mirrors + 1];
1373 case STATUSTYPE_INFO:
1374 DMEMIT("%d ", ms->nr_mirrors);
1375 for (m = 0; m < ms->nr_mirrors; m++) {
1376 DMEMIT("%s ", ms->mirror[m].dev->name);
1377 buffer[m] = device_status_char(&(ms->mirror[m]));
1381 DMEMIT("%llu/%llu 1 %s ",
1382 (unsigned long long)log->type->get_sync_count(log),
1383 (unsigned long long)ms->nr_regions, buffer);
1385 sz += log->type->status(log, type, result+sz, maxlen-sz);
1389 case STATUSTYPE_TABLE:
1390 sz = log->type->status(log, type, result, maxlen);
1392 DMEMIT("%d", ms->nr_mirrors);
1393 for (m = 0; m < ms->nr_mirrors; m++)
1394 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1395 (unsigned long long)ms->mirror[m].offset);
1397 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1398 DMEMIT(" 1 handle_errors");
1402 static int mirror_iterate_devices(struct dm_target *ti,
1403 iterate_devices_callout_fn fn, void *data)
1405 struct mirror_set *ms = ti->private;
1409 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1410 ret = fn(ti, ms->mirror[i].dev,
1411 ms->mirror[i].offset, ti->len, data);
1416 static struct target_type mirror_target = {
1418 .version = {1, 13, 2},
1419 .module = THIS_MODULE,
1423 .end_io = mirror_end_io,
1424 .presuspend = mirror_presuspend,
1425 .postsuspend = mirror_postsuspend,
1426 .resume = mirror_resume,
1427 .status = mirror_status,
1428 .iterate_devices = mirror_iterate_devices,
1431 static int __init dm_mirror_init(void)
1435 r = dm_register_target(&mirror_target);
1437 DMERR("Failed to register mirror target");
1447 static void __exit dm_mirror_exit(void)
1449 dm_unregister_target(&mirror_target);
1453 module_init(dm_mirror_init);
1454 module_exit(dm_mirror_exit);
1456 MODULE_DESCRIPTION(DM_NAME " mirror target");
1457 MODULE_AUTHOR("Joe Thornber");
1458 MODULE_LICENSE("GPL");