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;
75 atomic_t default_mirror; /* Default mirror */
77 struct workqueue_struct *kmirrord_wq;
78 struct work_struct kmirrord_work;
79 struct timer_list timer;
80 unsigned long timer_pending;
82 struct work_struct trigger_event;
85 struct mirror mirror[0];
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 #define MIN_READ_RECORDS 20
145 struct dm_raid1_read_record {
147 struct dm_bio_details details;
150 static struct kmem_cache *_dm_raid1_read_record_cache;
153 * Every mirror should look like this one.
155 #define DEFAULT_MIRROR 0
158 * This is yucky. We squirrel the mirror struct away inside
159 * bi_next for read/write buffers. This is safe since the bh
160 * doesn't get submitted to the lower levels of block layer.
162 static struct mirror *bio_get_m(struct bio *bio)
164 return (struct mirror *) bio->bi_next;
167 static void bio_set_m(struct bio *bio, struct mirror *m)
169 bio->bi_next = (struct bio *) m;
172 static struct mirror *get_default_mirror(struct mirror_set *ms)
174 return &ms->mirror[atomic_read(&ms->default_mirror)];
177 static void set_default_mirror(struct mirror *m)
179 struct mirror_set *ms = m->ms;
180 struct mirror *m0 = &(ms->mirror[0]);
182 atomic_set(&ms->default_mirror, m - m0);
185 static struct mirror *get_valid_mirror(struct mirror_set *ms)
189 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
190 if (!atomic_read(&m->error_count))
197 * @m: mirror device to fail
198 * @error_type: one of the enum's, DM_RAID1_*_ERROR
200 * If errors are being handled, record the type of
201 * error encountered for this device. If this type
202 * of error has already been recorded, we can return;
203 * otherwise, we must signal userspace by triggering
204 * an event. Additionally, if the device is the
205 * primary device, we must choose a new primary, but
206 * only if the mirror is in-sync.
208 * This function must not block.
210 static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
212 struct mirror_set *ms = m->ms;
218 * error_count is used for nothing more than a
219 * simple way to tell if a device has encountered
222 atomic_inc(&m->error_count);
224 if (test_and_set_bit(error_type, &m->error_type))
227 if (!errors_handled(ms))
230 if (m != get_default_mirror(ms))
235 * Better to issue requests to same failing device
236 * than to risk returning corrupt data.
238 DMERR("Primary mirror (%s) failed while out-of-sync: "
239 "Reads may fail.", m->dev->name);
243 new = get_valid_mirror(ms);
245 set_default_mirror(new);
247 DMWARN("All sides of mirror have failed.");
250 schedule_work(&ms->trigger_event);
253 static int mirror_flush(struct dm_target *ti)
255 struct mirror_set *ms = ti->private;
256 unsigned long error_bits;
259 struct dm_io_region io[ms->nr_mirrors];
261 struct dm_io_request io_req = {
262 .bi_rw = WRITE_BARRIER,
263 .mem.type = DM_IO_KMEM,
264 .mem.ptr.bvec = NULL,
265 .client = ms->io_client,
268 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
269 io[i].bdev = m->dev->bdev;
275 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
276 if (unlikely(error_bits != 0)) {
277 for (i = 0; i < ms->nr_mirrors; i++)
278 if (test_bit(i, &error_bits))
279 fail_mirror(ms->mirror + i,
280 DM_RAID1_FLUSH_ERROR);
287 /*-----------------------------------------------------------------
290 * When a mirror is first activated we may find that some regions
291 * are in the no-sync state. We have to recover these by
292 * recopying from the default mirror to all the others.
293 *---------------------------------------------------------------*/
294 static void recovery_complete(int read_err, unsigned long write_err,
297 struct dm_region *reg = context;
298 struct mirror_set *ms = dm_rh_region_context(reg);
302 /* Read error means the failure of default mirror. */
303 DMERR_LIMIT("Unable to read primary mirror during recovery");
304 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
308 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
311 * Bits correspond to devices (excluding default mirror).
312 * The default mirror cannot change during recovery.
314 for (m = 0; m < ms->nr_mirrors; m++) {
315 if (&ms->mirror[m] == get_default_mirror(ms))
317 if (test_bit(bit, &write_err))
318 fail_mirror(ms->mirror + m,
319 DM_RAID1_SYNC_ERROR);
324 dm_rh_recovery_end(reg, !(read_err || write_err));
327 static int recover(struct mirror_set *ms, struct dm_region *reg)
331 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
333 unsigned long flags = 0;
334 region_t key = dm_rh_get_region_key(reg);
335 sector_t region_size = dm_rh_get_region_size(ms->rh);
337 /* fill in the source */
338 m = get_default_mirror(ms);
339 from.bdev = m->dev->bdev;
340 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
341 if (key == (ms->nr_regions - 1)) {
343 * The final region may be smaller than
346 from.count = ms->ti->len & (region_size - 1);
348 from.count = region_size;
350 from.count = region_size;
352 /* fill in the destinations */
353 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
354 if (&ms->mirror[i] == get_default_mirror(ms))
358 dest->bdev = m->dev->bdev;
359 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
360 dest->count = from.count;
365 if (!errors_handled(ms))
366 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
368 r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
369 flags, recovery_complete, reg);
374 static void do_recovery(struct mirror_set *ms)
376 struct dm_region *reg;
377 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
381 * Start quiescing some regions.
383 dm_rh_recovery_prepare(ms->rh);
386 * Copy any already quiesced regions.
388 while ((reg = dm_rh_recovery_start(ms->rh))) {
389 r = recover(ms, reg);
391 dm_rh_recovery_end(reg, 0);
395 * Update the in sync flag.
398 (log->type->get_sync_count(log) == ms->nr_regions)) {
399 /* the sync is complete */
400 dm_table_event(ms->ti->table);
405 /*-----------------------------------------------------------------
407 *---------------------------------------------------------------*/
408 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
410 struct mirror *m = get_default_mirror(ms);
413 if (likely(!atomic_read(&m->error_count)))
416 if (m-- == ms->mirror)
418 } while (m != get_default_mirror(ms));
423 static int default_ok(struct mirror *m)
425 struct mirror *default_mirror = get_default_mirror(m->ms);
427 return !atomic_read(&default_mirror->error_count);
430 static int mirror_available(struct mirror_set *ms, struct bio *bio)
432 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
433 region_t region = dm_rh_bio_to_region(ms->rh, bio);
435 if (log->type->in_sync(log, region, 0))
436 return choose_mirror(ms, bio->bi_sector) ? 1 : 0;
442 * remap a buffer to a particular mirror.
444 static sector_t map_sector(struct mirror *m, struct bio *bio)
446 if (unlikely(!bio->bi_size))
448 return m->offset + (bio->bi_sector - m->ms->ti->begin);
451 static void map_bio(struct mirror *m, struct bio *bio)
453 bio->bi_bdev = m->dev->bdev;
454 bio->bi_sector = map_sector(m, bio);
457 static void map_region(struct dm_io_region *io, struct mirror *m,
460 io->bdev = m->dev->bdev;
461 io->sector = map_sector(m, bio);
462 io->count = bio->bi_size >> 9;
465 static void hold_bio(struct mirror_set *ms, struct bio *bio)
468 * If device is suspended, complete the bio.
470 if (atomic_read(&ms->suspend)) {
471 if (dm_noflush_suspending(ms->ti))
472 bio_endio(bio, DM_ENDIO_REQUEUE);
474 bio_endio(bio, -EIO);
479 * Hold bio until the suspend is complete.
481 spin_lock_irq(&ms->lock);
482 bio_list_add(&ms->holds, bio);
483 spin_unlock_irq(&ms->lock);
486 /*-----------------------------------------------------------------
488 *---------------------------------------------------------------*/
489 static void read_callback(unsigned long error, void *context)
491 struct bio *bio = context;
495 bio_set_m(bio, NULL);
497 if (likely(!error)) {
502 fail_mirror(m, DM_RAID1_READ_ERROR);
504 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
505 DMWARN_LIMIT("Read failure on mirror device %s. "
506 "Trying alternative device.",
508 queue_bio(m->ms, bio, bio_rw(bio));
512 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
514 bio_endio(bio, -EIO);
517 /* Asynchronous read. */
518 static void read_async_bio(struct mirror *m, struct bio *bio)
520 struct dm_io_region io;
521 struct dm_io_request io_req = {
523 .mem.type = DM_IO_BVEC,
524 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
525 .notify.fn = read_callback,
526 .notify.context = bio,
527 .client = m->ms->io_client,
530 map_region(&io, m, bio);
532 BUG_ON(dm_io(&io_req, 1, &io, NULL));
535 static inline int region_in_sync(struct mirror_set *ms, region_t region,
538 int state = dm_rh_get_state(ms->rh, region, may_block);
539 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
542 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
548 while ((bio = bio_list_pop(reads))) {
549 region = dm_rh_bio_to_region(ms->rh, bio);
550 m = get_default_mirror(ms);
553 * We can only read balance if the region is in sync.
555 if (likely(region_in_sync(ms, region, 1)))
556 m = choose_mirror(ms, bio->bi_sector);
557 else if (m && atomic_read(&m->error_count))
561 read_async_bio(m, bio);
563 bio_endio(bio, -EIO);
567 /*-----------------------------------------------------------------
570 * We do different things with the write io depending on the
571 * state of the region that it's in:
573 * SYNC: increment pending, use kcopyd to write to *all* mirrors
574 * RECOVERING: delay the io until recovery completes
575 * NOSYNC: increment pending, just write to the default mirror
576 *---------------------------------------------------------------*/
579 static void write_callback(unsigned long error, void *context)
582 struct bio *bio = (struct bio *) context;
583 struct mirror_set *ms;
587 ms = bio_get_m(bio)->ms;
588 bio_set_m(bio, NULL);
591 * NOTE: We don't decrement the pending count here,
592 * instead it is done by the targets endio function.
593 * This way we handle both writes to SYNC and NOSYNC
594 * regions with the same code.
596 if (likely(!error)) {
601 for (i = 0; i < ms->nr_mirrors; i++)
602 if (test_bit(i, &error))
603 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
606 * Need to raise event. Since raising
607 * events can block, we need to do it in
610 spin_lock_irqsave(&ms->lock, flags);
611 if (!ms->failures.head)
613 bio_list_add(&ms->failures, bio);
614 spin_unlock_irqrestore(&ms->lock, flags);
619 static void do_write(struct mirror_set *ms, struct bio *bio)
622 struct dm_io_region io[ms->nr_mirrors], *dest = io;
624 struct dm_io_request io_req = {
625 .bi_rw = WRITE | (bio->bi_rw & WRITE_BARRIER),
626 .mem.type = DM_IO_BVEC,
627 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
628 .notify.fn = write_callback,
629 .notify.context = bio,
630 .client = ms->io_client,
633 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
634 map_region(dest++, m, bio);
637 * Use default mirror because we only need it to retrieve the reference
638 * to the mirror set in write_callback().
640 bio_set_m(bio, get_default_mirror(ms));
642 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
645 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
649 struct bio_list sync, nosync, recover, *this_list = NULL;
650 struct bio_list requeue;
651 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
658 * Classify each write.
660 bio_list_init(&sync);
661 bio_list_init(&nosync);
662 bio_list_init(&recover);
663 bio_list_init(&requeue);
665 while ((bio = bio_list_pop(writes))) {
666 if (unlikely(bio_empty_barrier(bio))) {
667 bio_list_add(&sync, bio);
671 region = dm_rh_bio_to_region(ms->rh, bio);
673 if (log->type->is_remote_recovering &&
674 log->type->is_remote_recovering(log, region)) {
675 bio_list_add(&requeue, bio);
679 state = dm_rh_get_state(ms->rh, region, 1);
690 case DM_RH_RECOVERING:
691 this_list = &recover;
695 bio_list_add(this_list, bio);
699 * Add bios that are delayed due to remote recovery
700 * back on to the write queue
702 if (unlikely(requeue.head)) {
703 spin_lock_irq(&ms->lock);
704 bio_list_merge(&ms->writes, &requeue);
705 spin_unlock_irq(&ms->lock);
710 * Increment the pending counts for any regions that will
711 * be written to (writes to recover regions are going to
714 dm_rh_inc_pending(ms->rh, &sync);
715 dm_rh_inc_pending(ms->rh, &nosync);
718 * If the flush fails on a previous call and succeeds here,
719 * we must not reset the log_failure variable. We need
720 * userspace interaction to do that.
722 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
727 if (unlikely(ms->log_failure) && errors_handled(ms)) {
728 spin_lock_irq(&ms->lock);
729 bio_list_merge(&ms->failures, &sync);
730 spin_unlock_irq(&ms->lock);
733 while ((bio = bio_list_pop(&sync)))
736 while ((bio = bio_list_pop(&recover)))
737 dm_rh_delay(ms->rh, bio);
739 while ((bio = bio_list_pop(&nosync))) {
740 if (unlikely(ms->leg_failure) && errors_handled(ms)) {
741 spin_lock_irq(&ms->lock);
742 bio_list_add(&ms->failures, bio);
743 spin_unlock_irq(&ms->lock);
746 map_bio(get_default_mirror(ms), bio);
747 generic_make_request(bio);
752 static void do_failures(struct mirror_set *ms, struct bio_list *failures)
756 if (likely(!failures->head))
760 * If the log has failed, unattempted writes are being
761 * put on the holds 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 while ((bio = bio_list_pop(failures))) {
777 if (!ms->log_failure) {
779 dm_rh_mark_nosync(ms->rh, bio);
783 * If all the legs are dead, fail the I/O.
784 * If we have been told to handle errors, hold the bio
785 * and wait for userspace to deal with the problem.
786 * Otherwise pretend that the I/O succeeded. (This would
787 * be wrong if the failed leg returned after reboot and
788 * got replicated back to the good legs.)
790 if (!get_valid_mirror(ms))
791 bio_endio(bio, -EIO);
792 else if (errors_handled(ms))
799 static void trigger_event(struct work_struct *work)
801 struct mirror_set *ms =
802 container_of(work, struct mirror_set, trigger_event);
804 dm_table_event(ms->ti->table);
807 /*-----------------------------------------------------------------
809 *---------------------------------------------------------------*/
810 static void do_mirror(struct work_struct *work)
812 struct mirror_set *ms = container_of(work, struct mirror_set,
814 struct bio_list reads, writes, failures;
817 spin_lock_irqsave(&ms->lock, flags);
820 failures = ms->failures;
821 bio_list_init(&ms->reads);
822 bio_list_init(&ms->writes);
823 bio_list_init(&ms->failures);
824 spin_unlock_irqrestore(&ms->lock, flags);
826 dm_rh_update_states(ms->rh, errors_handled(ms));
828 do_reads(ms, &reads);
829 do_writes(ms, &writes);
830 do_failures(ms, &failures);
832 dm_table_unplug_all(ms->ti->table);
835 /*-----------------------------------------------------------------
837 *---------------------------------------------------------------*/
838 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
839 uint32_t region_size,
840 struct dm_target *ti,
841 struct dm_dirty_log *dl)
844 struct mirror_set *ms = NULL;
846 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
848 ms = kzalloc(len, GFP_KERNEL);
850 ti->error = "Cannot allocate mirror context";
854 spin_lock_init(&ms->lock);
855 bio_list_init(&ms->reads);
856 bio_list_init(&ms->writes);
857 bio_list_init(&ms->failures);
858 bio_list_init(&ms->holds);
861 ms->nr_mirrors = nr_mirrors;
862 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
866 atomic_set(&ms->suspend, 0);
867 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
869 ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
870 _dm_raid1_read_record_cache);
872 if (!ms->read_record_pool) {
873 ti->error = "Error creating mirror read_record_pool";
878 ms->io_client = dm_io_client_create(DM_IO_PAGES);
879 if (IS_ERR(ms->io_client)) {
880 ti->error = "Error creating dm_io client";
881 mempool_destroy(ms->read_record_pool);
886 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
887 wakeup_all_recovery_waiters,
888 ms->ti->begin, MAX_RECOVERY,
889 dl, region_size, ms->nr_regions);
890 if (IS_ERR(ms->rh)) {
891 ti->error = "Error creating dirty region hash";
892 dm_io_client_destroy(ms->io_client);
893 mempool_destroy(ms->read_record_pool);
901 static void free_context(struct mirror_set *ms, struct dm_target *ti,
905 dm_put_device(ti, ms->mirror[m].dev);
907 dm_io_client_destroy(ms->io_client);
908 dm_region_hash_destroy(ms->rh);
909 mempool_destroy(ms->read_record_pool);
913 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
914 unsigned int mirror, char **argv)
916 unsigned long long offset;
918 if (sscanf(argv[1], "%llu", &offset) != 1) {
919 ti->error = "Invalid offset";
923 if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
924 &ms->mirror[mirror].dev)) {
925 ti->error = "Device lookup failure";
929 ms->mirror[mirror].ms = ms;
930 atomic_set(&(ms->mirror[mirror].error_count), 0);
931 ms->mirror[mirror].error_type = 0;
932 ms->mirror[mirror].offset = offset;
938 * Create dirty log: log_type #log_params <log_params>
940 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
941 unsigned argc, char **argv,
944 unsigned param_count;
945 struct dm_dirty_log *dl;
948 ti->error = "Insufficient mirror log arguments";
952 if (sscanf(argv[1], "%u", ¶m_count) != 1) {
953 ti->error = "Invalid mirror log argument count";
957 *args_used = 2 + param_count;
959 if (argc < *args_used) {
960 ti->error = "Insufficient mirror log arguments";
964 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
967 ti->error = "Error creating mirror dirty log";
974 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
977 unsigned num_features;
978 struct dm_target *ti = ms->ti;
985 if (sscanf(argv[0], "%u", &num_features) != 1) {
986 ti->error = "Invalid number of features";
994 if (num_features > argc) {
995 ti->error = "Not enough arguments to support feature count";
999 if (!strcmp("handle_errors", argv[0]))
1000 ms->features |= DM_RAID1_HANDLE_ERRORS;
1002 ti->error = "Unrecognised feature requested";
1012 * Construct a mirror mapping:
1014 * log_type #log_params <log_params>
1015 * #mirrors [mirror_path offset]{2,}
1016 * [#features <features>]
1018 * log_type is "core" or "disk"
1019 * #log_params is between 1 and 3
1021 * If present, features must be "handle_errors".
1023 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1026 unsigned int nr_mirrors, m, args_used;
1027 struct mirror_set *ms;
1028 struct dm_dirty_log *dl;
1030 dl = create_dirty_log(ti, argc, argv, &args_used);
1037 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1038 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
1039 ti->error = "Invalid number of mirrors";
1040 dm_dirty_log_destroy(dl);
1046 if (argc < nr_mirrors * 2) {
1047 ti->error = "Too few mirror arguments";
1048 dm_dirty_log_destroy(dl);
1052 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1054 dm_dirty_log_destroy(dl);
1058 /* Get the mirror parameter sets */
1059 for (m = 0; m < nr_mirrors; m++) {
1060 r = get_mirror(ms, ti, m, argv);
1062 free_context(ms, ti, m);
1070 ti->split_io = dm_rh_get_region_size(ms->rh);
1071 ti->num_flush_requests = 1;
1073 ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
1074 if (!ms->kmirrord_wq) {
1075 DMERR("couldn't start kmirrord");
1077 goto err_free_context;
1079 INIT_WORK(&ms->kmirrord_work, do_mirror);
1080 init_timer(&ms->timer);
1081 ms->timer_pending = 0;
1082 INIT_WORK(&ms->trigger_event, trigger_event);
1084 r = parse_features(ms, argc, argv, &args_used);
1086 goto err_destroy_wq;
1092 * Any read-balancing addition depends on the
1093 * DM_RAID1_HANDLE_ERRORS flag being present.
1094 * This is because the decision to balance depends
1095 * on the sync state of a region. If the above
1096 * flag is not present, we ignore errors; and
1097 * the sync state may be inaccurate.
1101 ti->error = "Too many mirror arguments";
1103 goto err_destroy_wq;
1106 r = dm_kcopyd_client_create(DM_KCOPYD_PAGES, &ms->kcopyd_client);
1108 goto err_destroy_wq;
1114 destroy_workqueue(ms->kmirrord_wq);
1116 free_context(ms, ti, ms->nr_mirrors);
1120 static void mirror_dtr(struct dm_target *ti)
1122 struct mirror_set *ms = (struct mirror_set *) ti->private;
1124 del_timer_sync(&ms->timer);
1125 flush_workqueue(ms->kmirrord_wq);
1126 flush_scheduled_work();
1127 dm_kcopyd_client_destroy(ms->kcopyd_client);
1128 destroy_workqueue(ms->kmirrord_wq);
1129 free_context(ms, ti, ms->nr_mirrors);
1133 * Mirror mapping function
1135 static int mirror_map(struct dm_target *ti, struct bio *bio,
1136 union map_info *map_context)
1138 int r, rw = bio_rw(bio);
1140 struct mirror_set *ms = ti->private;
1141 struct dm_raid1_read_record *read_record = NULL;
1142 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1145 /* Save region for mirror_end_io() handler */
1146 map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1147 queue_bio(ms, bio, rw);
1148 return DM_MAPIO_SUBMITTED;
1151 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1152 if (r < 0 && r != -EWOULDBLOCK)
1156 * If region is not in-sync queue the bio.
1158 if (!r || (r == -EWOULDBLOCK)) {
1160 return -EWOULDBLOCK;
1162 queue_bio(ms, bio, rw);
1163 return DM_MAPIO_SUBMITTED;
1167 * The region is in-sync and we can perform reads directly.
1168 * Store enough information so we can retry if it fails.
1170 m = choose_mirror(ms, bio->bi_sector);
1174 read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1175 if (likely(read_record)) {
1176 dm_bio_record(&read_record->details, bio);
1177 map_context->ptr = read_record;
1183 return DM_MAPIO_REMAPPED;
1186 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1187 int error, union map_info *map_context)
1189 int rw = bio_rw(bio);
1190 struct mirror_set *ms = (struct mirror_set *) ti->private;
1191 struct mirror *m = NULL;
1192 struct dm_bio_details *bd = NULL;
1193 struct dm_raid1_read_record *read_record = map_context->ptr;
1196 * We need to dec pending if this was a write.
1199 if (likely(!bio_empty_barrier(bio)))
1200 dm_rh_dec(ms->rh, map_context->ll);
1204 if (error == -EOPNOTSUPP)
1207 if ((error == -EWOULDBLOCK) && bio_rw_flagged(bio, BIO_RW_AHEAD))
1210 if (unlikely(error)) {
1213 * There wasn't enough memory to record necessary
1214 * information for a retry or there was no other
1217 DMERR_LIMIT("Mirror read failed.");
1223 DMERR("Mirror read failed from %s. Trying alternative device.",
1226 fail_mirror(m, DM_RAID1_READ_ERROR);
1229 * A failed read is requeued for another attempt using an intact
1232 if (default_ok(m) || mirror_available(ms, bio)) {
1233 bd = &read_record->details;
1235 dm_bio_restore(bd, bio);
1236 mempool_free(read_record, ms->read_record_pool);
1237 map_context->ptr = NULL;
1238 queue_bio(ms, bio, rw);
1241 DMERR("All replicated volumes dead, failing I/O");
1246 mempool_free(read_record, ms->read_record_pool);
1247 map_context->ptr = NULL;
1253 static void mirror_presuspend(struct dm_target *ti)
1255 struct mirror_set *ms = (struct mirror_set *) ti->private;
1256 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1258 struct bio_list holds;
1261 atomic_set(&ms->suspend, 1);
1264 * We must finish up all the work that we've
1265 * generated (i.e. recovery work).
1267 dm_rh_stop_recovery(ms->rh);
1269 wait_event(_kmirrord_recovery_stopped,
1270 !dm_rh_recovery_in_flight(ms->rh));
1272 if (log->type->presuspend && log->type->presuspend(log))
1273 /* FIXME: need better error handling */
1274 DMWARN("log presuspend failed");
1277 * Now that recovery is complete/stopped and the
1278 * delayed bios are queued, we need to wait for
1279 * the worker thread to complete. This way,
1280 * we know that all of our I/O has been pushed.
1282 flush_workqueue(ms->kmirrord_wq);
1285 * Now set ms->suspend is set and the workqueue flushed, no more
1286 * entries can be added to ms->hold list, so process it.
1288 * Bios can still arrive concurrently with or after this
1289 * presuspend function, but they cannot join the hold list
1290 * because ms->suspend is set.
1292 spin_lock_irq(&ms->lock);
1294 bio_list_init(&ms->holds);
1295 spin_unlock_irq(&ms->lock);
1297 while ((bio = bio_list_pop(&holds)))
1301 static void mirror_postsuspend(struct dm_target *ti)
1303 struct mirror_set *ms = ti->private;
1304 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1306 if (log->type->postsuspend && log->type->postsuspend(log))
1307 /* FIXME: need better error handling */
1308 DMWARN("log postsuspend failed");
1311 static void mirror_resume(struct dm_target *ti)
1313 struct mirror_set *ms = ti->private;
1314 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1316 atomic_set(&ms->suspend, 0);
1317 if (log->type->resume && log->type->resume(log))
1318 /* FIXME: need better error handling */
1319 DMWARN("log resume failed");
1320 dm_rh_start_recovery(ms->rh);
1324 * device_status_char
1325 * @m: mirror device/leg we want the status of
1327 * We return one character representing the most severe error
1328 * we have encountered.
1329 * A => Alive - No failures
1330 * D => Dead - A write failure occurred leaving mirror out-of-sync
1331 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1332 * R => Read - A read failure occurred, mirror data unaffected
1336 static char device_status_char(struct mirror *m)
1338 if (!atomic_read(&(m->error_count)))
1341 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1342 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1343 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1344 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1348 static int mirror_status(struct dm_target *ti, status_type_t type,
1349 char *result, unsigned int maxlen)
1351 unsigned int m, sz = 0;
1352 struct mirror_set *ms = (struct mirror_set *) ti->private;
1353 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1354 char buffer[ms->nr_mirrors + 1];
1357 case STATUSTYPE_INFO:
1358 DMEMIT("%d ", ms->nr_mirrors);
1359 for (m = 0; m < ms->nr_mirrors; m++) {
1360 DMEMIT("%s ", ms->mirror[m].dev->name);
1361 buffer[m] = device_status_char(&(ms->mirror[m]));
1365 DMEMIT("%llu/%llu 1 %s ",
1366 (unsigned long long)log->type->get_sync_count(log),
1367 (unsigned long long)ms->nr_regions, buffer);
1369 sz += log->type->status(log, type, result+sz, maxlen-sz);
1373 case STATUSTYPE_TABLE:
1374 sz = log->type->status(log, type, result, maxlen);
1376 DMEMIT("%d", ms->nr_mirrors);
1377 for (m = 0; m < ms->nr_mirrors; m++)
1378 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1379 (unsigned long long)ms->mirror[m].offset);
1381 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1382 DMEMIT(" 1 handle_errors");
1388 static int mirror_iterate_devices(struct dm_target *ti,
1389 iterate_devices_callout_fn fn, void *data)
1391 struct mirror_set *ms = ti->private;
1395 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1396 ret = fn(ti, ms->mirror[i].dev,
1397 ms->mirror[i].offset, ti->len, data);
1402 static struct target_type mirror_target = {
1404 .version = {1, 12, 0},
1405 .module = THIS_MODULE,
1409 .end_io = mirror_end_io,
1410 .presuspend = mirror_presuspend,
1411 .postsuspend = mirror_postsuspend,
1412 .resume = mirror_resume,
1413 .status = mirror_status,
1414 .iterate_devices = mirror_iterate_devices,
1417 static int __init dm_mirror_init(void)
1421 _dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1422 if (!_dm_raid1_read_record_cache) {
1423 DMERR("Can't allocate dm_raid1_read_record cache");
1428 r = dm_register_target(&mirror_target);
1430 DMERR("Failed to register mirror target");
1437 kmem_cache_destroy(_dm_raid1_read_record_cache);
1442 static void __exit dm_mirror_exit(void)
1444 dm_unregister_target(&mirror_target);
1445 kmem_cache_destroy(_dm_raid1_read_record_cache);
1449 module_init(dm_mirror_init);
1450 module_exit(dm_mirror_exit);
1452 MODULE_DESCRIPTION(DM_NAME " mirror target");
1453 MODULE_AUTHOR("Joe Thornber");
1454 MODULE_LICENSE("GPL");