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_FLUSH,
263 .mem.type = DM_IO_KMEM,
264 .mem.ptr.addr = 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 + dm_target_offset(m->ms->ti, bio->bi_sector);
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 * Lock is required to avoid race condition during suspend
471 spin_lock_irq(&ms->lock);
473 if (atomic_read(&ms->suspend)) {
474 spin_unlock_irq(&ms->lock);
477 * If device is suspended, complete the bio.
479 if (dm_noflush_suspending(ms->ti))
480 bio_endio(bio, DM_ENDIO_REQUEUE);
482 bio_endio(bio, -EIO);
487 * Hold bio until the suspend is complete.
489 bio_list_add(&ms->holds, bio);
490 spin_unlock_irq(&ms->lock);
493 /*-----------------------------------------------------------------
495 *---------------------------------------------------------------*/
496 static void read_callback(unsigned long error, void *context)
498 struct bio *bio = context;
502 bio_set_m(bio, NULL);
504 if (likely(!error)) {
509 fail_mirror(m, DM_RAID1_READ_ERROR);
511 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
512 DMWARN_LIMIT("Read failure on mirror device %s. "
513 "Trying alternative device.",
515 queue_bio(m->ms, bio, bio_rw(bio));
519 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
521 bio_endio(bio, -EIO);
524 /* Asynchronous read. */
525 static void read_async_bio(struct mirror *m, struct bio *bio)
527 struct dm_io_region io;
528 struct dm_io_request io_req = {
530 .mem.type = DM_IO_BVEC,
531 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
532 .notify.fn = read_callback,
533 .notify.context = bio,
534 .client = m->ms->io_client,
537 map_region(&io, m, bio);
539 BUG_ON(dm_io(&io_req, 1, &io, NULL));
542 static inline int region_in_sync(struct mirror_set *ms, region_t region,
545 int state = dm_rh_get_state(ms->rh, region, may_block);
546 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
549 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
555 while ((bio = bio_list_pop(reads))) {
556 region = dm_rh_bio_to_region(ms->rh, bio);
557 m = get_default_mirror(ms);
560 * We can only read balance if the region is in sync.
562 if (likely(region_in_sync(ms, region, 1)))
563 m = choose_mirror(ms, bio->bi_sector);
564 else if (m && atomic_read(&m->error_count))
568 read_async_bio(m, bio);
570 bio_endio(bio, -EIO);
574 /*-----------------------------------------------------------------
577 * We do different things with the write io depending on the
578 * state of the region that it's in:
580 * SYNC: increment pending, use kcopyd to write to *all* mirrors
581 * RECOVERING: delay the io until recovery completes
582 * NOSYNC: increment pending, just write to the default mirror
583 *---------------------------------------------------------------*/
586 static void write_callback(unsigned long error, void *context)
589 struct bio *bio = (struct bio *) context;
590 struct mirror_set *ms;
594 ms = bio_get_m(bio)->ms;
595 bio_set_m(bio, NULL);
598 * NOTE: We don't decrement the pending count here,
599 * instead it is done by the targets endio function.
600 * This way we handle both writes to SYNC and NOSYNC
601 * regions with the same code.
603 if (likely(!error)) {
608 for (i = 0; i < ms->nr_mirrors; i++)
609 if (test_bit(i, &error))
610 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
613 * Need to raise event. Since raising
614 * events can block, we need to do it in
617 spin_lock_irqsave(&ms->lock, flags);
618 if (!ms->failures.head)
620 bio_list_add(&ms->failures, bio);
621 spin_unlock_irqrestore(&ms->lock, flags);
626 static void do_write(struct mirror_set *ms, struct bio *bio)
629 struct dm_io_region io[ms->nr_mirrors], *dest = io;
631 struct dm_io_request io_req = {
632 .bi_rw = WRITE | (bio->bi_rw & WRITE_FLUSH_FUA),
633 .mem.type = DM_IO_BVEC,
634 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
635 .notify.fn = write_callback,
636 .notify.context = bio,
637 .client = ms->io_client,
640 if (bio->bi_rw & REQ_DISCARD) {
641 io_req.bi_rw |= REQ_DISCARD;
642 io_req.mem.type = DM_IO_KMEM;
643 io_req.mem.ptr.addr = NULL;
646 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
647 map_region(dest++, m, bio);
650 * Use default mirror because we only need it to retrieve the reference
651 * to the mirror set in write_callback().
653 bio_set_m(bio, get_default_mirror(ms));
655 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
658 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
662 struct bio_list sync, nosync, recover, *this_list = NULL;
663 struct bio_list requeue;
664 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
671 * Classify each write.
673 bio_list_init(&sync);
674 bio_list_init(&nosync);
675 bio_list_init(&recover);
676 bio_list_init(&requeue);
678 while ((bio = bio_list_pop(writes))) {
679 if ((bio->bi_rw & REQ_FLUSH) ||
680 (bio->bi_rw & REQ_DISCARD)) {
681 bio_list_add(&sync, bio);
685 region = dm_rh_bio_to_region(ms->rh, bio);
687 if (log->type->is_remote_recovering &&
688 log->type->is_remote_recovering(log, region)) {
689 bio_list_add(&requeue, bio);
693 state = dm_rh_get_state(ms->rh, region, 1);
704 case DM_RH_RECOVERING:
705 this_list = &recover;
709 bio_list_add(this_list, bio);
713 * Add bios that are delayed due to remote recovery
714 * back on to the write queue
716 if (unlikely(requeue.head)) {
717 spin_lock_irq(&ms->lock);
718 bio_list_merge(&ms->writes, &requeue);
719 spin_unlock_irq(&ms->lock);
724 * Increment the pending counts for any regions that will
725 * be written to (writes to recover regions are going to
728 dm_rh_inc_pending(ms->rh, &sync);
729 dm_rh_inc_pending(ms->rh, &nosync);
732 * If the flush fails on a previous call and succeeds here,
733 * we must not reset the log_failure variable. We need
734 * userspace interaction to do that.
736 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
741 if (unlikely(ms->log_failure) && errors_handled(ms)) {
742 spin_lock_irq(&ms->lock);
743 bio_list_merge(&ms->failures, &sync);
744 spin_unlock_irq(&ms->lock);
747 while ((bio = bio_list_pop(&sync)))
750 while ((bio = bio_list_pop(&recover)))
751 dm_rh_delay(ms->rh, bio);
753 while ((bio = bio_list_pop(&nosync))) {
754 if (unlikely(ms->leg_failure) && errors_handled(ms)) {
755 spin_lock_irq(&ms->lock);
756 bio_list_add(&ms->failures, bio);
757 spin_unlock_irq(&ms->lock);
760 map_bio(get_default_mirror(ms), bio);
761 generic_make_request(bio);
766 static void do_failures(struct mirror_set *ms, struct bio_list *failures)
770 if (likely(!failures->head))
774 * If the log has failed, unattempted writes are being
775 * put on the holds list. We can't issue those writes
776 * until a log has been marked, so we must store them.
778 * If a 'noflush' suspend is in progress, we can requeue
779 * the I/O's to the core. This give userspace a chance
780 * to reconfigure the mirror, at which point the core
781 * will reissue the writes. If the 'noflush' flag is
782 * not set, we have no choice but to return errors.
784 * Some writes on the failures list may have been
785 * submitted before the log failure and represent a
786 * failure to write to one of the devices. It is ok
787 * for us to treat them the same and requeue them
790 while ((bio = bio_list_pop(failures))) {
791 if (!ms->log_failure) {
793 dm_rh_mark_nosync(ms->rh, bio);
797 * If all the legs are dead, fail the I/O.
798 * If we have been told to handle errors, hold the bio
799 * and wait for userspace to deal with the problem.
800 * Otherwise pretend that the I/O succeeded. (This would
801 * be wrong if the failed leg returned after reboot and
802 * got replicated back to the good legs.)
804 if (!get_valid_mirror(ms))
805 bio_endio(bio, -EIO);
806 else if (errors_handled(ms))
813 static void trigger_event(struct work_struct *work)
815 struct mirror_set *ms =
816 container_of(work, struct mirror_set, trigger_event);
818 dm_table_event(ms->ti->table);
821 /*-----------------------------------------------------------------
823 *---------------------------------------------------------------*/
824 static void do_mirror(struct work_struct *work)
826 struct mirror_set *ms = container_of(work, struct mirror_set,
828 struct bio_list reads, writes, failures;
831 spin_lock_irqsave(&ms->lock, flags);
834 failures = ms->failures;
835 bio_list_init(&ms->reads);
836 bio_list_init(&ms->writes);
837 bio_list_init(&ms->failures);
838 spin_unlock_irqrestore(&ms->lock, flags);
840 dm_rh_update_states(ms->rh, errors_handled(ms));
842 do_reads(ms, &reads);
843 do_writes(ms, &writes);
844 do_failures(ms, &failures);
846 dm_table_unplug_all(ms->ti->table);
849 /*-----------------------------------------------------------------
851 *---------------------------------------------------------------*/
852 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
853 uint32_t region_size,
854 struct dm_target *ti,
855 struct dm_dirty_log *dl)
858 struct mirror_set *ms = NULL;
860 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
862 ms = kzalloc(len, GFP_KERNEL);
864 ti->error = "Cannot allocate mirror context";
868 spin_lock_init(&ms->lock);
869 bio_list_init(&ms->reads);
870 bio_list_init(&ms->writes);
871 bio_list_init(&ms->failures);
872 bio_list_init(&ms->holds);
875 ms->nr_mirrors = nr_mirrors;
876 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
880 atomic_set(&ms->suspend, 0);
881 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
883 ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
884 _dm_raid1_read_record_cache);
886 if (!ms->read_record_pool) {
887 ti->error = "Error creating mirror read_record_pool";
892 ms->io_client = dm_io_client_create(DM_IO_PAGES);
893 if (IS_ERR(ms->io_client)) {
894 ti->error = "Error creating dm_io client";
895 mempool_destroy(ms->read_record_pool);
900 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
901 wakeup_all_recovery_waiters,
902 ms->ti->begin, MAX_RECOVERY,
903 dl, region_size, ms->nr_regions);
904 if (IS_ERR(ms->rh)) {
905 ti->error = "Error creating dirty region hash";
906 dm_io_client_destroy(ms->io_client);
907 mempool_destroy(ms->read_record_pool);
915 static void free_context(struct mirror_set *ms, struct dm_target *ti,
919 dm_put_device(ti, ms->mirror[m].dev);
921 dm_io_client_destroy(ms->io_client);
922 dm_region_hash_destroy(ms->rh);
923 mempool_destroy(ms->read_record_pool);
927 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
928 unsigned int mirror, char **argv)
930 unsigned long long offset;
932 if (sscanf(argv[1], "%llu", &offset) != 1) {
933 ti->error = "Invalid offset";
937 if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
938 &ms->mirror[mirror].dev)) {
939 ti->error = "Device lookup failure";
943 ms->mirror[mirror].ms = ms;
944 atomic_set(&(ms->mirror[mirror].error_count), 0);
945 ms->mirror[mirror].error_type = 0;
946 ms->mirror[mirror].offset = offset;
952 * Create dirty log: log_type #log_params <log_params>
954 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
955 unsigned argc, char **argv,
958 unsigned param_count;
959 struct dm_dirty_log *dl;
962 ti->error = "Insufficient mirror log arguments";
966 if (sscanf(argv[1], "%u", ¶m_count) != 1) {
967 ti->error = "Invalid mirror log argument count";
971 *args_used = 2 + param_count;
973 if (argc < *args_used) {
974 ti->error = "Insufficient mirror log arguments";
978 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
981 ti->error = "Error creating mirror dirty log";
988 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
991 unsigned num_features;
992 struct dm_target *ti = ms->ti;
999 if (sscanf(argv[0], "%u", &num_features) != 1) {
1000 ti->error = "Invalid number of features";
1008 if (num_features > argc) {
1009 ti->error = "Not enough arguments to support feature count";
1013 if (!strcmp("handle_errors", argv[0]))
1014 ms->features |= DM_RAID1_HANDLE_ERRORS;
1016 ti->error = "Unrecognised feature requested";
1026 * Construct a mirror mapping:
1028 * log_type #log_params <log_params>
1029 * #mirrors [mirror_path offset]{2,}
1030 * [#features <features>]
1032 * log_type is "core" or "disk"
1033 * #log_params is between 1 and 3
1035 * If present, features must be "handle_errors".
1037 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1040 unsigned int nr_mirrors, m, args_used;
1041 struct mirror_set *ms;
1042 struct dm_dirty_log *dl;
1044 dl = create_dirty_log(ti, argc, argv, &args_used);
1051 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1052 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
1053 ti->error = "Invalid number of mirrors";
1054 dm_dirty_log_destroy(dl);
1060 if (argc < nr_mirrors * 2) {
1061 ti->error = "Too few mirror arguments";
1062 dm_dirty_log_destroy(dl);
1066 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1068 dm_dirty_log_destroy(dl);
1072 /* Get the mirror parameter sets */
1073 for (m = 0; m < nr_mirrors; m++) {
1074 r = get_mirror(ms, ti, m, argv);
1076 free_context(ms, ti, m);
1084 ti->split_io = dm_rh_get_region_size(ms->rh);
1085 ti->num_flush_requests = 1;
1086 ti->num_discard_requests = 1;
1088 ms->kmirrord_wq = alloc_ordered_workqueue("kmirrord", WQ_MEM_RECLAIM);
1089 if (!ms->kmirrord_wq) {
1090 DMERR("couldn't start kmirrord");
1092 goto err_free_context;
1094 INIT_WORK(&ms->kmirrord_work, do_mirror);
1095 init_timer(&ms->timer);
1096 ms->timer_pending = 0;
1097 INIT_WORK(&ms->trigger_event, trigger_event);
1099 r = parse_features(ms, argc, argv, &args_used);
1101 goto err_destroy_wq;
1107 * Any read-balancing addition depends on the
1108 * DM_RAID1_HANDLE_ERRORS flag being present.
1109 * This is because the decision to balance depends
1110 * on the sync state of a region. If the above
1111 * flag is not present, we ignore errors; and
1112 * the sync state may be inaccurate.
1116 ti->error = "Too many mirror arguments";
1118 goto err_destroy_wq;
1121 r = dm_kcopyd_client_create(DM_KCOPYD_PAGES, &ms->kcopyd_client);
1123 goto err_destroy_wq;
1129 destroy_workqueue(ms->kmirrord_wq);
1131 free_context(ms, ti, ms->nr_mirrors);
1135 static void mirror_dtr(struct dm_target *ti)
1137 struct mirror_set *ms = (struct mirror_set *) ti->private;
1139 del_timer_sync(&ms->timer);
1140 flush_workqueue(ms->kmirrord_wq);
1141 flush_work_sync(&ms->trigger_event);
1142 dm_kcopyd_client_destroy(ms->kcopyd_client);
1143 destroy_workqueue(ms->kmirrord_wq);
1144 free_context(ms, ti, ms->nr_mirrors);
1148 * Mirror mapping function
1150 static int mirror_map(struct dm_target *ti, struct bio *bio,
1151 union map_info *map_context)
1153 int r, rw = bio_rw(bio);
1155 struct mirror_set *ms = ti->private;
1156 struct dm_raid1_read_record *read_record = NULL;
1157 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1160 /* Save region for mirror_end_io() handler */
1161 map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1162 queue_bio(ms, bio, rw);
1163 return DM_MAPIO_SUBMITTED;
1166 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1167 if (r < 0 && r != -EWOULDBLOCK)
1171 * If region is not in-sync queue the bio.
1173 if (!r || (r == -EWOULDBLOCK)) {
1175 return -EWOULDBLOCK;
1177 queue_bio(ms, bio, rw);
1178 return DM_MAPIO_SUBMITTED;
1182 * The region is in-sync and we can perform reads directly.
1183 * Store enough information so we can retry if it fails.
1185 m = choose_mirror(ms, bio->bi_sector);
1189 read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1190 if (likely(read_record)) {
1191 dm_bio_record(&read_record->details, bio);
1192 map_context->ptr = read_record;
1198 return DM_MAPIO_REMAPPED;
1201 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1202 int error, union map_info *map_context)
1204 int rw = bio_rw(bio);
1205 struct mirror_set *ms = (struct mirror_set *) ti->private;
1206 struct mirror *m = NULL;
1207 struct dm_bio_details *bd = NULL;
1208 struct dm_raid1_read_record *read_record = map_context->ptr;
1211 * We need to dec pending if this was a write.
1214 if (!(bio->bi_rw & REQ_FLUSH))
1215 dm_rh_dec(ms->rh, map_context->ll);
1219 if (error == -EOPNOTSUPP)
1222 if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
1225 if (unlikely(error)) {
1228 * There wasn't enough memory to record necessary
1229 * information for a retry or there was no other
1232 DMERR_LIMIT("Mirror read failed.");
1238 DMERR("Mirror read failed from %s. Trying alternative device.",
1241 fail_mirror(m, DM_RAID1_READ_ERROR);
1244 * A failed read is requeued for another attempt using an intact
1247 if (default_ok(m) || mirror_available(ms, bio)) {
1248 bd = &read_record->details;
1250 dm_bio_restore(bd, bio);
1251 mempool_free(read_record, ms->read_record_pool);
1252 map_context->ptr = NULL;
1253 queue_bio(ms, bio, rw);
1256 DMERR("All replicated volumes dead, failing I/O");
1261 mempool_free(read_record, ms->read_record_pool);
1262 map_context->ptr = NULL;
1268 static void mirror_presuspend(struct dm_target *ti)
1270 struct mirror_set *ms = (struct mirror_set *) ti->private;
1271 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1273 struct bio_list holds;
1276 atomic_set(&ms->suspend, 1);
1279 * Process bios in the hold list to start recovery waiting
1280 * for bios in the hold list. After the process, no bio has
1281 * a chance to be added in the hold list because ms->suspend
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 * We must finish up all the work that we've
1294 * generated (i.e. recovery work).
1296 dm_rh_stop_recovery(ms->rh);
1298 wait_event(_kmirrord_recovery_stopped,
1299 !dm_rh_recovery_in_flight(ms->rh));
1301 if (log->type->presuspend && log->type->presuspend(log))
1302 /* FIXME: need better error handling */
1303 DMWARN("log presuspend failed");
1306 * Now that recovery is complete/stopped and the
1307 * delayed bios are queued, we need to wait for
1308 * the worker thread to complete. This way,
1309 * we know that all of our I/O has been pushed.
1311 flush_workqueue(ms->kmirrord_wq);
1314 static void mirror_postsuspend(struct dm_target *ti)
1316 struct mirror_set *ms = ti->private;
1317 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1319 if (log->type->postsuspend && log->type->postsuspend(log))
1320 /* FIXME: need better error handling */
1321 DMWARN("log postsuspend failed");
1324 static void mirror_resume(struct dm_target *ti)
1326 struct mirror_set *ms = ti->private;
1327 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1329 atomic_set(&ms->suspend, 0);
1330 if (log->type->resume && log->type->resume(log))
1331 /* FIXME: need better error handling */
1332 DMWARN("log resume failed");
1333 dm_rh_start_recovery(ms->rh);
1337 * device_status_char
1338 * @m: mirror device/leg we want the status of
1340 * We return one character representing the most severe error
1341 * we have encountered.
1342 * A => Alive - No failures
1343 * D => Dead - A write failure occurred leaving mirror out-of-sync
1344 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1345 * R => Read - A read failure occurred, mirror data unaffected
1349 static char device_status_char(struct mirror *m)
1351 if (!atomic_read(&(m->error_count)))
1354 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1355 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1356 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1357 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1361 static int mirror_status(struct dm_target *ti, status_type_t type,
1362 char *result, unsigned int maxlen)
1364 unsigned int m, sz = 0;
1365 struct mirror_set *ms = (struct mirror_set *) ti->private;
1366 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1367 char buffer[ms->nr_mirrors + 1];
1370 case STATUSTYPE_INFO:
1371 DMEMIT("%d ", ms->nr_mirrors);
1372 for (m = 0; m < ms->nr_mirrors; m++) {
1373 DMEMIT("%s ", ms->mirror[m].dev->name);
1374 buffer[m] = device_status_char(&(ms->mirror[m]));
1378 DMEMIT("%llu/%llu 1 %s ",
1379 (unsigned long long)log->type->get_sync_count(log),
1380 (unsigned long long)ms->nr_regions, buffer);
1382 sz += log->type->status(log, type, result+sz, maxlen-sz);
1386 case STATUSTYPE_TABLE:
1387 sz = log->type->status(log, type, result, maxlen);
1389 DMEMIT("%d", ms->nr_mirrors);
1390 for (m = 0; m < ms->nr_mirrors; m++)
1391 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1392 (unsigned long long)ms->mirror[m].offset);
1394 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1395 DMEMIT(" 1 handle_errors");
1401 static int mirror_iterate_devices(struct dm_target *ti,
1402 iterate_devices_callout_fn fn, void *data)
1404 struct mirror_set *ms = ti->private;
1408 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1409 ret = fn(ti, ms->mirror[i].dev,
1410 ms->mirror[i].offset, ti->len, data);
1415 static struct target_type mirror_target = {
1417 .version = {1, 12, 0},
1418 .module = THIS_MODULE,
1422 .end_io = mirror_end_io,
1423 .presuspend = mirror_presuspend,
1424 .postsuspend = mirror_postsuspend,
1425 .resume = mirror_resume,
1426 .status = mirror_status,
1427 .iterate_devices = mirror_iterate_devices,
1430 static int __init dm_mirror_init(void)
1434 _dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1435 if (!_dm_raid1_read_record_cache) {
1436 DMERR("Can't allocate dm_raid1_read_record cache");
1441 r = dm_register_target(&mirror_target);
1443 DMERR("Failed to register mirror target");
1450 kmem_cache_destroy(_dm_raid1_read_record_cache);
1455 static void __exit dm_mirror_exit(void)
1457 dm_unregister_target(&mirror_target);
1458 kmem_cache_destroy(_dm_raid1_read_record_cache);
1462 module_init(dm_mirror_init);
1463 module_exit(dm_mirror_exit);
1465 MODULE_DESCRIPTION(DM_NAME " mirror target");
1466 MODULE_AUTHOR("Joe Thornber");
1467 MODULE_LICENSE("GPL");