Merge master.kernel.org:/pub/scm/linux/kernel/git/davem/sparc-2.6
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / md / dm-raid1.c
blobd12cf3e5e0763cfd9f3d8acb3771f8ee4197f283
1 /*
2 * Copyright (C) 2003 Sistina Software Limited.
4 * This file is released under the GPL.
5 */
7 #include "dm.h"
8 #include "dm-bio-list.h"
9 #include "dm-io.h"
10 #include "dm-log.h"
11 #include "kcopyd.h"
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/vmalloc.h>
21 #include <linux/workqueue.h>
23 static struct workqueue_struct *_kmirrord_wq;
24 static struct work_struct _kmirrord_work;
26 static inline void wake(void)
28 queue_work(_kmirrord_wq, &_kmirrord_work);
31 /*-----------------------------------------------------------------
32 * Region hash
34 * The mirror splits itself up into discrete regions. Each
35 * region can be in one of three states: clean, dirty,
36 * nosync. There is no need to put clean regions in the hash.
38 * In addition to being present in the hash table a region _may_
39 * be present on one of three lists.
41 * clean_regions: Regions on this list have no io pending to
42 * them, they are in sync, we are no longer interested in them,
43 * they are dull. rh_update_states() will remove them from the
44 * hash table.
46 * quiesced_regions: These regions have been spun down, ready
47 * for recovery. rh_recovery_start() will remove regions from
48 * this list and hand them to kmirrord, which will schedule the
49 * recovery io with kcopyd.
51 * recovered_regions: Regions that kcopyd has successfully
52 * recovered. rh_update_states() will now schedule any delayed
53 * io, up the recovery_count, and remove the region from the
54 * hash.
56 * There are 2 locks:
57 * A rw spin lock 'hash_lock' protects just the hash table,
58 * this is never held in write mode from interrupt context,
59 * which I believe means that we only have to disable irqs when
60 * doing a write lock.
62 * An ordinary spin lock 'region_lock' that protects the three
63 * lists in the region_hash, with the 'state', 'list' and
64 * 'bhs_delayed' fields of the regions. This is used from irq
65 * context, so all other uses will have to suspend local irqs.
66 *---------------------------------------------------------------*/
67 struct mirror_set;
68 struct region_hash {
69 struct mirror_set *ms;
70 uint32_t region_size;
71 unsigned region_shift;
73 /* holds persistent region state */
74 struct dirty_log *log;
76 /* hash table */
77 rwlock_t hash_lock;
78 mempool_t *region_pool;
79 unsigned int mask;
80 unsigned int nr_buckets;
81 struct list_head *buckets;
83 spinlock_t region_lock;
84 struct semaphore recovery_count;
85 struct list_head clean_regions;
86 struct list_head quiesced_regions;
87 struct list_head recovered_regions;
90 enum {
91 RH_CLEAN,
92 RH_DIRTY,
93 RH_NOSYNC,
94 RH_RECOVERING
97 struct region {
98 struct region_hash *rh; /* FIXME: can we get rid of this ? */
99 region_t key;
100 int state;
102 struct list_head hash_list;
103 struct list_head list;
105 atomic_t pending;
106 struct bio_list delayed_bios;
110 * Conversion fns
112 static inline region_t bio_to_region(struct region_hash *rh, struct bio *bio)
114 return bio->bi_sector >> rh->region_shift;
117 static inline sector_t region_to_sector(struct region_hash *rh, region_t region)
119 return region << rh->region_shift;
122 /* FIXME move this */
123 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw);
125 #define MIN_REGIONS 64
126 #define MAX_RECOVERY 1
127 static int rh_init(struct region_hash *rh, struct mirror_set *ms,
128 struct dirty_log *log, uint32_t region_size,
129 region_t nr_regions)
131 unsigned int nr_buckets, max_buckets;
132 size_t i;
135 * Calculate a suitable number of buckets for our hash
136 * table.
138 max_buckets = nr_regions >> 6;
139 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
141 nr_buckets >>= 1;
143 rh->ms = ms;
144 rh->log = log;
145 rh->region_size = region_size;
146 rh->region_shift = ffs(region_size) - 1;
147 rwlock_init(&rh->hash_lock);
148 rh->mask = nr_buckets - 1;
149 rh->nr_buckets = nr_buckets;
151 rh->buckets = vmalloc(nr_buckets * sizeof(*rh->buckets));
152 if (!rh->buckets) {
153 DMERR("unable to allocate region hash memory");
154 return -ENOMEM;
157 for (i = 0; i < nr_buckets; i++)
158 INIT_LIST_HEAD(rh->buckets + i);
160 spin_lock_init(&rh->region_lock);
161 sema_init(&rh->recovery_count, 0);
162 INIT_LIST_HEAD(&rh->clean_regions);
163 INIT_LIST_HEAD(&rh->quiesced_regions);
164 INIT_LIST_HEAD(&rh->recovered_regions);
166 rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
167 sizeof(struct region));
168 if (!rh->region_pool) {
169 vfree(rh->buckets);
170 rh->buckets = NULL;
171 return -ENOMEM;
174 return 0;
177 static void rh_exit(struct region_hash *rh)
179 unsigned int h;
180 struct region *reg, *nreg;
182 BUG_ON(!list_empty(&rh->quiesced_regions));
183 for (h = 0; h < rh->nr_buckets; h++) {
184 list_for_each_entry_safe(reg, nreg, rh->buckets + h, hash_list) {
185 BUG_ON(atomic_read(&reg->pending));
186 mempool_free(reg, rh->region_pool);
190 if (rh->log)
191 dm_destroy_dirty_log(rh->log);
192 if (rh->region_pool)
193 mempool_destroy(rh->region_pool);
194 vfree(rh->buckets);
197 #define RH_HASH_MULT 2654435387U
199 static inline unsigned int rh_hash(struct region_hash *rh, region_t region)
201 return (unsigned int) ((region * RH_HASH_MULT) >> 12) & rh->mask;
204 static struct region *__rh_lookup(struct region_hash *rh, region_t region)
206 struct region *reg;
208 list_for_each_entry (reg, rh->buckets + rh_hash(rh, region), hash_list)
209 if (reg->key == region)
210 return reg;
212 return NULL;
215 static void __rh_insert(struct region_hash *rh, struct region *reg)
217 unsigned int h = rh_hash(rh, reg->key);
218 list_add(&reg->hash_list, rh->buckets + h);
221 static struct region *__rh_alloc(struct region_hash *rh, region_t region)
223 struct region *reg, *nreg;
225 read_unlock(&rh->hash_lock);
226 nreg = mempool_alloc(rh->region_pool, GFP_NOIO);
227 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
228 RH_CLEAN : RH_NOSYNC;
229 nreg->rh = rh;
230 nreg->key = region;
232 INIT_LIST_HEAD(&nreg->list);
234 atomic_set(&nreg->pending, 0);
235 bio_list_init(&nreg->delayed_bios);
236 write_lock_irq(&rh->hash_lock);
238 reg = __rh_lookup(rh, region);
239 if (reg)
240 /* we lost the race */
241 mempool_free(nreg, rh->region_pool);
243 else {
244 __rh_insert(rh, nreg);
245 if (nreg->state == RH_CLEAN) {
246 spin_lock(&rh->region_lock);
247 list_add(&nreg->list, &rh->clean_regions);
248 spin_unlock(&rh->region_lock);
250 reg = nreg;
252 write_unlock_irq(&rh->hash_lock);
253 read_lock(&rh->hash_lock);
255 return reg;
258 static inline struct region *__rh_find(struct region_hash *rh, region_t region)
260 struct region *reg;
262 reg = __rh_lookup(rh, region);
263 if (!reg)
264 reg = __rh_alloc(rh, region);
266 return reg;
269 static int rh_state(struct region_hash *rh, region_t region, int may_block)
271 int r;
272 struct region *reg;
274 read_lock(&rh->hash_lock);
275 reg = __rh_lookup(rh, region);
276 read_unlock(&rh->hash_lock);
278 if (reg)
279 return reg->state;
282 * The region wasn't in the hash, so we fall back to the
283 * dirty log.
285 r = rh->log->type->in_sync(rh->log, region, may_block);
288 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
289 * taken as a RH_NOSYNC
291 return r == 1 ? RH_CLEAN : RH_NOSYNC;
294 static inline int rh_in_sync(struct region_hash *rh,
295 region_t region, int may_block)
297 int state = rh_state(rh, region, may_block);
298 return state == RH_CLEAN || state == RH_DIRTY;
301 static void dispatch_bios(struct mirror_set *ms, struct bio_list *bio_list)
303 struct bio *bio;
305 while ((bio = bio_list_pop(bio_list))) {
306 queue_bio(ms, bio, WRITE);
310 static void rh_update_states(struct region_hash *rh)
312 struct region *reg, *next;
314 LIST_HEAD(clean);
315 LIST_HEAD(recovered);
318 * Quickly grab the lists.
320 write_lock_irq(&rh->hash_lock);
321 spin_lock(&rh->region_lock);
322 if (!list_empty(&rh->clean_regions)) {
323 list_splice(&rh->clean_regions, &clean);
324 INIT_LIST_HEAD(&rh->clean_regions);
326 list_for_each_entry (reg, &clean, list) {
327 rh->log->type->clear_region(rh->log, reg->key);
328 list_del(&reg->hash_list);
332 if (!list_empty(&rh->recovered_regions)) {
333 list_splice(&rh->recovered_regions, &recovered);
334 INIT_LIST_HEAD(&rh->recovered_regions);
336 list_for_each_entry (reg, &recovered, list)
337 list_del(&reg->hash_list);
339 spin_unlock(&rh->region_lock);
340 write_unlock_irq(&rh->hash_lock);
343 * All the regions on the recovered and clean lists have
344 * now been pulled out of the system, so no need to do
345 * any more locking.
347 list_for_each_entry_safe (reg, next, &recovered, list) {
348 rh->log->type->clear_region(rh->log, reg->key);
349 rh->log->type->complete_resync_work(rh->log, reg->key, 1);
350 dispatch_bios(rh->ms, &reg->delayed_bios);
351 up(&rh->recovery_count);
352 mempool_free(reg, rh->region_pool);
355 if (!list_empty(&recovered))
356 rh->log->type->flush(rh->log);
358 list_for_each_entry_safe (reg, next, &clean, list)
359 mempool_free(reg, rh->region_pool);
362 static void rh_inc(struct region_hash *rh, region_t region)
364 struct region *reg;
366 read_lock(&rh->hash_lock);
367 reg = __rh_find(rh, region);
369 spin_lock_irq(&rh->region_lock);
370 atomic_inc(&reg->pending);
372 if (reg->state == RH_CLEAN) {
373 reg->state = RH_DIRTY;
374 list_del_init(&reg->list); /* take off the clean list */
375 spin_unlock_irq(&rh->region_lock);
377 rh->log->type->mark_region(rh->log, reg->key);
378 } else
379 spin_unlock_irq(&rh->region_lock);
382 read_unlock(&rh->hash_lock);
385 static void rh_inc_pending(struct region_hash *rh, struct bio_list *bios)
387 struct bio *bio;
389 for (bio = bios->head; bio; bio = bio->bi_next)
390 rh_inc(rh, bio_to_region(rh, bio));
393 static void rh_dec(struct region_hash *rh, region_t region)
395 unsigned long flags;
396 struct region *reg;
397 int should_wake = 0;
399 read_lock(&rh->hash_lock);
400 reg = __rh_lookup(rh, region);
401 read_unlock(&rh->hash_lock);
403 spin_lock_irqsave(&rh->region_lock, flags);
404 if (atomic_dec_and_test(&reg->pending)) {
406 * There is no pending I/O for this region.
407 * We can move the region to corresponding list for next action.
408 * At this point, the region is not yet connected to any list.
410 * If the state is RH_NOSYNC, the region should be kept off
411 * from clean list.
412 * The hash entry for RH_NOSYNC will remain in memory
413 * until the region is recovered or the map is reloaded.
416 /* do nothing for RH_NOSYNC */
417 if (reg->state == RH_RECOVERING) {
418 list_add_tail(&reg->list, &rh->quiesced_regions);
419 } else if (reg->state == RH_DIRTY) {
420 reg->state = RH_CLEAN;
421 list_add(&reg->list, &rh->clean_regions);
423 should_wake = 1;
425 spin_unlock_irqrestore(&rh->region_lock, flags);
427 if (should_wake)
428 wake();
432 * Starts quiescing a region in preparation for recovery.
434 static int __rh_recovery_prepare(struct region_hash *rh)
436 int r;
437 struct region *reg;
438 region_t region;
441 * Ask the dirty log what's next.
443 r = rh->log->type->get_resync_work(rh->log, &region);
444 if (r <= 0)
445 return r;
448 * Get this region, and start it quiescing by setting the
449 * recovering flag.
451 read_lock(&rh->hash_lock);
452 reg = __rh_find(rh, region);
453 read_unlock(&rh->hash_lock);
455 spin_lock_irq(&rh->region_lock);
456 reg->state = RH_RECOVERING;
458 /* Already quiesced ? */
459 if (atomic_read(&reg->pending))
460 list_del_init(&reg->list);
462 else {
463 list_del_init(&reg->list);
464 list_add(&reg->list, &rh->quiesced_regions);
466 spin_unlock_irq(&rh->region_lock);
468 return 1;
471 static void rh_recovery_prepare(struct region_hash *rh)
473 while (!down_trylock(&rh->recovery_count))
474 if (__rh_recovery_prepare(rh) <= 0) {
475 up(&rh->recovery_count);
476 break;
481 * Returns any quiesced regions.
483 static struct region *rh_recovery_start(struct region_hash *rh)
485 struct region *reg = NULL;
487 spin_lock_irq(&rh->region_lock);
488 if (!list_empty(&rh->quiesced_regions)) {
489 reg = list_entry(rh->quiesced_regions.next,
490 struct region, list);
491 list_del_init(&reg->list); /* remove from the quiesced list */
493 spin_unlock_irq(&rh->region_lock);
495 return reg;
498 /* FIXME: success ignored for now */
499 static void rh_recovery_end(struct region *reg, int success)
501 struct region_hash *rh = reg->rh;
503 spin_lock_irq(&rh->region_lock);
504 list_add(&reg->list, &reg->rh->recovered_regions);
505 spin_unlock_irq(&rh->region_lock);
507 wake();
510 static void rh_flush(struct region_hash *rh)
512 rh->log->type->flush(rh->log);
515 static void rh_delay(struct region_hash *rh, struct bio *bio)
517 struct region *reg;
519 read_lock(&rh->hash_lock);
520 reg = __rh_find(rh, bio_to_region(rh, bio));
521 bio_list_add(&reg->delayed_bios, bio);
522 read_unlock(&rh->hash_lock);
525 static void rh_stop_recovery(struct region_hash *rh)
527 int i;
529 /* wait for any recovering regions */
530 for (i = 0; i < MAX_RECOVERY; i++)
531 down(&rh->recovery_count);
534 static void rh_start_recovery(struct region_hash *rh)
536 int i;
538 for (i = 0; i < MAX_RECOVERY; i++)
539 up(&rh->recovery_count);
541 wake();
544 /*-----------------------------------------------------------------
545 * Mirror set structures.
546 *---------------------------------------------------------------*/
547 struct mirror {
548 atomic_t error_count;
549 struct dm_dev *dev;
550 sector_t offset;
553 struct mirror_set {
554 struct dm_target *ti;
555 struct list_head list;
556 struct region_hash rh;
557 struct kcopyd_client *kcopyd_client;
559 spinlock_t lock; /* protects the next two lists */
560 struct bio_list reads;
561 struct bio_list writes;
563 /* recovery */
564 region_t nr_regions;
565 int in_sync;
567 struct mirror *default_mirror; /* Default mirror */
569 unsigned int nr_mirrors;
570 struct mirror mirror[0];
574 * Every mirror should look like this one.
576 #define DEFAULT_MIRROR 0
579 * This is yucky. We squirrel the mirror_set struct away inside
580 * bi_next for write buffers. This is safe since the bh
581 * doesn't get submitted to the lower levels of block layer.
583 static struct mirror_set *bio_get_ms(struct bio *bio)
585 return (struct mirror_set *) bio->bi_next;
588 static void bio_set_ms(struct bio *bio, struct mirror_set *ms)
590 bio->bi_next = (struct bio *) ms;
593 /*-----------------------------------------------------------------
594 * Recovery.
596 * When a mirror is first activated we may find that some regions
597 * are in the no-sync state. We have to recover these by
598 * recopying from the default mirror to all the others.
599 *---------------------------------------------------------------*/
600 static void recovery_complete(int read_err, unsigned int write_err,
601 void *context)
603 struct region *reg = (struct region *) context;
605 /* FIXME: better error handling */
606 rh_recovery_end(reg, read_err || write_err);
609 static int recover(struct mirror_set *ms, struct region *reg)
611 int r;
612 unsigned int i;
613 struct io_region from, to[KCOPYD_MAX_REGIONS], *dest;
614 struct mirror *m;
615 unsigned long flags = 0;
617 /* fill in the source */
618 m = ms->default_mirror;
619 from.bdev = m->dev->bdev;
620 from.sector = m->offset + region_to_sector(reg->rh, reg->key);
621 if (reg->key == (ms->nr_regions - 1)) {
623 * The final region may be smaller than
624 * region_size.
626 from.count = ms->ti->len & (reg->rh->region_size - 1);
627 if (!from.count)
628 from.count = reg->rh->region_size;
629 } else
630 from.count = reg->rh->region_size;
632 /* fill in the destinations */
633 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
634 if (&ms->mirror[i] == ms->default_mirror)
635 continue;
637 m = ms->mirror + i;
638 dest->bdev = m->dev->bdev;
639 dest->sector = m->offset + region_to_sector(reg->rh, reg->key);
640 dest->count = from.count;
641 dest++;
644 /* hand to kcopyd */
645 set_bit(KCOPYD_IGNORE_ERROR, &flags);
646 r = kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to, flags,
647 recovery_complete, reg);
649 return r;
652 static void do_recovery(struct mirror_set *ms)
654 int r;
655 struct region *reg;
656 struct dirty_log *log = ms->rh.log;
659 * Start quiescing some regions.
661 rh_recovery_prepare(&ms->rh);
664 * Copy any already quiesced regions.
666 while ((reg = rh_recovery_start(&ms->rh))) {
667 r = recover(ms, reg);
668 if (r)
669 rh_recovery_end(reg, 0);
673 * Update the in sync flag.
675 if (!ms->in_sync &&
676 (log->type->get_sync_count(log) == ms->nr_regions)) {
677 /* the sync is complete */
678 dm_table_event(ms->ti->table);
679 ms->in_sync = 1;
683 /*-----------------------------------------------------------------
684 * Reads
685 *---------------------------------------------------------------*/
686 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
688 /* FIXME: add read balancing */
689 return ms->default_mirror;
693 * remap a buffer to a particular mirror.
695 static void map_bio(struct mirror_set *ms, struct mirror *m, struct bio *bio)
697 bio->bi_bdev = m->dev->bdev;
698 bio->bi_sector = m->offset + (bio->bi_sector - ms->ti->begin);
701 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
703 region_t region;
704 struct bio *bio;
705 struct mirror *m;
707 while ((bio = bio_list_pop(reads))) {
708 region = bio_to_region(&ms->rh, bio);
711 * We can only read balance if the region is in sync.
713 if (rh_in_sync(&ms->rh, region, 0))
714 m = choose_mirror(ms, bio->bi_sector);
715 else
716 m = ms->default_mirror;
718 map_bio(ms, m, bio);
719 generic_make_request(bio);
723 /*-----------------------------------------------------------------
724 * Writes.
726 * We do different things with the write io depending on the
727 * state of the region that it's in:
729 * SYNC: increment pending, use kcopyd to write to *all* mirrors
730 * RECOVERING: delay the io until recovery completes
731 * NOSYNC: increment pending, just write to the default mirror
732 *---------------------------------------------------------------*/
733 static void write_callback(unsigned long error, void *context)
735 unsigned int i;
736 int uptodate = 1;
737 struct bio *bio = (struct bio *) context;
738 struct mirror_set *ms;
740 ms = bio_get_ms(bio);
741 bio_set_ms(bio, NULL);
744 * NOTE: We don't decrement the pending count here,
745 * instead it is done by the targets endio function.
746 * This way we handle both writes to SYNC and NOSYNC
747 * regions with the same code.
750 if (error) {
752 * only error the io if all mirrors failed.
753 * FIXME: bogus
755 uptodate = 0;
756 for (i = 0; i < ms->nr_mirrors; i++)
757 if (!test_bit(i, &error)) {
758 uptodate = 1;
759 break;
762 bio_endio(bio, bio->bi_size, 0);
765 static void do_write(struct mirror_set *ms, struct bio *bio)
767 unsigned int i;
768 struct io_region io[KCOPYD_MAX_REGIONS+1];
769 struct mirror *m;
771 for (i = 0; i < ms->nr_mirrors; i++) {
772 m = ms->mirror + i;
774 io[i].bdev = m->dev->bdev;
775 io[i].sector = m->offset + (bio->bi_sector - ms->ti->begin);
776 io[i].count = bio->bi_size >> 9;
779 bio_set_ms(bio, ms);
780 dm_io_async_bvec(ms->nr_mirrors, io, WRITE,
781 bio->bi_io_vec + bio->bi_idx,
782 write_callback, bio);
785 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
787 int state;
788 struct bio *bio;
789 struct bio_list sync, nosync, recover, *this_list = NULL;
791 if (!writes->head)
792 return;
795 * Classify each write.
797 bio_list_init(&sync);
798 bio_list_init(&nosync);
799 bio_list_init(&recover);
801 while ((bio = bio_list_pop(writes))) {
802 state = rh_state(&ms->rh, bio_to_region(&ms->rh, bio), 1);
803 switch (state) {
804 case RH_CLEAN:
805 case RH_DIRTY:
806 this_list = &sync;
807 break;
809 case RH_NOSYNC:
810 this_list = &nosync;
811 break;
813 case RH_RECOVERING:
814 this_list = &recover;
815 break;
818 bio_list_add(this_list, bio);
822 * Increment the pending counts for any regions that will
823 * be written to (writes to recover regions are going to
824 * be delayed).
826 rh_inc_pending(&ms->rh, &sync);
827 rh_inc_pending(&ms->rh, &nosync);
828 rh_flush(&ms->rh);
831 * Dispatch io.
833 while ((bio = bio_list_pop(&sync)))
834 do_write(ms, bio);
836 while ((bio = bio_list_pop(&recover)))
837 rh_delay(&ms->rh, bio);
839 while ((bio = bio_list_pop(&nosync))) {
840 map_bio(ms, ms->default_mirror, bio);
841 generic_make_request(bio);
845 /*-----------------------------------------------------------------
846 * kmirrord
847 *---------------------------------------------------------------*/
848 static LIST_HEAD(_mirror_sets);
849 static DECLARE_RWSEM(_mirror_sets_lock);
851 static void do_mirror(struct mirror_set *ms)
853 struct bio_list reads, writes;
855 spin_lock(&ms->lock);
856 reads = ms->reads;
857 writes = ms->writes;
858 bio_list_init(&ms->reads);
859 bio_list_init(&ms->writes);
860 spin_unlock(&ms->lock);
862 rh_update_states(&ms->rh);
863 do_recovery(ms);
864 do_reads(ms, &reads);
865 do_writes(ms, &writes);
868 static void do_work(void *ignored)
870 struct mirror_set *ms;
872 down_read(&_mirror_sets_lock);
873 list_for_each_entry (ms, &_mirror_sets, list)
874 do_mirror(ms);
875 up_read(&_mirror_sets_lock);
878 /*-----------------------------------------------------------------
879 * Target functions
880 *---------------------------------------------------------------*/
881 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
882 uint32_t region_size,
883 struct dm_target *ti,
884 struct dirty_log *dl)
886 size_t len;
887 struct mirror_set *ms = NULL;
889 if (array_too_big(sizeof(*ms), sizeof(ms->mirror[0]), nr_mirrors))
890 return NULL;
892 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
894 ms = kmalloc(len, GFP_KERNEL);
895 if (!ms) {
896 ti->error = "dm-mirror: Cannot allocate mirror context";
897 return NULL;
900 memset(ms, 0, len);
901 spin_lock_init(&ms->lock);
903 ms->ti = ti;
904 ms->nr_mirrors = nr_mirrors;
905 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
906 ms->in_sync = 0;
907 ms->default_mirror = &ms->mirror[DEFAULT_MIRROR];
909 if (rh_init(&ms->rh, ms, dl, region_size, ms->nr_regions)) {
910 ti->error = "dm-mirror: Error creating dirty region hash";
911 kfree(ms);
912 return NULL;
915 return ms;
918 static void free_context(struct mirror_set *ms, struct dm_target *ti,
919 unsigned int m)
921 while (m--)
922 dm_put_device(ti, ms->mirror[m].dev);
924 rh_exit(&ms->rh);
925 kfree(ms);
928 static inline int _check_region_size(struct dm_target *ti, uint32_t size)
930 return !(size % (PAGE_SIZE >> 9) || (size & (size - 1)) ||
931 size > ti->len);
934 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
935 unsigned int mirror, char **argv)
937 unsigned long long offset;
939 if (sscanf(argv[1], "%llu", &offset) != 1) {
940 ti->error = "dm-mirror: Invalid offset";
941 return -EINVAL;
944 if (dm_get_device(ti, argv[0], offset, ti->len,
945 dm_table_get_mode(ti->table),
946 &ms->mirror[mirror].dev)) {
947 ti->error = "dm-mirror: Device lookup failure";
948 return -ENXIO;
951 ms->mirror[mirror].offset = offset;
953 return 0;
956 static int add_mirror_set(struct mirror_set *ms)
958 down_write(&_mirror_sets_lock);
959 list_add_tail(&ms->list, &_mirror_sets);
960 up_write(&_mirror_sets_lock);
961 wake();
963 return 0;
966 static void del_mirror_set(struct mirror_set *ms)
968 down_write(&_mirror_sets_lock);
969 list_del(&ms->list);
970 up_write(&_mirror_sets_lock);
974 * Create dirty log: log_type #log_params <log_params>
976 static struct dirty_log *create_dirty_log(struct dm_target *ti,
977 unsigned int argc, char **argv,
978 unsigned int *args_used)
980 unsigned int param_count;
981 struct dirty_log *dl;
983 if (argc < 2) {
984 ti->error = "dm-mirror: Insufficient mirror log arguments";
985 return NULL;
988 if (sscanf(argv[1], "%u", &param_count) != 1) {
989 ti->error = "dm-mirror: Invalid mirror log argument count";
990 return NULL;
993 *args_used = 2 + param_count;
995 if (argc < *args_used) {
996 ti->error = "dm-mirror: Insufficient mirror log arguments";
997 return NULL;
1000 dl = dm_create_dirty_log(argv[0], ti, param_count, argv + 2);
1001 if (!dl) {
1002 ti->error = "dm-mirror: Error creating mirror dirty log";
1003 return NULL;
1006 if (!_check_region_size(ti, dl->type->get_region_size(dl))) {
1007 ti->error = "dm-mirror: Invalid region size";
1008 dm_destroy_dirty_log(dl);
1009 return NULL;
1012 return dl;
1016 * Construct a mirror mapping:
1018 * log_type #log_params <log_params>
1019 * #mirrors [mirror_path offset]{2,}
1021 * log_type is "core" or "disk"
1022 * #log_params is between 1 and 3
1024 #define DM_IO_PAGES 64
1025 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1027 int r;
1028 unsigned int nr_mirrors, m, args_used;
1029 struct mirror_set *ms;
1030 struct dirty_log *dl;
1032 dl = create_dirty_log(ti, argc, argv, &args_used);
1033 if (!dl)
1034 return -EINVAL;
1036 argv += args_used;
1037 argc -= args_used;
1039 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1040 nr_mirrors < 2 || nr_mirrors > KCOPYD_MAX_REGIONS + 1) {
1041 ti->error = "dm-mirror: Invalid number of mirrors";
1042 dm_destroy_dirty_log(dl);
1043 return -EINVAL;
1046 argv++, argc--;
1048 if (argc != nr_mirrors * 2) {
1049 ti->error = "dm-mirror: Wrong number of mirror arguments";
1050 dm_destroy_dirty_log(dl);
1051 return -EINVAL;
1054 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1055 if (!ms) {
1056 dm_destroy_dirty_log(dl);
1057 return -ENOMEM;
1060 /* Get the mirror parameter sets */
1061 for (m = 0; m < nr_mirrors; m++) {
1062 r = get_mirror(ms, ti, m, argv);
1063 if (r) {
1064 free_context(ms, ti, m);
1065 return r;
1067 argv += 2;
1068 argc -= 2;
1071 ti->private = ms;
1072 ti->split_io = ms->rh.region_size;
1074 r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
1075 if (r) {
1076 free_context(ms, ti, ms->nr_mirrors);
1077 return r;
1080 add_mirror_set(ms);
1081 return 0;
1084 static void mirror_dtr(struct dm_target *ti)
1086 struct mirror_set *ms = (struct mirror_set *) ti->private;
1088 del_mirror_set(ms);
1089 kcopyd_client_destroy(ms->kcopyd_client);
1090 free_context(ms, ti, ms->nr_mirrors);
1093 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
1095 int should_wake = 0;
1096 struct bio_list *bl;
1098 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
1099 spin_lock(&ms->lock);
1100 should_wake = !(bl->head);
1101 bio_list_add(bl, bio);
1102 spin_unlock(&ms->lock);
1104 if (should_wake)
1105 wake();
1109 * Mirror mapping function
1111 static int mirror_map(struct dm_target *ti, struct bio *bio,
1112 union map_info *map_context)
1114 int r, rw = bio_rw(bio);
1115 struct mirror *m;
1116 struct mirror_set *ms = ti->private;
1118 map_context->ll = bio->bi_sector >> ms->rh.region_shift;
1120 if (rw == WRITE) {
1121 queue_bio(ms, bio, rw);
1122 return 0;
1125 r = ms->rh.log->type->in_sync(ms->rh.log,
1126 bio_to_region(&ms->rh, bio), 0);
1127 if (r < 0 && r != -EWOULDBLOCK)
1128 return r;
1130 if (r == -EWOULDBLOCK) /* FIXME: ugly */
1131 r = 0;
1134 * We don't want to fast track a recovery just for a read
1135 * ahead. So we just let it silently fail.
1136 * FIXME: get rid of this.
1138 if (!r && rw == READA)
1139 return -EIO;
1141 if (!r) {
1142 /* Pass this io over to the daemon */
1143 queue_bio(ms, bio, rw);
1144 return 0;
1147 m = choose_mirror(ms, bio->bi_sector);
1148 if (!m)
1149 return -EIO;
1151 map_bio(ms, m, bio);
1152 return 1;
1155 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1156 int error, union map_info *map_context)
1158 int rw = bio_rw(bio);
1159 struct mirror_set *ms = (struct mirror_set *) ti->private;
1160 region_t region = map_context->ll;
1163 * We need to dec pending if this was a write.
1165 if (rw == WRITE)
1166 rh_dec(&ms->rh, region);
1168 return 0;
1171 static void mirror_postsuspend(struct dm_target *ti)
1173 struct mirror_set *ms = (struct mirror_set *) ti->private;
1174 struct dirty_log *log = ms->rh.log;
1176 rh_stop_recovery(&ms->rh);
1177 if (log->type->suspend && log->type->suspend(log))
1178 /* FIXME: need better error handling */
1179 DMWARN("log suspend failed");
1182 static void mirror_resume(struct dm_target *ti)
1184 struct mirror_set *ms = (struct mirror_set *) ti->private;
1185 struct dirty_log *log = ms->rh.log;
1186 if (log->type->resume && log->type->resume(log))
1187 /* FIXME: need better error handling */
1188 DMWARN("log resume failed");
1189 rh_start_recovery(&ms->rh);
1192 static int mirror_status(struct dm_target *ti, status_type_t type,
1193 char *result, unsigned int maxlen)
1195 unsigned int m, sz;
1196 struct mirror_set *ms = (struct mirror_set *) ti->private;
1198 sz = ms->rh.log->type->status(ms->rh.log, type, result, maxlen);
1200 switch (type) {
1201 case STATUSTYPE_INFO:
1202 DMEMIT("%d ", ms->nr_mirrors);
1203 for (m = 0; m < ms->nr_mirrors; m++)
1204 DMEMIT("%s ", ms->mirror[m].dev->name);
1206 DMEMIT("%llu/%llu",
1207 (unsigned long long)ms->rh.log->type->
1208 get_sync_count(ms->rh.log),
1209 (unsigned long long)ms->nr_regions);
1210 break;
1212 case STATUSTYPE_TABLE:
1213 DMEMIT("%d ", ms->nr_mirrors);
1214 for (m = 0; m < ms->nr_mirrors; m++)
1215 DMEMIT("%s %llu ", ms->mirror[m].dev->name,
1216 (unsigned long long)ms->mirror[m].offset);
1219 return 0;
1222 static struct target_type mirror_target = {
1223 .name = "mirror",
1224 .version = {1, 0, 1},
1225 .module = THIS_MODULE,
1226 .ctr = mirror_ctr,
1227 .dtr = mirror_dtr,
1228 .map = mirror_map,
1229 .end_io = mirror_end_io,
1230 .postsuspend = mirror_postsuspend,
1231 .resume = mirror_resume,
1232 .status = mirror_status,
1235 static int __init dm_mirror_init(void)
1237 int r;
1239 r = dm_dirty_log_init();
1240 if (r)
1241 return r;
1243 _kmirrord_wq = create_singlethread_workqueue("kmirrord");
1244 if (!_kmirrord_wq) {
1245 DMERR("couldn't start kmirrord");
1246 dm_dirty_log_exit();
1247 return r;
1249 INIT_WORK(&_kmirrord_work, do_work, NULL);
1251 r = dm_register_target(&mirror_target);
1252 if (r < 0) {
1253 DMERR("%s: Failed to register mirror target",
1254 mirror_target.name);
1255 dm_dirty_log_exit();
1256 destroy_workqueue(_kmirrord_wq);
1259 return r;
1262 static void __exit dm_mirror_exit(void)
1264 int r;
1266 r = dm_unregister_target(&mirror_target);
1267 if (r < 0)
1268 DMERR("%s: unregister failed %d", mirror_target.name, r);
1270 destroy_workqueue(_kmirrord_wq);
1271 dm_dirty_log_exit();
1274 /* Module hooks */
1275 module_init(dm_mirror_init);
1276 module_exit(dm_mirror_exit);
1278 MODULE_DESCRIPTION(DM_NAME " mirror target");
1279 MODULE_AUTHOR("Joe Thornber");
1280 MODULE_LICENSE("GPL");