dm-raid1.c revision 1da177e4c3f41524e886b7f1b8a0c1fc7321cac 
1/*
2 * Copyright (C) 2003 Sistina Software Limited.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
8#include "dm-bio-list.h"
9#include "dm-io.h"
10#include "dm-log.h"
11#include "kcopyd.h"
12
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>
22
23static struct workqueue_struct *_kmirrord_wq;
24static struct work_struct _kmirrord_work;
25
26static inline void wake(void)
27{
28	queue_work(_kmirrord_wq, &_kmirrord_work);
29}
30
31/*-----------------------------------------------------------------
32 * Region hash
33 *
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.
37 *
38 * In addition to being present in the hash table a region _may_
39 * be present on one of three lists.
40 *
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.
45 *
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.
50 *
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.
55 *
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.
61 *
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 *---------------------------------------------------------------*/
67struct mirror_set;
68struct region_hash {
69	struct mirror_set *ms;
70	uint32_t region_size;
71	unsigned region_shift;
72
73	/* holds persistent region state */
74	struct dirty_log *log;
75
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;
82
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;
88};
89
90enum {
91	RH_CLEAN,
92	RH_DIRTY,
93	RH_NOSYNC,
94	RH_RECOVERING
95};
96
97struct region {
98	struct region_hash *rh;	/* FIXME: can we get rid of this ? */
99	region_t key;
100	int state;
101
102	struct list_head hash_list;
103	struct list_head list;
104
105	atomic_t pending;
106	struct bio_list delayed_bios;
107};
108
109/*
110 * Conversion fns
111 */
112static inline region_t bio_to_region(struct region_hash *rh, struct bio *bio)
113{
114	return bio->bi_sector >> rh->region_shift;
115}
116
117static inline sector_t region_to_sector(struct region_hash *rh, region_t region)
118{
119	return region << rh->region_shift;
120}
121
122/* FIXME move this */
123static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw);
124
125static void *region_alloc(unsigned int __nocast gfp_mask, void *pool_data)
126{
127	return kmalloc(sizeof(struct region), gfp_mask);
128}
129
130static void region_free(void *element, void *pool_data)
131{
132	kfree(element);
133}
134
135#define MIN_REGIONS 64
136#define MAX_RECOVERY 1
137static int rh_init(struct region_hash *rh, struct mirror_set *ms,
138		   struct dirty_log *log, uint32_t region_size,
139		   region_t nr_regions)
140{
141	unsigned int nr_buckets, max_buckets;
142	size_t i;
143
144	/*
145	 * Calculate a suitable number of buckets for our hash
146	 * table.
147	 */
148	max_buckets = nr_regions >> 6;
149	for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
150		;
151	nr_buckets >>= 1;
152
153	rh->ms = ms;
154	rh->log = log;
155	rh->region_size = region_size;
156	rh->region_shift = ffs(region_size) - 1;
157	rwlock_init(&rh->hash_lock);
158	rh->mask = nr_buckets - 1;
159	rh->nr_buckets = nr_buckets;
160
161	rh->buckets = vmalloc(nr_buckets * sizeof(*rh->buckets));
162	if (!rh->buckets) {
163		DMERR("unable to allocate region hash memory");
164		return -ENOMEM;
165	}
166
167	for (i = 0; i < nr_buckets; i++)
168		INIT_LIST_HEAD(rh->buckets + i);
169
170	spin_lock_init(&rh->region_lock);
171	sema_init(&rh->recovery_count, 0);
172	INIT_LIST_HEAD(&rh->clean_regions);
173	INIT_LIST_HEAD(&rh->quiesced_regions);
174	INIT_LIST_HEAD(&rh->recovered_regions);
175
176	rh->region_pool = mempool_create(MIN_REGIONS, region_alloc,
177					 region_free, NULL);
178	if (!rh->region_pool) {
179		vfree(rh->buckets);
180		rh->buckets = NULL;
181		return -ENOMEM;
182	}
183
184	return 0;
185}
186
187static void rh_exit(struct region_hash *rh)
188{
189	unsigned int h;
190	struct region *reg, *nreg;
191
192	BUG_ON(!list_empty(&rh->quiesced_regions));
193	for (h = 0; h < rh->nr_buckets; h++) {
194		list_for_each_entry_safe(reg, nreg, rh->buckets + h, hash_list) {
195			BUG_ON(atomic_read(&reg->pending));
196			mempool_free(reg, rh->region_pool);
197		}
198	}
199
200	if (rh->log)
201		dm_destroy_dirty_log(rh->log);
202	if (rh->region_pool)
203		mempool_destroy(rh->region_pool);
204	vfree(rh->buckets);
205}
206
207#define RH_HASH_MULT 2654435387U
208
209static inline unsigned int rh_hash(struct region_hash *rh, region_t region)
210{
211	return (unsigned int) ((region * RH_HASH_MULT) >> 12) & rh->mask;
212}
213
214static struct region *__rh_lookup(struct region_hash *rh, region_t region)
215{
216	struct region *reg;
217
218	list_for_each_entry (reg, rh->buckets + rh_hash(rh, region), hash_list)
219		if (reg->key == region)
220			return reg;
221
222	return NULL;
223}
224
225static void __rh_insert(struct region_hash *rh, struct region *reg)
226{
227	unsigned int h = rh_hash(rh, reg->key);
228	list_add(&reg->hash_list, rh->buckets + h);
229}
230
231static struct region *__rh_alloc(struct region_hash *rh, region_t region)
232{
233	struct region *reg, *nreg;
234
235	read_unlock(&rh->hash_lock);
236	nreg = mempool_alloc(rh->region_pool, GFP_NOIO);
237	nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
238		RH_CLEAN : RH_NOSYNC;
239	nreg->rh = rh;
240	nreg->key = region;
241
242	INIT_LIST_HEAD(&nreg->list);
243
244	atomic_set(&nreg->pending, 0);
245	bio_list_init(&nreg->delayed_bios);
246	write_lock_irq(&rh->hash_lock);
247
248	reg = __rh_lookup(rh, region);
249	if (reg)
250		/* we lost the race */
251		mempool_free(nreg, rh->region_pool);
252
253	else {
254		__rh_insert(rh, nreg);
255		if (nreg->state == RH_CLEAN) {
256			spin_lock(&rh->region_lock);
257			list_add(&nreg->list, &rh->clean_regions);
258			spin_unlock(&rh->region_lock);
259		}
260		reg = nreg;
261	}
262	write_unlock_irq(&rh->hash_lock);
263	read_lock(&rh->hash_lock);
264
265	return reg;
266}
267
268static inline struct region *__rh_find(struct region_hash *rh, region_t region)
269{
270	struct region *reg;
271
272	reg = __rh_lookup(rh, region);
273	if (!reg)
274		reg = __rh_alloc(rh, region);
275
276	return reg;
277}
278
279static int rh_state(struct region_hash *rh, region_t region, int may_block)
280{
281	int r;
282	struct region *reg;
283
284	read_lock(&rh->hash_lock);
285	reg = __rh_lookup(rh, region);
286	read_unlock(&rh->hash_lock);
287
288	if (reg)
289		return reg->state;
290
291	/*
292	 * The region wasn't in the hash, so we fall back to the
293	 * dirty log.
294	 */
295	r = rh->log->type->in_sync(rh->log, region, may_block);
296
297	/*
298	 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
299	 * taken as a RH_NOSYNC
300	 */
301	return r == 1 ? RH_CLEAN : RH_NOSYNC;
302}
303
304static inline int rh_in_sync(struct region_hash *rh,
305			     region_t region, int may_block)
306{
307	int state = rh_state(rh, region, may_block);
308	return state == RH_CLEAN || state == RH_DIRTY;
309}
310
311static void dispatch_bios(struct mirror_set *ms, struct bio_list *bio_list)
312{
313	struct bio *bio;
314
315	while ((bio = bio_list_pop(bio_list))) {
316		queue_bio(ms, bio, WRITE);
317	}
318}
319
320static void rh_update_states(struct region_hash *rh)
321{
322	struct region *reg, *next;
323
324	LIST_HEAD(clean);
325	LIST_HEAD(recovered);
326
327	/*
328	 * Quickly grab the lists.
329	 */
330	write_lock_irq(&rh->hash_lock);
331	spin_lock(&rh->region_lock);
332	if (!list_empty(&rh->clean_regions)) {
333		list_splice(&rh->clean_regions, &clean);
334		INIT_LIST_HEAD(&rh->clean_regions);
335
336		list_for_each_entry (reg, &clean, list) {
337			rh->log->type->clear_region(rh->log, reg->key);
338			list_del(&reg->hash_list);
339		}
340	}
341
342	if (!list_empty(&rh->recovered_regions)) {
343		list_splice(&rh->recovered_regions, &recovered);
344		INIT_LIST_HEAD(&rh->recovered_regions);
345
346		list_for_each_entry (reg, &recovered, list)
347			list_del(&reg->hash_list);
348	}
349	spin_unlock(&rh->region_lock);
350	write_unlock_irq(&rh->hash_lock);
351
352	/*
353	 * All the regions on the recovered and clean lists have
354	 * now been pulled out of the system, so no need to do
355	 * any more locking.
356	 */
357	list_for_each_entry_safe (reg, next, &recovered, list) {
358		rh->log->type->clear_region(rh->log, reg->key);
359		rh->log->type->complete_resync_work(rh->log, reg->key, 1);
360		dispatch_bios(rh->ms, &reg->delayed_bios);
361		up(&rh->recovery_count);
362		mempool_free(reg, rh->region_pool);
363	}
364
365	if (!list_empty(&recovered))
366		rh->log->type->flush(rh->log);
367
368	list_for_each_entry_safe (reg, next, &clean, list)
369		mempool_free(reg, rh->region_pool);
370}
371
372static void rh_inc(struct region_hash *rh, region_t region)
373{
374	struct region *reg;
375
376	read_lock(&rh->hash_lock);
377	reg = __rh_find(rh, region);
378	if (reg->state == RH_CLEAN) {
379		rh->log->type->mark_region(rh->log, reg->key);
380
381		spin_lock_irq(&rh->region_lock);
382		reg->state = RH_DIRTY;
383		list_del_init(&reg->list);	/* take off the clean list */
384		spin_unlock_irq(&rh->region_lock);
385	}
386
387	atomic_inc(&reg->pending);
388	read_unlock(&rh->hash_lock);
389}
390
391static void rh_inc_pending(struct region_hash *rh, struct bio_list *bios)
392{
393	struct bio *bio;
394
395	for (bio = bios->head; bio; bio = bio->bi_next)
396		rh_inc(rh, bio_to_region(rh, bio));
397}
398
399static void rh_dec(struct region_hash *rh, region_t region)
400{
401	unsigned long flags;
402	struct region *reg;
403	int should_wake = 0;
404
405	read_lock(&rh->hash_lock);
406	reg = __rh_lookup(rh, region);
407	read_unlock(&rh->hash_lock);
408
409	if (atomic_dec_and_test(&reg->pending)) {
410		spin_lock_irqsave(&rh->region_lock, flags);
411		if (reg->state == RH_RECOVERING) {
412			list_add_tail(&reg->list, &rh->quiesced_regions);
413		} else {
414			reg->state = RH_CLEAN;
415			list_add(&reg->list, &rh->clean_regions);
416		}
417		spin_unlock_irqrestore(&rh->region_lock, flags);
418		should_wake = 1;
419	}
420
421	if (should_wake)
422		wake();
423}
424
425/*
426 * Starts quiescing a region in preparation for recovery.
427 */
428static int __rh_recovery_prepare(struct region_hash *rh)
429{
430	int r;
431	struct region *reg;
432	region_t region;
433
434	/*
435	 * Ask the dirty log what's next.
436	 */
437	r = rh->log->type->get_resync_work(rh->log, &region);
438	if (r <= 0)
439		return r;
440
441	/*
442	 * Get this region, and start it quiescing by setting the
443	 * recovering flag.
444	 */
445	read_lock(&rh->hash_lock);
446	reg = __rh_find(rh, region);
447	read_unlock(&rh->hash_lock);
448
449	spin_lock_irq(&rh->region_lock);
450	reg->state = RH_RECOVERING;
451
452	/* Already quiesced ? */
453	if (atomic_read(&reg->pending))
454		list_del_init(&reg->list);
455
456	else {
457		list_del_init(&reg->list);
458		list_add(&reg->list, &rh->quiesced_regions);
459	}
460	spin_unlock_irq(&rh->region_lock);
461
462	return 1;
463}
464
465static void rh_recovery_prepare(struct region_hash *rh)
466{
467	while (!down_trylock(&rh->recovery_count))
468		if (__rh_recovery_prepare(rh) <= 0) {
469			up(&rh->recovery_count);
470			break;
471		}
472}
473
474/*
475 * Returns any quiesced regions.
476 */
477static struct region *rh_recovery_start(struct region_hash *rh)
478{
479	struct region *reg = NULL;
480
481	spin_lock_irq(&rh->region_lock);
482	if (!list_empty(&rh->quiesced_regions)) {
483		reg = list_entry(rh->quiesced_regions.next,
484				 struct region, list);
485		list_del_init(&reg->list);	/* remove from the quiesced list */
486	}
487	spin_unlock_irq(&rh->region_lock);
488
489	return reg;
490}
491
492/* FIXME: success ignored for now */
493static void rh_recovery_end(struct region *reg, int success)
494{
495	struct region_hash *rh = reg->rh;
496
497	spin_lock_irq(&rh->region_lock);
498	list_add(&reg->list, &reg->rh->recovered_regions);
499	spin_unlock_irq(&rh->region_lock);
500
501	wake();
502}
503
504static void rh_flush(struct region_hash *rh)
505{
506	rh->log->type->flush(rh->log);
507}
508
509static void rh_delay(struct region_hash *rh, struct bio *bio)
510{
511	struct region *reg;
512
513	read_lock(&rh->hash_lock);
514	reg = __rh_find(rh, bio_to_region(rh, bio));
515	bio_list_add(&reg->delayed_bios, bio);
516	read_unlock(&rh->hash_lock);
517}
518
519static void rh_stop_recovery(struct region_hash *rh)
520{
521	int i;
522
523	/* wait for any recovering regions */
524	for (i = 0; i < MAX_RECOVERY; i++)
525		down(&rh->recovery_count);
526}
527
528static void rh_start_recovery(struct region_hash *rh)
529{
530	int i;
531
532	for (i = 0; i < MAX_RECOVERY; i++)
533		up(&rh->recovery_count);
534
535	wake();
536}
537
538/*-----------------------------------------------------------------
539 * Mirror set structures.
540 *---------------------------------------------------------------*/
541struct mirror {
542	atomic_t error_count;
543	struct dm_dev *dev;
544	sector_t offset;
545};
546
547struct mirror_set {
548	struct dm_target *ti;
549	struct list_head list;
550	struct region_hash rh;
551	struct kcopyd_client *kcopyd_client;
552
553	spinlock_t lock;	/* protects the next two lists */
554	struct bio_list reads;
555	struct bio_list writes;
556
557	/* recovery */
558	region_t nr_regions;
559	int in_sync;
560
561	unsigned int nr_mirrors;
562	struct mirror mirror[0];
563};
564
565/*
566 * Every mirror should look like this one.
567 */
568#define DEFAULT_MIRROR 0
569
570/*
571 * This is yucky.  We squirrel the mirror_set struct away inside
572 * bi_next for write buffers.  This is safe since the bh
573 * doesn't get submitted to the lower levels of block layer.
574 */
575static struct mirror_set *bio_get_ms(struct bio *bio)
576{
577	return (struct mirror_set *) bio->bi_next;
578}
579
580static void bio_set_ms(struct bio *bio, struct mirror_set *ms)
581{
582	bio->bi_next = (struct bio *) ms;
583}
584
585/*-----------------------------------------------------------------
586 * Recovery.
587 *
588 * When a mirror is first activated we may find that some regions
589 * are in the no-sync state.  We have to recover these by
590 * recopying from the default mirror to all the others.
591 *---------------------------------------------------------------*/
592static void recovery_complete(int read_err, unsigned int write_err,
593			      void *context)
594{
595	struct region *reg = (struct region *) context;
596
597	/* FIXME: better error handling */
598	rh_recovery_end(reg, read_err || write_err);
599}
600
601static int recover(struct mirror_set *ms, struct region *reg)
602{
603	int r;
604	unsigned int i;
605	struct io_region from, to[KCOPYD_MAX_REGIONS], *dest;
606	struct mirror *m;
607	unsigned long flags = 0;
608
609	/* fill in the source */
610	m = ms->mirror + DEFAULT_MIRROR;
611	from.bdev = m->dev->bdev;
612	from.sector = m->offset + region_to_sector(reg->rh, reg->key);
613	if (reg->key == (ms->nr_regions - 1)) {
614		/*
615		 * The final region may be smaller than
616		 * region_size.
617		 */
618		from.count = ms->ti->len & (reg->rh->region_size - 1);
619		if (!from.count)
620			from.count = reg->rh->region_size;
621	} else
622		from.count = reg->rh->region_size;
623
624	/* fill in the destinations */
625	for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
626		if (i == DEFAULT_MIRROR)
627			continue;
628
629		m = ms->mirror + i;
630		dest->bdev = m->dev->bdev;
631		dest->sector = m->offset + region_to_sector(reg->rh, reg->key);
632		dest->count = from.count;
633		dest++;
634	}
635
636	/* hand to kcopyd */
637	set_bit(KCOPYD_IGNORE_ERROR, &flags);
638	r = kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to, flags,
639			recovery_complete, reg);
640
641	return r;
642}
643
644static void do_recovery(struct mirror_set *ms)
645{
646	int r;
647	struct region *reg;
648	struct dirty_log *log = ms->rh.log;
649
650	/*
651	 * Start quiescing some regions.
652	 */
653	rh_recovery_prepare(&ms->rh);
654
655	/*
656	 * Copy any already quiesced regions.
657	 */
658	while ((reg = rh_recovery_start(&ms->rh))) {
659		r = recover(ms, reg);
660		if (r)
661			rh_recovery_end(reg, 0);
662	}
663
664	/*
665	 * Update the in sync flag.
666	 */
667	if (!ms->in_sync &&
668	    (log->type->get_sync_count(log) == ms->nr_regions)) {
669		/* the sync is complete */
670		dm_table_event(ms->ti->table);
671		ms->in_sync = 1;
672	}
673}
674
675/*-----------------------------------------------------------------
676 * Reads
677 *---------------------------------------------------------------*/
678static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
679{
680	/* FIXME: add read balancing */
681	return ms->mirror + DEFAULT_MIRROR;
682}
683
684/*
685 * remap a buffer to a particular mirror.
686 */
687static void map_bio(struct mirror_set *ms, struct mirror *m, struct bio *bio)
688{
689	bio->bi_bdev = m->dev->bdev;
690	bio->bi_sector = m->offset + (bio->bi_sector - ms->ti->begin);
691}
692
693static void do_reads(struct mirror_set *ms, struct bio_list *reads)
694{
695	region_t region;
696	struct bio *bio;
697	struct mirror *m;
698
699	while ((bio = bio_list_pop(reads))) {
700		region = bio_to_region(&ms->rh, bio);
701
702		/*
703		 * We can only read balance if the region is in sync.
704		 */
705		if (rh_in_sync(&ms->rh, region, 0))
706			m = choose_mirror(ms, bio->bi_sector);
707		else
708			m = ms->mirror + DEFAULT_MIRROR;
709
710		map_bio(ms, m, bio);
711		generic_make_request(bio);
712	}
713}
714
715/*-----------------------------------------------------------------
716 * Writes.
717 *
718 * We do different things with the write io depending on the
719 * state of the region that it's in:
720 *
721 * SYNC: 	increment pending, use kcopyd to write to *all* mirrors
722 * RECOVERING:	delay the io until recovery completes
723 * NOSYNC:	increment pending, just write to the default mirror
724 *---------------------------------------------------------------*/
725static void write_callback(unsigned long error, void *context)
726{
727	unsigned int i;
728	int uptodate = 1;
729	struct bio *bio = (struct bio *) context;
730	struct mirror_set *ms;
731
732	ms = bio_get_ms(bio);
733	bio_set_ms(bio, NULL);
734
735	/*
736	 * NOTE: We don't decrement the pending count here,
737	 * instead it is done by the targets endio function.
738	 * This way we handle both writes to SYNC and NOSYNC
739	 * regions with the same code.
740	 */
741
742	if (error) {
743		/*
744		 * only error the io if all mirrors failed.
745		 * FIXME: bogus
746		 */
747		uptodate = 0;
748		for (i = 0; i < ms->nr_mirrors; i++)
749			if (!test_bit(i, &error)) {
750				uptodate = 1;
751				break;
752			}
753	}
754	bio_endio(bio, bio->bi_size, 0);
755}
756
757static void do_write(struct mirror_set *ms, struct bio *bio)
758{
759	unsigned int i;
760	struct io_region io[KCOPYD_MAX_REGIONS+1];
761	struct mirror *m;
762
763	for (i = 0; i < ms->nr_mirrors; i++) {
764		m = ms->mirror + i;
765
766		io[i].bdev = m->dev->bdev;
767		io[i].sector = m->offset + (bio->bi_sector - ms->ti->begin);
768		io[i].count = bio->bi_size >> 9;
769	}
770
771	bio_set_ms(bio, ms);
772	dm_io_async_bvec(ms->nr_mirrors, io, WRITE,
773			 bio->bi_io_vec + bio->bi_idx,
774			 write_callback, bio);
775}
776
777static void do_writes(struct mirror_set *ms, struct bio_list *writes)
778{
779	int state;
780	struct bio *bio;
781	struct bio_list sync, nosync, recover, *this_list = NULL;
782
783	if (!writes->head)
784		return;
785
786	/*
787	 * Classify each write.
788	 */
789	bio_list_init(&sync);
790	bio_list_init(&nosync);
791	bio_list_init(&recover);
792
793	while ((bio = bio_list_pop(writes))) {
794		state = rh_state(&ms->rh, bio_to_region(&ms->rh, bio), 1);
795		switch (state) {
796		case RH_CLEAN:
797		case RH_DIRTY:
798			this_list = &sync;
799			break;
800
801		case RH_NOSYNC:
802			this_list = &nosync;
803			break;
804
805		case RH_RECOVERING:
806			this_list = &recover;
807			break;
808		}
809
810		bio_list_add(this_list, bio);
811	}
812
813	/*
814	 * Increment the pending counts for any regions that will
815	 * be written to (writes to recover regions are going to
816	 * be delayed).
817	 */
818	rh_inc_pending(&ms->rh, &sync);
819	rh_inc_pending(&ms->rh, &nosync);
820	rh_flush(&ms->rh);
821
822	/*
823	 * Dispatch io.
824	 */
825	while ((bio = bio_list_pop(&sync)))
826		do_write(ms, bio);
827
828	while ((bio = bio_list_pop(&recover)))
829		rh_delay(&ms->rh, bio);
830
831	while ((bio = bio_list_pop(&nosync))) {
832		map_bio(ms, ms->mirror + DEFAULT_MIRROR, bio);
833		generic_make_request(bio);
834	}
835}
836
837/*-----------------------------------------------------------------
838 * kmirrord
839 *---------------------------------------------------------------*/
840static LIST_HEAD(_mirror_sets);
841static DECLARE_RWSEM(_mirror_sets_lock);
842
843static void do_mirror(struct mirror_set *ms)
844{
845	struct bio_list reads, writes;
846
847	spin_lock(&ms->lock);
848	reads = ms->reads;
849	writes = ms->writes;
850	bio_list_init(&ms->reads);
851	bio_list_init(&ms->writes);
852	spin_unlock(&ms->lock);
853
854	rh_update_states(&ms->rh);
855	do_recovery(ms);
856	do_reads(ms, &reads);
857	do_writes(ms, &writes);
858}
859
860static void do_work(void *ignored)
861{
862	struct mirror_set *ms;
863
864	down_read(&_mirror_sets_lock);
865	list_for_each_entry (ms, &_mirror_sets, list)
866		do_mirror(ms);
867	up_read(&_mirror_sets_lock);
868}
869
870/*-----------------------------------------------------------------
871 * Target functions
872 *---------------------------------------------------------------*/
873static struct mirror_set *alloc_context(unsigned int nr_mirrors,
874					uint32_t region_size,
875					struct dm_target *ti,
876					struct dirty_log *dl)
877{
878	size_t len;
879	struct mirror_set *ms = NULL;
880
881	if (array_too_big(sizeof(*ms), sizeof(ms->mirror[0]), nr_mirrors))
882		return NULL;
883
884	len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
885
886	ms = kmalloc(len, GFP_KERNEL);
887	if (!ms) {
888		ti->error = "dm-mirror: Cannot allocate mirror context";
889		return NULL;
890	}
891
892	memset(ms, 0, len);
893	spin_lock_init(&ms->lock);
894
895	ms->ti = ti;
896	ms->nr_mirrors = nr_mirrors;
897	ms->nr_regions = dm_sector_div_up(ti->len, region_size);
898	ms->in_sync = 0;
899
900	if (rh_init(&ms->rh, ms, dl, region_size, ms->nr_regions)) {
901		ti->error = "dm-mirror: Error creating dirty region hash";
902		kfree(ms);
903		return NULL;
904	}
905
906	return ms;
907}
908
909static void free_context(struct mirror_set *ms, struct dm_target *ti,
910			 unsigned int m)
911{
912	while (m--)
913		dm_put_device(ti, ms->mirror[m].dev);
914
915	rh_exit(&ms->rh);
916	kfree(ms);
917}
918
919static inline int _check_region_size(struct dm_target *ti, uint32_t size)
920{
921	return !(size % (PAGE_SIZE >> 9) || (size & (size - 1)) ||
922		 size > ti->len);
923}
924
925static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
926		      unsigned int mirror, char **argv)
927{
928	sector_t offset;
929
930	if (sscanf(argv[1], SECTOR_FORMAT, &offset) != 1) {
931		ti->error = "dm-mirror: Invalid offset";
932		return -EINVAL;
933	}
934
935	if (dm_get_device(ti, argv[0], offset, ti->len,
936			  dm_table_get_mode(ti->table),
937			  &ms->mirror[mirror].dev)) {
938		ti->error = "dm-mirror: Device lookup failure";
939		return -ENXIO;
940	}
941
942	ms->mirror[mirror].offset = offset;
943
944	return 0;
945}
946
947static int add_mirror_set(struct mirror_set *ms)
948{
949	down_write(&_mirror_sets_lock);
950	list_add_tail(&ms->list, &_mirror_sets);
951	up_write(&_mirror_sets_lock);
952	wake();
953
954	return 0;
955}
956
957static void del_mirror_set(struct mirror_set *ms)
958{
959	down_write(&_mirror_sets_lock);
960	list_del(&ms->list);
961	up_write(&_mirror_sets_lock);
962}
963
964/*
965 * Create dirty log: log_type #log_params <log_params>
966 */
967static struct dirty_log *create_dirty_log(struct dm_target *ti,
968					  unsigned int argc, char **argv,
969					  unsigned int *args_used)
970{
971	unsigned int param_count;
972	struct dirty_log *dl;
973
974	if (argc < 2) {
975		ti->error = "dm-mirror: Insufficient mirror log arguments";
976		return NULL;
977	}
978
979	if (sscanf(argv[1], "%u", &param_count) != 1) {
980		ti->error = "dm-mirror: Invalid mirror log argument count";
981		return NULL;
982	}
983
984	*args_used = 2 + param_count;
985
986	if (argc < *args_used) {
987		ti->error = "dm-mirror: Insufficient mirror log arguments";
988		return NULL;
989	}
990
991	dl = dm_create_dirty_log(argv[0], ti, param_count, argv + 2);
992	if (!dl) {
993		ti->error = "dm-mirror: Error creating mirror dirty log";
994		return NULL;
995	}
996
997	if (!_check_region_size(ti, dl->type->get_region_size(dl))) {
998		ti->error = "dm-mirror: Invalid region size";
999		dm_destroy_dirty_log(dl);
1000		return NULL;
1001	}
1002
1003	return dl;
1004}
1005
1006/*
1007 * Construct a mirror mapping:
1008 *
1009 * log_type #log_params <log_params>
1010 * #mirrors [mirror_path offset]{2,}
1011 *
1012 * log_type is "core" or "disk"
1013 * #log_params is between 1 and 3
1014 */
1015#define DM_IO_PAGES 64
1016static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1017{
1018	int r;
1019	unsigned int nr_mirrors, m, args_used;
1020	struct mirror_set *ms;
1021	struct dirty_log *dl;
1022
1023	dl = create_dirty_log(ti, argc, argv, &args_used);
1024	if (!dl)
1025		return -EINVAL;
1026
1027	argv += args_used;
1028	argc -= args_used;
1029
1030	if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1031	    nr_mirrors < 2 || nr_mirrors > KCOPYD_MAX_REGIONS + 1) {
1032		ti->error = "dm-mirror: Invalid number of mirrors";
1033		dm_destroy_dirty_log(dl);
1034		return -EINVAL;
1035	}
1036
1037	argv++, argc--;
1038
1039	if (argc != nr_mirrors * 2) {
1040		ti->error = "dm-mirror: Wrong number of mirror arguments";
1041		dm_destroy_dirty_log(dl);
1042		return -EINVAL;
1043	}
1044
1045	ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1046	if (!ms) {
1047		dm_destroy_dirty_log(dl);
1048		return -ENOMEM;
1049	}
1050
1051	/* Get the mirror parameter sets */
1052	for (m = 0; m < nr_mirrors; m++) {
1053		r = get_mirror(ms, ti, m, argv);
1054		if (r) {
1055			free_context(ms, ti, m);
1056			return r;
1057		}
1058		argv += 2;
1059		argc -= 2;
1060	}
1061
1062	ti->private = ms;
1063
1064	r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
1065	if (r) {
1066		free_context(ms, ti, ms->nr_mirrors);
1067		return r;
1068	}
1069
1070	add_mirror_set(ms);
1071	return 0;
1072}
1073
1074static void mirror_dtr(struct dm_target *ti)
1075{
1076	struct mirror_set *ms = (struct mirror_set *) ti->private;
1077
1078	del_mirror_set(ms);
1079	kcopyd_client_destroy(ms->kcopyd_client);
1080	free_context(ms, ti, ms->nr_mirrors);
1081}
1082
1083static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
1084{
1085	int should_wake = 0;
1086	struct bio_list *bl;
1087
1088	bl = (rw == WRITE) ? &ms->writes : &ms->reads;
1089	spin_lock(&ms->lock);
1090	should_wake = !(bl->head);
1091	bio_list_add(bl, bio);
1092	spin_unlock(&ms->lock);
1093
1094	if (should_wake)
1095		wake();
1096}
1097
1098/*
1099 * Mirror mapping function
1100 */
1101static int mirror_map(struct dm_target *ti, struct bio *bio,
1102		      union map_info *map_context)
1103{
1104	int r, rw = bio_rw(bio);
1105	struct mirror *m;
1106	struct mirror_set *ms = ti->private;
1107
1108	map_context->ll = bio->bi_sector >> ms->rh.region_shift;
1109
1110	if (rw == WRITE) {
1111		queue_bio(ms, bio, rw);
1112		return 0;
1113	}
1114
1115	r = ms->rh.log->type->in_sync(ms->rh.log,
1116				      bio_to_region(&ms->rh, bio), 0);
1117	if (r < 0 && r != -EWOULDBLOCK)
1118		return r;
1119
1120	if (r == -EWOULDBLOCK)	/* FIXME: ugly */
1121		r = 0;
1122
1123	/*
1124	 * We don't want to fast track a recovery just for a read
1125	 * ahead.  So we just let it silently fail.
1126	 * FIXME: get rid of this.
1127	 */
1128	if (!r && rw == READA)
1129		return -EIO;
1130
1131	if (!r) {
1132		/* Pass this io over to the daemon */
1133		queue_bio(ms, bio, rw);
1134		return 0;
1135	}
1136
1137	m = choose_mirror(ms, bio->bi_sector);
1138	if (!m)
1139		return -EIO;
1140
1141	map_bio(ms, m, bio);
1142	return 1;
1143}
1144
1145static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1146			 int error, union map_info *map_context)
1147{
1148	int rw = bio_rw(bio);
1149	struct mirror_set *ms = (struct mirror_set *) ti->private;
1150	region_t region = map_context->ll;
1151
1152	/*
1153	 * We need to dec pending if this was a write.
1154	 */
1155	if (rw == WRITE)
1156		rh_dec(&ms->rh, region);
1157
1158	return 0;
1159}
1160
1161static void mirror_postsuspend(struct dm_target *ti)
1162{
1163	struct mirror_set *ms = (struct mirror_set *) ti->private;
1164	struct dirty_log *log = ms->rh.log;
1165
1166	rh_stop_recovery(&ms->rh);
1167	if (log->type->suspend && log->type->suspend(log))
1168		/* FIXME: need better error handling */
1169		DMWARN("log suspend failed");
1170}
1171
1172static void mirror_resume(struct dm_target *ti)
1173{
1174	struct mirror_set *ms = (struct mirror_set *) ti->private;
1175	struct dirty_log *log = ms->rh.log;
1176	if (log->type->resume && log->type->resume(log))
1177		/* FIXME: need better error handling */
1178		DMWARN("log resume failed");
1179	rh_start_recovery(&ms->rh);
1180}
1181
1182static int mirror_status(struct dm_target *ti, status_type_t type,
1183			 char *result, unsigned int maxlen)
1184{
1185	unsigned int m, sz;
1186	struct mirror_set *ms = (struct mirror_set *) ti->private;
1187
1188	sz = ms->rh.log->type->status(ms->rh.log, type, result, maxlen);
1189
1190	switch (type) {
1191	case STATUSTYPE_INFO:
1192		DMEMIT("%d ", ms->nr_mirrors);
1193		for (m = 0; m < ms->nr_mirrors; m++)
1194			DMEMIT("%s ", ms->mirror[m].dev->name);
1195
1196		DMEMIT(SECTOR_FORMAT "/" SECTOR_FORMAT,
1197		       ms->rh.log->type->get_sync_count(ms->rh.log),
1198		       ms->nr_regions);
1199		break;
1200
1201	case STATUSTYPE_TABLE:
1202		DMEMIT("%d ", ms->nr_mirrors);
1203		for (m = 0; m < ms->nr_mirrors; m++)
1204			DMEMIT("%s " SECTOR_FORMAT " ",
1205			       ms->mirror[m].dev->name, ms->mirror[m].offset);
1206	}
1207
1208	return 0;
1209}
1210
1211static struct target_type mirror_target = {
1212	.name	 = "mirror",
1213	.version = {1, 0, 1},
1214	.module	 = THIS_MODULE,
1215	.ctr	 = mirror_ctr,
1216	.dtr	 = mirror_dtr,
1217	.map	 = mirror_map,
1218	.end_io	 = mirror_end_io,
1219	.postsuspend = mirror_postsuspend,
1220	.resume	 = mirror_resume,
1221	.status	 = mirror_status,
1222};
1223
1224static int __init dm_mirror_init(void)
1225{
1226	int r;
1227
1228	r = dm_dirty_log_init();
1229	if (r)
1230		return r;
1231
1232	_kmirrord_wq = create_workqueue("kmirrord");
1233	if (!_kmirrord_wq) {
1234		DMERR("couldn't start kmirrord");
1235		dm_dirty_log_exit();
1236		return r;
1237	}
1238	INIT_WORK(&_kmirrord_work, do_work, NULL);
1239
1240	r = dm_register_target(&mirror_target);
1241	if (r < 0) {
1242		DMERR("%s: Failed to register mirror target",
1243		      mirror_target.name);
1244		dm_dirty_log_exit();
1245		destroy_workqueue(_kmirrord_wq);
1246	}
1247
1248	return r;
1249}
1250
1251static void __exit dm_mirror_exit(void)
1252{
1253	int r;
1254
1255	r = dm_unregister_target(&mirror_target);
1256	if (r < 0)
1257		DMERR("%s: unregister failed %d", mirror_target.name, r);
1258
1259	destroy_workqueue(_kmirrord_wq);
1260	dm_dirty_log_exit();
1261}
1262
1263/* Module hooks */
1264module_init(dm_mirror_init);
1265module_exit(dm_mirror_exit);
1266
1267MODULE_DESCRIPTION(DM_NAME " mirror target");
1268MODULE_AUTHOR("Joe Thornber");
1269MODULE_LICENSE("GPL");
1270