blob: 6cfa8d435d55733b7b8f0c9671b8791d090b7643 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Al Virodd0fc662005-10-07 07:46:04 +0100125static void *region_alloc(gfp_t gfp_mask, void *pool_data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126{
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);
Jun'ichi Nomura844e8d92005-09-09 16:23:42 -0700378
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800379 spin_lock_irq(&rh->region_lock);
Jun'ichi Nomura844e8d92005-09-09 16:23:42 -0700380 atomic_inc(&reg->pending);
381
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 if (reg->state == RH_CLEAN) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 reg->state = RH_DIRTY;
384 list_del_init(&reg->list); /* take off the clean list */
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800385 spin_unlock_irq(&rh->region_lock);
386
387 rh->log->type->mark_region(rh->log, reg->key);
388 } else
389 spin_unlock_irq(&rh->region_lock);
390
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 read_unlock(&rh->hash_lock);
393}
394
395static void rh_inc_pending(struct region_hash *rh, struct bio_list *bios)
396{
397 struct bio *bio;
398
399 for (bio = bios->head; bio; bio = bio->bi_next)
400 rh_inc(rh, bio_to_region(rh, bio));
401}
402
403static void rh_dec(struct region_hash *rh, region_t region)
404{
405 unsigned long flags;
406 struct region *reg;
407 int should_wake = 0;
408
409 read_lock(&rh->hash_lock);
410 reg = __rh_lookup(rh, region);
411 read_unlock(&rh->hash_lock);
412
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800413 spin_lock_irqsave(&rh->region_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 if (atomic_dec_and_test(&reg->pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 if (reg->state == RH_RECOVERING) {
416 list_add_tail(&reg->list, &rh->quiesced_regions);
417 } else {
418 reg->state = RH_CLEAN;
419 list_add(&reg->list, &rh->clean_regions);
420 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 should_wake = 1;
422 }
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800423 spin_unlock_irqrestore(&rh->region_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424
425 if (should_wake)
426 wake();
427}
428
429/*
430 * Starts quiescing a region in preparation for recovery.
431 */
432static int __rh_recovery_prepare(struct region_hash *rh)
433{
434 int r;
435 struct region *reg;
436 region_t region;
437
438 /*
439 * Ask the dirty log what's next.
440 */
441 r = rh->log->type->get_resync_work(rh->log, &region);
442 if (r <= 0)
443 return r;
444
445 /*
446 * Get this region, and start it quiescing by setting the
447 * recovering flag.
448 */
449 read_lock(&rh->hash_lock);
450 reg = __rh_find(rh, region);
451 read_unlock(&rh->hash_lock);
452
453 spin_lock_irq(&rh->region_lock);
454 reg->state = RH_RECOVERING;
455
456 /* Already quiesced ? */
457 if (atomic_read(&reg->pending))
458 list_del_init(&reg->list);
459
460 else {
461 list_del_init(&reg->list);
462 list_add(&reg->list, &rh->quiesced_regions);
463 }
464 spin_unlock_irq(&rh->region_lock);
465
466 return 1;
467}
468
469static void rh_recovery_prepare(struct region_hash *rh)
470{
471 while (!down_trylock(&rh->recovery_count))
472 if (__rh_recovery_prepare(rh) <= 0) {
473 up(&rh->recovery_count);
474 break;
475 }
476}
477
478/*
479 * Returns any quiesced regions.
480 */
481static struct region *rh_recovery_start(struct region_hash *rh)
482{
483 struct region *reg = NULL;
484
485 spin_lock_irq(&rh->region_lock);
486 if (!list_empty(&rh->quiesced_regions)) {
487 reg = list_entry(rh->quiesced_regions.next,
488 struct region, list);
489 list_del_init(&reg->list); /* remove from the quiesced list */
490 }
491 spin_unlock_irq(&rh->region_lock);
492
493 return reg;
494}
495
496/* FIXME: success ignored for now */
497static void rh_recovery_end(struct region *reg, int success)
498{
499 struct region_hash *rh = reg->rh;
500
501 spin_lock_irq(&rh->region_lock);
502 list_add(&reg->list, &reg->rh->recovered_regions);
503 spin_unlock_irq(&rh->region_lock);
504
505 wake();
506}
507
508static void rh_flush(struct region_hash *rh)
509{
510 rh->log->type->flush(rh->log);
511}
512
513static void rh_delay(struct region_hash *rh, struct bio *bio)
514{
515 struct region *reg;
516
517 read_lock(&rh->hash_lock);
518 reg = __rh_find(rh, bio_to_region(rh, bio));
519 bio_list_add(&reg->delayed_bios, bio);
520 read_unlock(&rh->hash_lock);
521}
522
523static void rh_stop_recovery(struct region_hash *rh)
524{
525 int i;
526
527 /* wait for any recovering regions */
528 for (i = 0; i < MAX_RECOVERY; i++)
529 down(&rh->recovery_count);
530}
531
532static void rh_start_recovery(struct region_hash *rh)
533{
534 int i;
535
536 for (i = 0; i < MAX_RECOVERY; i++)
537 up(&rh->recovery_count);
538
539 wake();
540}
541
542/*-----------------------------------------------------------------
543 * Mirror set structures.
544 *---------------------------------------------------------------*/
545struct mirror {
546 atomic_t error_count;
547 struct dm_dev *dev;
548 sector_t offset;
549};
550
551struct mirror_set {
552 struct dm_target *ti;
553 struct list_head list;
554 struct region_hash rh;
555 struct kcopyd_client *kcopyd_client;
556
557 spinlock_t lock; /* protects the next two lists */
558 struct bio_list reads;
559 struct bio_list writes;
560
561 /* recovery */
562 region_t nr_regions;
563 int in_sync;
564
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800565 struct mirror *default_mirror; /* Default mirror */
566
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 unsigned int nr_mirrors;
568 struct mirror mirror[0];
569};
570
571/*
572 * Every mirror should look like this one.
573 */
574#define DEFAULT_MIRROR 0
575
576/*
577 * This is yucky. We squirrel the mirror_set struct away inside
578 * bi_next for write buffers. This is safe since the bh
579 * doesn't get submitted to the lower levels of block layer.
580 */
581static struct mirror_set *bio_get_ms(struct bio *bio)
582{
583 return (struct mirror_set *) bio->bi_next;
584}
585
586static void bio_set_ms(struct bio *bio, struct mirror_set *ms)
587{
588 bio->bi_next = (struct bio *) ms;
589}
590
591/*-----------------------------------------------------------------
592 * Recovery.
593 *
594 * When a mirror is first activated we may find that some regions
595 * are in the no-sync state. We have to recover these by
596 * recopying from the default mirror to all the others.
597 *---------------------------------------------------------------*/
598static void recovery_complete(int read_err, unsigned int write_err,
599 void *context)
600{
601 struct region *reg = (struct region *) context;
602
603 /* FIXME: better error handling */
604 rh_recovery_end(reg, read_err || write_err);
605}
606
607static int recover(struct mirror_set *ms, struct region *reg)
608{
609 int r;
610 unsigned int i;
611 struct io_region from, to[KCOPYD_MAX_REGIONS], *dest;
612 struct mirror *m;
613 unsigned long flags = 0;
614
615 /* fill in the source */
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800616 m = ms->default_mirror;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 from.bdev = m->dev->bdev;
618 from.sector = m->offset + region_to_sector(reg->rh, reg->key);
619 if (reg->key == (ms->nr_regions - 1)) {
620 /*
621 * The final region may be smaller than
622 * region_size.
623 */
624 from.count = ms->ti->len & (reg->rh->region_size - 1);
625 if (!from.count)
626 from.count = reg->rh->region_size;
627 } else
628 from.count = reg->rh->region_size;
629
630 /* fill in the destinations */
631 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800632 if (&ms->mirror[i] == ms->default_mirror)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 continue;
634
635 m = ms->mirror + i;
636 dest->bdev = m->dev->bdev;
637 dest->sector = m->offset + region_to_sector(reg->rh, reg->key);
638 dest->count = from.count;
639 dest++;
640 }
641
642 /* hand to kcopyd */
643 set_bit(KCOPYD_IGNORE_ERROR, &flags);
644 r = kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to, flags,
645 recovery_complete, reg);
646
647 return r;
648}
649
650static void do_recovery(struct mirror_set *ms)
651{
652 int r;
653 struct region *reg;
654 struct dirty_log *log = ms->rh.log;
655
656 /*
657 * Start quiescing some regions.
658 */
659 rh_recovery_prepare(&ms->rh);
660
661 /*
662 * Copy any already quiesced regions.
663 */
664 while ((reg = rh_recovery_start(&ms->rh))) {
665 r = recover(ms, reg);
666 if (r)
667 rh_recovery_end(reg, 0);
668 }
669
670 /*
671 * Update the in sync flag.
672 */
673 if (!ms->in_sync &&
674 (log->type->get_sync_count(log) == ms->nr_regions)) {
675 /* the sync is complete */
676 dm_table_event(ms->ti->table);
677 ms->in_sync = 1;
678 }
679}
680
681/*-----------------------------------------------------------------
682 * Reads
683 *---------------------------------------------------------------*/
684static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
685{
686 /* FIXME: add read balancing */
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800687 return ms->default_mirror;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
689
690/*
691 * remap a buffer to a particular mirror.
692 */
693static void map_bio(struct mirror_set *ms, struct mirror *m, struct bio *bio)
694{
695 bio->bi_bdev = m->dev->bdev;
696 bio->bi_sector = m->offset + (bio->bi_sector - ms->ti->begin);
697}
698
699static void do_reads(struct mirror_set *ms, struct bio_list *reads)
700{
701 region_t region;
702 struct bio *bio;
703 struct mirror *m;
704
705 while ((bio = bio_list_pop(reads))) {
706 region = bio_to_region(&ms->rh, bio);
707
708 /*
709 * We can only read balance if the region is in sync.
710 */
711 if (rh_in_sync(&ms->rh, region, 0))
712 m = choose_mirror(ms, bio->bi_sector);
713 else
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800714 m = ms->default_mirror;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
716 map_bio(ms, m, bio);
717 generic_make_request(bio);
718 }
719}
720
721/*-----------------------------------------------------------------
722 * Writes.
723 *
724 * We do different things with the write io depending on the
725 * state of the region that it's in:
726 *
727 * SYNC: increment pending, use kcopyd to write to *all* mirrors
728 * RECOVERING: delay the io until recovery completes
729 * NOSYNC: increment pending, just write to the default mirror
730 *---------------------------------------------------------------*/
731static void write_callback(unsigned long error, void *context)
732{
733 unsigned int i;
734 int uptodate = 1;
735 struct bio *bio = (struct bio *) context;
736 struct mirror_set *ms;
737
738 ms = bio_get_ms(bio);
739 bio_set_ms(bio, NULL);
740
741 /*
742 * NOTE: We don't decrement the pending count here,
743 * instead it is done by the targets endio function.
744 * This way we handle both writes to SYNC and NOSYNC
745 * regions with the same code.
746 */
747
748 if (error) {
749 /*
750 * only error the io if all mirrors failed.
751 * FIXME: bogus
752 */
753 uptodate = 0;
754 for (i = 0; i < ms->nr_mirrors; i++)
755 if (!test_bit(i, &error)) {
756 uptodate = 1;
757 break;
758 }
759 }
760 bio_endio(bio, bio->bi_size, 0);
761}
762
763static void do_write(struct mirror_set *ms, struct bio *bio)
764{
765 unsigned int i;
766 struct io_region io[KCOPYD_MAX_REGIONS+1];
767 struct mirror *m;
768
769 for (i = 0; i < ms->nr_mirrors; i++) {
770 m = ms->mirror + i;
771
772 io[i].bdev = m->dev->bdev;
773 io[i].sector = m->offset + (bio->bi_sector - ms->ti->begin);
774 io[i].count = bio->bi_size >> 9;
775 }
776
777 bio_set_ms(bio, ms);
778 dm_io_async_bvec(ms->nr_mirrors, io, WRITE,
779 bio->bi_io_vec + bio->bi_idx,
780 write_callback, bio);
781}
782
783static void do_writes(struct mirror_set *ms, struct bio_list *writes)
784{
785 int state;
786 struct bio *bio;
787 struct bio_list sync, nosync, recover, *this_list = NULL;
788
789 if (!writes->head)
790 return;
791
792 /*
793 * Classify each write.
794 */
795 bio_list_init(&sync);
796 bio_list_init(&nosync);
797 bio_list_init(&recover);
798
799 while ((bio = bio_list_pop(writes))) {
800 state = rh_state(&ms->rh, bio_to_region(&ms->rh, bio), 1);
801 switch (state) {
802 case RH_CLEAN:
803 case RH_DIRTY:
804 this_list = &sync;
805 break;
806
807 case RH_NOSYNC:
808 this_list = &nosync;
809 break;
810
811 case RH_RECOVERING:
812 this_list = &recover;
813 break;
814 }
815
816 bio_list_add(this_list, bio);
817 }
818
819 /*
820 * Increment the pending counts for any regions that will
821 * be written to (writes to recover regions are going to
822 * be delayed).
823 */
824 rh_inc_pending(&ms->rh, &sync);
825 rh_inc_pending(&ms->rh, &nosync);
826 rh_flush(&ms->rh);
827
828 /*
829 * Dispatch io.
830 */
831 while ((bio = bio_list_pop(&sync)))
832 do_write(ms, bio);
833
834 while ((bio = bio_list_pop(&recover)))
835 rh_delay(&ms->rh, bio);
836
837 while ((bio = bio_list_pop(&nosync))) {
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800838 map_bio(ms, ms->default_mirror, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 generic_make_request(bio);
840 }
841}
842
843/*-----------------------------------------------------------------
844 * kmirrord
845 *---------------------------------------------------------------*/
846static LIST_HEAD(_mirror_sets);
847static DECLARE_RWSEM(_mirror_sets_lock);
848
849static void do_mirror(struct mirror_set *ms)
850{
851 struct bio_list reads, writes;
852
853 spin_lock(&ms->lock);
854 reads = ms->reads;
855 writes = ms->writes;
856 bio_list_init(&ms->reads);
857 bio_list_init(&ms->writes);
858 spin_unlock(&ms->lock);
859
860 rh_update_states(&ms->rh);
861 do_recovery(ms);
862 do_reads(ms, &reads);
863 do_writes(ms, &writes);
864}
865
866static void do_work(void *ignored)
867{
868 struct mirror_set *ms;
869
870 down_read(&_mirror_sets_lock);
871 list_for_each_entry (ms, &_mirror_sets, list)
872 do_mirror(ms);
873 up_read(&_mirror_sets_lock);
874}
875
876/*-----------------------------------------------------------------
877 * Target functions
878 *---------------------------------------------------------------*/
879static struct mirror_set *alloc_context(unsigned int nr_mirrors,
880 uint32_t region_size,
881 struct dm_target *ti,
882 struct dirty_log *dl)
883{
884 size_t len;
885 struct mirror_set *ms = NULL;
886
887 if (array_too_big(sizeof(*ms), sizeof(ms->mirror[0]), nr_mirrors))
888 return NULL;
889
890 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
891
892 ms = kmalloc(len, GFP_KERNEL);
893 if (!ms) {
894 ti->error = "dm-mirror: Cannot allocate mirror context";
895 return NULL;
896 }
897
898 memset(ms, 0, len);
899 spin_lock_init(&ms->lock);
900
901 ms->ti = ti;
902 ms->nr_mirrors = nr_mirrors;
903 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
904 ms->in_sync = 0;
Jonathan E Brassowa1a19082006-01-06 00:20:05 -0800905 ms->default_mirror = &ms->mirror[DEFAULT_MIRROR];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
907 if (rh_init(&ms->rh, ms, dl, region_size, ms->nr_regions)) {
908 ti->error = "dm-mirror: Error creating dirty region hash";
909 kfree(ms);
910 return NULL;
911 }
912
913 return ms;
914}
915
916static void free_context(struct mirror_set *ms, struct dm_target *ti,
917 unsigned int m)
918{
919 while (m--)
920 dm_put_device(ti, ms->mirror[m].dev);
921
922 rh_exit(&ms->rh);
923 kfree(ms);
924}
925
926static inline int _check_region_size(struct dm_target *ti, uint32_t size)
927{
928 return !(size % (PAGE_SIZE >> 9) || (size & (size - 1)) ||
929 size > ti->len);
930}
931
932static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
933 unsigned int mirror, char **argv)
934{
935 sector_t offset;
936
937 if (sscanf(argv[1], SECTOR_FORMAT, &offset) != 1) {
938 ti->error = "dm-mirror: Invalid offset";
939 return -EINVAL;
940 }
941
942 if (dm_get_device(ti, argv[0], offset, ti->len,
943 dm_table_get_mode(ti->table),
944 &ms->mirror[mirror].dev)) {
945 ti->error = "dm-mirror: Device lookup failure";
946 return -ENXIO;
947 }
948
949 ms->mirror[mirror].offset = offset;
950
951 return 0;
952}
953
954static int add_mirror_set(struct mirror_set *ms)
955{
956 down_write(&_mirror_sets_lock);
957 list_add_tail(&ms->list, &_mirror_sets);
958 up_write(&_mirror_sets_lock);
959 wake();
960
961 return 0;
962}
963
964static void del_mirror_set(struct mirror_set *ms)
965{
966 down_write(&_mirror_sets_lock);
967 list_del(&ms->list);
968 up_write(&_mirror_sets_lock);
969}
970
971/*
972 * Create dirty log: log_type #log_params <log_params>
973 */
974static struct dirty_log *create_dirty_log(struct dm_target *ti,
975 unsigned int argc, char **argv,
976 unsigned int *args_used)
977{
978 unsigned int param_count;
979 struct dirty_log *dl;
980
981 if (argc < 2) {
982 ti->error = "dm-mirror: Insufficient mirror log arguments";
983 return NULL;
984 }
985
986 if (sscanf(argv[1], "%u", &param_count) != 1) {
987 ti->error = "dm-mirror: Invalid mirror log argument count";
988 return NULL;
989 }
990
991 *args_used = 2 + param_count;
992
993 if (argc < *args_used) {
994 ti->error = "dm-mirror: Insufficient mirror log arguments";
995 return NULL;
996 }
997
998 dl = dm_create_dirty_log(argv[0], ti, param_count, argv + 2);
999 if (!dl) {
1000 ti->error = "dm-mirror: Error creating mirror dirty log";
1001 return NULL;
1002 }
1003
1004 if (!_check_region_size(ti, dl->type->get_region_size(dl))) {
1005 ti->error = "dm-mirror: Invalid region size";
1006 dm_destroy_dirty_log(dl);
1007 return NULL;
1008 }
1009
1010 return dl;
1011}
1012
1013/*
1014 * Construct a mirror mapping:
1015 *
1016 * log_type #log_params <log_params>
1017 * #mirrors [mirror_path offset]{2,}
1018 *
1019 * log_type is "core" or "disk"
1020 * #log_params is between 1 and 3
1021 */
1022#define DM_IO_PAGES 64
1023static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1024{
1025 int r;
1026 unsigned int nr_mirrors, m, args_used;
1027 struct mirror_set *ms;
1028 struct dirty_log *dl;
1029
1030 dl = create_dirty_log(ti, argc, argv, &args_used);
1031 if (!dl)
1032 return -EINVAL;
1033
1034 argv += args_used;
1035 argc -= args_used;
1036
1037 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1038 nr_mirrors < 2 || nr_mirrors > KCOPYD_MAX_REGIONS + 1) {
1039 ti->error = "dm-mirror: Invalid number of mirrors";
1040 dm_destroy_dirty_log(dl);
1041 return -EINVAL;
1042 }
1043
1044 argv++, argc--;
1045
1046 if (argc != nr_mirrors * 2) {
1047 ti->error = "dm-mirror: Wrong number of mirror arguments";
1048 dm_destroy_dirty_log(dl);
1049 return -EINVAL;
1050 }
1051
1052 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1053 if (!ms) {
1054 dm_destroy_dirty_log(dl);
1055 return -ENOMEM;
1056 }
1057
1058 /* Get the mirror parameter sets */
1059 for (m = 0; m < nr_mirrors; m++) {
1060 r = get_mirror(ms, ti, m, argv);
1061 if (r) {
1062 free_context(ms, ti, m);
1063 return r;
1064 }
1065 argv += 2;
1066 argc -= 2;
1067 }
1068
1069 ti->private = ms;
Alasdair G Kergond88854f2005-07-07 17:59:34 -07001070 ti->split_io = ms->rh.region_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
1072 r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
1073 if (r) {
1074 free_context(ms, ti, ms->nr_mirrors);
1075 return r;
1076 }
1077
1078 add_mirror_set(ms);
1079 return 0;
1080}
1081
1082static void mirror_dtr(struct dm_target *ti)
1083{
1084 struct mirror_set *ms = (struct mirror_set *) ti->private;
1085
1086 del_mirror_set(ms);
1087 kcopyd_client_destroy(ms->kcopyd_client);
1088 free_context(ms, ti, ms->nr_mirrors);
1089}
1090
1091static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
1092{
1093 int should_wake = 0;
1094 struct bio_list *bl;
1095
1096 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
1097 spin_lock(&ms->lock);
1098 should_wake = !(bl->head);
1099 bio_list_add(bl, bio);
1100 spin_unlock(&ms->lock);
1101
1102 if (should_wake)
1103 wake();
1104}
1105
1106/*
1107 * Mirror mapping function
1108 */
1109static int mirror_map(struct dm_target *ti, struct bio *bio,
1110 union map_info *map_context)
1111{
1112 int r, rw = bio_rw(bio);
1113 struct mirror *m;
1114 struct mirror_set *ms = ti->private;
1115
1116 map_context->ll = bio->bi_sector >> ms->rh.region_shift;
1117
1118 if (rw == WRITE) {
1119 queue_bio(ms, bio, rw);
1120 return 0;
1121 }
1122
1123 r = ms->rh.log->type->in_sync(ms->rh.log,
1124 bio_to_region(&ms->rh, bio), 0);
1125 if (r < 0 && r != -EWOULDBLOCK)
1126 return r;
1127
1128 if (r == -EWOULDBLOCK) /* FIXME: ugly */
1129 r = 0;
1130
1131 /*
1132 * We don't want to fast track a recovery just for a read
1133 * ahead. So we just let it silently fail.
1134 * FIXME: get rid of this.
1135 */
1136 if (!r && rw == READA)
1137 return -EIO;
1138
1139 if (!r) {
1140 /* Pass this io over to the daemon */
1141 queue_bio(ms, bio, rw);
1142 return 0;
1143 }
1144
1145 m = choose_mirror(ms, bio->bi_sector);
1146 if (!m)
1147 return -EIO;
1148
1149 map_bio(ms, m, bio);
1150 return 1;
1151}
1152
1153static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1154 int error, union map_info *map_context)
1155{
1156 int rw = bio_rw(bio);
1157 struct mirror_set *ms = (struct mirror_set *) ti->private;
1158 region_t region = map_context->ll;
1159
1160 /*
1161 * We need to dec pending if this was a write.
1162 */
1163 if (rw == WRITE)
1164 rh_dec(&ms->rh, region);
1165
1166 return 0;
1167}
1168
1169static void mirror_postsuspend(struct dm_target *ti)
1170{
1171 struct mirror_set *ms = (struct mirror_set *) ti->private;
1172 struct dirty_log *log = ms->rh.log;
1173
1174 rh_stop_recovery(&ms->rh);
1175 if (log->type->suspend && log->type->suspend(log))
1176 /* FIXME: need better error handling */
1177 DMWARN("log suspend failed");
1178}
1179
1180static void mirror_resume(struct dm_target *ti)
1181{
1182 struct mirror_set *ms = (struct mirror_set *) ti->private;
1183 struct dirty_log *log = ms->rh.log;
1184 if (log->type->resume && log->type->resume(log))
1185 /* FIXME: need better error handling */
1186 DMWARN("log resume failed");
1187 rh_start_recovery(&ms->rh);
1188}
1189
1190static int mirror_status(struct dm_target *ti, status_type_t type,
1191 char *result, unsigned int maxlen)
1192{
1193 unsigned int m, sz;
1194 struct mirror_set *ms = (struct mirror_set *) ti->private;
1195
1196 sz = ms->rh.log->type->status(ms->rh.log, type, result, maxlen);
1197
1198 switch (type) {
1199 case STATUSTYPE_INFO:
1200 DMEMIT("%d ", ms->nr_mirrors);
1201 for (m = 0; m < ms->nr_mirrors; m++)
1202 DMEMIT("%s ", ms->mirror[m].dev->name);
1203
1204 DMEMIT(SECTOR_FORMAT "/" SECTOR_FORMAT,
1205 ms->rh.log->type->get_sync_count(ms->rh.log),
1206 ms->nr_regions);
1207 break;
1208
1209 case STATUSTYPE_TABLE:
1210 DMEMIT("%d ", ms->nr_mirrors);
1211 for (m = 0; m < ms->nr_mirrors; m++)
1212 DMEMIT("%s " SECTOR_FORMAT " ",
1213 ms->mirror[m].dev->name, ms->mirror[m].offset);
1214 }
1215
1216 return 0;
1217}
1218
1219static struct target_type mirror_target = {
1220 .name = "mirror",
1221 .version = {1, 0, 1},
1222 .module = THIS_MODULE,
1223 .ctr = mirror_ctr,
1224 .dtr = mirror_dtr,
1225 .map = mirror_map,
1226 .end_io = mirror_end_io,
1227 .postsuspend = mirror_postsuspend,
1228 .resume = mirror_resume,
1229 .status = mirror_status,
1230};
1231
1232static int __init dm_mirror_init(void)
1233{
1234 int r;
1235
1236 r = dm_dirty_log_init();
1237 if (r)
1238 return r;
1239
Alasdair G Kergon48f1f532005-08-04 12:53:37 -07001240 _kmirrord_wq = create_singlethread_workqueue("kmirrord");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 if (!_kmirrord_wq) {
1242 DMERR("couldn't start kmirrord");
1243 dm_dirty_log_exit();
1244 return r;
1245 }
1246 INIT_WORK(&_kmirrord_work, do_work, NULL);
1247
1248 r = dm_register_target(&mirror_target);
1249 if (r < 0) {
1250 DMERR("%s: Failed to register mirror target",
1251 mirror_target.name);
1252 dm_dirty_log_exit();
1253 destroy_workqueue(_kmirrord_wq);
1254 }
1255
1256 return r;
1257}
1258
1259static void __exit dm_mirror_exit(void)
1260{
1261 int r;
1262
1263 r = dm_unregister_target(&mirror_target);
1264 if (r < 0)
1265 DMERR("%s: unregister failed %d", mirror_target.name, r);
1266
1267 destroy_workqueue(_kmirrord_wq);
1268 dm_dirty_log_exit();
1269}
1270
1271/* Module hooks */
1272module_init(dm_mirror_init);
1273module_exit(dm_mirror_exit);
1274
1275MODULE_DESCRIPTION(DM_NAME " mirror target");
1276MODULE_AUTHOR("Joe Thornber");
1277MODULE_LICENSE("GPL");