?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
kbuild: fix silentoldconfig with make O=
[linux-2.6/verdex.git] / drivers / md / dm-raid1.c
blob8632825137538bbb050eadf6cf0c13b101cdc05d
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
23 static struct workqueue_struct *_kmirrord_wq;
24 static struct work_struct _kmirrord_work;
25
26 static 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 *---------------------------------------------------------------*/
67 struct mirror_set;
68 struct 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
90 enum {
91 RH_CLEAN,
92 RH_DIRTY,
93 RH_NOSYNC,
94 RH_RECOVERING
95 };
96
97 struct 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 */
112 static inline region_t bio_to_region(struct region_hash *rh, struct bio *bio)
113 {
114 return bio->bi_sector >> rh->region_shift;
115 }
116
117 static 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 */
123 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw);
124
125 static void *region_alloc(unsigned int __nocast gfp_mask, void *pool_data)
126 {
127 return kmalloc(sizeof(struct region), gfp_mask);
128 }
129
130 static void region_free(void *element, void *pool_data)
131 {
132 kfree(element);
133 }
134
135 #define MIN_REGIONS 64
136 #define MAX_RECOVERY 1
137 static 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
187 static 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
209 static 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
214 static 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
225 static 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
231 static 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
268 static 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
279 static 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
304 static 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
311 static 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
320 static 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
372 static 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
379 atomic_inc(&reg->pending);
380
381 spin_lock_irq(&rh->region_lock);
382 if (reg->state == RH_CLEAN) {
383 rh->log->type->mark_region(rh->log, reg->key);
384
385 reg->state = RH_DIRTY;
386 list_del_init(&reg->list); /* take off the clean list */
387 }
388 spin_unlock_irq(&rh->region_lock);
389
390 read_unlock(&rh->hash_lock);
391 }
392
393 static void rh_inc_pending(struct region_hash *rh, struct bio_list *bios)
394 {
395 struct bio *bio;
396
397 for (bio = bios->head; bio; bio = bio->bi_next)
398 rh_inc(rh, bio_to_region(rh, bio));
399 }
400
401 static void rh_dec(struct region_hash *rh, region_t region)
402 {
403 unsigned long flags;
404 struct region *reg;
405 int should_wake = 0;
406
407 read_lock(&rh->hash_lock);
408 reg = __rh_lookup(rh, region);
409 read_unlock(&rh->hash_lock);
410
411 if (atomic_dec_and_test(&reg->pending)) {
412 spin_lock_irqsave(&rh->region_lock, flags);
413 if (atomic_read(&reg->pending)) { /* check race */
414 spin_unlock_irqrestore(&rh->region_lock, flags);
415 return;
416 }
417 if (reg->state == RH_RECOVERING) {
418 list_add_tail(&reg->list, &rh->quiesced_regions);
419 } else {
420 reg->state = RH_CLEAN;
421 list_add(&reg->list, &rh->clean_regions);
422 }
423 spin_unlock_irqrestore(&rh->region_lock, flags);
424 should_wake = 1;
425 }
426
427 if (should_wake)
428 wake();
429 }
430
431 /*
432 * Starts quiescing a region in preparation for recovery.
433 */
434 static int __rh_recovery_prepare(struct region_hash *rh)
435 {
436 int r;
437 struct region *reg;
438 region_t region;
439
440 /*
441 * Ask the dirty log what's next.
442 */
443 r = rh->log->type->get_resync_work(rh->log, &region);
444 if (r <= 0)
445 return r;
446
447 /*
448 * Get this region, and start it quiescing by setting the
449 * recovering flag.
450 */
451 read_lock(&rh->hash_lock);
452 reg = __rh_find(rh, region);
453 read_unlock(&rh->hash_lock);
454
455 spin_lock_irq(&rh->region_lock);
456 reg->state = RH_RECOVERING;
457
458 /* Already quiesced ? */
459 if (atomic_read(&reg->pending))
460 list_del_init(&reg->list);
461
462 else {
463 list_del_init(&reg->list);
464 list_add(&reg->list, &rh->quiesced_regions);
465 }
466 spin_unlock_irq(&rh->region_lock);
467
468 return 1;
469 }
470
471 static void rh_recovery_prepare(struct region_hash *rh)
472 {
473 while (!down_trylock(&rh->recovery_count))
474 if (__rh_recovery_prepare(rh) <= 0) {
475 up(&rh->recovery_count);
476 break;
477 }
478 }
479
480 /*
481 * Returns any quiesced regions.
482 */
483 static struct region *rh_recovery_start(struct region_hash *rh)
484 {
485 struct region *reg = NULL;
486
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 */
492 }
493 spin_unlock_irq(&rh->region_lock);
494
495 return reg;
496 }
497
498 /* FIXME: success ignored for now */
499 static void rh_recovery_end(struct region *reg, int success)
500 {
501 struct region_hash *rh = reg->rh;
502
503 spin_lock_irq(&rh->region_lock);
504 list_add(&reg->list, &reg->rh->recovered_regions);
505 spin_unlock_irq(&rh->region_lock);
506
507 wake();
508 }
509
510 static void rh_flush(struct region_hash *rh)
511 {
512 rh->log->type->flush(rh->log);
513 }
514
515 static void rh_delay(struct region_hash *rh, struct bio *bio)
516 {
517 struct region *reg;
518
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);
523 }
524
525 static void rh_stop_recovery(struct region_hash *rh)
526 {
527 int i;
528
529 /* wait for any recovering regions */
530 for (i = 0; i < MAX_RECOVERY; i++)
531 down(&rh->recovery_count);
532 }
533
534 static void rh_start_recovery(struct region_hash *rh)
535 {
536 int i;
537
538 for (i = 0; i < MAX_RECOVERY; i++)
539 up(&rh->recovery_count);
540
541 wake();
542 }
543
544 /*-----------------------------------------------------------------
545 * Mirror set structures.
546 *---------------------------------------------------------------*/
547 struct mirror {
548 atomic_t error_count;
549 struct dm_dev *dev;
550 sector_t offset;
551 };
552
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;
558
559 spinlock_t lock; /* protects the next two lists */
560 struct bio_list reads;
561 struct bio_list writes;
562
563 /* recovery */
564 region_t nr_regions;
565 int in_sync;
566
567 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 */
581 static struct mirror_set *bio_get_ms(struct bio *bio)
582 {
583 return (struct mirror_set *) bio->bi_next;
584 }
585
586 static 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 *---------------------------------------------------------------*/
598 static 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
607 static 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 */
616 m = ms->mirror + DEFAULT_MIRROR;
617 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++) {
632 if (i == DEFAULT_MIRROR)
633 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
650 static 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 *---------------------------------------------------------------*/
684 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
685 {
686 /* FIXME: add read balancing */
687 return ms->mirror + DEFAULT_MIRROR;
688 }
689
690 /*
691 * remap a buffer to a particular mirror.
692 */
693 static 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
699 static 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
714 m = ms->mirror + DEFAULT_MIRROR;
715
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 *---------------------------------------------------------------*/
731 static 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
763 static 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
783 static 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))) {
838 map_bio(ms, ms->mirror + DEFAULT_MIRROR, bio);
839 generic_make_request(bio);
840 }
841 }
842
843 /*-----------------------------------------------------------------
844 * kmirrord
845 *---------------------------------------------------------------*/
846 static LIST_HEAD(_mirror_sets);
847 static DECLARE_RWSEM(_mirror_sets_lock);
848
849 static 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
866 static 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 *---------------------------------------------------------------*/
879 static 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;
905
906 if (rh_init(&ms->rh, ms, dl, region_size, ms->nr_regions)) {
907 ti->error = "dm-mirror: Error creating dirty region hash";
908 kfree(ms);
909 return NULL;
910 }
911
912 return ms;
913 }
914
915 static void free_context(struct mirror_set *ms, struct dm_target *ti,
916 unsigned int m)
917 {
918 while (m--)
919 dm_put_device(ti, ms->mirror[m].dev);
920
921 rh_exit(&ms->rh);
922 kfree(ms);
923 }
924
925 static inline int _check_region_size(struct dm_target *ti, uint32_t size)
926 {
927 return !(size % (PAGE_SIZE >> 9) || (size & (size - 1)) ||
928 size > ti->len);
929 }
930
931 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
932 unsigned int mirror, char **argv)
933 {
934 sector_t offset;
935
936 if (sscanf(argv[1], SECTOR_FORMAT, &offset) != 1) {
937 ti->error = "dm-mirror: Invalid offset";
938 return -EINVAL;
939 }
940
941 if (dm_get_device(ti, argv[0], offset, ti->len,
942 dm_table_get_mode(ti->table),
943 &ms->mirror[mirror].dev)) {
944 ti->error = "dm-mirror: Device lookup failure";
945 return -ENXIO;
946 }
947
948 ms->mirror[mirror].offset = offset;
949
950 return 0;
951 }
952
953 static int add_mirror_set(struct mirror_set *ms)
954 {
955 down_write(&_mirror_sets_lock);
956 list_add_tail(&ms->list, &_mirror_sets);
957 up_write(&_mirror_sets_lock);
958 wake();
959
960 return 0;
961 }
962
963 static void del_mirror_set(struct mirror_set *ms)
964 {
965 down_write(&_mirror_sets_lock);
966 list_del(&ms->list);
967 up_write(&_mirror_sets_lock);
968 }
969
970 /*
971 * Create dirty log: log_type #log_params <log_params>
972 */
973 static struct dirty_log *create_dirty_log(struct dm_target *ti,
974 unsigned int argc, char **argv,
975 unsigned int *args_used)
976 {
977 unsigned int param_count;
978 struct dirty_log *dl;
979
980 if (argc < 2) {
981 ti->error = "dm-mirror: Insufficient mirror log arguments";
982 return NULL;
983 }
984
985 if (sscanf(argv[1], "%u", &param_count) != 1) {
986 ti->error = "dm-mirror: Invalid mirror log argument count";
987 return NULL;
988 }
989
990 *args_used = 2 + param_count;
991
992 if (argc < *args_used) {
993 ti->error = "dm-mirror: Insufficient mirror log arguments";
994 return NULL;
995 }
996
997 dl = dm_create_dirty_log(argv[0], ti, param_count, argv + 2);
998 if (!dl) {
999 ti->error = "dm-mirror: Error creating mirror dirty log";
1000 return NULL;
1001 }
1002
1003 if (!_check_region_size(ti, dl->type->get_region_size(dl))) {
1004 ti->error = "dm-mirror: Invalid region size";
1005 dm_destroy_dirty_log(dl);
1006 return NULL;
1007 }
1008
1009 return dl;
1010 }
1011
1012 /*
1013 * Construct a mirror mapping:
1014 *
1015 * log_type #log_params <log_params>
1016 * #mirrors [mirror_path offset]{2,}
1017 *
1018 * log_type is "core" or "disk"
1019 * #log_params is between 1 and 3
1020 */
1021 #define DM_IO_PAGES 64
1022 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1023 {
1024 int r;
1025 unsigned int nr_mirrors, m, args_used;
1026 struct mirror_set *ms;
1027 struct dirty_log *dl;
1028
1029 dl = create_dirty_log(ti, argc, argv, &args_used);
1030 if (!dl)
1031 return -EINVAL;
1032
1033 argv += args_used;
1034 argc -= args_used;
1035
1036 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1037 nr_mirrors < 2 || nr_mirrors > KCOPYD_MAX_REGIONS + 1) {
1038 ti->error = "dm-mirror: Invalid number of mirrors";
1039 dm_destroy_dirty_log(dl);
1040 return -EINVAL;
1041 }
1042
1043 argv++, argc--;
1044
1045 if (argc != nr_mirrors * 2) {
1046 ti->error = "dm-mirror: Wrong number of mirror arguments";
1047 dm_destroy_dirty_log(dl);
1048 return -EINVAL;
1049 }
1050
1051 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1052 if (!ms) {
1053 dm_destroy_dirty_log(dl);
1054 return -ENOMEM;
1055 }
1056
1057 /* Get the mirror parameter sets */
1058 for (m = 0; m < nr_mirrors; m++) {
1059 r = get_mirror(ms, ti, m, argv);
1060 if (r) {
1061 free_context(ms, ti, m);
1062 return r;
1063 }
1064 argv += 2;
1065 argc -= 2;
1066 }
1067
1068 ti->private = ms;
1069 ti->split_io = ms->rh.region_size;
1070
1071 r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
1072 if (r) {
1073 free_context(ms, ti, ms->nr_mirrors);
1074 return r;
1075 }
1076
1077 add_mirror_set(ms);
1078 return 0;
1079 }
1080
1081 static void mirror_dtr(struct dm_target *ti)
1082 {
1083 struct mirror_set *ms = (struct mirror_set *) ti->private;
1084
1085 del_mirror_set(ms);
1086 kcopyd_client_destroy(ms->kcopyd_client);
1087 free_context(ms, ti, ms->nr_mirrors);
1088 }
1089
1090 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
1091 {
1092 int should_wake = 0;
1093 struct bio_list *bl;
1094
1095 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
1096 spin_lock(&ms->lock);
1097 should_wake = !(bl->head);
1098 bio_list_add(bl, bio);
1099 spin_unlock(&ms->lock);
1100
1101 if (should_wake)
1102 wake();
1103 }
1104
1105 /*
1106 * Mirror mapping function
1107 */
1108 static int mirror_map(struct dm_target *ti, struct bio *bio,
1109 union map_info *map_context)
1110 {
1111 int r, rw = bio_rw(bio);
1112 struct mirror *m;
1113 struct mirror_set *ms = ti->private;
1114
1115 map_context->ll = bio->bi_sector >> ms->rh.region_shift;
1116
1117 if (rw == WRITE) {
1118 queue_bio(ms, bio, rw);
1119 return 0;
1120 }
1121
1122 r = ms->rh.log->type->in_sync(ms->rh.log,
1123 bio_to_region(&ms->rh, bio), 0);
1124 if (r < 0 && r != -EWOULDBLOCK)
1125 return r;
1126
1127 if (r == -EWOULDBLOCK) /* FIXME: ugly */
1128 r = 0;
1129
1130 /*
1131 * We don't want to fast track a recovery just for a read
1132 * ahead. So we just let it silently fail.
1133 * FIXME: get rid of this.
1134 */
1135 if (!r && rw == READA)
1136 return -EIO;
1137
1138 if (!r) {
1139 /* Pass this io over to the daemon */
1140 queue_bio(ms, bio, rw);
1141 return 0;
1142 }
1143
1144 m = choose_mirror(ms, bio->bi_sector);
1145 if (!m)
1146 return -EIO;
1147
1148 map_bio(ms, m, bio);
1149 return 1;
1150 }
1151
1152 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1153 int error, union map_info *map_context)
1154 {
1155 int rw = bio_rw(bio);
1156 struct mirror_set *ms = (struct mirror_set *) ti->private;
1157 region_t region = map_context->ll;
1158
1159 /*
1160 * We need to dec pending if this was a write.
1161 */
1162 if (rw == WRITE)
1163 rh_dec(&ms->rh, region);
1164
1165 return 0;
1166 }
1167
1168 static void mirror_postsuspend(struct dm_target *ti)
1169 {
1170 struct mirror_set *ms = (struct mirror_set *) ti->private;
1171 struct dirty_log *log = ms->rh.log;
1172
1173 rh_stop_recovery(&ms->rh);
1174 if (log->type->suspend && log->type->suspend(log))
1175 /* FIXME: need better error handling */
1176 DMWARN("log suspend failed");
1177 }
1178
1179 static void mirror_resume(struct dm_target *ti)
1180 {
1181 struct mirror_set *ms = (struct mirror_set *) ti->private;
1182 struct dirty_log *log = ms->rh.log;
1183 if (log->type->resume && log->type->resume(log))
1184 /* FIXME: need better error handling */
1185 DMWARN("log resume failed");
1186 rh_start_recovery(&ms->rh);
1187 }
1188
1189 static int mirror_status(struct dm_target *ti, status_type_t type,
1190 char *result, unsigned int maxlen)
1191 {
1192 unsigned int m, sz;
1193 struct mirror_set *ms = (struct mirror_set *) ti->private;
1194
1195 sz = ms->rh.log->type->status(ms->rh.log, type, result, maxlen);
1196
1197 switch (type) {
1198 case STATUSTYPE_INFO:
1199 DMEMIT("%d ", ms->nr_mirrors);
1200 for (m = 0; m < ms->nr_mirrors; m++)
1201 DMEMIT("%s ", ms->mirror[m].dev->name);
1202
1203 DMEMIT(SECTOR_FORMAT "/" SECTOR_FORMAT,
1204 ms->rh.log->type->get_sync_count(ms->rh.log),
1205 ms->nr_regions);
1206 break;
1207
1208 case STATUSTYPE_TABLE:
1209 DMEMIT("%d ", ms->nr_mirrors);
1210 for (m = 0; m < ms->nr_mirrors; m++)
1211 DMEMIT("%s " SECTOR_FORMAT " ",
1212 ms->mirror[m].dev->name, ms->mirror[m].offset);
1213 }
1214
1215 return 0;
1216 }
1217
1218 static struct target_type mirror_target = {
1219 .name = "mirror",
1220 .version = {1, 0, 1},
1221 .module = THIS_MODULE,
1222 .ctr = mirror_ctr,
1223 .dtr = mirror_dtr,
1224 .map = mirror_map,
1225 .end_io = mirror_end_io,
1226 .postsuspend = mirror_postsuspend,
1227 .resume = mirror_resume,
1228 .status = mirror_status,
1229 };
1230
1231 static int __init dm_mirror_init(void)
1232 {
1233 int r;
1234
1235 r = dm_dirty_log_init();
1236 if (r)
1237 return r;
1238
1239 _kmirrord_wq = create_singlethread_workqueue("kmirrord");
1240 if (!_kmirrord_wq) {
1241 DMERR("couldn't start kmirrord");
1242 dm_dirty_log_exit();
1243 return r;
1244 }
1245 INIT_WORK(&_kmirrord_work, do_work, NULL);
1246
1247 r = dm_register_target(&mirror_target);
1248 if (r < 0) {
1249 DMERR("%s: Failed to register mirror target",
1250 mirror_target.name);
1251 dm_dirty_log_exit();
1252 destroy_workqueue(_kmirrord_wq);
1253 }
1254
1255 return r;
1256 }
1257
1258 static void __exit dm_mirror_exit(void)
1259 {
1260 int r;
1261
1262 r = dm_unregister_target(&mirror_target);
1263 if (r < 0)
1264 DMERR("%s: unregister failed %d", mirror_target.name, r);
1265
1266 destroy_workqueue(_kmirrord_wq);
1267 dm_dirty_log_exit();
1268 }
1269
1270 /* Module hooks */
1271 module_init(dm_mirror_init);
1272 module_exit(dm_mirror_exit);
1273
1274 MODULE_DESCRIPTION(DM_NAME " mirror target");
1275 MODULE_AUTHOR("Joe Thornber");
1276 MODULE_LICENSE("GPL");
Verdex Pro support