blob: 59f8d9df9e1ac41c3149f146a1a42b129be111ec [file] [log] [blame]
Heinz Mauelshagen1f965b12008-10-21 17:45:06 +01001/*
2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include <linux/dm-dirty-log.h>
9#include <linux/dm-region-hash.h>
10
11#include <linux/ctype.h>
12#include <linux/init.h>
13#include <linux/module.h>
14#include <linux/vmalloc.h>
15
16#include "dm.h"
17#include "dm-bio-list.h"
18
19#define DM_MSG_PREFIX "region hash"
20
21/*-----------------------------------------------------------------
22 * Region hash
23 *
24 * The mirror splits itself up into discrete regions. Each
25 * region can be in one of three states: clean, dirty,
26 * nosync. There is no need to put clean regions in the hash.
27 *
28 * In addition to being present in the hash table a region _may_
29 * be present on one of three lists.
30 *
31 * clean_regions: Regions on this list have no io pending to
32 * them, they are in sync, we are no longer interested in them,
33 * they are dull. dm_rh_update_states() will remove them from the
34 * hash table.
35 *
36 * quiesced_regions: These regions have been spun down, ready
37 * for recovery. rh_recovery_start() will remove regions from
38 * this list and hand them to kmirrord, which will schedule the
39 * recovery io with kcopyd.
40 *
41 * recovered_regions: Regions that kcopyd has successfully
42 * recovered. dm_rh_update_states() will now schedule any delayed
43 * io, up the recovery_count, and remove the region from the
44 * hash.
45 *
46 * There are 2 locks:
47 * A rw spin lock 'hash_lock' protects just the hash table,
48 * this is never held in write mode from interrupt context,
49 * which I believe means that we only have to disable irqs when
50 * doing a write lock.
51 *
52 * An ordinary spin lock 'region_lock' that protects the three
53 * lists in the region_hash, with the 'state', 'list' and
54 * 'delayed_bios' fields of the regions. This is used from irq
55 * context, so all other uses will have to suspend local irqs.
56 *---------------------------------------------------------------*/
57struct dm_region_hash {
58 uint32_t region_size;
59 unsigned region_shift;
60
61 /* holds persistent region state */
62 struct dm_dirty_log *log;
63
64 /* hash table */
65 rwlock_t hash_lock;
66 mempool_t *region_pool;
67 unsigned mask;
68 unsigned nr_buckets;
69 unsigned prime;
70 unsigned shift;
71 struct list_head *buckets;
72
73 unsigned max_recovery; /* Max # of regions to recover in parallel */
74
75 spinlock_t region_lock;
76 atomic_t recovery_in_flight;
77 struct semaphore recovery_count;
78 struct list_head clean_regions;
79 struct list_head quiesced_regions;
80 struct list_head recovered_regions;
81 struct list_head failed_recovered_regions;
82
83 void *context;
84 sector_t target_begin;
85
86 /* Callback function to schedule bios writes */
87 void (*dispatch_bios)(void *context, struct bio_list *bios);
88
89 /* Callback function to wakeup callers worker thread. */
90 void (*wakeup_workers)(void *context);
91
92 /* Callback function to wakeup callers recovery waiters. */
93 void (*wakeup_all_recovery_waiters)(void *context);
94};
95
96struct dm_region {
97 struct dm_region_hash *rh; /* FIXME: can we get rid of this ? */
98 region_t key;
99 int state;
100
101 struct list_head hash_list;
102 struct list_head list;
103
104 atomic_t pending;
105 struct bio_list delayed_bios;
106};
107
108/*
109 * Conversion fns
110 */
111static region_t dm_rh_sector_to_region(struct dm_region_hash *rh, sector_t sector)
112{
113 return sector >> rh->region_shift;
114}
115
116sector_t dm_rh_region_to_sector(struct dm_region_hash *rh, region_t region)
117{
118 return region << rh->region_shift;
119}
120EXPORT_SYMBOL_GPL(dm_rh_region_to_sector);
121
122region_t dm_rh_bio_to_region(struct dm_region_hash *rh, struct bio *bio)
123{
124 return dm_rh_sector_to_region(rh, bio->bi_sector - rh->target_begin);
125}
126EXPORT_SYMBOL_GPL(dm_rh_bio_to_region);
127
128void *dm_rh_region_context(struct dm_region *reg)
129{
130 return reg->rh->context;
131}
132EXPORT_SYMBOL_GPL(dm_rh_region_context);
133
134region_t dm_rh_get_region_key(struct dm_region *reg)
135{
136 return reg->key;
137}
138EXPORT_SYMBOL_GPL(dm_rh_get_region_key);
139
140sector_t dm_rh_get_region_size(struct dm_region_hash *rh)
141{
142 return rh->region_size;
143}
144EXPORT_SYMBOL_GPL(dm_rh_get_region_size);
145
146/*
147 * FIXME: shall we pass in a structure instead of all these args to
148 * dm_region_hash_create()????
149 */
150#define RH_HASH_MULT 2654435387U
151#define RH_HASH_SHIFT 12
152
153#define MIN_REGIONS 64
154struct dm_region_hash *dm_region_hash_create(
155 void *context, void (*dispatch_bios)(void *context,
156 struct bio_list *bios),
157 void (*wakeup_workers)(void *context),
158 void (*wakeup_all_recovery_waiters)(void *context),
159 sector_t target_begin, unsigned max_recovery,
160 struct dm_dirty_log *log, uint32_t region_size,
161 region_t nr_regions)
162{
163 struct dm_region_hash *rh;
164 unsigned nr_buckets, max_buckets;
165 size_t i;
166
167 /*
168 * Calculate a suitable number of buckets for our hash
169 * table.
170 */
171 max_buckets = nr_regions >> 6;
172 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
173 ;
174 nr_buckets >>= 1;
175
176 rh = kmalloc(sizeof(*rh), GFP_KERNEL);
177 if (!rh) {
178 DMERR("unable to allocate region hash memory");
179 return ERR_PTR(-ENOMEM);
180 }
181
182 rh->context = context;
183 rh->dispatch_bios = dispatch_bios;
184 rh->wakeup_workers = wakeup_workers;
185 rh->wakeup_all_recovery_waiters = wakeup_all_recovery_waiters;
186 rh->target_begin = target_begin;
187 rh->max_recovery = max_recovery;
188 rh->log = log;
189 rh->region_size = region_size;
190 rh->region_shift = ffs(region_size) - 1;
191 rwlock_init(&rh->hash_lock);
192 rh->mask = nr_buckets - 1;
193 rh->nr_buckets = nr_buckets;
194
195 rh->shift = RH_HASH_SHIFT;
196 rh->prime = RH_HASH_MULT;
197
198 rh->buckets = vmalloc(nr_buckets * sizeof(*rh->buckets));
199 if (!rh->buckets) {
200 DMERR("unable to allocate region hash bucket memory");
201 kfree(rh);
202 return ERR_PTR(-ENOMEM);
203 }
204
205 for (i = 0; i < nr_buckets; i++)
206 INIT_LIST_HEAD(rh->buckets + i);
207
208 spin_lock_init(&rh->region_lock);
209 sema_init(&rh->recovery_count, 0);
210 atomic_set(&rh->recovery_in_flight, 0);
211 INIT_LIST_HEAD(&rh->clean_regions);
212 INIT_LIST_HEAD(&rh->quiesced_regions);
213 INIT_LIST_HEAD(&rh->recovered_regions);
214 INIT_LIST_HEAD(&rh->failed_recovered_regions);
215
216 rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
217 sizeof(struct dm_region));
218 if (!rh->region_pool) {
219 vfree(rh->buckets);
220 kfree(rh);
221 rh = ERR_PTR(-ENOMEM);
222 }
223
224 return rh;
225}
226EXPORT_SYMBOL_GPL(dm_region_hash_create);
227
228void dm_region_hash_destroy(struct dm_region_hash *rh)
229{
230 unsigned h;
231 struct dm_region *reg, *nreg;
232
233 BUG_ON(!list_empty(&rh->quiesced_regions));
234 for (h = 0; h < rh->nr_buckets; h++) {
235 list_for_each_entry_safe(reg, nreg, rh->buckets + h,
236 hash_list) {
237 BUG_ON(atomic_read(&reg->pending));
238 mempool_free(reg, rh->region_pool);
239 }
240 }
241
242 if (rh->log)
243 dm_dirty_log_destroy(rh->log);
244
245 if (rh->region_pool)
246 mempool_destroy(rh->region_pool);
247
248 vfree(rh->buckets);
249 kfree(rh);
250}
251EXPORT_SYMBOL_GPL(dm_region_hash_destroy);
252
253struct dm_dirty_log *dm_rh_dirty_log(struct dm_region_hash *rh)
254{
255 return rh->log;
256}
257EXPORT_SYMBOL_GPL(dm_rh_dirty_log);
258
259static unsigned rh_hash(struct dm_region_hash *rh, region_t region)
260{
261 return (unsigned) ((region * rh->prime) >> rh->shift) & rh->mask;
262}
263
264static struct dm_region *__rh_lookup(struct dm_region_hash *rh, region_t region)
265{
266 struct dm_region *reg;
267 struct list_head *bucket = rh->buckets + rh_hash(rh, region);
268
269 list_for_each_entry(reg, bucket, hash_list)
270 if (reg->key == region)
271 return reg;
272
273 return NULL;
274}
275
276static void __rh_insert(struct dm_region_hash *rh, struct dm_region *reg)
277{
278 list_add(&reg->hash_list, rh->buckets + rh_hash(rh, reg->key));
279}
280
281static struct dm_region *__rh_alloc(struct dm_region_hash *rh, region_t region)
282{
283 struct dm_region *reg, *nreg;
284
285 nreg = mempool_alloc(rh->region_pool, GFP_ATOMIC);
286 if (unlikely(!nreg))
287 nreg = kmalloc(sizeof(*nreg), GFP_NOIO);
288
289 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
290 DM_RH_CLEAN : DM_RH_NOSYNC;
291 nreg->rh = rh;
292 nreg->key = region;
293 INIT_LIST_HEAD(&nreg->list);
294 atomic_set(&nreg->pending, 0);
295 bio_list_init(&nreg->delayed_bios);
296
297 write_lock_irq(&rh->hash_lock);
298 reg = __rh_lookup(rh, region);
299 if (reg)
300 /* We lost the race. */
301 mempool_free(nreg, rh->region_pool);
302 else {
303 __rh_insert(rh, nreg);
304 if (nreg->state == DM_RH_CLEAN) {
305 spin_lock(&rh->region_lock);
306 list_add(&nreg->list, &rh->clean_regions);
307 spin_unlock(&rh->region_lock);
308 }
309
310 reg = nreg;
311 }
312 write_unlock_irq(&rh->hash_lock);
313
314 return reg;
315}
316
317static struct dm_region *__rh_find(struct dm_region_hash *rh, region_t region)
318{
319 struct dm_region *reg;
320
321 reg = __rh_lookup(rh, region);
322 if (!reg) {
323 read_unlock(&rh->hash_lock);
324 reg = __rh_alloc(rh, region);
325 read_lock(&rh->hash_lock);
326 }
327
328 return reg;
329}
330
331int dm_rh_get_state(struct dm_region_hash *rh, region_t region, int may_block)
332{
333 int r;
334 struct dm_region *reg;
335
336 read_lock(&rh->hash_lock);
337 reg = __rh_lookup(rh, region);
338 read_unlock(&rh->hash_lock);
339
340 if (reg)
341 return reg->state;
342
343 /*
344 * The region wasn't in the hash, so we fall back to the
345 * dirty log.
346 */
347 r = rh->log->type->in_sync(rh->log, region, may_block);
348
349 /*
350 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
351 * taken as a DM_RH_NOSYNC
352 */
353 return r == 1 ? DM_RH_CLEAN : DM_RH_NOSYNC;
354}
355EXPORT_SYMBOL_GPL(dm_rh_get_state);
356
357static void complete_resync_work(struct dm_region *reg, int success)
358{
359 struct dm_region_hash *rh = reg->rh;
360
361 rh->log->type->set_region_sync(rh->log, reg->key, success);
362
363 /*
364 * Dispatch the bios before we call 'wake_up_all'.
365 * This is important because if we are suspending,
366 * we want to know that recovery is complete and
367 * the work queue is flushed. If we wake_up_all
368 * before we dispatch_bios (queue bios and call wake()),
369 * then we risk suspending before the work queue
370 * has been properly flushed.
371 */
372 rh->dispatch_bios(rh->context, &reg->delayed_bios);
373 if (atomic_dec_and_test(&rh->recovery_in_flight))
374 rh->wakeup_all_recovery_waiters(rh->context);
375 up(&rh->recovery_count);
376}
377
378/* dm_rh_mark_nosync
379 * @ms
380 * @bio
381 * @done
382 * @error
383 *
384 * The bio was written on some mirror(s) but failed on other mirror(s).
385 * We can successfully endio the bio but should avoid the region being
386 * marked clean by setting the state DM_RH_NOSYNC.
387 *
388 * This function is _not_ safe in interrupt context!
389 */
390void dm_rh_mark_nosync(struct dm_region_hash *rh,
391 struct bio *bio, unsigned done, int error)
392{
393 unsigned long flags;
394 struct dm_dirty_log *log = rh->log;
395 struct dm_region *reg;
396 region_t region = dm_rh_bio_to_region(rh, bio);
397 int recovering = 0;
398
399 /* We must inform the log that the sync count has changed. */
400 log->type->set_region_sync(log, region, 0);
401
402 read_lock(&rh->hash_lock);
403 reg = __rh_find(rh, region);
404 read_unlock(&rh->hash_lock);
405
406 /* region hash entry should exist because write was in-flight */
407 BUG_ON(!reg);
408 BUG_ON(!list_empty(&reg->list));
409
410 spin_lock_irqsave(&rh->region_lock, flags);
411 /*
412 * Possible cases:
413 * 1) DM_RH_DIRTY
414 * 2) DM_RH_NOSYNC: was dirty, other preceeding writes failed
415 * 3) DM_RH_RECOVERING: flushing pending writes
416 * Either case, the region should have not been connected to list.
417 */
418 recovering = (reg->state == DM_RH_RECOVERING);
419 reg->state = DM_RH_NOSYNC;
420 BUG_ON(!list_empty(&reg->list));
421 spin_unlock_irqrestore(&rh->region_lock, flags);
422
423 bio_endio(bio, error);
424 if (recovering)
425 complete_resync_work(reg, 0);
426}
427EXPORT_SYMBOL_GPL(dm_rh_mark_nosync);
428
429void dm_rh_update_states(struct dm_region_hash *rh, int errors_handled)
430{
431 struct dm_region *reg, *next;
432
433 LIST_HEAD(clean);
434 LIST_HEAD(recovered);
435 LIST_HEAD(failed_recovered);
436
437 /*
438 * Quickly grab the lists.
439 */
440 write_lock_irq(&rh->hash_lock);
441 spin_lock(&rh->region_lock);
442 if (!list_empty(&rh->clean_regions)) {
443 list_splice_init(&rh->clean_regions, &clean);
444
445 list_for_each_entry(reg, &clean, list)
446 list_del(&reg->hash_list);
447 }
448
449 if (!list_empty(&rh->recovered_regions)) {
450 list_splice_init(&rh->recovered_regions, &recovered);
451
452 list_for_each_entry(reg, &recovered, list)
453 list_del(&reg->hash_list);
454 }
455
456 if (!list_empty(&rh->failed_recovered_regions)) {
457 list_splice_init(&rh->failed_recovered_regions,
458 &failed_recovered);
459
460 list_for_each_entry(reg, &failed_recovered, list)
461 list_del(&reg->hash_list);
462 }
463
464 spin_unlock(&rh->region_lock);
465 write_unlock_irq(&rh->hash_lock);
466
467 /*
468 * All the regions on the recovered and clean lists have
469 * now been pulled out of the system, so no need to do
470 * any more locking.
471 */
472 list_for_each_entry_safe(reg, next, &recovered, list) {
473 rh->log->type->clear_region(rh->log, reg->key);
474 complete_resync_work(reg, 1);
475 mempool_free(reg, rh->region_pool);
476 }
477
478 list_for_each_entry_safe(reg, next, &failed_recovered, list) {
479 complete_resync_work(reg, errors_handled ? 0 : 1);
480 mempool_free(reg, rh->region_pool);
481 }
482
483 list_for_each_entry_safe(reg, next, &clean, list) {
484 rh->log->type->clear_region(rh->log, reg->key);
485 mempool_free(reg, rh->region_pool);
486 }
487
488 rh->log->type->flush(rh->log);
489}
490EXPORT_SYMBOL_GPL(dm_rh_update_states);
491
492static void rh_inc(struct dm_region_hash *rh, region_t region)
493{
494 struct dm_region *reg;
495
496 read_lock(&rh->hash_lock);
497 reg = __rh_find(rh, region);
498
499 spin_lock_irq(&rh->region_lock);
500 atomic_inc(&reg->pending);
501
502 if (reg->state == DM_RH_CLEAN) {
503 reg->state = DM_RH_DIRTY;
504 list_del_init(&reg->list); /* take off the clean list */
505 spin_unlock_irq(&rh->region_lock);
506
507 rh->log->type->mark_region(rh->log, reg->key);
508 } else
509 spin_unlock_irq(&rh->region_lock);
510
511
512 read_unlock(&rh->hash_lock);
513}
514
515void dm_rh_inc_pending(struct dm_region_hash *rh, struct bio_list *bios)
516{
517 struct bio *bio;
518
519 for (bio = bios->head; bio; bio = bio->bi_next)
520 rh_inc(rh, dm_rh_bio_to_region(rh, bio));
521}
522EXPORT_SYMBOL_GPL(dm_rh_inc_pending);
523
524void dm_rh_dec(struct dm_region_hash *rh, region_t region)
525{
526 unsigned long flags;
527 struct dm_region *reg;
528 int should_wake = 0;
529
530 read_lock(&rh->hash_lock);
531 reg = __rh_lookup(rh, region);
532 read_unlock(&rh->hash_lock);
533
534 spin_lock_irqsave(&rh->region_lock, flags);
535 if (atomic_dec_and_test(&reg->pending)) {
536 /*
537 * There is no pending I/O for this region.
538 * We can move the region to corresponding list for next action.
539 * At this point, the region is not yet connected to any list.
540 *
541 * If the state is DM_RH_NOSYNC, the region should be kept off
542 * from clean list.
543 * The hash entry for DM_RH_NOSYNC will remain in memory
544 * until the region is recovered or the map is reloaded.
545 */
546
547 /* do nothing for DM_RH_NOSYNC */
548 if (reg->state == DM_RH_RECOVERING) {
549 list_add_tail(&reg->list, &rh->quiesced_regions);
550 } else if (reg->state == DM_RH_DIRTY) {
551 reg->state = DM_RH_CLEAN;
552 list_add(&reg->list, &rh->clean_regions);
553 }
554 should_wake = 1;
555 }
556 spin_unlock_irqrestore(&rh->region_lock, flags);
557
558 if (should_wake)
559 rh->wakeup_workers(rh->context);
560}
561EXPORT_SYMBOL_GPL(dm_rh_dec);
562
563/*
564 * Starts quiescing a region in preparation for recovery.
565 */
566static int __rh_recovery_prepare(struct dm_region_hash *rh)
567{
568 int r;
569 region_t region;
570 struct dm_region *reg;
571
572 /*
573 * Ask the dirty log what's next.
574 */
575 r = rh->log->type->get_resync_work(rh->log, &region);
576 if (r <= 0)
577 return r;
578
579 /*
580 * Get this region, and start it quiescing by setting the
581 * recovering flag.
582 */
583 read_lock(&rh->hash_lock);
584 reg = __rh_find(rh, region);
585 read_unlock(&rh->hash_lock);
586
587 spin_lock_irq(&rh->region_lock);
588 reg->state = DM_RH_RECOVERING;
589
590 /* Already quiesced ? */
591 if (atomic_read(&reg->pending))
592 list_del_init(&reg->list);
593 else
594 list_move(&reg->list, &rh->quiesced_regions);
595
596 spin_unlock_irq(&rh->region_lock);
597
598 return 1;
599}
600
601void dm_rh_recovery_prepare(struct dm_region_hash *rh)
602{
603 /* Extra reference to avoid race with dm_rh_stop_recovery */
604 atomic_inc(&rh->recovery_in_flight);
605
606 while (!down_trylock(&rh->recovery_count)) {
607 atomic_inc(&rh->recovery_in_flight);
608 if (__rh_recovery_prepare(rh) <= 0) {
609 atomic_dec(&rh->recovery_in_flight);
610 up(&rh->recovery_count);
611 break;
612 }
613 }
614
615 /* Drop the extra reference */
616 if (atomic_dec_and_test(&rh->recovery_in_flight))
617 rh->wakeup_all_recovery_waiters(rh->context);
618}
619EXPORT_SYMBOL_GPL(dm_rh_recovery_prepare);
620
621/*
622 * Returns any quiesced regions.
623 */
624struct dm_region *dm_rh_recovery_start(struct dm_region_hash *rh)
625{
626 struct dm_region *reg = NULL;
627
628 spin_lock_irq(&rh->region_lock);
629 if (!list_empty(&rh->quiesced_regions)) {
630 reg = list_entry(rh->quiesced_regions.next,
631 struct dm_region, list);
632 list_del_init(&reg->list); /* remove from the quiesced list */
633 }
634 spin_unlock_irq(&rh->region_lock);
635
636 return reg;
637}
638EXPORT_SYMBOL_GPL(dm_rh_recovery_start);
639
640void dm_rh_recovery_end(struct dm_region *reg, int success)
641{
642 struct dm_region_hash *rh = reg->rh;
643
644 spin_lock_irq(&rh->region_lock);
645 if (success)
646 list_add(&reg->list, &reg->rh->recovered_regions);
647 else {
648 reg->state = DM_RH_NOSYNC;
649 list_add(&reg->list, &reg->rh->failed_recovered_regions);
650 }
651 spin_unlock_irq(&rh->region_lock);
652
653 rh->wakeup_workers(rh->context);
654}
655EXPORT_SYMBOL_GPL(dm_rh_recovery_end);
656
657/* Return recovery in flight count. */
658int dm_rh_recovery_in_flight(struct dm_region_hash *rh)
659{
660 return atomic_read(&rh->recovery_in_flight);
661}
662EXPORT_SYMBOL_GPL(dm_rh_recovery_in_flight);
663
664int dm_rh_flush(struct dm_region_hash *rh)
665{
666 return rh->log->type->flush(rh->log);
667}
668EXPORT_SYMBOL_GPL(dm_rh_flush);
669
670void dm_rh_delay(struct dm_region_hash *rh, struct bio *bio)
671{
672 struct dm_region *reg;
673
674 read_lock(&rh->hash_lock);
675 reg = __rh_find(rh, dm_rh_bio_to_region(rh, bio));
676 bio_list_add(&reg->delayed_bios, bio);
677 read_unlock(&rh->hash_lock);
678}
679EXPORT_SYMBOL_GPL(dm_rh_delay);
680
681void dm_rh_stop_recovery(struct dm_region_hash *rh)
682{
683 int i;
684
685 /* wait for any recovering regions */
686 for (i = 0; i < rh->max_recovery; i++)
687 down(&rh->recovery_count);
688}
689EXPORT_SYMBOL_GPL(dm_rh_stop_recovery);
690
691void dm_rh_start_recovery(struct dm_region_hash *rh)
692{
693 int i;
694
695 for (i = 0; i < rh->max_recovery; i++)
696 up(&rh->recovery_count);
697
698 rh->wakeup_workers(rh->context);
699}
700EXPORT_SYMBOL_GPL(dm_rh_start_recovery);
701
702MODULE_DESCRIPTION(DM_NAME " region hash");
703MODULE_AUTHOR("Joe Thornber/Heinz Mauelshagen <dm-devel@redhat.com>");
704MODULE_LICENSE("GPL");