2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/dm-dirty-log.h>
9 #include <linux/dm-region-hash.h>
11 #include <linux/ctype.h>
12 #include <linux/init.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/vmalloc.h>
19 #define DM_MSG_PREFIX "region hash"
21 /*-----------------------------------------------------------------
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.
28 * In addition to being present in the hash table a region _may_
29 * be present on one of three lists.
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
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.
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
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
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 *---------------------------------------------------------------*/
57 struct dm_region_hash {
59 unsigned region_shift;
61 /* holds persistent region state */
62 struct dm_dirty_log *log;
70 struct list_head *buckets;
73 * If there was a flush failure no regions can be marked clean.
77 unsigned max_recovery; /* Max # of regions to recover in parallel */
79 spinlock_t region_lock;
80 atomic_t recovery_in_flight;
81 struct list_head clean_regions;
82 struct list_head quiesced_regions;
83 struct list_head recovered_regions;
84 struct list_head failed_recovered_regions;
85 struct semaphore recovery_count;
87 mempool_t region_pool;
90 sector_t target_begin;
92 /* Callback function to schedule bios writes */
93 void (*dispatch_bios)(void *context, struct bio_list *bios);
95 /* Callback function to wakeup callers worker thread. */
96 void (*wakeup_workers)(void *context);
98 /* Callback function to wakeup callers recovery waiters. */
99 void (*wakeup_all_recovery_waiters)(void *context);
103 struct dm_region_hash *rh; /* FIXME: can we get rid of this ? */
107 struct list_head hash_list;
108 struct list_head list;
111 struct bio_list delayed_bios;
117 static region_t dm_rh_sector_to_region(struct dm_region_hash *rh, sector_t sector)
119 return sector >> rh->region_shift;
122 sector_t dm_rh_region_to_sector(struct dm_region_hash *rh, region_t region)
124 return region << rh->region_shift;
126 EXPORT_SYMBOL_GPL(dm_rh_region_to_sector);
128 region_t dm_rh_bio_to_region(struct dm_region_hash *rh, struct bio *bio)
130 return dm_rh_sector_to_region(rh, bio->bi_iter.bi_sector -
133 EXPORT_SYMBOL_GPL(dm_rh_bio_to_region);
135 void *dm_rh_region_context(struct dm_region *reg)
137 return reg->rh->context;
139 EXPORT_SYMBOL_GPL(dm_rh_region_context);
141 region_t dm_rh_get_region_key(struct dm_region *reg)
145 EXPORT_SYMBOL_GPL(dm_rh_get_region_key);
147 sector_t dm_rh_get_region_size(struct dm_region_hash *rh)
149 return rh->region_size;
151 EXPORT_SYMBOL_GPL(dm_rh_get_region_size);
154 * FIXME: shall we pass in a structure instead of all these args to
155 * dm_region_hash_create()????
157 #define RH_HASH_MULT 2654435387U
158 #define RH_HASH_SHIFT 12
160 #define MIN_REGIONS 64
161 struct dm_region_hash *dm_region_hash_create(
162 void *context, void (*dispatch_bios)(void *context,
163 struct bio_list *bios),
164 void (*wakeup_workers)(void *context),
165 void (*wakeup_all_recovery_waiters)(void *context),
166 sector_t target_begin, unsigned max_recovery,
167 struct dm_dirty_log *log, uint32_t region_size,
170 struct dm_region_hash *rh;
171 unsigned nr_buckets, max_buckets;
176 * Calculate a suitable number of buckets for our hash
179 max_buckets = nr_regions >> 6;
180 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
184 rh = kzalloc(sizeof(*rh), GFP_KERNEL);
186 DMERR("unable to allocate region hash memory");
187 return ERR_PTR(-ENOMEM);
190 rh->context = context;
191 rh->dispatch_bios = dispatch_bios;
192 rh->wakeup_workers = wakeup_workers;
193 rh->wakeup_all_recovery_waiters = wakeup_all_recovery_waiters;
194 rh->target_begin = target_begin;
195 rh->max_recovery = max_recovery;
197 rh->region_size = region_size;
198 rh->region_shift = __ffs(region_size);
199 rwlock_init(&rh->hash_lock);
200 rh->mask = nr_buckets - 1;
201 rh->nr_buckets = nr_buckets;
203 rh->shift = RH_HASH_SHIFT;
204 rh->prime = RH_HASH_MULT;
206 rh->buckets = vmalloc(array_size(nr_buckets, sizeof(*rh->buckets)));
208 DMERR("unable to allocate region hash bucket memory");
210 return ERR_PTR(-ENOMEM);
213 for (i = 0; i < nr_buckets; i++)
214 INIT_LIST_HEAD(rh->buckets + i);
216 spin_lock_init(&rh->region_lock);
217 sema_init(&rh->recovery_count, 0);
218 atomic_set(&rh->recovery_in_flight, 0);
219 INIT_LIST_HEAD(&rh->clean_regions);
220 INIT_LIST_HEAD(&rh->quiesced_regions);
221 INIT_LIST_HEAD(&rh->recovered_regions);
222 INIT_LIST_HEAD(&rh->failed_recovered_regions);
223 rh->flush_failure = 0;
225 ret = mempool_init_kmalloc_pool(&rh->region_pool, MIN_REGIONS,
226 sizeof(struct dm_region));
230 rh = ERR_PTR(-ENOMEM);
235 EXPORT_SYMBOL_GPL(dm_region_hash_create);
237 void dm_region_hash_destroy(struct dm_region_hash *rh)
240 struct dm_region *reg, *nreg;
242 BUG_ON(!list_empty(&rh->quiesced_regions));
243 for (h = 0; h < rh->nr_buckets; h++) {
244 list_for_each_entry_safe(reg, nreg, rh->buckets + h,
246 BUG_ON(atomic_read(®->pending));
247 mempool_free(reg, &rh->region_pool);
252 dm_dirty_log_destroy(rh->log);
254 mempool_exit(&rh->region_pool);
258 EXPORT_SYMBOL_GPL(dm_region_hash_destroy);
260 struct dm_dirty_log *dm_rh_dirty_log(struct dm_region_hash *rh)
264 EXPORT_SYMBOL_GPL(dm_rh_dirty_log);
266 static unsigned rh_hash(struct dm_region_hash *rh, region_t region)
268 return (unsigned) ((region * rh->prime) >> rh->shift) & rh->mask;
271 static struct dm_region *__rh_lookup(struct dm_region_hash *rh, region_t region)
273 struct dm_region *reg;
274 struct list_head *bucket = rh->buckets + rh_hash(rh, region);
276 list_for_each_entry(reg, bucket, hash_list)
277 if (reg->key == region)
283 static void __rh_insert(struct dm_region_hash *rh, struct dm_region *reg)
285 list_add(®->hash_list, rh->buckets + rh_hash(rh, reg->key));
288 static struct dm_region *__rh_alloc(struct dm_region_hash *rh, region_t region)
290 struct dm_region *reg, *nreg;
292 nreg = mempool_alloc(&rh->region_pool, GFP_ATOMIC);
294 nreg = kmalloc(sizeof(*nreg), GFP_NOIO | __GFP_NOFAIL);
296 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
297 DM_RH_CLEAN : DM_RH_NOSYNC;
300 INIT_LIST_HEAD(&nreg->list);
301 atomic_set(&nreg->pending, 0);
302 bio_list_init(&nreg->delayed_bios);
304 write_lock_irq(&rh->hash_lock);
305 reg = __rh_lookup(rh, region);
307 /* We lost the race. */
308 mempool_free(nreg, &rh->region_pool);
310 __rh_insert(rh, nreg);
311 if (nreg->state == DM_RH_CLEAN) {
312 spin_lock(&rh->region_lock);
313 list_add(&nreg->list, &rh->clean_regions);
314 spin_unlock(&rh->region_lock);
319 write_unlock_irq(&rh->hash_lock);
324 static struct dm_region *__rh_find(struct dm_region_hash *rh, region_t region)
326 struct dm_region *reg;
328 reg = __rh_lookup(rh, region);
330 read_unlock(&rh->hash_lock);
331 reg = __rh_alloc(rh, region);
332 read_lock(&rh->hash_lock);
338 int dm_rh_get_state(struct dm_region_hash *rh, region_t region, int may_block)
341 struct dm_region *reg;
343 read_lock(&rh->hash_lock);
344 reg = __rh_lookup(rh, region);
345 read_unlock(&rh->hash_lock);
351 * The region wasn't in the hash, so we fall back to the
354 r = rh->log->type->in_sync(rh->log, region, may_block);
357 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
358 * taken as a DM_RH_NOSYNC
360 return r == 1 ? DM_RH_CLEAN : DM_RH_NOSYNC;
362 EXPORT_SYMBOL_GPL(dm_rh_get_state);
364 static void complete_resync_work(struct dm_region *reg, int success)
366 struct dm_region_hash *rh = reg->rh;
368 rh->log->type->set_region_sync(rh->log, reg->key, success);
371 * Dispatch the bios before we call 'wake_up_all'.
372 * This is important because if we are suspending,
373 * we want to know that recovery is complete and
374 * the work queue is flushed. If we wake_up_all
375 * before we dispatch_bios (queue bios and call wake()),
376 * then we risk suspending before the work queue
377 * has been properly flushed.
379 rh->dispatch_bios(rh->context, ®->delayed_bios);
380 if (atomic_dec_and_test(&rh->recovery_in_flight))
381 rh->wakeup_all_recovery_waiters(rh->context);
382 up(&rh->recovery_count);
389 * The bio was written on some mirror(s) but failed on other mirror(s).
390 * We can successfully endio the bio but should avoid the region being
391 * marked clean by setting the state DM_RH_NOSYNC.
393 * This function is _not_ safe in interrupt context!
395 void dm_rh_mark_nosync(struct dm_region_hash *rh, struct bio *bio)
398 struct dm_dirty_log *log = rh->log;
399 struct dm_region *reg;
400 region_t region = dm_rh_bio_to_region(rh, bio);
403 if (bio->bi_opf & REQ_PREFLUSH) {
404 rh->flush_failure = 1;
408 if (bio_op(bio) == REQ_OP_DISCARD)
411 /* We must inform the log that the sync count has changed. */
412 log->type->set_region_sync(log, region, 0);
414 read_lock(&rh->hash_lock);
415 reg = __rh_find(rh, region);
416 read_unlock(&rh->hash_lock);
418 /* region hash entry should exist because write was in-flight */
420 BUG_ON(!list_empty(®->list));
422 spin_lock_irqsave(&rh->region_lock, flags);
426 * 2) DM_RH_NOSYNC: was dirty, other preceding writes failed
427 * 3) DM_RH_RECOVERING: flushing pending writes
428 * Either case, the region should have not been connected to list.
430 recovering = (reg->state == DM_RH_RECOVERING);
431 reg->state = DM_RH_NOSYNC;
432 BUG_ON(!list_empty(®->list));
433 spin_unlock_irqrestore(&rh->region_lock, flags);
436 complete_resync_work(reg, 0);
438 EXPORT_SYMBOL_GPL(dm_rh_mark_nosync);
440 void dm_rh_update_states(struct dm_region_hash *rh, int errors_handled)
442 struct dm_region *reg, *next;
445 LIST_HEAD(recovered);
446 LIST_HEAD(failed_recovered);
449 * Quickly grab the lists.
451 write_lock_irq(&rh->hash_lock);
452 spin_lock(&rh->region_lock);
453 if (!list_empty(&rh->clean_regions)) {
454 list_splice_init(&rh->clean_regions, &clean);
456 list_for_each_entry(reg, &clean, list)
457 list_del(®->hash_list);
460 if (!list_empty(&rh->recovered_regions)) {
461 list_splice_init(&rh->recovered_regions, &recovered);
463 list_for_each_entry(reg, &recovered, list)
464 list_del(®->hash_list);
467 if (!list_empty(&rh->failed_recovered_regions)) {
468 list_splice_init(&rh->failed_recovered_regions,
471 list_for_each_entry(reg, &failed_recovered, list)
472 list_del(®->hash_list);
475 spin_unlock(&rh->region_lock);
476 write_unlock_irq(&rh->hash_lock);
479 * All the regions on the recovered and clean lists have
480 * now been pulled out of the system, so no need to do
483 list_for_each_entry_safe(reg, next, &recovered, list) {
484 rh->log->type->clear_region(rh->log, reg->key);
485 complete_resync_work(reg, 1);
486 mempool_free(reg, &rh->region_pool);
489 list_for_each_entry_safe(reg, next, &failed_recovered, list) {
490 complete_resync_work(reg, errors_handled ? 0 : 1);
491 mempool_free(reg, &rh->region_pool);
494 list_for_each_entry_safe(reg, next, &clean, list) {
495 rh->log->type->clear_region(rh->log, reg->key);
496 mempool_free(reg, &rh->region_pool);
499 rh->log->type->flush(rh->log);
501 EXPORT_SYMBOL_GPL(dm_rh_update_states);
503 static void rh_inc(struct dm_region_hash *rh, region_t region)
505 struct dm_region *reg;
507 read_lock(&rh->hash_lock);
508 reg = __rh_find(rh, region);
510 spin_lock_irq(&rh->region_lock);
511 atomic_inc(®->pending);
513 if (reg->state == DM_RH_CLEAN) {
514 reg->state = DM_RH_DIRTY;
515 list_del_init(®->list); /* take off the clean list */
516 spin_unlock_irq(&rh->region_lock);
518 rh->log->type->mark_region(rh->log, reg->key);
520 spin_unlock_irq(&rh->region_lock);
523 read_unlock(&rh->hash_lock);
526 void dm_rh_inc_pending(struct dm_region_hash *rh, struct bio_list *bios)
530 for (bio = bios->head; bio; bio = bio->bi_next) {
531 if (bio->bi_opf & REQ_PREFLUSH || bio_op(bio) == REQ_OP_DISCARD)
533 rh_inc(rh, dm_rh_bio_to_region(rh, bio));
536 EXPORT_SYMBOL_GPL(dm_rh_inc_pending);
538 void dm_rh_dec(struct dm_region_hash *rh, region_t region)
541 struct dm_region *reg;
544 read_lock(&rh->hash_lock);
545 reg = __rh_lookup(rh, region);
546 read_unlock(&rh->hash_lock);
548 spin_lock_irqsave(&rh->region_lock, flags);
549 if (atomic_dec_and_test(®->pending)) {
551 * There is no pending I/O for this region.
552 * We can move the region to corresponding list for next action.
553 * At this point, the region is not yet connected to any list.
555 * If the state is DM_RH_NOSYNC, the region should be kept off
557 * The hash entry for DM_RH_NOSYNC will remain in memory
558 * until the region is recovered or the map is reloaded.
561 /* do nothing for DM_RH_NOSYNC */
562 if (unlikely(rh->flush_failure)) {
564 * If a write flush failed some time ago, we
565 * don't know whether or not this write made it
566 * to the disk, so we must resync the device.
568 reg->state = DM_RH_NOSYNC;
569 } else if (reg->state == DM_RH_RECOVERING) {
570 list_add_tail(®->list, &rh->quiesced_regions);
571 } else if (reg->state == DM_RH_DIRTY) {
572 reg->state = DM_RH_CLEAN;
573 list_add(®->list, &rh->clean_regions);
577 spin_unlock_irqrestore(&rh->region_lock, flags);
580 rh->wakeup_workers(rh->context);
582 EXPORT_SYMBOL_GPL(dm_rh_dec);
585 * Starts quiescing a region in preparation for recovery.
587 static int __rh_recovery_prepare(struct dm_region_hash *rh)
591 struct dm_region *reg;
594 * Ask the dirty log what's next.
596 r = rh->log->type->get_resync_work(rh->log, ®ion);
601 * Get this region, and start it quiescing by setting the
604 read_lock(&rh->hash_lock);
605 reg = __rh_find(rh, region);
606 read_unlock(&rh->hash_lock);
608 spin_lock_irq(&rh->region_lock);
609 reg->state = DM_RH_RECOVERING;
611 /* Already quiesced ? */
612 if (atomic_read(®->pending))
613 list_del_init(®->list);
615 list_move(®->list, &rh->quiesced_regions);
617 spin_unlock_irq(&rh->region_lock);
622 void dm_rh_recovery_prepare(struct dm_region_hash *rh)
624 /* Extra reference to avoid race with dm_rh_stop_recovery */
625 atomic_inc(&rh->recovery_in_flight);
627 while (!down_trylock(&rh->recovery_count)) {
628 atomic_inc(&rh->recovery_in_flight);
629 if (__rh_recovery_prepare(rh) <= 0) {
630 atomic_dec(&rh->recovery_in_flight);
631 up(&rh->recovery_count);
636 /* Drop the extra reference */
637 if (atomic_dec_and_test(&rh->recovery_in_flight))
638 rh->wakeup_all_recovery_waiters(rh->context);
640 EXPORT_SYMBOL_GPL(dm_rh_recovery_prepare);
643 * Returns any quiesced regions.
645 struct dm_region *dm_rh_recovery_start(struct dm_region_hash *rh)
647 struct dm_region *reg = NULL;
649 spin_lock_irq(&rh->region_lock);
650 if (!list_empty(&rh->quiesced_regions)) {
651 reg = list_entry(rh->quiesced_regions.next,
652 struct dm_region, list);
653 list_del_init(®->list); /* remove from the quiesced list */
655 spin_unlock_irq(&rh->region_lock);
659 EXPORT_SYMBOL_GPL(dm_rh_recovery_start);
661 void dm_rh_recovery_end(struct dm_region *reg, int success)
663 struct dm_region_hash *rh = reg->rh;
665 spin_lock_irq(&rh->region_lock);
667 list_add(®->list, ®->rh->recovered_regions);
669 list_add(®->list, ®->rh->failed_recovered_regions);
671 spin_unlock_irq(&rh->region_lock);
673 rh->wakeup_workers(rh->context);
675 EXPORT_SYMBOL_GPL(dm_rh_recovery_end);
677 /* Return recovery in flight count. */
678 int dm_rh_recovery_in_flight(struct dm_region_hash *rh)
680 return atomic_read(&rh->recovery_in_flight);
682 EXPORT_SYMBOL_GPL(dm_rh_recovery_in_flight);
684 int dm_rh_flush(struct dm_region_hash *rh)
686 return rh->log->type->flush(rh->log);
688 EXPORT_SYMBOL_GPL(dm_rh_flush);
690 void dm_rh_delay(struct dm_region_hash *rh, struct bio *bio)
692 struct dm_region *reg;
694 read_lock(&rh->hash_lock);
695 reg = __rh_find(rh, dm_rh_bio_to_region(rh, bio));
696 bio_list_add(®->delayed_bios, bio);
697 read_unlock(&rh->hash_lock);
699 EXPORT_SYMBOL_GPL(dm_rh_delay);
701 void dm_rh_stop_recovery(struct dm_region_hash *rh)
705 /* wait for any recovering regions */
706 for (i = 0; i < rh->max_recovery; i++)
707 down(&rh->recovery_count);
709 EXPORT_SYMBOL_GPL(dm_rh_stop_recovery);
711 void dm_rh_start_recovery(struct dm_region_hash *rh)
715 for (i = 0; i < rh->max_recovery; i++)
716 up(&rh->recovery_count);
718 rh->wakeup_workers(rh->context);
720 EXPORT_SYMBOL_GPL(dm_rh_start_recovery);
722 MODULE_DESCRIPTION(DM_NAME " region hash");
723 MODULE_AUTHOR("Joe Thornber/Heinz Mauelshagen <dm-devel@redhat.com>");
724 MODULE_LICENSE("GPL");