1/*
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/slab.h>
15#include <linux/vmalloc.h>
16
17#include "dm.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 unsigned mask;
67 unsigned nr_buckets;
68 unsigned prime;
69 unsigned shift;
70 struct list_head *buckets;
71
72 /*
73 * If there was a flush failure no regions can be marked clean.
74 */
75 int flush_failure;
76
77 unsigned max_recovery; /* Max # of regions to recover in parallel */
78
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;
86
87 mempool_t region_pool;
88
89 void *context;
90 sector_t target_begin;
91
92 /* Callback function to schedule bios writes */
93 void (*dispatch_bios)(void *context, struct bio_list *bios);
94
95 /* Callback function to wakeup callers worker thread. */
96 void (*wakeup_workers)(void *context);
97
98 /* Callback function to wakeup callers recovery waiters. */
99 void (*wakeup_all_recovery_waiters)(void *context);
100};
101
102struct dm_region {
103 struct dm_region_hash *rh; /* FIXME: can we get rid of this ? */
104 region_t key;
105 int state;
106
107 struct list_head hash_list;
108 struct list_head list;
109
110 atomic_t pending;
111 struct bio_list delayed_bios;
112};
113
114/*
115 * Conversion fns
116 */
117static region_t dm_rh_sector_to_region(struct dm_region_hash *rh, sector_t sector)
118{
119 return sector >> rh->region_shift;
120}
121
122sector_t dm_rh_region_to_sector(struct dm_region_hash *rh, region_t region)
123{
124 return region << rh->region_shift;
125}
126EXPORT_SYMBOL_GPL(dm_rh_region_to_sector);
127
128region_t dm_rh_bio_to_region(struct dm_region_hash *rh, struct bio *bio)
129{
130 return dm_rh_sector_to_region(rh, bio->bi_iter.bi_sector -
131 rh->target_begin);
132}
133EXPORT_SYMBOL_GPL(dm_rh_bio_to_region);
134
135void *dm_rh_region_context(struct dm_region *reg)
136{
137 return reg->rh->context;
138}
139EXPORT_SYMBOL_GPL(dm_rh_region_context);
140
141region_t dm_rh_get_region_key(struct dm_region *reg)
142{
143 return reg->key;
144}
145EXPORT_SYMBOL_GPL(dm_rh_get_region_key);
146
147sector_t dm_rh_get_region_size(struct dm_region_hash *rh)
148{
149 return rh->region_size;
150}
151EXPORT_SYMBOL_GPL(dm_rh_get_region_size);
152
153/*
154 * FIXME: shall we pass in a structure instead of all these args to
155 * dm_region_hash_create()????
156 */
157#define RH_HASH_MULT 2654435387U
158#define RH_HASH_SHIFT 12
159
160#define MIN_REGIONS 64
161struct 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,
168 region_t nr_regions)
169{
170 struct dm_region_hash *rh;
171 unsigned nr_buckets, max_buckets;
172 size_t i;
173 int ret;
174
175 /*
176 * Calculate a suitable number of buckets for our hash
177 * table.
178 */
179 max_buckets = nr_regions >> 6;
180 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
181 ;
182 nr_buckets >>= 1;
183
184 rh = kzalloc(sizeof(*rh), GFP_KERNEL);
185 if (!rh) {
186 DMERR("unable to allocate region hash memory");
187 return ERR_PTR(-ENOMEM);
188 }
189
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;
196 rh->log = log;
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;
202
203 rh->shift = RH_HASH_SHIFT;
204 rh->prime = RH_HASH_MULT;
205
206 rh->buckets = vmalloc(array_size(nr_buckets, sizeof(*rh->buckets)));
207 if (!rh->buckets) {
208 DMERR("unable to allocate region hash bucket memory");
209 kfree(rh);
210 return ERR_PTR(-ENOMEM);
211 }
212
213 for (i = 0; i < nr_buckets; i++)
214 INIT_LIST_HEAD(rh->buckets + i);
215
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;
224
225 ret = mempool_init_kmalloc_pool(&rh->region_pool, MIN_REGIONS,
226 sizeof(struct dm_region));
227 if (ret) {
228 vfree(rh->buckets);
229 kfree(rh);
230 rh = ERR_PTR(-ENOMEM);
231 }
232
233 return rh;
234}
235EXPORT_SYMBOL_GPL(dm_region_hash_create);
236
237void dm_region_hash_destroy(struct dm_region_hash *rh)
238{
239 unsigned h;
240 struct dm_region *reg, *nreg;
241
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,
245 hash_list) {
246 BUG_ON(atomic_read(&reg->pending));
247 mempool_free(reg, &rh->region_pool);
248 }
249 }
250
251 if (rh->log)
252 dm_dirty_log_destroy(rh->log);
253
254 mempool_exit(&rh->region_pool);
255 vfree(rh->buckets);
256 kfree(rh);
257}
258EXPORT_SYMBOL_GPL(dm_region_hash_destroy);
259
260struct dm_dirty_log *dm_rh_dirty_log(struct dm_region_hash *rh)
261{
262 return rh->log;
263}
264EXPORT_SYMBOL_GPL(dm_rh_dirty_log);
265
266static unsigned rh_hash(struct dm_region_hash *rh, region_t region)
267{
268 return (unsigned) ((region * rh->prime) >> rh->shift) & rh->mask;
269}
270
271static struct dm_region *__rh_lookup(struct dm_region_hash *rh, region_t region)
272{
273 struct dm_region *reg;
274 struct list_head *bucket = rh->buckets + rh_hash(rh, region);
275
276 list_for_each_entry(reg, bucket, hash_list)
277 if (reg->key == region)
278 return reg;
279
280 return NULL;
281}
282
283static void __rh_insert(struct dm_region_hash *rh, struct dm_region *reg)
284{
285 list_add(&reg->hash_list, rh->buckets + rh_hash(rh, reg->key));
286}
287
288static struct dm_region *__rh_alloc(struct dm_region_hash *rh, region_t region)
289{
290 struct dm_region *reg, *nreg;
291
292 nreg = mempool_alloc(&rh->region_pool, GFP_ATOMIC);
293 if (unlikely(!nreg))
294 nreg = kmalloc(sizeof(*nreg), GFP_NOIO | __GFP_NOFAIL);
295
296 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
297 DM_RH_CLEAN : DM_RH_NOSYNC;
298 nreg->rh = rh;
299 nreg->key = region;
300 INIT_LIST_HEAD(&nreg->list);
301 atomic_set(&nreg->pending, 0);
302 bio_list_init(&nreg->delayed_bios);
303
304 write_lock_irq(&rh->hash_lock);
305 reg = __rh_lookup(rh, region);
306 if (reg)
307 /* We lost the race. */
308 mempool_free(nreg, &rh->region_pool);
309 else {
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);
315 }
316
317 reg = nreg;
318 }
319 write_unlock_irq(&rh->hash_lock);
320
321 return reg;
322}
323
324static struct dm_region *__rh_find(struct dm_region_hash *rh, region_t region)
325{
326 struct dm_region *reg;
327
328 reg = __rh_lookup(rh, region);
329 if (!reg) {
330 read_unlock(&rh->hash_lock);
331 reg = __rh_alloc(rh, region);
332 read_lock(&rh->hash_lock);
333 }
334
335 return reg;
336}
337
338int dm_rh_get_state(struct dm_region_hash *rh, region_t region, int may_block)
339{
340 int r;
341 struct dm_region *reg;
342
343 read_lock(&rh->hash_lock);
344 reg = __rh_lookup(rh, region);
345 read_unlock(&rh->hash_lock);
346
347 if (reg)
348 return reg->state;
349
350 /*
351 * The region wasn't in the hash, so we fall back to the
352 * dirty log.
353 */
354 r = rh->log->type->in_sync(rh->log, region, may_block);
355
356 /*
357 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
358 * taken as a DM_RH_NOSYNC
359 */
360 return r == 1 ? DM_RH_CLEAN : DM_RH_NOSYNC;
361}
362EXPORT_SYMBOL_GPL(dm_rh_get_state);
363
364static void complete_resync_work(struct dm_region *reg, int success)
365{
366 struct dm_region_hash *rh = reg->rh;
367
368 rh->log->type->set_region_sync(rh->log, reg->key, success);
369
370 /*
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.
378 */
379 rh->dispatch_bios(rh->context, &reg->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);
383}
384
385/* dm_rh_mark_nosync
386 * @ms
387 * @bio
388 *
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.
392 *
393 * This function is _not_ safe in interrupt context!
394 */
395void dm_rh_mark_nosync(struct dm_region_hash *rh, struct bio *bio)
396{
397 unsigned long flags;
398 struct dm_dirty_log *log = rh->log;
399 struct dm_region *reg;
400 region_t region = dm_rh_bio_to_region(rh, bio);
401 int recovering = 0;
402
403 if (bio->bi_opf & REQ_PREFLUSH) {
404 rh->flush_failure = 1;
405 return;
406 }
407
408 if (bio_op(bio) == REQ_OP_DISCARD)
409 return;
410
411 /* We must inform the log that the sync count has changed. */
412 log->type->set_region_sync(log, region, 0);
413
414 read_lock(&rh->hash_lock);
415 reg = __rh_find(rh, region);
416 read_unlock(&rh->hash_lock);
417
418 /* region hash entry should exist because write was in-flight */
419 BUG_ON(!reg);
420 BUG_ON(!list_empty(&reg->list));
421
422 spin_lock_irqsave(&rh->region_lock, flags);
423 /*
424 * Possible cases:
425 * 1) DM_RH_DIRTY
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.
429 */
430 recovering = (reg->state == DM_RH_RECOVERING);
431 reg->state = DM_RH_NOSYNC;
432 BUG_ON(!list_empty(&reg->list));
433 spin_unlock_irqrestore(&rh->region_lock, flags);
434
435 if (recovering)
436 complete_resync_work(reg, 0);
437}
438EXPORT_SYMBOL_GPL(dm_rh_mark_nosync);
439
440void dm_rh_update_states(struct dm_region_hash *rh, int errors_handled)
441{
442 struct dm_region *reg, *next;
443
444 LIST_HEAD(clean);
445 LIST_HEAD(recovered);
446 LIST_HEAD(failed_recovered);
447
448 /*
449 * Quickly grab the lists.
450 */
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);
455
456 list_for_each_entry(reg, &clean, list)
457 list_del(&reg->hash_list);
458 }
459
460 if (!list_empty(&rh->recovered_regions)) {
461 list_splice_init(&rh->recovered_regions, &recovered);
462
463 list_for_each_entry(reg, &recovered, list)
464 list_del(&reg->hash_list);
465 }
466
467 if (!list_empty(&rh->failed_recovered_regions)) {
468 list_splice_init(&rh->failed_recovered_regions,
469 &failed_recovered);
470
471 list_for_each_entry(reg, &failed_recovered, list)
472 list_del(&reg->hash_list);
473 }
474
475 spin_unlock(&rh->region_lock);
476 write_unlock_irq(&rh->hash_lock);
477
478 /*
479 * All the regions on the recovered and clean lists have
480 * now been pulled out of the system, so no need to do
481 * any more locking.
482 */
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);
487 }
488
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);
492 }
493
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);
497 }
498
499 rh->log->type->flush(rh->log);
500}
501EXPORT_SYMBOL_GPL(dm_rh_update_states);
502
503static void rh_inc(struct dm_region_hash *rh, region_t region)
504{
505 struct dm_region *reg;
506
507 read_lock(&rh->hash_lock);
508 reg = __rh_find(rh, region);
509
510 spin_lock_irq(&rh->region_lock);
511 atomic_inc(&reg->pending);
512
513 if (reg->state == DM_RH_CLEAN) {
514 reg->state = DM_RH_DIRTY;
515 list_del_init(&reg->list); /* take off the clean list */
516 spin_unlock_irq(&rh->region_lock);
517
518 rh->log->type->mark_region(rh->log, reg->key);
519 } else
520 spin_unlock_irq(&rh->region_lock);
521
522
523 read_unlock(&rh->hash_lock);
524}
525
526void dm_rh_inc_pending(struct dm_region_hash *rh, struct bio_list *bios)
527{
528 struct bio *bio;
529
530 for (bio = bios->head; bio; bio = bio->bi_next) {
531 if (bio->bi_opf & REQ_PREFLUSH || bio_op(bio) == REQ_OP_DISCARD)
532 continue;
533 rh_inc(rh, dm_rh_bio_to_region(rh, bio));
534 }
535}
536EXPORT_SYMBOL_GPL(dm_rh_inc_pending);
537
538void dm_rh_dec(struct dm_region_hash *rh, region_t region)
539{
540 unsigned long flags;
541 struct dm_region *reg;
542 int should_wake = 0;
543
544 read_lock(&rh->hash_lock);
545 reg = __rh_lookup(rh, region);
546 read_unlock(&rh->hash_lock);
547
548 spin_lock_irqsave(&rh->region_lock, flags);
549 if (atomic_dec_and_test(&reg->pending)) {
550 /*
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.
554 *
555 * If the state is DM_RH_NOSYNC, the region should be kept off
556 * from clean list.
557 * The hash entry for DM_RH_NOSYNC will remain in memory
558 * until the region is recovered or the map is reloaded.
559 */
560
561 /* do nothing for DM_RH_NOSYNC */
562 if (unlikely(rh->flush_failure)) {
563 /*
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.
567 */
568 reg->state = DM_RH_NOSYNC;
569 } else if (reg->state == DM_RH_RECOVERING) {
570 list_add_tail(&reg->list, &rh->quiesced_regions);
571 } else if (reg->state == DM_RH_DIRTY) {
572 reg->state = DM_RH_CLEAN;
573 list_add(&reg->list, &rh->clean_regions);
574 }
575 should_wake = 1;
576 }
577 spin_unlock_irqrestore(&rh->region_lock, flags);
578
579 if (should_wake)
580 rh->wakeup_workers(rh->context);
581}
582EXPORT_SYMBOL_GPL(dm_rh_dec);
583
584/*
585 * Starts quiescing a region in preparation for recovery.
586 */
587static int __rh_recovery_prepare(struct dm_region_hash *rh)
588{
589 int r;
590 region_t region;
591 struct dm_region *reg;
592
593 /*
594 * Ask the dirty log what's next.
595 */
596 r = rh->log->type->get_resync_work(rh->log, &region);
597 if (r <= 0)
598 return r;
599
600 /*
601 * Get this region, and start it quiescing by setting the
602 * recovering flag.
603 */
604 read_lock(&rh->hash_lock);
605 reg = __rh_find(rh, region);
606 read_unlock(&rh->hash_lock);
607
608 spin_lock_irq(&rh->region_lock);
609 reg->state = DM_RH_RECOVERING;
610
611 /* Already quiesced ? */
612 if (atomic_read(&reg->pending))
613 list_del_init(&reg->list);
614 else
615 list_move(&reg->list, &rh->quiesced_regions);
616
617 spin_unlock_irq(&rh->region_lock);
618
619 return 1;
620}
621
622void dm_rh_recovery_prepare(struct dm_region_hash *rh)
623{
624 /* Extra reference to avoid race with dm_rh_stop_recovery */
625 atomic_inc(&rh->recovery_in_flight);
626
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);
632 break;
633 }
634 }
635
636 /* Drop the extra reference */
637 if (atomic_dec_and_test(&rh->recovery_in_flight))
638 rh->wakeup_all_recovery_waiters(rh->context);
639}
640EXPORT_SYMBOL_GPL(dm_rh_recovery_prepare);
641
642/*
643 * Returns any quiesced regions.
644 */
645struct dm_region *dm_rh_recovery_start(struct dm_region_hash *rh)
646{
647 struct dm_region *reg = NULL;
648
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(&reg->list); /* remove from the quiesced list */
654 }
655 spin_unlock_irq(&rh->region_lock);
656
657 return reg;
658}
659EXPORT_SYMBOL_GPL(dm_rh_recovery_start);
660
661void dm_rh_recovery_end(struct dm_region *reg, int success)
662{
663 struct dm_region_hash *rh = reg->rh;
664
665 spin_lock_irq(&rh->region_lock);
666 if (success)
667 list_add(&reg->list, &reg->rh->recovered_regions);
668 else
669 list_add(&reg->list, &reg->rh->failed_recovered_regions);
670
671 spin_unlock_irq(&rh->region_lock);
672
673 rh->wakeup_workers(rh->context);
674}
675EXPORT_SYMBOL_GPL(dm_rh_recovery_end);
676
677/* Return recovery in flight count. */
678int dm_rh_recovery_in_flight(struct dm_region_hash *rh)
679{
680 return atomic_read(&rh->recovery_in_flight);
681}
682EXPORT_SYMBOL_GPL(dm_rh_recovery_in_flight);
683
684int dm_rh_flush(struct dm_region_hash *rh)
685{
686 return rh->log->type->flush(rh->log);
687}
688EXPORT_SYMBOL_GPL(dm_rh_flush);
689
690void dm_rh_delay(struct dm_region_hash *rh, struct bio *bio)
691{
692 struct dm_region *reg;
693
694 read_lock(&rh->hash_lock);
695 reg = __rh_find(rh, dm_rh_bio_to_region(rh, bio));
696 bio_list_add(&reg->delayed_bios, bio);
697 read_unlock(&rh->hash_lock);
698}
699EXPORT_SYMBOL_GPL(dm_rh_delay);
700
701void dm_rh_stop_recovery(struct dm_region_hash *rh)
702{
703 int i;
704
705 /* wait for any recovering regions */
706 for (i = 0; i < rh->max_recovery; i++)
707 down(&rh->recovery_count);
708}
709EXPORT_SYMBOL_GPL(dm_rh_stop_recovery);
710
711void dm_rh_start_recovery(struct dm_region_hash *rh)
712{
713 int i;
714
715 for (i = 0; i < rh->max_recovery; i++)
716 up(&rh->recovery_count);
717
718 rh->wakeup_workers(rh->context);
719}
720EXPORT_SYMBOL_GPL(dm_rh_start_recovery);
721
722MODULE_DESCRIPTION(DM_NAME " region hash");
723MODULE_AUTHOR("Joe Thornber/Heinz Mauelshagen <dm-devel@redhat.com>");
724MODULE_LICENSE("GPL");
725