Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2003 Sistina Software Limited.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 *
6 * This file is released under the GPL.
7 */
8
9#include <linux/dm-dirty-log.h>
10#include <linux/dm-region-hash.h>
11
12#include <linux/ctype.h>
13#include <linux/init.h>
14#include <linux/module.h>
15#include <linux/slab.h>
16#include <linux/vmalloc.h>
17
18#include "dm.h"
19
20#define DM_MSG_PREFIX "region hash"
21
22/*
23 *------------------------------------------------------------------
24 * Region hash
25 *
26 * The mirror splits itself up into discrete regions. Each
27 * region can be in one of three states: clean, dirty,
28 * nosync. There is no need to put clean regions in the hash.
29 *
30 * In addition to being present in the hash table a region _may_
31 * be present on one of three lists.
32 *
33 * clean_regions: Regions on this list have no io pending to
34 * them, they are in sync, we are no longer interested in them,
35 * they are dull. dm_rh_update_states() will remove them from the
36 * hash table.
37 *
38 * quiesced_regions: These regions have been spun down, ready
39 * for recovery. rh_recovery_start() will remove regions from
40 * this list and hand them to kmirrord, which will schedule the
41 * recovery io with kcopyd.
42 *
43 * recovered_regions: Regions that kcopyd has successfully
44 * recovered. dm_rh_update_states() will now schedule any delayed
45 * io, up the recovery_count, and remove the region from the
46 * hash.
47 *
48 * There are 2 locks:
49 * A rw spin lock 'hash_lock' protects just the hash table,
50 * this is never held in write mode from interrupt context,
51 * which I believe means that we only have to disable irqs when
52 * doing a write lock.
53 *
54 * An ordinary spin lock 'region_lock' that protects the three
55 * lists in the region_hash, with the 'state', 'list' and
56 * 'delayed_bios' fields of the regions. This is used from irq
57 * context, so all other uses will have to suspend local irqs.
58 *------------------------------------------------------------------
59 */
60struct dm_region_hash {
61 uint32_t region_size;
62 unsigned int region_shift;
63
64 /* holds persistent region state */
65 struct dm_dirty_log *log;
66
67 /* hash table */
68 rwlock_t hash_lock;
69 unsigned int mask;
70 unsigned int nr_buckets;
71 unsigned int prime;
72 unsigned int shift;
73 struct list_head *buckets;
74
75 /*
76 * If there was a flush failure no regions can be marked clean.
77 */
78 int flush_failure;
79
80 unsigned int max_recovery; /* Max # of regions to recover in parallel */
81
82 spinlock_t region_lock;
83 atomic_t recovery_in_flight;
84 struct list_head clean_regions;
85 struct list_head quiesced_regions;
86 struct list_head recovered_regions;
87 struct list_head failed_recovered_regions;
88 struct semaphore recovery_count;
89
90 mempool_t region_pool;
91
92 void *context;
93 sector_t target_begin;
94
95 /* Callback function to schedule bios writes */
96 void (*dispatch_bios)(void *context, struct bio_list *bios);
97
98 /* Callback function to wakeup callers worker thread. */
99 void (*wakeup_workers)(void *context);
100
101 /* Callback function to wakeup callers recovery waiters. */
102 void (*wakeup_all_recovery_waiters)(void *context);
103};
104
105struct dm_region {
106 struct dm_region_hash *rh; /* FIXME: can we get rid of this ? */
107 region_t key;
108 int state;
109
110 struct list_head hash_list;
111 struct list_head list;
112
113 atomic_t pending;
114 struct bio_list delayed_bios;
115};
116
117/*
118 * Conversion fns
119 */
120static region_t dm_rh_sector_to_region(struct dm_region_hash *rh, sector_t sector)
121{
122 return sector >> rh->region_shift;
123}
124
125sector_t dm_rh_region_to_sector(struct dm_region_hash *rh, region_t region)
126{
127 return region << rh->region_shift;
128}
129EXPORT_SYMBOL_GPL(dm_rh_region_to_sector);
130
131region_t dm_rh_bio_to_region(struct dm_region_hash *rh, struct bio *bio)
132{
133 return dm_rh_sector_to_region(rh, bio->bi_iter.bi_sector -
134 rh->target_begin);
135}
136EXPORT_SYMBOL_GPL(dm_rh_bio_to_region);
137
138void *dm_rh_region_context(struct dm_region *reg)
139{
140 return reg->rh->context;
141}
142EXPORT_SYMBOL_GPL(dm_rh_region_context);
143
144region_t dm_rh_get_region_key(struct dm_region *reg)
145{
146 return reg->key;
147}
148EXPORT_SYMBOL_GPL(dm_rh_get_region_key);
149
150sector_t dm_rh_get_region_size(struct dm_region_hash *rh)
151{
152 return rh->region_size;
153}
154EXPORT_SYMBOL_GPL(dm_rh_get_region_size);
155
156/*
157 * FIXME: shall we pass in a structure instead of all these args to
158 * dm_region_hash_create()????
159 */
160#define RH_HASH_MULT 2654435387U
161#define RH_HASH_SHIFT 12
162
163#define MIN_REGIONS 64
164struct dm_region_hash *dm_region_hash_create(
165 void *context, void (*dispatch_bios)(void *context,
166 struct bio_list *bios),
167 void (*wakeup_workers)(void *context),
168 void (*wakeup_all_recovery_waiters)(void *context),
169 sector_t target_begin, unsigned int max_recovery,
170 struct dm_dirty_log *log, uint32_t region_size,
171 region_t nr_regions)
172{
173 struct dm_region_hash *rh;
174 unsigned int nr_buckets, max_buckets;
175 size_t i;
176 int ret;
177
178 /*
179 * Calculate a suitable number of buckets for our hash
180 * table.
181 */
182 max_buckets = nr_regions >> 6;
183 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
184 ;
185 nr_buckets >>= 1;
186
187 rh = kzalloc(sizeof(*rh), GFP_KERNEL);
188 if (!rh) {
189 DMERR("unable to allocate region hash memory");
190 return ERR_PTR(-ENOMEM);
191 }
192
193 rh->context = context;
194 rh->dispatch_bios = dispatch_bios;
195 rh->wakeup_workers = wakeup_workers;
196 rh->wakeup_all_recovery_waiters = wakeup_all_recovery_waiters;
197 rh->target_begin = target_begin;
198 rh->max_recovery = max_recovery;
199 rh->log = log;
200 rh->region_size = region_size;
201 rh->region_shift = __ffs(region_size);
202 rwlock_init(&rh->hash_lock);
203 rh->mask = nr_buckets - 1;
204 rh->nr_buckets = nr_buckets;
205
206 rh->shift = RH_HASH_SHIFT;
207 rh->prime = RH_HASH_MULT;
208
209 rh->buckets = vmalloc(array_size(nr_buckets, sizeof(*rh->buckets)));
210 if (!rh->buckets) {
211 DMERR("unable to allocate region hash bucket memory");
212 kfree(rh);
213 return ERR_PTR(-ENOMEM);
214 }
215
216 for (i = 0; i < nr_buckets; i++)
217 INIT_LIST_HEAD(rh->buckets + i);
218
219 spin_lock_init(&rh->region_lock);
220 sema_init(&rh->recovery_count, 0);
221 atomic_set(&rh->recovery_in_flight, 0);
222 INIT_LIST_HEAD(&rh->clean_regions);
223 INIT_LIST_HEAD(&rh->quiesced_regions);
224 INIT_LIST_HEAD(&rh->recovered_regions);
225 INIT_LIST_HEAD(&rh->failed_recovered_regions);
226 rh->flush_failure = 0;
227
228 ret = mempool_init_kmalloc_pool(&rh->region_pool, MIN_REGIONS,
229 sizeof(struct dm_region));
230 if (ret) {
231 vfree(rh->buckets);
232 kfree(rh);
233 rh = ERR_PTR(-ENOMEM);
234 }
235
236 return rh;
237}
238EXPORT_SYMBOL_GPL(dm_region_hash_create);
239
240void dm_region_hash_destroy(struct dm_region_hash *rh)
241{
242 unsigned int h;
243 struct dm_region *reg, *nreg;
244
245 BUG_ON(!list_empty(&rh->quiesced_regions));
246 for (h = 0; h < rh->nr_buckets; h++) {
247 list_for_each_entry_safe(reg, nreg, rh->buckets + h,
248 hash_list) {
249 BUG_ON(atomic_read(®->pending));
250 mempool_free(reg, &rh->region_pool);
251 }
252 }
253
254 if (rh->log)
255 dm_dirty_log_destroy(rh->log);
256
257 mempool_exit(&rh->region_pool);
258 vfree(rh->buckets);
259 kfree(rh);
260}
261EXPORT_SYMBOL_GPL(dm_region_hash_destroy);
262
263struct dm_dirty_log *dm_rh_dirty_log(struct dm_region_hash *rh)
264{
265 return rh->log;
266}
267EXPORT_SYMBOL_GPL(dm_rh_dirty_log);
268
269static unsigned int rh_hash(struct dm_region_hash *rh, region_t region)
270{
271 return (unsigned int) ((region * rh->prime) >> rh->shift) & rh->mask;
272}
273
274static struct dm_region *__rh_lookup(struct dm_region_hash *rh, region_t region)
275{
276 struct dm_region *reg;
277 struct list_head *bucket = rh->buckets + rh_hash(rh, region);
278
279 list_for_each_entry(reg, bucket, hash_list)
280 if (reg->key == region)
281 return reg;
282
283 return NULL;
284}
285
286static void __rh_insert(struct dm_region_hash *rh, struct dm_region *reg)
287{
288 list_add(®->hash_list, rh->buckets + rh_hash(rh, reg->key));
289}
290
291static struct dm_region *__rh_alloc(struct dm_region_hash *rh, region_t region)
292{
293 struct dm_region *reg, *nreg;
294
295 nreg = mempool_alloc(&rh->region_pool, GFP_ATOMIC);
296 if (unlikely(!nreg))
297 nreg = kmalloc(sizeof(*nreg), GFP_NOIO | __GFP_NOFAIL);
298
299 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
300 DM_RH_CLEAN : DM_RH_NOSYNC;
301 nreg->rh = rh;
302 nreg->key = region;
303 INIT_LIST_HEAD(&nreg->list);
304 atomic_set(&nreg->pending, 0);
305 bio_list_init(&nreg->delayed_bios);
306
307 write_lock_irq(&rh->hash_lock);
308 reg = __rh_lookup(rh, region);
309 if (reg)
310 /* We lost the race. */
311 mempool_free(nreg, &rh->region_pool);
312 else {
313 __rh_insert(rh, nreg);
314 if (nreg->state == DM_RH_CLEAN) {
315 spin_lock(&rh->region_lock);
316 list_add(&nreg->list, &rh->clean_regions);
317 spin_unlock(&rh->region_lock);
318 }
319
320 reg = nreg;
321 }
322 write_unlock_irq(&rh->hash_lock);
323
324 return reg;
325}
326
327static struct dm_region *__rh_find(struct dm_region_hash *rh, region_t region)
328{
329 struct dm_region *reg;
330
331 reg = __rh_lookup(rh, region);
332 if (!reg) {
333 read_unlock(&rh->hash_lock);
334 reg = __rh_alloc(rh, region);
335 read_lock(&rh->hash_lock);
336 }
337
338 return reg;
339}
340
341int dm_rh_get_state(struct dm_region_hash *rh, region_t region, int may_block)
342{
343 int r;
344 struct dm_region *reg;
345
346 read_lock(&rh->hash_lock);
347 reg = __rh_lookup(rh, region);
348 read_unlock(&rh->hash_lock);
349
350 if (reg)
351 return reg->state;
352
353 /*
354 * The region wasn't in the hash, so we fall back to the
355 * dirty log.
356 */
357 r = rh->log->type->in_sync(rh->log, region, may_block);
358
359 /*
360 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
361 * taken as a DM_RH_NOSYNC
362 */
363 return r == 1 ? DM_RH_CLEAN : DM_RH_NOSYNC;
364}
365EXPORT_SYMBOL_GPL(dm_rh_get_state);
366
367static void complete_resync_work(struct dm_region *reg, int success)
368{
369 struct dm_region_hash *rh = reg->rh;
370
371 rh->log->type->set_region_sync(rh->log, reg->key, success);
372
373 /*
374 * Dispatch the bios before we call 'wake_up_all'.
375 * This is important because if we are suspending,
376 * we want to know that recovery is complete and
377 * the work queue is flushed. If we wake_up_all
378 * before we dispatch_bios (queue bios and call wake()),
379 * then we risk suspending before the work queue
380 * has been properly flushed.
381 */
382 rh->dispatch_bios(rh->context, ®->delayed_bios);
383 if (atomic_dec_and_test(&rh->recovery_in_flight))
384 rh->wakeup_all_recovery_waiters(rh->context);
385 up(&rh->recovery_count);
386}
387
388/* dm_rh_mark_nosync
389 * @ms
390 * @bio
391 *
392 * The bio was written on some mirror(s) but failed on other mirror(s).
393 * We can successfully endio the bio but should avoid the region being
394 * marked clean by setting the state DM_RH_NOSYNC.
395 *
396 * This function is _not_ safe in interrupt context!
397 */
398void dm_rh_mark_nosync(struct dm_region_hash *rh, struct bio *bio)
399{
400 unsigned long flags;
401 struct dm_dirty_log *log = rh->log;
402 struct dm_region *reg;
403 region_t region = dm_rh_bio_to_region(rh, bio);
404 int recovering = 0;
405
406 if (bio->bi_opf & REQ_PREFLUSH) {
407 rh->flush_failure = 1;
408 return;
409 }
410
411 if (bio_op(bio) == REQ_OP_DISCARD)
412 return;
413
414 /* We must inform the log that the sync count has changed. */
415 log->type->set_region_sync(log, region, 0);
416
417 read_lock(&rh->hash_lock);
418 reg = __rh_find(rh, region);
419 read_unlock(&rh->hash_lock);
420
421 /* region hash entry should exist because write was in-flight */
422 BUG_ON(!reg);
423 BUG_ON(!list_empty(®->list));
424
425 spin_lock_irqsave(&rh->region_lock, flags);
426 /*
427 * Possible cases:
428 * 1) DM_RH_DIRTY
429 * 2) DM_RH_NOSYNC: was dirty, other preceding writes failed
430 * 3) DM_RH_RECOVERING: flushing pending writes
431 * Either case, the region should have not been connected to list.
432 */
433 recovering = (reg->state == DM_RH_RECOVERING);
434 reg->state = DM_RH_NOSYNC;
435 BUG_ON(!list_empty(®->list));
436 spin_unlock_irqrestore(&rh->region_lock, flags);
437
438 if (recovering)
439 complete_resync_work(reg, 0);
440}
441EXPORT_SYMBOL_GPL(dm_rh_mark_nosync);
442
443void dm_rh_update_states(struct dm_region_hash *rh, int errors_handled)
444{
445 struct dm_region *reg, *next;
446
447 LIST_HEAD(clean);
448 LIST_HEAD(recovered);
449 LIST_HEAD(failed_recovered);
450
451 /*
452 * Quickly grab the lists.
453 */
454 write_lock_irq(&rh->hash_lock);
455 spin_lock(&rh->region_lock);
456 if (!list_empty(&rh->clean_regions)) {
457 list_splice_init(&rh->clean_regions, &clean);
458
459 list_for_each_entry(reg, &clean, list)
460 list_del(®->hash_list);
461 }
462
463 if (!list_empty(&rh->recovered_regions)) {
464 list_splice_init(&rh->recovered_regions, &recovered);
465
466 list_for_each_entry(reg, &recovered, list)
467 list_del(®->hash_list);
468 }
469
470 if (!list_empty(&rh->failed_recovered_regions)) {
471 list_splice_init(&rh->failed_recovered_regions,
472 &failed_recovered);
473
474 list_for_each_entry(reg, &failed_recovered, list)
475 list_del(®->hash_list);
476 }
477
478 spin_unlock(&rh->region_lock);
479 write_unlock_irq(&rh->hash_lock);
480
481 /*
482 * All the regions on the recovered and clean lists have
483 * now been pulled out of the system, so no need to do
484 * any more locking.
485 */
486 list_for_each_entry_safe(reg, next, &recovered, list) {
487 rh->log->type->clear_region(rh->log, reg->key);
488 complete_resync_work(reg, 1);
489 mempool_free(reg, &rh->region_pool);
490 }
491
492 list_for_each_entry_safe(reg, next, &failed_recovered, list) {
493 complete_resync_work(reg, errors_handled ? 0 : 1);
494 mempool_free(reg, &rh->region_pool);
495 }
496
497 list_for_each_entry_safe(reg, next, &clean, list) {
498 rh->log->type->clear_region(rh->log, reg->key);
499 mempool_free(reg, &rh->region_pool);
500 }
501
502 rh->log->type->flush(rh->log);
503}
504EXPORT_SYMBOL_GPL(dm_rh_update_states);
505
506static void rh_inc(struct dm_region_hash *rh, region_t region)
507{
508 struct dm_region *reg;
509
510 read_lock(&rh->hash_lock);
511 reg = __rh_find(rh, region);
512
513 spin_lock_irq(&rh->region_lock);
514 atomic_inc(®->pending);
515
516 if (reg->state == DM_RH_CLEAN) {
517 reg->state = DM_RH_DIRTY;
518 list_del_init(®->list); /* take off the clean list */
519 spin_unlock_irq(&rh->region_lock);
520
521 rh->log->type->mark_region(rh->log, reg->key);
522 } else
523 spin_unlock_irq(&rh->region_lock);
524
525
526 read_unlock(&rh->hash_lock);
527}
528
529void dm_rh_inc_pending(struct dm_region_hash *rh, struct bio_list *bios)
530{
531 struct bio *bio;
532
533 for (bio = bios->head; bio; bio = bio->bi_next) {
534 if (bio->bi_opf & REQ_PREFLUSH || bio_op(bio) == REQ_OP_DISCARD)
535 continue;
536 rh_inc(rh, dm_rh_bio_to_region(rh, bio));
537 }
538}
539EXPORT_SYMBOL_GPL(dm_rh_inc_pending);
540
541void dm_rh_dec(struct dm_region_hash *rh, region_t region)
542{
543 unsigned long flags;
544 struct dm_region *reg;
545 int should_wake = 0;
546
547 read_lock(&rh->hash_lock);
548 reg = __rh_lookup(rh, region);
549 read_unlock(&rh->hash_lock);
550
551 spin_lock_irqsave(&rh->region_lock, flags);
552 if (atomic_dec_and_test(®->pending)) {
553 /*
554 * There is no pending I/O for this region.
555 * We can move the region to corresponding list for next action.
556 * At this point, the region is not yet connected to any list.
557 *
558 * If the state is DM_RH_NOSYNC, the region should be kept off
559 * from clean list.
560 * The hash entry for DM_RH_NOSYNC will remain in memory
561 * until the region is recovered or the map is reloaded.
562 */
563
564 /* do nothing for DM_RH_NOSYNC */
565 if (unlikely(rh->flush_failure)) {
566 /*
567 * If a write flush failed some time ago, we
568 * don't know whether or not this write made it
569 * to the disk, so we must resync the device.
570 */
571 reg->state = DM_RH_NOSYNC;
572 } else if (reg->state == DM_RH_RECOVERING) {
573 list_add_tail(®->list, &rh->quiesced_regions);
574 } else if (reg->state == DM_RH_DIRTY) {
575 reg->state = DM_RH_CLEAN;
576 list_add(®->list, &rh->clean_regions);
577 }
578 should_wake = 1;
579 }
580 spin_unlock_irqrestore(&rh->region_lock, flags);
581
582 if (should_wake)
583 rh->wakeup_workers(rh->context);
584}
585EXPORT_SYMBOL_GPL(dm_rh_dec);
586
587/*
588 * Starts quiescing a region in preparation for recovery.
589 */
590static int __rh_recovery_prepare(struct dm_region_hash *rh)
591{
592 int r;
593 region_t region;
594 struct dm_region *reg;
595
596 /*
597 * Ask the dirty log what's next.
598 */
599 r = rh->log->type->get_resync_work(rh->log, ®ion);
600 if (r <= 0)
601 return r;
602
603 /*
604 * Get this region, and start it quiescing by setting the
605 * recovering flag.
606 */
607 read_lock(&rh->hash_lock);
608 reg = __rh_find(rh, region);
609 read_unlock(&rh->hash_lock);
610
611 spin_lock_irq(&rh->region_lock);
612 reg->state = DM_RH_RECOVERING;
613
614 /* Already quiesced ? */
615 if (atomic_read(®->pending))
616 list_del_init(®->list);
617 else
618 list_move(®->list, &rh->quiesced_regions);
619
620 spin_unlock_irq(&rh->region_lock);
621
622 return 1;
623}
624
625void dm_rh_recovery_prepare(struct dm_region_hash *rh)
626{
627 /* Extra reference to avoid race with dm_rh_stop_recovery */
628 atomic_inc(&rh->recovery_in_flight);
629
630 while (!down_trylock(&rh->recovery_count)) {
631 atomic_inc(&rh->recovery_in_flight);
632 if (__rh_recovery_prepare(rh) <= 0) {
633 atomic_dec(&rh->recovery_in_flight);
634 up(&rh->recovery_count);
635 break;
636 }
637 }
638
639 /* Drop the extra reference */
640 if (atomic_dec_and_test(&rh->recovery_in_flight))
641 rh->wakeup_all_recovery_waiters(rh->context);
642}
643EXPORT_SYMBOL_GPL(dm_rh_recovery_prepare);
644
645/*
646 * Returns any quiesced regions.
647 */
648struct dm_region *dm_rh_recovery_start(struct dm_region_hash *rh)
649{
650 struct dm_region *reg = NULL;
651
652 spin_lock_irq(&rh->region_lock);
653 if (!list_empty(&rh->quiesced_regions)) {
654 reg = list_entry(rh->quiesced_regions.next,
655 struct dm_region, list);
656 list_del_init(®->list); /* remove from the quiesced list */
657 }
658 spin_unlock_irq(&rh->region_lock);
659
660 return reg;
661}
662EXPORT_SYMBOL_GPL(dm_rh_recovery_start);
663
664void dm_rh_recovery_end(struct dm_region *reg, int success)
665{
666 struct dm_region_hash *rh = reg->rh;
667
668 spin_lock_irq(&rh->region_lock);
669 if (success)
670 list_add(®->list, ®->rh->recovered_regions);
671 else
672 list_add(®->list, ®->rh->failed_recovered_regions);
673
674 spin_unlock_irq(&rh->region_lock);
675
676 rh->wakeup_workers(rh->context);
677}
678EXPORT_SYMBOL_GPL(dm_rh_recovery_end);
679
680/* Return recovery in flight count. */
681int dm_rh_recovery_in_flight(struct dm_region_hash *rh)
682{
683 return atomic_read(&rh->recovery_in_flight);
684}
685EXPORT_SYMBOL_GPL(dm_rh_recovery_in_flight);
686
687int dm_rh_flush(struct dm_region_hash *rh)
688{
689 return rh->log->type->flush(rh->log);
690}
691EXPORT_SYMBOL_GPL(dm_rh_flush);
692
693void dm_rh_delay(struct dm_region_hash *rh, struct bio *bio)
694{
695 struct dm_region *reg;
696
697 read_lock(&rh->hash_lock);
698 reg = __rh_find(rh, dm_rh_bio_to_region(rh, bio));
699 bio_list_add(®->delayed_bios, bio);
700 read_unlock(&rh->hash_lock);
701}
702EXPORT_SYMBOL_GPL(dm_rh_delay);
703
704void dm_rh_stop_recovery(struct dm_region_hash *rh)
705{
706 int i;
707
708 /* wait for any recovering regions */
709 for (i = 0; i < rh->max_recovery; i++)
710 down(&rh->recovery_count);
711}
712EXPORT_SYMBOL_GPL(dm_rh_stop_recovery);
713
714void dm_rh_start_recovery(struct dm_region_hash *rh)
715{
716 int i;
717
718 for (i = 0; i < rh->max_recovery; i++)
719 up(&rh->recovery_count);
720
721 rh->wakeup_workers(rh->context);
722}
723EXPORT_SYMBOL_GPL(dm_rh_start_recovery);
724
725MODULE_DESCRIPTION(DM_NAME " region hash");
726MODULE_AUTHOR("Joe Thornber/Heinz Mauelshagen <dm-devel@redhat.com>");
727MODULE_LICENSE("GPL");
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 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 /*
84 * If there was a flush failure no regions can be marked clean.
85 */
86 int flush_failure;
87
88 void *context;
89 sector_t target_begin;
90
91 /* Callback function to schedule bios writes */
92 void (*dispatch_bios)(void *context, struct bio_list *bios);
93
94 /* Callback function to wakeup callers worker thread. */
95 void (*wakeup_workers)(void *context);
96
97 /* Callback function to wakeup callers recovery waiters. */
98 void (*wakeup_all_recovery_waiters)(void *context);
99};
100
101struct dm_region {
102 struct dm_region_hash *rh; /* FIXME: can we get rid of this ? */
103 region_t key;
104 int state;
105
106 struct list_head hash_list;
107 struct list_head list;
108
109 atomic_t pending;
110 struct bio_list delayed_bios;
111};
112
113/*
114 * Conversion fns
115 */
116static region_t dm_rh_sector_to_region(struct dm_region_hash *rh, sector_t sector)
117{
118 return sector >> rh->region_shift;
119}
120
121sector_t dm_rh_region_to_sector(struct dm_region_hash *rh, region_t region)
122{
123 return region << rh->region_shift;
124}
125EXPORT_SYMBOL_GPL(dm_rh_region_to_sector);
126
127region_t dm_rh_bio_to_region(struct dm_region_hash *rh, struct bio *bio)
128{
129 return dm_rh_sector_to_region(rh, bio->bi_iter.bi_sector -
130 rh->target_begin);
131}
132EXPORT_SYMBOL_GPL(dm_rh_bio_to_region);
133
134void *dm_rh_region_context(struct dm_region *reg)
135{
136 return reg->rh->context;
137}
138EXPORT_SYMBOL_GPL(dm_rh_region_context);
139
140region_t dm_rh_get_region_key(struct dm_region *reg)
141{
142 return reg->key;
143}
144EXPORT_SYMBOL_GPL(dm_rh_get_region_key);
145
146sector_t dm_rh_get_region_size(struct dm_region_hash *rh)
147{
148 return rh->region_size;
149}
150EXPORT_SYMBOL_GPL(dm_rh_get_region_size);
151
152/*
153 * FIXME: shall we pass in a structure instead of all these args to
154 * dm_region_hash_create()????
155 */
156#define RH_HASH_MULT 2654435387U
157#define RH_HASH_SHIFT 12
158
159#define MIN_REGIONS 64
160struct dm_region_hash *dm_region_hash_create(
161 void *context, void (*dispatch_bios)(void *context,
162 struct bio_list *bios),
163 void (*wakeup_workers)(void *context),
164 void (*wakeup_all_recovery_waiters)(void *context),
165 sector_t target_begin, unsigned max_recovery,
166 struct dm_dirty_log *log, uint32_t region_size,
167 region_t nr_regions)
168{
169 struct dm_region_hash *rh;
170 unsigned nr_buckets, max_buckets;
171 size_t i;
172
173 /*
174 * Calculate a suitable number of buckets for our hash
175 * table.
176 */
177 max_buckets = nr_regions >> 6;
178 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
179 ;
180 nr_buckets >>= 1;
181
182 rh = kmalloc(sizeof(*rh), GFP_KERNEL);
183 if (!rh) {
184 DMERR("unable to allocate region hash memory");
185 return ERR_PTR(-ENOMEM);
186 }
187
188 rh->context = context;
189 rh->dispatch_bios = dispatch_bios;
190 rh->wakeup_workers = wakeup_workers;
191 rh->wakeup_all_recovery_waiters = wakeup_all_recovery_waiters;
192 rh->target_begin = target_begin;
193 rh->max_recovery = max_recovery;
194 rh->log = log;
195 rh->region_size = region_size;
196 rh->region_shift = __ffs(region_size);
197 rwlock_init(&rh->hash_lock);
198 rh->mask = nr_buckets - 1;
199 rh->nr_buckets = nr_buckets;
200
201 rh->shift = RH_HASH_SHIFT;
202 rh->prime = RH_HASH_MULT;
203
204 rh->buckets = vmalloc(nr_buckets * sizeof(*rh->buckets));
205 if (!rh->buckets) {
206 DMERR("unable to allocate region hash bucket memory");
207 kfree(rh);
208 return ERR_PTR(-ENOMEM);
209 }
210
211 for (i = 0; i < nr_buckets; i++)
212 INIT_LIST_HEAD(rh->buckets + i);
213
214 spin_lock_init(&rh->region_lock);
215 sema_init(&rh->recovery_count, 0);
216 atomic_set(&rh->recovery_in_flight, 0);
217 INIT_LIST_HEAD(&rh->clean_regions);
218 INIT_LIST_HEAD(&rh->quiesced_regions);
219 INIT_LIST_HEAD(&rh->recovered_regions);
220 INIT_LIST_HEAD(&rh->failed_recovered_regions);
221 rh->flush_failure = 0;
222
223 rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
224 sizeof(struct dm_region));
225 if (!rh->region_pool) {
226 vfree(rh->buckets);
227 kfree(rh);
228 rh = ERR_PTR(-ENOMEM);
229 }
230
231 return rh;
232}
233EXPORT_SYMBOL_GPL(dm_region_hash_create);
234
235void dm_region_hash_destroy(struct dm_region_hash *rh)
236{
237 unsigned h;
238 struct dm_region *reg, *nreg;
239
240 BUG_ON(!list_empty(&rh->quiesced_regions));
241 for (h = 0; h < rh->nr_buckets; h++) {
242 list_for_each_entry_safe(reg, nreg, rh->buckets + h,
243 hash_list) {
244 BUG_ON(atomic_read(®->pending));
245 mempool_free(reg, rh->region_pool);
246 }
247 }
248
249 if (rh->log)
250 dm_dirty_log_destroy(rh->log);
251
252 mempool_destroy(rh->region_pool);
253 vfree(rh->buckets);
254 kfree(rh);
255}
256EXPORT_SYMBOL_GPL(dm_region_hash_destroy);
257
258struct dm_dirty_log *dm_rh_dirty_log(struct dm_region_hash *rh)
259{
260 return rh->log;
261}
262EXPORT_SYMBOL_GPL(dm_rh_dirty_log);
263
264static unsigned rh_hash(struct dm_region_hash *rh, region_t region)
265{
266 return (unsigned) ((region * rh->prime) >> rh->shift) & rh->mask;
267}
268
269static struct dm_region *__rh_lookup(struct dm_region_hash *rh, region_t region)
270{
271 struct dm_region *reg;
272 struct list_head *bucket = rh->buckets + rh_hash(rh, region);
273
274 list_for_each_entry(reg, bucket, hash_list)
275 if (reg->key == region)
276 return reg;
277
278 return NULL;
279}
280
281static void __rh_insert(struct dm_region_hash *rh, struct dm_region *reg)
282{
283 list_add(®->hash_list, rh->buckets + rh_hash(rh, reg->key));
284}
285
286static struct dm_region *__rh_alloc(struct dm_region_hash *rh, region_t region)
287{
288 struct dm_region *reg, *nreg;
289
290 nreg = mempool_alloc(rh->region_pool, GFP_ATOMIC);
291 if (unlikely(!nreg))
292 nreg = kmalloc(sizeof(*nreg), GFP_NOIO | __GFP_NOFAIL);
293
294 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
295 DM_RH_CLEAN : DM_RH_NOSYNC;
296 nreg->rh = rh;
297 nreg->key = region;
298 INIT_LIST_HEAD(&nreg->list);
299 atomic_set(&nreg->pending, 0);
300 bio_list_init(&nreg->delayed_bios);
301
302 write_lock_irq(&rh->hash_lock);
303 reg = __rh_lookup(rh, region);
304 if (reg)
305 /* We lost the race. */
306 mempool_free(nreg, rh->region_pool);
307 else {
308 __rh_insert(rh, nreg);
309 if (nreg->state == DM_RH_CLEAN) {
310 spin_lock(&rh->region_lock);
311 list_add(&nreg->list, &rh->clean_regions);
312 spin_unlock(&rh->region_lock);
313 }
314
315 reg = nreg;
316 }
317 write_unlock_irq(&rh->hash_lock);
318
319 return reg;
320}
321
322static struct dm_region *__rh_find(struct dm_region_hash *rh, region_t region)
323{
324 struct dm_region *reg;
325
326 reg = __rh_lookup(rh, region);
327 if (!reg) {
328 read_unlock(&rh->hash_lock);
329 reg = __rh_alloc(rh, region);
330 read_lock(&rh->hash_lock);
331 }
332
333 return reg;
334}
335
336int dm_rh_get_state(struct dm_region_hash *rh, region_t region, int may_block)
337{
338 int r;
339 struct dm_region *reg;
340
341 read_lock(&rh->hash_lock);
342 reg = __rh_lookup(rh, region);
343 read_unlock(&rh->hash_lock);
344
345 if (reg)
346 return reg->state;
347
348 /*
349 * The region wasn't in the hash, so we fall back to the
350 * dirty log.
351 */
352 r = rh->log->type->in_sync(rh->log, region, may_block);
353
354 /*
355 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
356 * taken as a DM_RH_NOSYNC
357 */
358 return r == 1 ? DM_RH_CLEAN : DM_RH_NOSYNC;
359}
360EXPORT_SYMBOL_GPL(dm_rh_get_state);
361
362static void complete_resync_work(struct dm_region *reg, int success)
363{
364 struct dm_region_hash *rh = reg->rh;
365
366 rh->log->type->set_region_sync(rh->log, reg->key, success);
367
368 /*
369 * Dispatch the bios before we call 'wake_up_all'.
370 * This is important because if we are suspending,
371 * we want to know that recovery is complete and
372 * the work queue is flushed. If we wake_up_all
373 * before we dispatch_bios (queue bios and call wake()),
374 * then we risk suspending before the work queue
375 * has been properly flushed.
376 */
377 rh->dispatch_bios(rh->context, ®->delayed_bios);
378 if (atomic_dec_and_test(&rh->recovery_in_flight))
379 rh->wakeup_all_recovery_waiters(rh->context);
380 up(&rh->recovery_count);
381}
382
383/* dm_rh_mark_nosync
384 * @ms
385 * @bio
386 *
387 * The bio was written on some mirror(s) but failed on other mirror(s).
388 * We can successfully endio the bio but should avoid the region being
389 * marked clean by setting the state DM_RH_NOSYNC.
390 *
391 * This function is _not_ safe in interrupt context!
392 */
393void dm_rh_mark_nosync(struct dm_region_hash *rh, struct bio *bio)
394{
395 unsigned long flags;
396 struct dm_dirty_log *log = rh->log;
397 struct dm_region *reg;
398 region_t region = dm_rh_bio_to_region(rh, bio);
399 int recovering = 0;
400
401 if (bio->bi_opf & REQ_PREFLUSH) {
402 rh->flush_failure = 1;
403 return;
404 }
405
406 if (bio_op(bio) == REQ_OP_DISCARD)
407 return;
408
409 /* We must inform the log that the sync count has changed. */
410 log->type->set_region_sync(log, region, 0);
411
412 read_lock(&rh->hash_lock);
413 reg = __rh_find(rh, region);
414 read_unlock(&rh->hash_lock);
415
416 /* region hash entry should exist because write was in-flight */
417 BUG_ON(!reg);
418 BUG_ON(!list_empty(®->list));
419
420 spin_lock_irqsave(&rh->region_lock, flags);
421 /*
422 * Possible cases:
423 * 1) DM_RH_DIRTY
424 * 2) DM_RH_NOSYNC: was dirty, other preceding writes failed
425 * 3) DM_RH_RECOVERING: flushing pending writes
426 * Either case, the region should have not been connected to list.
427 */
428 recovering = (reg->state == DM_RH_RECOVERING);
429 reg->state = DM_RH_NOSYNC;
430 BUG_ON(!list_empty(®->list));
431 spin_unlock_irqrestore(&rh->region_lock, flags);
432
433 if (recovering)
434 complete_resync_work(reg, 0);
435}
436EXPORT_SYMBOL_GPL(dm_rh_mark_nosync);
437
438void dm_rh_update_states(struct dm_region_hash *rh, int errors_handled)
439{
440 struct dm_region *reg, *next;
441
442 LIST_HEAD(clean);
443 LIST_HEAD(recovered);
444 LIST_HEAD(failed_recovered);
445
446 /*
447 * Quickly grab the lists.
448 */
449 write_lock_irq(&rh->hash_lock);
450 spin_lock(&rh->region_lock);
451 if (!list_empty(&rh->clean_regions)) {
452 list_splice_init(&rh->clean_regions, &clean);
453
454 list_for_each_entry(reg, &clean, list)
455 list_del(®->hash_list);
456 }
457
458 if (!list_empty(&rh->recovered_regions)) {
459 list_splice_init(&rh->recovered_regions, &recovered);
460
461 list_for_each_entry(reg, &recovered, list)
462 list_del(®->hash_list);
463 }
464
465 if (!list_empty(&rh->failed_recovered_regions)) {
466 list_splice_init(&rh->failed_recovered_regions,
467 &failed_recovered);
468
469 list_for_each_entry(reg, &failed_recovered, list)
470 list_del(®->hash_list);
471 }
472
473 spin_unlock(&rh->region_lock);
474 write_unlock_irq(&rh->hash_lock);
475
476 /*
477 * All the regions on the recovered and clean lists have
478 * now been pulled out of the system, so no need to do
479 * any more locking.
480 */
481 list_for_each_entry_safe(reg, next, &recovered, list) {
482 rh->log->type->clear_region(rh->log, reg->key);
483 complete_resync_work(reg, 1);
484 mempool_free(reg, rh->region_pool);
485 }
486
487 list_for_each_entry_safe(reg, next, &failed_recovered, list) {
488 complete_resync_work(reg, errors_handled ? 0 : 1);
489 mempool_free(reg, rh->region_pool);
490 }
491
492 list_for_each_entry_safe(reg, next, &clean, list) {
493 rh->log->type->clear_region(rh->log, reg->key);
494 mempool_free(reg, rh->region_pool);
495 }
496
497 rh->log->type->flush(rh->log);
498}
499EXPORT_SYMBOL_GPL(dm_rh_update_states);
500
501static void rh_inc(struct dm_region_hash *rh, region_t region)
502{
503 struct dm_region *reg;
504
505 read_lock(&rh->hash_lock);
506 reg = __rh_find(rh, region);
507
508 spin_lock_irq(&rh->region_lock);
509 atomic_inc(®->pending);
510
511 if (reg->state == DM_RH_CLEAN) {
512 reg->state = DM_RH_DIRTY;
513 list_del_init(®->list); /* take off the clean list */
514 spin_unlock_irq(&rh->region_lock);
515
516 rh->log->type->mark_region(rh->log, reg->key);
517 } else
518 spin_unlock_irq(&rh->region_lock);
519
520
521 read_unlock(&rh->hash_lock);
522}
523
524void dm_rh_inc_pending(struct dm_region_hash *rh, struct bio_list *bios)
525{
526 struct bio *bio;
527
528 for (bio = bios->head; bio; bio = bio->bi_next) {
529 if (bio->bi_opf & REQ_PREFLUSH || bio_op(bio) == REQ_OP_DISCARD)
530 continue;
531 rh_inc(rh, dm_rh_bio_to_region(rh, bio));
532 }
533}
534EXPORT_SYMBOL_GPL(dm_rh_inc_pending);
535
536void dm_rh_dec(struct dm_region_hash *rh, region_t region)
537{
538 unsigned long flags;
539 struct dm_region *reg;
540 int should_wake = 0;
541
542 read_lock(&rh->hash_lock);
543 reg = __rh_lookup(rh, region);
544 read_unlock(&rh->hash_lock);
545
546 spin_lock_irqsave(&rh->region_lock, flags);
547 if (atomic_dec_and_test(®->pending)) {
548 /*
549 * There is no pending I/O for this region.
550 * We can move the region to corresponding list for next action.
551 * At this point, the region is not yet connected to any list.
552 *
553 * If the state is DM_RH_NOSYNC, the region should be kept off
554 * from clean list.
555 * The hash entry for DM_RH_NOSYNC will remain in memory
556 * until the region is recovered or the map is reloaded.
557 */
558
559 /* do nothing for DM_RH_NOSYNC */
560 if (unlikely(rh->flush_failure)) {
561 /*
562 * If a write flush failed some time ago, we
563 * don't know whether or not this write made it
564 * to the disk, so we must resync the device.
565 */
566 reg->state = DM_RH_NOSYNC;
567 } else if (reg->state == DM_RH_RECOVERING) {
568 list_add_tail(®->list, &rh->quiesced_regions);
569 } else if (reg->state == DM_RH_DIRTY) {
570 reg->state = DM_RH_CLEAN;
571 list_add(®->list, &rh->clean_regions);
572 }
573 should_wake = 1;
574 }
575 spin_unlock_irqrestore(&rh->region_lock, flags);
576
577 if (should_wake)
578 rh->wakeup_workers(rh->context);
579}
580EXPORT_SYMBOL_GPL(dm_rh_dec);
581
582/*
583 * Starts quiescing a region in preparation for recovery.
584 */
585static int __rh_recovery_prepare(struct dm_region_hash *rh)
586{
587 int r;
588 region_t region;
589 struct dm_region *reg;
590
591 /*
592 * Ask the dirty log what's next.
593 */
594 r = rh->log->type->get_resync_work(rh->log, ®ion);
595 if (r <= 0)
596 return r;
597
598 /*
599 * Get this region, and start it quiescing by setting the
600 * recovering flag.
601 */
602 read_lock(&rh->hash_lock);
603 reg = __rh_find(rh, region);
604 read_unlock(&rh->hash_lock);
605
606 spin_lock_irq(&rh->region_lock);
607 reg->state = DM_RH_RECOVERING;
608
609 /* Already quiesced ? */
610 if (atomic_read(®->pending))
611 list_del_init(®->list);
612 else
613 list_move(®->list, &rh->quiesced_regions);
614
615 spin_unlock_irq(&rh->region_lock);
616
617 return 1;
618}
619
620void dm_rh_recovery_prepare(struct dm_region_hash *rh)
621{
622 /* Extra reference to avoid race with dm_rh_stop_recovery */
623 atomic_inc(&rh->recovery_in_flight);
624
625 while (!down_trylock(&rh->recovery_count)) {
626 atomic_inc(&rh->recovery_in_flight);
627 if (__rh_recovery_prepare(rh) <= 0) {
628 atomic_dec(&rh->recovery_in_flight);
629 up(&rh->recovery_count);
630 break;
631 }
632 }
633
634 /* Drop the extra reference */
635 if (atomic_dec_and_test(&rh->recovery_in_flight))
636 rh->wakeup_all_recovery_waiters(rh->context);
637}
638EXPORT_SYMBOL_GPL(dm_rh_recovery_prepare);
639
640/*
641 * Returns any quiesced regions.
642 */
643struct dm_region *dm_rh_recovery_start(struct dm_region_hash *rh)
644{
645 struct dm_region *reg = NULL;
646
647 spin_lock_irq(&rh->region_lock);
648 if (!list_empty(&rh->quiesced_regions)) {
649 reg = list_entry(rh->quiesced_regions.next,
650 struct dm_region, list);
651 list_del_init(®->list); /* remove from the quiesced list */
652 }
653 spin_unlock_irq(&rh->region_lock);
654
655 return reg;
656}
657EXPORT_SYMBOL_GPL(dm_rh_recovery_start);
658
659void dm_rh_recovery_end(struct dm_region *reg, int success)
660{
661 struct dm_region_hash *rh = reg->rh;
662
663 spin_lock_irq(&rh->region_lock);
664 if (success)
665 list_add(®->list, ®->rh->recovered_regions);
666 else
667 list_add(®->list, ®->rh->failed_recovered_regions);
668
669 spin_unlock_irq(&rh->region_lock);
670
671 rh->wakeup_workers(rh->context);
672}
673EXPORT_SYMBOL_GPL(dm_rh_recovery_end);
674
675/* Return recovery in flight count. */
676int dm_rh_recovery_in_flight(struct dm_region_hash *rh)
677{
678 return atomic_read(&rh->recovery_in_flight);
679}
680EXPORT_SYMBOL_GPL(dm_rh_recovery_in_flight);
681
682int dm_rh_flush(struct dm_region_hash *rh)
683{
684 return rh->log->type->flush(rh->log);
685}
686EXPORT_SYMBOL_GPL(dm_rh_flush);
687
688void dm_rh_delay(struct dm_region_hash *rh, struct bio *bio)
689{
690 struct dm_region *reg;
691
692 read_lock(&rh->hash_lock);
693 reg = __rh_find(rh, dm_rh_bio_to_region(rh, bio));
694 bio_list_add(®->delayed_bios, bio);
695 read_unlock(&rh->hash_lock);
696}
697EXPORT_SYMBOL_GPL(dm_rh_delay);
698
699void dm_rh_stop_recovery(struct dm_region_hash *rh)
700{
701 int i;
702
703 /* wait for any recovering regions */
704 for (i = 0; i < rh->max_recovery; i++)
705 down(&rh->recovery_count);
706}
707EXPORT_SYMBOL_GPL(dm_rh_stop_recovery);
708
709void dm_rh_start_recovery(struct dm_region_hash *rh)
710{
711 int i;
712
713 for (i = 0; i < rh->max_recovery; i++)
714 up(&rh->recovery_count);
715
716 rh->wakeup_workers(rh->context);
717}
718EXPORT_SYMBOL_GPL(dm_rh_start_recovery);
719
720MODULE_DESCRIPTION(DM_NAME " region hash");
721MODULE_AUTHOR("Joe Thornber/Heinz Mauelshagen <dm-devel@redhat.com>");
722MODULE_LICENSE("GPL");