Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2009-2011 Red Hat, Inc.
4 *
5 * Author: Mikulas Patocka <mpatocka@redhat.com>
6 *
7 * This file is released under the GPL.
8 */
9
10#include <linux/dm-bufio.h>
11
12#include <linux/device-mapper.h>
13#include <linux/dm-io.h>
14#include <linux/slab.h>
15#include <linux/sched/mm.h>
16#include <linux/jiffies.h>
17#include <linux/vmalloc.h>
18#include <linux/shrinker.h>
19#include <linux/module.h>
20#include <linux/rbtree.h>
21#include <linux/stacktrace.h>
22#include <linux/jump_label.h>
23
24#include "dm.h"
25
26#define DM_MSG_PREFIX "bufio"
27
28/*
29 * Memory management policy:
30 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
31 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
32 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
33 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
34 * dirty buffers.
35 */
36#define DM_BUFIO_MIN_BUFFERS 8
37
38#define DM_BUFIO_MEMORY_PERCENT 2
39#define DM_BUFIO_VMALLOC_PERCENT 25
40#define DM_BUFIO_WRITEBACK_RATIO 3
41#define DM_BUFIO_LOW_WATERMARK_RATIO 16
42
43/*
44 * Check buffer ages in this interval (seconds)
45 */
46#define DM_BUFIO_WORK_TIMER_SECS 30
47
48/*
49 * Free buffers when they are older than this (seconds)
50 */
51#define DM_BUFIO_DEFAULT_AGE_SECS 300
52
53/*
54 * The nr of bytes of cached data to keep around.
55 */
56#define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
57
58/*
59 * Align buffer writes to this boundary.
60 * Tests show that SSDs have the highest IOPS when using 4k writes.
61 */
62#define DM_BUFIO_WRITE_ALIGN 4096
63
64/*
65 * dm_buffer->list_mode
66 */
67#define LIST_CLEAN 0
68#define LIST_DIRTY 1
69#define LIST_SIZE 2
70
71/*--------------------------------------------------------------*/
72
73/*
74 * Rather than use an LRU list, we use a clock algorithm where entries
75 * are held in a circular list. When an entry is 'hit' a reference bit
76 * is set. The least recently used entry is approximated by running a
77 * cursor around the list selecting unreferenced entries. Referenced
78 * entries have their reference bit cleared as the cursor passes them.
79 */
80struct lru_entry {
81 struct list_head list;
82 atomic_t referenced;
83};
84
85struct lru_iter {
86 struct lru *lru;
87 struct list_head list;
88 struct lru_entry *stop;
89 struct lru_entry *e;
90};
91
92struct lru {
93 struct list_head *cursor;
94 unsigned long count;
95
96 struct list_head iterators;
97};
98
99/*--------------*/
100
101static void lru_init(struct lru *lru)
102{
103 lru->cursor = NULL;
104 lru->count = 0;
105 INIT_LIST_HEAD(&lru->iterators);
106}
107
108static void lru_destroy(struct lru *lru)
109{
110 WARN_ON_ONCE(lru->cursor);
111 WARN_ON_ONCE(!list_empty(&lru->iterators));
112}
113
114/*
115 * Insert a new entry into the lru.
116 */
117static void lru_insert(struct lru *lru, struct lru_entry *le)
118{
119 /*
120 * Don't be tempted to set to 1, makes the lru aspect
121 * perform poorly.
122 */
123 atomic_set(&le->referenced, 0);
124
125 if (lru->cursor) {
126 list_add_tail(&le->list, lru->cursor);
127 } else {
128 INIT_LIST_HEAD(&le->list);
129 lru->cursor = &le->list;
130 }
131 lru->count++;
132}
133
134/*--------------*/
135
136/*
137 * Convert a list_head pointer to an lru_entry pointer.
138 */
139static inline struct lru_entry *to_le(struct list_head *l)
140{
141 return container_of(l, struct lru_entry, list);
142}
143
144/*
145 * Initialize an lru_iter and add it to the list of cursors in the lru.
146 */
147static void lru_iter_begin(struct lru *lru, struct lru_iter *it)
148{
149 it->lru = lru;
150 it->stop = lru->cursor ? to_le(lru->cursor->prev) : NULL;
151 it->e = lru->cursor ? to_le(lru->cursor) : NULL;
152 list_add(&it->list, &lru->iterators);
153}
154
155/*
156 * Remove an lru_iter from the list of cursors in the lru.
157 */
158static inline void lru_iter_end(struct lru_iter *it)
159{
160 list_del(&it->list);
161}
162
163/* Predicate function type to be used with lru_iter_next */
164typedef bool (*iter_predicate)(struct lru_entry *le, void *context);
165
166/*
167 * Advance the cursor to the next entry that passes the
168 * predicate, and return that entry. Returns NULL if the
169 * iteration is complete.
170 */
171static struct lru_entry *lru_iter_next(struct lru_iter *it,
172 iter_predicate pred, void *context)
173{
174 struct lru_entry *e;
175
176 while (it->e) {
177 e = it->e;
178
179 /* advance the cursor */
180 if (it->e == it->stop)
181 it->e = NULL;
182 else
183 it->e = to_le(it->e->list.next);
184
185 if (pred(e, context))
186 return e;
187 }
188
189 return NULL;
190}
191
192/*
193 * Invalidate a specific lru_entry and update all cursors in
194 * the lru accordingly.
195 */
196static void lru_iter_invalidate(struct lru *lru, struct lru_entry *e)
197{
198 struct lru_iter *it;
199
200 list_for_each_entry(it, &lru->iterators, list) {
201 /* Move c->e forwards if necc. */
202 if (it->e == e) {
203 it->e = to_le(it->e->list.next);
204 if (it->e == e)
205 it->e = NULL;
206 }
207
208 /* Move it->stop backwards if necc. */
209 if (it->stop == e) {
210 it->stop = to_le(it->stop->list.prev);
211 if (it->stop == e)
212 it->stop = NULL;
213 }
214 }
215}
216
217/*--------------*/
218
219/*
220 * Remove a specific entry from the lru.
221 */
222static void lru_remove(struct lru *lru, struct lru_entry *le)
223{
224 lru_iter_invalidate(lru, le);
225 if (lru->count == 1) {
226 lru->cursor = NULL;
227 } else {
228 if (lru->cursor == &le->list)
229 lru->cursor = lru->cursor->next;
230 list_del(&le->list);
231 }
232 lru->count--;
233}
234
235/*
236 * Mark as referenced.
237 */
238static inline void lru_reference(struct lru_entry *le)
239{
240 atomic_set(&le->referenced, 1);
241}
242
243/*--------------*/
244
245/*
246 * Remove the least recently used entry (approx), that passes the predicate.
247 * Returns NULL on failure.
248 */
249enum evict_result {
250 ER_EVICT,
251 ER_DONT_EVICT,
252 ER_STOP, /* stop looking for something to evict */
253};
254
255typedef enum evict_result (*le_predicate)(struct lru_entry *le, void *context);
256
257static struct lru_entry *lru_evict(struct lru *lru, le_predicate pred, void *context, bool no_sleep)
258{
259 unsigned long tested = 0;
260 struct list_head *h = lru->cursor;
261 struct lru_entry *le;
262
263 if (!h)
264 return NULL;
265 /*
266 * In the worst case we have to loop around twice. Once to clear
267 * the reference flags, and then again to discover the predicate
268 * fails for all entries.
269 */
270 while (tested < lru->count) {
271 le = container_of(h, struct lru_entry, list);
272
273 if (atomic_read(&le->referenced)) {
274 atomic_set(&le->referenced, 0);
275 } else {
276 tested++;
277 switch (pred(le, context)) {
278 case ER_EVICT:
279 /*
280 * Adjust the cursor, so we start the next
281 * search from here.
282 */
283 lru->cursor = le->list.next;
284 lru_remove(lru, le);
285 return le;
286
287 case ER_DONT_EVICT:
288 break;
289
290 case ER_STOP:
291 lru->cursor = le->list.next;
292 return NULL;
293 }
294 }
295
296 h = h->next;
297
298 if (!no_sleep)
299 cond_resched();
300 }
301
302 return NULL;
303}
304
305/*--------------------------------------------------------------*/
306
307/*
308 * Buffer state bits.
309 */
310#define B_READING 0
311#define B_WRITING 1
312#define B_DIRTY 2
313
314/*
315 * Describes how the block was allocated:
316 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
317 * See the comment at alloc_buffer_data.
318 */
319enum data_mode {
320 DATA_MODE_SLAB = 0,
321 DATA_MODE_GET_FREE_PAGES = 1,
322 DATA_MODE_VMALLOC = 2,
323 DATA_MODE_LIMIT = 3
324};
325
326struct dm_buffer {
327 /* protected by the locks in dm_buffer_cache */
328 struct rb_node node;
329
330 /* immutable, so don't need protecting */
331 sector_t block;
332 void *data;
333 unsigned char data_mode; /* DATA_MODE_* */
334
335 /*
336 * These two fields are used in isolation, so do not need
337 * a surrounding lock.
338 */
339 atomic_t hold_count;
340 unsigned long last_accessed;
341
342 /*
343 * Everything else is protected by the mutex in
344 * dm_bufio_client
345 */
346 unsigned long state;
347 struct lru_entry lru;
348 unsigned char list_mode; /* LIST_* */
349 blk_status_t read_error;
350 blk_status_t write_error;
351 unsigned int dirty_start;
352 unsigned int dirty_end;
353 unsigned int write_start;
354 unsigned int write_end;
355 struct list_head write_list;
356 struct dm_bufio_client *c;
357 void (*end_io)(struct dm_buffer *b, blk_status_t bs);
358#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
359#define MAX_STACK 10
360 unsigned int stack_len;
361 unsigned long stack_entries[MAX_STACK];
362#endif
363};
364
365/*--------------------------------------------------------------*/
366
367/*
368 * The buffer cache manages buffers, particularly:
369 * - inc/dec of holder count
370 * - setting the last_accessed field
371 * - maintains clean/dirty state along with lru
372 * - selecting buffers that match predicates
373 *
374 * It does *not* handle:
375 * - allocation/freeing of buffers.
376 * - IO
377 * - Eviction or cache sizing.
378 *
379 * cache_get() and cache_put() are threadsafe, you do not need to
380 * protect these calls with a surrounding mutex. All the other
381 * methods are not threadsafe; they do use locking primitives, but
382 * only enough to ensure get/put are threadsafe.
383 */
384
385struct buffer_tree {
386 union {
387 struct rw_semaphore lock;
388 rwlock_t spinlock;
389 } u;
390 struct rb_root root;
391} ____cacheline_aligned_in_smp;
392
393struct dm_buffer_cache {
394 struct lru lru[LIST_SIZE];
395 /*
396 * We spread entries across multiple trees to reduce contention
397 * on the locks.
398 */
399 unsigned int num_locks;
400 bool no_sleep;
401 struct buffer_tree trees[];
402};
403
404static DEFINE_STATIC_KEY_FALSE(no_sleep_enabled);
405
406static inline unsigned int cache_index(sector_t block, unsigned int num_locks)
407{
408 return dm_hash_locks_index(block, num_locks);
409}
410
411static inline void cache_read_lock(struct dm_buffer_cache *bc, sector_t block)
412{
413 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
414 read_lock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
415 else
416 down_read(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
417}
418
419static inline void cache_read_unlock(struct dm_buffer_cache *bc, sector_t block)
420{
421 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
422 read_unlock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
423 else
424 up_read(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
425}
426
427static inline void cache_write_lock(struct dm_buffer_cache *bc, sector_t block)
428{
429 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
430 write_lock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
431 else
432 down_write(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
433}
434
435static inline void cache_write_unlock(struct dm_buffer_cache *bc, sector_t block)
436{
437 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
438 write_unlock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
439 else
440 up_write(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
441}
442
443/*
444 * Sometimes we want to repeatedly get and drop locks as part of an iteration.
445 * This struct helps avoid redundant drop and gets of the same lock.
446 */
447struct lock_history {
448 struct dm_buffer_cache *cache;
449 bool write;
450 unsigned int previous;
451 unsigned int no_previous;
452};
453
454static void lh_init(struct lock_history *lh, struct dm_buffer_cache *cache, bool write)
455{
456 lh->cache = cache;
457 lh->write = write;
458 lh->no_previous = cache->num_locks;
459 lh->previous = lh->no_previous;
460}
461
462static void __lh_lock(struct lock_history *lh, unsigned int index)
463{
464 if (lh->write) {
465 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
466 write_lock_bh(&lh->cache->trees[index].u.spinlock);
467 else
468 down_write(&lh->cache->trees[index].u.lock);
469 } else {
470 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
471 read_lock_bh(&lh->cache->trees[index].u.spinlock);
472 else
473 down_read(&lh->cache->trees[index].u.lock);
474 }
475}
476
477static void __lh_unlock(struct lock_history *lh, unsigned int index)
478{
479 if (lh->write) {
480 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
481 write_unlock_bh(&lh->cache->trees[index].u.spinlock);
482 else
483 up_write(&lh->cache->trees[index].u.lock);
484 } else {
485 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
486 read_unlock_bh(&lh->cache->trees[index].u.spinlock);
487 else
488 up_read(&lh->cache->trees[index].u.lock);
489 }
490}
491
492/*
493 * Make sure you call this since it will unlock the final lock.
494 */
495static void lh_exit(struct lock_history *lh)
496{
497 if (lh->previous != lh->no_previous) {
498 __lh_unlock(lh, lh->previous);
499 lh->previous = lh->no_previous;
500 }
501}
502
503/*
504 * Named 'next' because there is no corresponding
505 * 'up/unlock' call since it's done automatically.
506 */
507static void lh_next(struct lock_history *lh, sector_t b)
508{
509 unsigned int index = cache_index(b, lh->no_previous); /* no_previous is num_locks */
510
511 if (lh->previous != lh->no_previous) {
512 if (lh->previous != index) {
513 __lh_unlock(lh, lh->previous);
514 __lh_lock(lh, index);
515 lh->previous = index;
516 }
517 } else {
518 __lh_lock(lh, index);
519 lh->previous = index;
520 }
521}
522
523static inline struct dm_buffer *le_to_buffer(struct lru_entry *le)
524{
525 return container_of(le, struct dm_buffer, lru);
526}
527
528static struct dm_buffer *list_to_buffer(struct list_head *l)
529{
530 struct lru_entry *le = list_entry(l, struct lru_entry, list);
531
532 if (!le)
533 return NULL;
534
535 return le_to_buffer(le);
536}
537
538static void cache_init(struct dm_buffer_cache *bc, unsigned int num_locks, bool no_sleep)
539{
540 unsigned int i;
541
542 bc->num_locks = num_locks;
543 bc->no_sleep = no_sleep;
544
545 for (i = 0; i < bc->num_locks; i++) {
546 if (no_sleep)
547 rwlock_init(&bc->trees[i].u.spinlock);
548 else
549 init_rwsem(&bc->trees[i].u.lock);
550 bc->trees[i].root = RB_ROOT;
551 }
552
553 lru_init(&bc->lru[LIST_CLEAN]);
554 lru_init(&bc->lru[LIST_DIRTY]);
555}
556
557static void cache_destroy(struct dm_buffer_cache *bc)
558{
559 unsigned int i;
560
561 for (i = 0; i < bc->num_locks; i++)
562 WARN_ON_ONCE(!RB_EMPTY_ROOT(&bc->trees[i].root));
563
564 lru_destroy(&bc->lru[LIST_CLEAN]);
565 lru_destroy(&bc->lru[LIST_DIRTY]);
566}
567
568/*--------------*/
569
570/*
571 * not threadsafe, or racey depending how you look at it
572 */
573static inline unsigned long cache_count(struct dm_buffer_cache *bc, int list_mode)
574{
575 return bc->lru[list_mode].count;
576}
577
578static inline unsigned long cache_total(struct dm_buffer_cache *bc)
579{
580 return cache_count(bc, LIST_CLEAN) + cache_count(bc, LIST_DIRTY);
581}
582
583/*--------------*/
584
585/*
586 * Gets a specific buffer, indexed by block.
587 * If the buffer is found then its holder count will be incremented and
588 * lru_reference will be called.
589 *
590 * threadsafe
591 */
592static struct dm_buffer *__cache_get(const struct rb_root *root, sector_t block)
593{
594 struct rb_node *n = root->rb_node;
595 struct dm_buffer *b;
596
597 while (n) {
598 b = container_of(n, struct dm_buffer, node);
599
600 if (b->block == block)
601 return b;
602
603 n = block < b->block ? n->rb_left : n->rb_right;
604 }
605
606 return NULL;
607}
608
609static void __cache_inc_buffer(struct dm_buffer *b)
610{
611 atomic_inc(&b->hold_count);
612 WRITE_ONCE(b->last_accessed, jiffies);
613}
614
615static struct dm_buffer *cache_get(struct dm_buffer_cache *bc, sector_t block)
616{
617 struct dm_buffer *b;
618
619 cache_read_lock(bc, block);
620 b = __cache_get(&bc->trees[cache_index(block, bc->num_locks)].root, block);
621 if (b) {
622 lru_reference(&b->lru);
623 __cache_inc_buffer(b);
624 }
625 cache_read_unlock(bc, block);
626
627 return b;
628}
629
630/*--------------*/
631
632/*
633 * Returns true if the hold count hits zero.
634 * threadsafe
635 */
636static bool cache_put(struct dm_buffer_cache *bc, struct dm_buffer *b)
637{
638 bool r;
639
640 cache_read_lock(bc, b->block);
641 BUG_ON(!atomic_read(&b->hold_count));
642 r = atomic_dec_and_test(&b->hold_count);
643 cache_read_unlock(bc, b->block);
644
645 return r;
646}
647
648/*--------------*/
649
650typedef enum evict_result (*b_predicate)(struct dm_buffer *, void *);
651
652/*
653 * Evicts a buffer based on a predicate. The oldest buffer that
654 * matches the predicate will be selected. In addition to the
655 * predicate the hold_count of the selected buffer will be zero.
656 */
657struct evict_wrapper {
658 struct lock_history *lh;
659 b_predicate pred;
660 void *context;
661};
662
663/*
664 * Wraps the buffer predicate turning it into an lru predicate. Adds
665 * extra test for hold_count.
666 */
667static enum evict_result __evict_pred(struct lru_entry *le, void *context)
668{
669 struct evict_wrapper *w = context;
670 struct dm_buffer *b = le_to_buffer(le);
671
672 lh_next(w->lh, b->block);
673
674 if (atomic_read(&b->hold_count))
675 return ER_DONT_EVICT;
676
677 return w->pred(b, w->context);
678}
679
680static struct dm_buffer *__cache_evict(struct dm_buffer_cache *bc, int list_mode,
681 b_predicate pred, void *context,
682 struct lock_history *lh)
683{
684 struct evict_wrapper w = {.lh = lh, .pred = pred, .context = context};
685 struct lru_entry *le;
686 struct dm_buffer *b;
687
688 le = lru_evict(&bc->lru[list_mode], __evict_pred, &w, bc->no_sleep);
689 if (!le)
690 return NULL;
691
692 b = le_to_buffer(le);
693 /* __evict_pred will have locked the appropriate tree. */
694 rb_erase(&b->node, &bc->trees[cache_index(b->block, bc->num_locks)].root);
695
696 return b;
697}
698
699static struct dm_buffer *cache_evict(struct dm_buffer_cache *bc, int list_mode,
700 b_predicate pred, void *context)
701{
702 struct dm_buffer *b;
703 struct lock_history lh;
704
705 lh_init(&lh, bc, true);
706 b = __cache_evict(bc, list_mode, pred, context, &lh);
707 lh_exit(&lh);
708
709 return b;
710}
711
712/*--------------*/
713
714/*
715 * Mark a buffer as clean or dirty. Not threadsafe.
716 */
717static void cache_mark(struct dm_buffer_cache *bc, struct dm_buffer *b, int list_mode)
718{
719 cache_write_lock(bc, b->block);
720 if (list_mode != b->list_mode) {
721 lru_remove(&bc->lru[b->list_mode], &b->lru);
722 b->list_mode = list_mode;
723 lru_insert(&bc->lru[b->list_mode], &b->lru);
724 }
725 cache_write_unlock(bc, b->block);
726}
727
728/*--------------*/
729
730/*
731 * Runs through the lru associated with 'old_mode', if the predicate matches then
732 * it moves them to 'new_mode'. Not threadsafe.
733 */
734static void __cache_mark_many(struct dm_buffer_cache *bc, int old_mode, int new_mode,
735 b_predicate pred, void *context, struct lock_history *lh)
736{
737 struct lru_entry *le;
738 struct dm_buffer *b;
739 struct evict_wrapper w = {.lh = lh, .pred = pred, .context = context};
740
741 while (true) {
742 le = lru_evict(&bc->lru[old_mode], __evict_pred, &w, bc->no_sleep);
743 if (!le)
744 break;
745
746 b = le_to_buffer(le);
747 b->list_mode = new_mode;
748 lru_insert(&bc->lru[b->list_mode], &b->lru);
749 }
750}
751
752static void cache_mark_many(struct dm_buffer_cache *bc, int old_mode, int new_mode,
753 b_predicate pred, void *context)
754{
755 struct lock_history lh;
756
757 lh_init(&lh, bc, true);
758 __cache_mark_many(bc, old_mode, new_mode, pred, context, &lh);
759 lh_exit(&lh);
760}
761
762/*--------------*/
763
764/*
765 * Iterates through all clean or dirty entries calling a function for each
766 * entry. The callback may terminate the iteration early. Not threadsafe.
767 */
768
769/*
770 * Iterator functions should return one of these actions to indicate
771 * how the iteration should proceed.
772 */
773enum it_action {
774 IT_NEXT,
775 IT_COMPLETE,
776};
777
778typedef enum it_action (*iter_fn)(struct dm_buffer *b, void *context);
779
780static void __cache_iterate(struct dm_buffer_cache *bc, int list_mode,
781 iter_fn fn, void *context, struct lock_history *lh)
782{
783 struct lru *lru = &bc->lru[list_mode];
784 struct lru_entry *le, *first;
785
786 if (!lru->cursor)
787 return;
788
789 first = le = to_le(lru->cursor);
790 do {
791 struct dm_buffer *b = le_to_buffer(le);
792
793 lh_next(lh, b->block);
794
795 switch (fn(b, context)) {
796 case IT_NEXT:
797 break;
798
799 case IT_COMPLETE:
800 return;
801 }
802 cond_resched();
803
804 le = to_le(le->list.next);
805 } while (le != first);
806}
807
808static void cache_iterate(struct dm_buffer_cache *bc, int list_mode,
809 iter_fn fn, void *context)
810{
811 struct lock_history lh;
812
813 lh_init(&lh, bc, false);
814 __cache_iterate(bc, list_mode, fn, context, &lh);
815 lh_exit(&lh);
816}
817
818/*--------------*/
819
820/*
821 * Passes ownership of the buffer to the cache. Returns false if the
822 * buffer was already present (in which case ownership does not pass).
823 * eg, a race with another thread.
824 *
825 * Holder count should be 1 on insertion.
826 *
827 * Not threadsafe.
828 */
829static bool __cache_insert(struct rb_root *root, struct dm_buffer *b)
830{
831 struct rb_node **new = &root->rb_node, *parent = NULL;
832 struct dm_buffer *found;
833
834 while (*new) {
835 found = container_of(*new, struct dm_buffer, node);
836
837 if (found->block == b->block)
838 return false;
839
840 parent = *new;
841 new = b->block < found->block ?
842 &found->node.rb_left : &found->node.rb_right;
843 }
844
845 rb_link_node(&b->node, parent, new);
846 rb_insert_color(&b->node, root);
847
848 return true;
849}
850
851static bool cache_insert(struct dm_buffer_cache *bc, struct dm_buffer *b)
852{
853 bool r;
854
855 if (WARN_ON_ONCE(b->list_mode >= LIST_SIZE))
856 return false;
857
858 cache_write_lock(bc, b->block);
859 BUG_ON(atomic_read(&b->hold_count) != 1);
860 r = __cache_insert(&bc->trees[cache_index(b->block, bc->num_locks)].root, b);
861 if (r)
862 lru_insert(&bc->lru[b->list_mode], &b->lru);
863 cache_write_unlock(bc, b->block);
864
865 return r;
866}
867
868/*--------------*/
869
870/*
871 * Removes buffer from cache, ownership of the buffer passes back to the caller.
872 * Fails if the hold_count is not one (ie. the caller holds the only reference).
873 *
874 * Not threadsafe.
875 */
876static bool cache_remove(struct dm_buffer_cache *bc, struct dm_buffer *b)
877{
878 bool r;
879
880 cache_write_lock(bc, b->block);
881
882 if (atomic_read(&b->hold_count) != 1) {
883 r = false;
884 } else {
885 r = true;
886 rb_erase(&b->node, &bc->trees[cache_index(b->block, bc->num_locks)].root);
887 lru_remove(&bc->lru[b->list_mode], &b->lru);
888 }
889
890 cache_write_unlock(bc, b->block);
891
892 return r;
893}
894
895/*--------------*/
896
897typedef void (*b_release)(struct dm_buffer *);
898
899static struct dm_buffer *__find_next(struct rb_root *root, sector_t block)
900{
901 struct rb_node *n = root->rb_node;
902 struct dm_buffer *b;
903 struct dm_buffer *best = NULL;
904
905 while (n) {
906 b = container_of(n, struct dm_buffer, node);
907
908 if (b->block == block)
909 return b;
910
911 if (block <= b->block) {
912 n = n->rb_left;
913 best = b;
914 } else {
915 n = n->rb_right;
916 }
917 }
918
919 return best;
920}
921
922static void __remove_range(struct dm_buffer_cache *bc,
923 struct rb_root *root,
924 sector_t begin, sector_t end,
925 b_predicate pred, b_release release)
926{
927 struct dm_buffer *b;
928
929 while (true) {
930 cond_resched();
931
932 b = __find_next(root, begin);
933 if (!b || (b->block >= end))
934 break;
935
936 begin = b->block + 1;
937
938 if (atomic_read(&b->hold_count))
939 continue;
940
941 if (pred(b, NULL) == ER_EVICT) {
942 rb_erase(&b->node, root);
943 lru_remove(&bc->lru[b->list_mode], &b->lru);
944 release(b);
945 }
946 }
947}
948
949static void cache_remove_range(struct dm_buffer_cache *bc,
950 sector_t begin, sector_t end,
951 b_predicate pred, b_release release)
952{
953 unsigned int i;
954
955 BUG_ON(bc->no_sleep);
956 for (i = 0; i < bc->num_locks; i++) {
957 down_write(&bc->trees[i].u.lock);
958 __remove_range(bc, &bc->trees[i].root, begin, end, pred, release);
959 up_write(&bc->trees[i].u.lock);
960 }
961}
962
963/*----------------------------------------------------------------*/
964
965/*
966 * Linking of buffers:
967 * All buffers are linked to buffer_cache with their node field.
968 *
969 * Clean buffers that are not being written (B_WRITING not set)
970 * are linked to lru[LIST_CLEAN] with their lru_list field.
971 *
972 * Dirty and clean buffers that are being written are linked to
973 * lru[LIST_DIRTY] with their lru_list field. When the write
974 * finishes, the buffer cannot be relinked immediately (because we
975 * are in an interrupt context and relinking requires process
976 * context), so some clean-not-writing buffers can be held on
977 * dirty_lru too. They are later added to lru in the process
978 * context.
979 */
980struct dm_bufio_client {
981 struct block_device *bdev;
982 unsigned int block_size;
983 s8 sectors_per_block_bits;
984
985 bool no_sleep;
986 struct mutex lock;
987 spinlock_t spinlock;
988
989 int async_write_error;
990
991 void (*alloc_callback)(struct dm_buffer *buf);
992 void (*write_callback)(struct dm_buffer *buf);
993 struct kmem_cache *slab_buffer;
994 struct kmem_cache *slab_cache;
995 struct dm_io_client *dm_io;
996
997 struct list_head reserved_buffers;
998 unsigned int need_reserved_buffers;
999
1000 unsigned int minimum_buffers;
1001
1002 sector_t start;
1003
1004 struct shrinker *shrinker;
1005 struct work_struct shrink_work;
1006 atomic_long_t need_shrink;
1007
1008 wait_queue_head_t free_buffer_wait;
1009
1010 struct list_head client_list;
1011
1012 /*
1013 * Used by global_cleanup to sort the clients list.
1014 */
1015 unsigned long oldest_buffer;
1016
1017 struct dm_buffer_cache cache; /* must be last member */
1018};
1019
1020/*----------------------------------------------------------------*/
1021
1022#define dm_bufio_in_request() (!!current->bio_list)
1023
1024static void dm_bufio_lock(struct dm_bufio_client *c)
1025{
1026 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1027 spin_lock_bh(&c->spinlock);
1028 else
1029 mutex_lock_nested(&c->lock, dm_bufio_in_request());
1030}
1031
1032static void dm_bufio_unlock(struct dm_bufio_client *c)
1033{
1034 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1035 spin_unlock_bh(&c->spinlock);
1036 else
1037 mutex_unlock(&c->lock);
1038}
1039
1040/*----------------------------------------------------------------*/
1041
1042/*
1043 * Default cache size: available memory divided by the ratio.
1044 */
1045static unsigned long dm_bufio_default_cache_size;
1046
1047/*
1048 * Total cache size set by the user.
1049 */
1050static unsigned long dm_bufio_cache_size;
1051
1052/*
1053 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
1054 * at any time. If it disagrees, the user has changed cache size.
1055 */
1056static unsigned long dm_bufio_cache_size_latch;
1057
1058static DEFINE_SPINLOCK(global_spinlock);
1059
1060/*
1061 * Buffers are freed after this timeout
1062 */
1063static unsigned int dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
1064static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
1065
1066static unsigned long dm_bufio_peak_allocated;
1067static unsigned long dm_bufio_allocated_kmem_cache;
1068static unsigned long dm_bufio_allocated_get_free_pages;
1069static unsigned long dm_bufio_allocated_vmalloc;
1070static unsigned long dm_bufio_current_allocated;
1071
1072/*----------------------------------------------------------------*/
1073
1074/*
1075 * The current number of clients.
1076 */
1077static int dm_bufio_client_count;
1078
1079/*
1080 * The list of all clients.
1081 */
1082static LIST_HEAD(dm_bufio_all_clients);
1083
1084/*
1085 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
1086 */
1087static DEFINE_MUTEX(dm_bufio_clients_lock);
1088
1089static struct workqueue_struct *dm_bufio_wq;
1090static struct delayed_work dm_bufio_cleanup_old_work;
1091static struct work_struct dm_bufio_replacement_work;
1092
1093
1094#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1095static void buffer_record_stack(struct dm_buffer *b)
1096{
1097 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
1098}
1099#endif
1100
1101/*----------------------------------------------------------------*/
1102
1103static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
1104{
1105 unsigned char data_mode;
1106 long diff;
1107
1108 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
1109 &dm_bufio_allocated_kmem_cache,
1110 &dm_bufio_allocated_get_free_pages,
1111 &dm_bufio_allocated_vmalloc,
1112 };
1113
1114 data_mode = b->data_mode;
1115 diff = (long)b->c->block_size;
1116 if (unlink)
1117 diff = -diff;
1118
1119 spin_lock(&global_spinlock);
1120
1121 *class_ptr[data_mode] += diff;
1122
1123 dm_bufio_current_allocated += diff;
1124
1125 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
1126 dm_bufio_peak_allocated = dm_bufio_current_allocated;
1127
1128 if (!unlink) {
1129 if (dm_bufio_current_allocated > dm_bufio_cache_size)
1130 queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
1131 }
1132
1133 spin_unlock(&global_spinlock);
1134}
1135
1136/*
1137 * Change the number of clients and recalculate per-client limit.
1138 */
1139static void __cache_size_refresh(void)
1140{
1141 if (WARN_ON(!mutex_is_locked(&dm_bufio_clients_lock)))
1142 return;
1143 if (WARN_ON(dm_bufio_client_count < 0))
1144 return;
1145
1146 dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
1147
1148 /*
1149 * Use default if set to 0 and report the actual cache size used.
1150 */
1151 if (!dm_bufio_cache_size_latch) {
1152 (void)cmpxchg(&dm_bufio_cache_size, 0,
1153 dm_bufio_default_cache_size);
1154 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
1155 }
1156}
1157
1158/*
1159 * Allocating buffer data.
1160 *
1161 * Small buffers are allocated with kmem_cache, to use space optimally.
1162 *
1163 * For large buffers, we choose between get_free_pages and vmalloc.
1164 * Each has advantages and disadvantages.
1165 *
1166 * __get_free_pages can randomly fail if the memory is fragmented.
1167 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
1168 * as low as 128M) so using it for caching is not appropriate.
1169 *
1170 * If the allocation may fail we use __get_free_pages. Memory fragmentation
1171 * won't have a fatal effect here, but it just causes flushes of some other
1172 * buffers and more I/O will be performed. Don't use __get_free_pages if it
1173 * always fails (i.e. order > MAX_PAGE_ORDER).
1174 *
1175 * If the allocation shouldn't fail we use __vmalloc. This is only for the
1176 * initial reserve allocation, so there's no risk of wasting all vmalloc
1177 * space.
1178 */
1179static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
1180 unsigned char *data_mode)
1181{
1182 if (unlikely(c->slab_cache != NULL)) {
1183 *data_mode = DATA_MODE_SLAB;
1184 return kmem_cache_alloc(c->slab_cache, gfp_mask);
1185 }
1186
1187 if (c->block_size <= KMALLOC_MAX_SIZE &&
1188 gfp_mask & __GFP_NORETRY) {
1189 *data_mode = DATA_MODE_GET_FREE_PAGES;
1190 return (void *)__get_free_pages(gfp_mask,
1191 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
1192 }
1193
1194 *data_mode = DATA_MODE_VMALLOC;
1195
1196 return __vmalloc(c->block_size, gfp_mask);
1197}
1198
1199/*
1200 * Free buffer's data.
1201 */
1202static void free_buffer_data(struct dm_bufio_client *c,
1203 void *data, unsigned char data_mode)
1204{
1205 switch (data_mode) {
1206 case DATA_MODE_SLAB:
1207 kmem_cache_free(c->slab_cache, data);
1208 break;
1209
1210 case DATA_MODE_GET_FREE_PAGES:
1211 free_pages((unsigned long)data,
1212 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
1213 break;
1214
1215 case DATA_MODE_VMALLOC:
1216 vfree(data);
1217 break;
1218
1219 default:
1220 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
1221 data_mode);
1222 BUG();
1223 }
1224}
1225
1226/*
1227 * Allocate buffer and its data.
1228 */
1229static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
1230{
1231 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
1232
1233 if (!b)
1234 return NULL;
1235
1236 b->c = c;
1237
1238 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
1239 if (!b->data) {
1240 kmem_cache_free(c->slab_buffer, b);
1241 return NULL;
1242 }
1243 adjust_total_allocated(b, false);
1244
1245#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1246 b->stack_len = 0;
1247#endif
1248 return b;
1249}
1250
1251/*
1252 * Free buffer and its data.
1253 */
1254static void free_buffer(struct dm_buffer *b)
1255{
1256 struct dm_bufio_client *c = b->c;
1257
1258 adjust_total_allocated(b, true);
1259 free_buffer_data(c, b->data, b->data_mode);
1260 kmem_cache_free(c->slab_buffer, b);
1261}
1262
1263/*
1264 *--------------------------------------------------------------------------
1265 * Submit I/O on the buffer.
1266 *
1267 * Bio interface is faster but it has some problems:
1268 * the vector list is limited (increasing this limit increases
1269 * memory-consumption per buffer, so it is not viable);
1270 *
1271 * the memory must be direct-mapped, not vmalloced;
1272 *
1273 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
1274 * it is not vmalloced, try using the bio interface.
1275 *
1276 * If the buffer is big, if it is vmalloced or if the underlying device
1277 * rejects the bio because it is too large, use dm-io layer to do the I/O.
1278 * The dm-io layer splits the I/O into multiple requests, avoiding the above
1279 * shortcomings.
1280 *--------------------------------------------------------------------------
1281 */
1282
1283/*
1284 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
1285 * that the request was handled directly with bio interface.
1286 */
1287static void dmio_complete(unsigned long error, void *context)
1288{
1289 struct dm_buffer *b = context;
1290
1291 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
1292}
1293
1294static void use_dmio(struct dm_buffer *b, enum req_op op, sector_t sector,
1295 unsigned int n_sectors, unsigned int offset)
1296{
1297 int r;
1298 struct dm_io_request io_req = {
1299 .bi_opf = op,
1300 .notify.fn = dmio_complete,
1301 .notify.context = b,
1302 .client = b->c->dm_io,
1303 };
1304 struct dm_io_region region = {
1305 .bdev = b->c->bdev,
1306 .sector = sector,
1307 .count = n_sectors,
1308 };
1309
1310 if (b->data_mode != DATA_MODE_VMALLOC) {
1311 io_req.mem.type = DM_IO_KMEM;
1312 io_req.mem.ptr.addr = (char *)b->data + offset;
1313 } else {
1314 io_req.mem.type = DM_IO_VMA;
1315 io_req.mem.ptr.vma = (char *)b->data + offset;
1316 }
1317
1318 r = dm_io(&io_req, 1, ®ion, NULL);
1319 if (unlikely(r))
1320 b->end_io(b, errno_to_blk_status(r));
1321}
1322
1323static void bio_complete(struct bio *bio)
1324{
1325 struct dm_buffer *b = bio->bi_private;
1326 blk_status_t status = bio->bi_status;
1327
1328 bio_uninit(bio);
1329 kfree(bio);
1330 b->end_io(b, status);
1331}
1332
1333static void use_bio(struct dm_buffer *b, enum req_op op, sector_t sector,
1334 unsigned int n_sectors, unsigned int offset)
1335{
1336 struct bio *bio;
1337 char *ptr;
1338 unsigned int len;
1339
1340 bio = bio_kmalloc(1, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN);
1341 if (!bio) {
1342 use_dmio(b, op, sector, n_sectors, offset);
1343 return;
1344 }
1345 bio_init(bio, b->c->bdev, bio->bi_inline_vecs, 1, op);
1346 bio->bi_iter.bi_sector = sector;
1347 bio->bi_end_io = bio_complete;
1348 bio->bi_private = b;
1349
1350 ptr = (char *)b->data + offset;
1351 len = n_sectors << SECTOR_SHIFT;
1352
1353 __bio_add_page(bio, virt_to_page(ptr), len, offset_in_page(ptr));
1354
1355 submit_bio(bio);
1356}
1357
1358static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
1359{
1360 sector_t sector;
1361
1362 if (likely(c->sectors_per_block_bits >= 0))
1363 sector = block << c->sectors_per_block_bits;
1364 else
1365 sector = block * (c->block_size >> SECTOR_SHIFT);
1366 sector += c->start;
1367
1368 return sector;
1369}
1370
1371static void submit_io(struct dm_buffer *b, enum req_op op,
1372 void (*end_io)(struct dm_buffer *, blk_status_t))
1373{
1374 unsigned int n_sectors;
1375 sector_t sector;
1376 unsigned int offset, end;
1377
1378 b->end_io = end_io;
1379
1380 sector = block_to_sector(b->c, b->block);
1381
1382 if (op != REQ_OP_WRITE) {
1383 n_sectors = b->c->block_size >> SECTOR_SHIFT;
1384 offset = 0;
1385 } else {
1386 if (b->c->write_callback)
1387 b->c->write_callback(b);
1388 offset = b->write_start;
1389 end = b->write_end;
1390 offset &= -DM_BUFIO_WRITE_ALIGN;
1391 end += DM_BUFIO_WRITE_ALIGN - 1;
1392 end &= -DM_BUFIO_WRITE_ALIGN;
1393 if (unlikely(end > b->c->block_size))
1394 end = b->c->block_size;
1395
1396 sector += offset >> SECTOR_SHIFT;
1397 n_sectors = (end - offset) >> SECTOR_SHIFT;
1398 }
1399
1400 if (b->data_mode != DATA_MODE_VMALLOC)
1401 use_bio(b, op, sector, n_sectors, offset);
1402 else
1403 use_dmio(b, op, sector, n_sectors, offset);
1404}
1405
1406/*
1407 *--------------------------------------------------------------
1408 * Writing dirty buffers
1409 *--------------------------------------------------------------
1410 */
1411
1412/*
1413 * The endio routine for write.
1414 *
1415 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
1416 * it.
1417 */
1418static void write_endio(struct dm_buffer *b, blk_status_t status)
1419{
1420 b->write_error = status;
1421 if (unlikely(status)) {
1422 struct dm_bufio_client *c = b->c;
1423
1424 (void)cmpxchg(&c->async_write_error, 0,
1425 blk_status_to_errno(status));
1426 }
1427
1428 BUG_ON(!test_bit(B_WRITING, &b->state));
1429
1430 smp_mb__before_atomic();
1431 clear_bit(B_WRITING, &b->state);
1432 smp_mb__after_atomic();
1433
1434 wake_up_bit(&b->state, B_WRITING);
1435}
1436
1437/*
1438 * Initiate a write on a dirty buffer, but don't wait for it.
1439 *
1440 * - If the buffer is not dirty, exit.
1441 * - If there some previous write going on, wait for it to finish (we can't
1442 * have two writes on the same buffer simultaneously).
1443 * - Submit our write and don't wait on it. We set B_WRITING indicating
1444 * that there is a write in progress.
1445 */
1446static void __write_dirty_buffer(struct dm_buffer *b,
1447 struct list_head *write_list)
1448{
1449 if (!test_bit(B_DIRTY, &b->state))
1450 return;
1451
1452 clear_bit(B_DIRTY, &b->state);
1453 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
1454
1455 b->write_start = b->dirty_start;
1456 b->write_end = b->dirty_end;
1457
1458 if (!write_list)
1459 submit_io(b, REQ_OP_WRITE, write_endio);
1460 else
1461 list_add_tail(&b->write_list, write_list);
1462}
1463
1464static void __flush_write_list(struct list_head *write_list)
1465{
1466 struct blk_plug plug;
1467
1468 blk_start_plug(&plug);
1469 while (!list_empty(write_list)) {
1470 struct dm_buffer *b =
1471 list_entry(write_list->next, struct dm_buffer, write_list);
1472 list_del(&b->write_list);
1473 submit_io(b, REQ_OP_WRITE, write_endio);
1474 cond_resched();
1475 }
1476 blk_finish_plug(&plug);
1477}
1478
1479/*
1480 * Wait until any activity on the buffer finishes. Possibly write the
1481 * buffer if it is dirty. When this function finishes, there is no I/O
1482 * running on the buffer and the buffer is not dirty.
1483 */
1484static void __make_buffer_clean(struct dm_buffer *b)
1485{
1486 BUG_ON(atomic_read(&b->hold_count));
1487
1488 /* smp_load_acquire() pairs with read_endio()'s smp_mb__before_atomic() */
1489 if (!smp_load_acquire(&b->state)) /* fast case */
1490 return;
1491
1492 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1493 __write_dirty_buffer(b, NULL);
1494 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
1495}
1496
1497static enum evict_result is_clean(struct dm_buffer *b, void *context)
1498{
1499 struct dm_bufio_client *c = context;
1500
1501 /* These should never happen */
1502 if (WARN_ON_ONCE(test_bit(B_WRITING, &b->state)))
1503 return ER_DONT_EVICT;
1504 if (WARN_ON_ONCE(test_bit(B_DIRTY, &b->state)))
1505 return ER_DONT_EVICT;
1506 if (WARN_ON_ONCE(b->list_mode != LIST_CLEAN))
1507 return ER_DONT_EVICT;
1508
1509 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep &&
1510 unlikely(test_bit(B_READING, &b->state)))
1511 return ER_DONT_EVICT;
1512
1513 return ER_EVICT;
1514}
1515
1516static enum evict_result is_dirty(struct dm_buffer *b, void *context)
1517{
1518 /* These should never happen */
1519 if (WARN_ON_ONCE(test_bit(B_READING, &b->state)))
1520 return ER_DONT_EVICT;
1521 if (WARN_ON_ONCE(b->list_mode != LIST_DIRTY))
1522 return ER_DONT_EVICT;
1523
1524 return ER_EVICT;
1525}
1526
1527/*
1528 * Find some buffer that is not held by anybody, clean it, unlink it and
1529 * return it.
1530 */
1531static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
1532{
1533 struct dm_buffer *b;
1534
1535 b = cache_evict(&c->cache, LIST_CLEAN, is_clean, c);
1536 if (b) {
1537 /* this also waits for pending reads */
1538 __make_buffer_clean(b);
1539 return b;
1540 }
1541
1542 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1543 return NULL;
1544
1545 b = cache_evict(&c->cache, LIST_DIRTY, is_dirty, NULL);
1546 if (b) {
1547 __make_buffer_clean(b);
1548 return b;
1549 }
1550
1551 return NULL;
1552}
1553
1554/*
1555 * Wait until some other threads free some buffer or release hold count on
1556 * some buffer.
1557 *
1558 * This function is entered with c->lock held, drops it and regains it
1559 * before exiting.
1560 */
1561static void __wait_for_free_buffer(struct dm_bufio_client *c)
1562{
1563 DECLARE_WAITQUEUE(wait, current);
1564
1565 add_wait_queue(&c->free_buffer_wait, &wait);
1566 set_current_state(TASK_UNINTERRUPTIBLE);
1567 dm_bufio_unlock(c);
1568
1569 /*
1570 * It's possible to miss a wake up event since we don't always
1571 * hold c->lock when wake_up is called. So we have a timeout here,
1572 * just in case.
1573 */
1574 io_schedule_timeout(5 * HZ);
1575
1576 remove_wait_queue(&c->free_buffer_wait, &wait);
1577
1578 dm_bufio_lock(c);
1579}
1580
1581enum new_flag {
1582 NF_FRESH = 0,
1583 NF_READ = 1,
1584 NF_GET = 2,
1585 NF_PREFETCH = 3
1586};
1587
1588/*
1589 * Allocate a new buffer. If the allocation is not possible, wait until
1590 * some other thread frees a buffer.
1591 *
1592 * May drop the lock and regain it.
1593 */
1594static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
1595{
1596 struct dm_buffer *b;
1597 bool tried_noio_alloc = false;
1598
1599 /*
1600 * dm-bufio is resistant to allocation failures (it just keeps
1601 * one buffer reserved in cases all the allocations fail).
1602 * So set flags to not try too hard:
1603 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
1604 * mutex and wait ourselves.
1605 * __GFP_NORETRY: don't retry and rather return failure
1606 * __GFP_NOMEMALLOC: don't use emergency reserves
1607 * __GFP_NOWARN: don't print a warning in case of failure
1608 *
1609 * For debugging, if we set the cache size to 1, no new buffers will
1610 * be allocated.
1611 */
1612 while (1) {
1613 if (dm_bufio_cache_size_latch != 1) {
1614 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
1615 if (b)
1616 return b;
1617 }
1618
1619 if (nf == NF_PREFETCH)
1620 return NULL;
1621
1622 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
1623 dm_bufio_unlock(c);
1624 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
1625 dm_bufio_lock(c);
1626 if (b)
1627 return b;
1628 tried_noio_alloc = true;
1629 }
1630
1631 if (!list_empty(&c->reserved_buffers)) {
1632 b = list_to_buffer(c->reserved_buffers.next);
1633 list_del(&b->lru.list);
1634 c->need_reserved_buffers++;
1635
1636 return b;
1637 }
1638
1639 b = __get_unclaimed_buffer(c);
1640 if (b)
1641 return b;
1642
1643 __wait_for_free_buffer(c);
1644 }
1645}
1646
1647static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
1648{
1649 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
1650
1651 if (!b)
1652 return NULL;
1653
1654 if (c->alloc_callback)
1655 c->alloc_callback(b);
1656
1657 return b;
1658}
1659
1660/*
1661 * Free a buffer and wake other threads waiting for free buffers.
1662 */
1663static void __free_buffer_wake(struct dm_buffer *b)
1664{
1665 struct dm_bufio_client *c = b->c;
1666
1667 b->block = -1;
1668 if (!c->need_reserved_buffers)
1669 free_buffer(b);
1670 else {
1671 list_add(&b->lru.list, &c->reserved_buffers);
1672 c->need_reserved_buffers--;
1673 }
1674
1675 /*
1676 * We hold the bufio lock here, so no one can add entries to the
1677 * wait queue anyway.
1678 */
1679 if (unlikely(waitqueue_active(&c->free_buffer_wait)))
1680 wake_up(&c->free_buffer_wait);
1681}
1682
1683static enum evict_result cleaned(struct dm_buffer *b, void *context)
1684{
1685 if (WARN_ON_ONCE(test_bit(B_READING, &b->state)))
1686 return ER_DONT_EVICT; /* should never happen */
1687
1688 if (test_bit(B_DIRTY, &b->state) || test_bit(B_WRITING, &b->state))
1689 return ER_DONT_EVICT;
1690 else
1691 return ER_EVICT;
1692}
1693
1694static void __move_clean_buffers(struct dm_bufio_client *c)
1695{
1696 cache_mark_many(&c->cache, LIST_DIRTY, LIST_CLEAN, cleaned, NULL);
1697}
1698
1699struct write_context {
1700 int no_wait;
1701 struct list_head *write_list;
1702};
1703
1704static enum it_action write_one(struct dm_buffer *b, void *context)
1705{
1706 struct write_context *wc = context;
1707
1708 if (wc->no_wait && test_bit(B_WRITING, &b->state))
1709 return IT_COMPLETE;
1710
1711 __write_dirty_buffer(b, wc->write_list);
1712 return IT_NEXT;
1713}
1714
1715static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
1716 struct list_head *write_list)
1717{
1718 struct write_context wc = {.no_wait = no_wait, .write_list = write_list};
1719
1720 __move_clean_buffers(c);
1721 cache_iterate(&c->cache, LIST_DIRTY, write_one, &wc);
1722}
1723
1724/*
1725 * Check if we're over watermark.
1726 * If we are over threshold_buffers, start freeing buffers.
1727 * If we're over "limit_buffers", block until we get under the limit.
1728 */
1729static void __check_watermark(struct dm_bufio_client *c,
1730 struct list_head *write_list)
1731{
1732 if (cache_count(&c->cache, LIST_DIRTY) >
1733 cache_count(&c->cache, LIST_CLEAN) * DM_BUFIO_WRITEBACK_RATIO)
1734 __write_dirty_buffers_async(c, 1, write_list);
1735}
1736
1737/*
1738 *--------------------------------------------------------------
1739 * Getting a buffer
1740 *--------------------------------------------------------------
1741 */
1742
1743static void cache_put_and_wake(struct dm_bufio_client *c, struct dm_buffer *b)
1744{
1745 /*
1746 * Relying on waitqueue_active() is racey, but we sleep
1747 * with schedule_timeout anyway.
1748 */
1749 if (cache_put(&c->cache, b) &&
1750 unlikely(waitqueue_active(&c->free_buffer_wait)))
1751 wake_up(&c->free_buffer_wait);
1752}
1753
1754/*
1755 * This assumes you have already checked the cache to see if the buffer
1756 * is already present (it will recheck after dropping the lock for allocation).
1757 */
1758static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
1759 enum new_flag nf, int *need_submit,
1760 struct list_head *write_list)
1761{
1762 struct dm_buffer *b, *new_b = NULL;
1763
1764 *need_submit = 0;
1765
1766 /* This can't be called with NF_GET */
1767 if (WARN_ON_ONCE(nf == NF_GET))
1768 return NULL;
1769
1770 new_b = __alloc_buffer_wait(c, nf);
1771 if (!new_b)
1772 return NULL;
1773
1774 /*
1775 * We've had a period where the mutex was unlocked, so need to
1776 * recheck the buffer tree.
1777 */
1778 b = cache_get(&c->cache, block);
1779 if (b) {
1780 __free_buffer_wake(new_b);
1781 goto found_buffer;
1782 }
1783
1784 __check_watermark(c, write_list);
1785
1786 b = new_b;
1787 atomic_set(&b->hold_count, 1);
1788 WRITE_ONCE(b->last_accessed, jiffies);
1789 b->block = block;
1790 b->read_error = 0;
1791 b->write_error = 0;
1792 b->list_mode = LIST_CLEAN;
1793
1794 if (nf == NF_FRESH)
1795 b->state = 0;
1796 else {
1797 b->state = 1 << B_READING;
1798 *need_submit = 1;
1799 }
1800
1801 /*
1802 * We mustn't insert into the cache until the B_READING state
1803 * is set. Otherwise another thread could get it and use
1804 * it before it had been read.
1805 */
1806 cache_insert(&c->cache, b);
1807
1808 return b;
1809
1810found_buffer:
1811 if (nf == NF_PREFETCH) {
1812 cache_put_and_wake(c, b);
1813 return NULL;
1814 }
1815
1816 /*
1817 * Note: it is essential that we don't wait for the buffer to be
1818 * read if dm_bufio_get function is used. Both dm_bufio_get and
1819 * dm_bufio_prefetch can be used in the driver request routine.
1820 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1821 * the same buffer, it would deadlock if we waited.
1822 */
1823 if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state))) {
1824 cache_put_and_wake(c, b);
1825 return NULL;
1826 }
1827
1828 return b;
1829}
1830
1831/*
1832 * The endio routine for reading: set the error, clear the bit and wake up
1833 * anyone waiting on the buffer.
1834 */
1835static void read_endio(struct dm_buffer *b, blk_status_t status)
1836{
1837 b->read_error = status;
1838
1839 BUG_ON(!test_bit(B_READING, &b->state));
1840
1841 smp_mb__before_atomic();
1842 clear_bit(B_READING, &b->state);
1843 smp_mb__after_atomic();
1844
1845 wake_up_bit(&b->state, B_READING);
1846}
1847
1848/*
1849 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1850 * functions is similar except that dm_bufio_new doesn't read the
1851 * buffer from the disk (assuming that the caller overwrites all the data
1852 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1853 */
1854static void *new_read(struct dm_bufio_client *c, sector_t block,
1855 enum new_flag nf, struct dm_buffer **bp)
1856{
1857 int need_submit = 0;
1858 struct dm_buffer *b;
1859
1860 LIST_HEAD(write_list);
1861
1862 *bp = NULL;
1863
1864 /*
1865 * Fast path, hopefully the block is already in the cache. No need
1866 * to get the client lock for this.
1867 */
1868 b = cache_get(&c->cache, block);
1869 if (b) {
1870 if (nf == NF_PREFETCH) {
1871 cache_put_and_wake(c, b);
1872 return NULL;
1873 }
1874
1875 /*
1876 * Note: it is essential that we don't wait for the buffer to be
1877 * read if dm_bufio_get function is used. Both dm_bufio_get and
1878 * dm_bufio_prefetch can be used in the driver request routine.
1879 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1880 * the same buffer, it would deadlock if we waited.
1881 */
1882 if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state))) {
1883 cache_put_and_wake(c, b);
1884 return NULL;
1885 }
1886 }
1887
1888 if (!b) {
1889 if (nf == NF_GET)
1890 return NULL;
1891
1892 dm_bufio_lock(c);
1893 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1894 dm_bufio_unlock(c);
1895 }
1896
1897#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1898 if (b && (atomic_read(&b->hold_count) == 1))
1899 buffer_record_stack(b);
1900#endif
1901
1902 __flush_write_list(&write_list);
1903
1904 if (!b)
1905 return NULL;
1906
1907 if (need_submit)
1908 submit_io(b, REQ_OP_READ, read_endio);
1909
1910 if (nf != NF_GET) /* we already tested this condition above */
1911 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1912
1913 if (b->read_error) {
1914 int error = blk_status_to_errno(b->read_error);
1915
1916 dm_bufio_release(b);
1917
1918 return ERR_PTR(error);
1919 }
1920
1921 *bp = b;
1922
1923 return b->data;
1924}
1925
1926void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1927 struct dm_buffer **bp)
1928{
1929 return new_read(c, block, NF_GET, bp);
1930}
1931EXPORT_SYMBOL_GPL(dm_bufio_get);
1932
1933void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1934 struct dm_buffer **bp)
1935{
1936 if (WARN_ON_ONCE(dm_bufio_in_request()))
1937 return ERR_PTR(-EINVAL);
1938
1939 return new_read(c, block, NF_READ, bp);
1940}
1941EXPORT_SYMBOL_GPL(dm_bufio_read);
1942
1943void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1944 struct dm_buffer **bp)
1945{
1946 if (WARN_ON_ONCE(dm_bufio_in_request()))
1947 return ERR_PTR(-EINVAL);
1948
1949 return new_read(c, block, NF_FRESH, bp);
1950}
1951EXPORT_SYMBOL_GPL(dm_bufio_new);
1952
1953void dm_bufio_prefetch(struct dm_bufio_client *c,
1954 sector_t block, unsigned int n_blocks)
1955{
1956 struct blk_plug plug;
1957
1958 LIST_HEAD(write_list);
1959
1960 if (WARN_ON_ONCE(dm_bufio_in_request()))
1961 return; /* should never happen */
1962
1963 blk_start_plug(&plug);
1964
1965 for (; n_blocks--; block++) {
1966 int need_submit;
1967 struct dm_buffer *b;
1968
1969 b = cache_get(&c->cache, block);
1970 if (b) {
1971 /* already in cache */
1972 cache_put_and_wake(c, b);
1973 continue;
1974 }
1975
1976 dm_bufio_lock(c);
1977 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1978 &write_list);
1979 if (unlikely(!list_empty(&write_list))) {
1980 dm_bufio_unlock(c);
1981 blk_finish_plug(&plug);
1982 __flush_write_list(&write_list);
1983 blk_start_plug(&plug);
1984 dm_bufio_lock(c);
1985 }
1986 if (unlikely(b != NULL)) {
1987 dm_bufio_unlock(c);
1988
1989 if (need_submit)
1990 submit_io(b, REQ_OP_READ, read_endio);
1991 dm_bufio_release(b);
1992
1993 cond_resched();
1994
1995 if (!n_blocks)
1996 goto flush_plug;
1997 dm_bufio_lock(c);
1998 }
1999 dm_bufio_unlock(c);
2000 }
2001
2002flush_plug:
2003 blk_finish_plug(&plug);
2004}
2005EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
2006
2007void dm_bufio_release(struct dm_buffer *b)
2008{
2009 struct dm_bufio_client *c = b->c;
2010
2011 /*
2012 * If there were errors on the buffer, and the buffer is not
2013 * to be written, free the buffer. There is no point in caching
2014 * invalid buffer.
2015 */
2016 if ((b->read_error || b->write_error) &&
2017 !test_bit_acquire(B_READING, &b->state) &&
2018 !test_bit(B_WRITING, &b->state) &&
2019 !test_bit(B_DIRTY, &b->state)) {
2020 dm_bufio_lock(c);
2021
2022 /* cache remove can fail if there are other holders */
2023 if (cache_remove(&c->cache, b)) {
2024 __free_buffer_wake(b);
2025 dm_bufio_unlock(c);
2026 return;
2027 }
2028
2029 dm_bufio_unlock(c);
2030 }
2031
2032 cache_put_and_wake(c, b);
2033}
2034EXPORT_SYMBOL_GPL(dm_bufio_release);
2035
2036void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
2037 unsigned int start, unsigned int end)
2038{
2039 struct dm_bufio_client *c = b->c;
2040
2041 BUG_ON(start >= end);
2042 BUG_ON(end > b->c->block_size);
2043
2044 dm_bufio_lock(c);
2045
2046 BUG_ON(test_bit(B_READING, &b->state));
2047
2048 if (!test_and_set_bit(B_DIRTY, &b->state)) {
2049 b->dirty_start = start;
2050 b->dirty_end = end;
2051 cache_mark(&c->cache, b, LIST_DIRTY);
2052 } else {
2053 if (start < b->dirty_start)
2054 b->dirty_start = start;
2055 if (end > b->dirty_end)
2056 b->dirty_end = end;
2057 }
2058
2059 dm_bufio_unlock(c);
2060}
2061EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
2062
2063void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
2064{
2065 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
2066}
2067EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
2068
2069void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
2070{
2071 LIST_HEAD(write_list);
2072
2073 if (WARN_ON_ONCE(dm_bufio_in_request()))
2074 return; /* should never happen */
2075
2076 dm_bufio_lock(c);
2077 __write_dirty_buffers_async(c, 0, &write_list);
2078 dm_bufio_unlock(c);
2079 __flush_write_list(&write_list);
2080}
2081EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
2082
2083/*
2084 * For performance, it is essential that the buffers are written asynchronously
2085 * and simultaneously (so that the block layer can merge the writes) and then
2086 * waited upon.
2087 *
2088 * Finally, we flush hardware disk cache.
2089 */
2090static bool is_writing(struct lru_entry *e, void *context)
2091{
2092 struct dm_buffer *b = le_to_buffer(e);
2093
2094 return test_bit(B_WRITING, &b->state);
2095}
2096
2097int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
2098{
2099 int a, f;
2100 unsigned long nr_buffers;
2101 struct lru_entry *e;
2102 struct lru_iter it;
2103
2104 LIST_HEAD(write_list);
2105
2106 dm_bufio_lock(c);
2107 __write_dirty_buffers_async(c, 0, &write_list);
2108 dm_bufio_unlock(c);
2109 __flush_write_list(&write_list);
2110 dm_bufio_lock(c);
2111
2112 nr_buffers = cache_count(&c->cache, LIST_DIRTY);
2113 lru_iter_begin(&c->cache.lru[LIST_DIRTY], &it);
2114 while ((e = lru_iter_next(&it, is_writing, c))) {
2115 struct dm_buffer *b = le_to_buffer(e);
2116 __cache_inc_buffer(b);
2117
2118 BUG_ON(test_bit(B_READING, &b->state));
2119
2120 if (nr_buffers) {
2121 nr_buffers--;
2122 dm_bufio_unlock(c);
2123 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
2124 dm_bufio_lock(c);
2125 } else {
2126 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
2127 }
2128
2129 if (!test_bit(B_DIRTY, &b->state) && !test_bit(B_WRITING, &b->state))
2130 cache_mark(&c->cache, b, LIST_CLEAN);
2131
2132 cache_put_and_wake(c, b);
2133
2134 cond_resched();
2135 }
2136 lru_iter_end(&it);
2137
2138 wake_up(&c->free_buffer_wait);
2139 dm_bufio_unlock(c);
2140
2141 a = xchg(&c->async_write_error, 0);
2142 f = dm_bufio_issue_flush(c);
2143 if (a)
2144 return a;
2145
2146 return f;
2147}
2148EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
2149
2150/*
2151 * Use dm-io to send an empty barrier to flush the device.
2152 */
2153int dm_bufio_issue_flush(struct dm_bufio_client *c)
2154{
2155 struct dm_io_request io_req = {
2156 .bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
2157 .mem.type = DM_IO_KMEM,
2158 .mem.ptr.addr = NULL,
2159 .client = c->dm_io,
2160 };
2161 struct dm_io_region io_reg = {
2162 .bdev = c->bdev,
2163 .sector = 0,
2164 .count = 0,
2165 };
2166
2167 if (WARN_ON_ONCE(dm_bufio_in_request()))
2168 return -EINVAL;
2169
2170 return dm_io(&io_req, 1, &io_reg, NULL);
2171}
2172EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
2173
2174/*
2175 * Use dm-io to send a discard request to flush the device.
2176 */
2177int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
2178{
2179 struct dm_io_request io_req = {
2180 .bi_opf = REQ_OP_DISCARD | REQ_SYNC,
2181 .mem.type = DM_IO_KMEM,
2182 .mem.ptr.addr = NULL,
2183 .client = c->dm_io,
2184 };
2185 struct dm_io_region io_reg = {
2186 .bdev = c->bdev,
2187 .sector = block_to_sector(c, block),
2188 .count = block_to_sector(c, count),
2189 };
2190
2191 if (WARN_ON_ONCE(dm_bufio_in_request()))
2192 return -EINVAL; /* discards are optional */
2193
2194 return dm_io(&io_req, 1, &io_reg, NULL);
2195}
2196EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
2197
2198static bool forget_buffer(struct dm_bufio_client *c, sector_t block)
2199{
2200 struct dm_buffer *b;
2201
2202 b = cache_get(&c->cache, block);
2203 if (b) {
2204 if (likely(!smp_load_acquire(&b->state))) {
2205 if (cache_remove(&c->cache, b))
2206 __free_buffer_wake(b);
2207 else
2208 cache_put_and_wake(c, b);
2209 } else {
2210 cache_put_and_wake(c, b);
2211 }
2212 }
2213
2214 return b ? true : false;
2215}
2216
2217/*
2218 * Free the given buffer.
2219 *
2220 * This is just a hint, if the buffer is in use or dirty, this function
2221 * does nothing.
2222 */
2223void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
2224{
2225 dm_bufio_lock(c);
2226 forget_buffer(c, block);
2227 dm_bufio_unlock(c);
2228}
2229EXPORT_SYMBOL_GPL(dm_bufio_forget);
2230
2231static enum evict_result idle(struct dm_buffer *b, void *context)
2232{
2233 return b->state ? ER_DONT_EVICT : ER_EVICT;
2234}
2235
2236void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
2237{
2238 dm_bufio_lock(c);
2239 cache_remove_range(&c->cache, block, block + n_blocks, idle, __free_buffer_wake);
2240 dm_bufio_unlock(c);
2241}
2242EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
2243
2244void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned int n)
2245{
2246 c->minimum_buffers = n;
2247}
2248EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
2249
2250unsigned int dm_bufio_get_block_size(struct dm_bufio_client *c)
2251{
2252 return c->block_size;
2253}
2254EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
2255
2256sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
2257{
2258 sector_t s = bdev_nr_sectors(c->bdev);
2259
2260 if (s >= c->start)
2261 s -= c->start;
2262 else
2263 s = 0;
2264 if (likely(c->sectors_per_block_bits >= 0))
2265 s >>= c->sectors_per_block_bits;
2266 else
2267 sector_div(s, c->block_size >> SECTOR_SHIFT);
2268 return s;
2269}
2270EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
2271
2272struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
2273{
2274 return c->dm_io;
2275}
2276EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
2277
2278sector_t dm_bufio_get_block_number(struct dm_buffer *b)
2279{
2280 return b->block;
2281}
2282EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
2283
2284void *dm_bufio_get_block_data(struct dm_buffer *b)
2285{
2286 return b->data;
2287}
2288EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
2289
2290void *dm_bufio_get_aux_data(struct dm_buffer *b)
2291{
2292 return b + 1;
2293}
2294EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
2295
2296struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
2297{
2298 return b->c;
2299}
2300EXPORT_SYMBOL_GPL(dm_bufio_get_client);
2301
2302static enum it_action warn_leak(struct dm_buffer *b, void *context)
2303{
2304 bool *warned = context;
2305
2306 WARN_ON(!(*warned));
2307 *warned = true;
2308 DMERR("leaked buffer %llx, hold count %u, list %d",
2309 (unsigned long long)b->block, atomic_read(&b->hold_count), b->list_mode);
2310#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
2311 stack_trace_print(b->stack_entries, b->stack_len, 1);
2312 /* mark unclaimed to avoid WARN_ON at end of drop_buffers() */
2313 atomic_set(&b->hold_count, 0);
2314#endif
2315 return IT_NEXT;
2316}
2317
2318static void drop_buffers(struct dm_bufio_client *c)
2319{
2320 int i;
2321 struct dm_buffer *b;
2322
2323 if (WARN_ON(dm_bufio_in_request()))
2324 return; /* should never happen */
2325
2326 /*
2327 * An optimization so that the buffers are not written one-by-one.
2328 */
2329 dm_bufio_write_dirty_buffers_async(c);
2330
2331 dm_bufio_lock(c);
2332
2333 while ((b = __get_unclaimed_buffer(c)))
2334 __free_buffer_wake(b);
2335
2336 for (i = 0; i < LIST_SIZE; i++) {
2337 bool warned = false;
2338
2339 cache_iterate(&c->cache, i, warn_leak, &warned);
2340 }
2341
2342#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
2343 while ((b = __get_unclaimed_buffer(c)))
2344 __free_buffer_wake(b);
2345#endif
2346
2347 for (i = 0; i < LIST_SIZE; i++)
2348 WARN_ON(cache_count(&c->cache, i));
2349
2350 dm_bufio_unlock(c);
2351}
2352
2353static unsigned long get_retain_buffers(struct dm_bufio_client *c)
2354{
2355 unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
2356
2357 if (likely(c->sectors_per_block_bits >= 0))
2358 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
2359 else
2360 retain_bytes /= c->block_size;
2361
2362 return retain_bytes;
2363}
2364
2365static void __scan(struct dm_bufio_client *c)
2366{
2367 int l;
2368 struct dm_buffer *b;
2369 unsigned long freed = 0;
2370 unsigned long retain_target = get_retain_buffers(c);
2371 unsigned long count = cache_total(&c->cache);
2372
2373 for (l = 0; l < LIST_SIZE; l++) {
2374 while (true) {
2375 if (count - freed <= retain_target)
2376 atomic_long_set(&c->need_shrink, 0);
2377 if (!atomic_long_read(&c->need_shrink))
2378 break;
2379
2380 b = cache_evict(&c->cache, l,
2381 l == LIST_CLEAN ? is_clean : is_dirty, c);
2382 if (!b)
2383 break;
2384
2385 __make_buffer_clean(b);
2386 __free_buffer_wake(b);
2387
2388 atomic_long_dec(&c->need_shrink);
2389 freed++;
2390 cond_resched();
2391 }
2392 }
2393}
2394
2395static void shrink_work(struct work_struct *w)
2396{
2397 struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
2398
2399 dm_bufio_lock(c);
2400 __scan(c);
2401 dm_bufio_unlock(c);
2402}
2403
2404static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
2405{
2406 struct dm_bufio_client *c;
2407
2408 c = shrink->private_data;
2409 atomic_long_add(sc->nr_to_scan, &c->need_shrink);
2410 queue_work(dm_bufio_wq, &c->shrink_work);
2411
2412 return sc->nr_to_scan;
2413}
2414
2415static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
2416{
2417 struct dm_bufio_client *c = shrink->private_data;
2418 unsigned long count = cache_total(&c->cache);
2419 unsigned long retain_target = get_retain_buffers(c);
2420 unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
2421
2422 if (unlikely(count < retain_target))
2423 count = 0;
2424 else
2425 count -= retain_target;
2426
2427 if (unlikely(count < queued_for_cleanup))
2428 count = 0;
2429 else
2430 count -= queued_for_cleanup;
2431
2432 return count;
2433}
2434
2435/*
2436 * Create the buffering interface
2437 */
2438struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned int block_size,
2439 unsigned int reserved_buffers, unsigned int aux_size,
2440 void (*alloc_callback)(struct dm_buffer *),
2441 void (*write_callback)(struct dm_buffer *),
2442 unsigned int flags)
2443{
2444 int r;
2445 unsigned int num_locks;
2446 struct dm_bufio_client *c;
2447 char slab_name[27];
2448
2449 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
2450 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
2451 r = -EINVAL;
2452 goto bad_client;
2453 }
2454
2455 num_locks = dm_num_hash_locks();
2456 c = kzalloc(sizeof(*c) + (num_locks * sizeof(struct buffer_tree)), GFP_KERNEL);
2457 if (!c) {
2458 r = -ENOMEM;
2459 goto bad_client;
2460 }
2461 cache_init(&c->cache, num_locks, (flags & DM_BUFIO_CLIENT_NO_SLEEP) != 0);
2462
2463 c->bdev = bdev;
2464 c->block_size = block_size;
2465 if (is_power_of_2(block_size))
2466 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
2467 else
2468 c->sectors_per_block_bits = -1;
2469
2470 c->alloc_callback = alloc_callback;
2471 c->write_callback = write_callback;
2472
2473 if (flags & DM_BUFIO_CLIENT_NO_SLEEP) {
2474 c->no_sleep = true;
2475 static_branch_inc(&no_sleep_enabled);
2476 }
2477
2478 mutex_init(&c->lock);
2479 spin_lock_init(&c->spinlock);
2480 INIT_LIST_HEAD(&c->reserved_buffers);
2481 c->need_reserved_buffers = reserved_buffers;
2482
2483 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
2484
2485 init_waitqueue_head(&c->free_buffer_wait);
2486 c->async_write_error = 0;
2487
2488 c->dm_io = dm_io_client_create();
2489 if (IS_ERR(c->dm_io)) {
2490 r = PTR_ERR(c->dm_io);
2491 goto bad_dm_io;
2492 }
2493
2494 if (block_size <= KMALLOC_MAX_SIZE &&
2495 (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
2496 unsigned int align = min(1U << __ffs(block_size), (unsigned int)PAGE_SIZE);
2497
2498 snprintf(slab_name, sizeof(slab_name), "dm_bufio_cache-%u", block_size);
2499 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
2500 SLAB_RECLAIM_ACCOUNT, NULL);
2501 if (!c->slab_cache) {
2502 r = -ENOMEM;
2503 goto bad;
2504 }
2505 }
2506 if (aux_size)
2507 snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer-%u", aux_size);
2508 else
2509 snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer");
2510 c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
2511 0, SLAB_RECLAIM_ACCOUNT, NULL);
2512 if (!c->slab_buffer) {
2513 r = -ENOMEM;
2514 goto bad;
2515 }
2516
2517 while (c->need_reserved_buffers) {
2518 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
2519
2520 if (!b) {
2521 r = -ENOMEM;
2522 goto bad;
2523 }
2524 __free_buffer_wake(b);
2525 }
2526
2527 INIT_WORK(&c->shrink_work, shrink_work);
2528 atomic_long_set(&c->need_shrink, 0);
2529
2530 c->shrinker = shrinker_alloc(0, "dm-bufio:(%u:%u)",
2531 MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev));
2532 if (!c->shrinker) {
2533 r = -ENOMEM;
2534 goto bad;
2535 }
2536
2537 c->shrinker->count_objects = dm_bufio_shrink_count;
2538 c->shrinker->scan_objects = dm_bufio_shrink_scan;
2539 c->shrinker->seeks = 1;
2540 c->shrinker->batch = 0;
2541 c->shrinker->private_data = c;
2542
2543 shrinker_register(c->shrinker);
2544
2545 mutex_lock(&dm_bufio_clients_lock);
2546 dm_bufio_client_count++;
2547 list_add(&c->client_list, &dm_bufio_all_clients);
2548 __cache_size_refresh();
2549 mutex_unlock(&dm_bufio_clients_lock);
2550
2551 return c;
2552
2553bad:
2554 while (!list_empty(&c->reserved_buffers)) {
2555 struct dm_buffer *b = list_to_buffer(c->reserved_buffers.next);
2556
2557 list_del(&b->lru.list);
2558 free_buffer(b);
2559 }
2560 kmem_cache_destroy(c->slab_cache);
2561 kmem_cache_destroy(c->slab_buffer);
2562 dm_io_client_destroy(c->dm_io);
2563bad_dm_io:
2564 mutex_destroy(&c->lock);
2565 if (c->no_sleep)
2566 static_branch_dec(&no_sleep_enabled);
2567 kfree(c);
2568bad_client:
2569 return ERR_PTR(r);
2570}
2571EXPORT_SYMBOL_GPL(dm_bufio_client_create);
2572
2573/*
2574 * Free the buffering interface.
2575 * It is required that there are no references on any buffers.
2576 */
2577void dm_bufio_client_destroy(struct dm_bufio_client *c)
2578{
2579 unsigned int i;
2580
2581 drop_buffers(c);
2582
2583 shrinker_free(c->shrinker);
2584 flush_work(&c->shrink_work);
2585
2586 mutex_lock(&dm_bufio_clients_lock);
2587
2588 list_del(&c->client_list);
2589 dm_bufio_client_count--;
2590 __cache_size_refresh();
2591
2592 mutex_unlock(&dm_bufio_clients_lock);
2593
2594 WARN_ON(c->need_reserved_buffers);
2595
2596 while (!list_empty(&c->reserved_buffers)) {
2597 struct dm_buffer *b = list_to_buffer(c->reserved_buffers.next);
2598
2599 list_del(&b->lru.list);
2600 free_buffer(b);
2601 }
2602
2603 for (i = 0; i < LIST_SIZE; i++)
2604 if (cache_count(&c->cache, i))
2605 DMERR("leaked buffer count %d: %lu", i, cache_count(&c->cache, i));
2606
2607 for (i = 0; i < LIST_SIZE; i++)
2608 WARN_ON(cache_count(&c->cache, i));
2609
2610 cache_destroy(&c->cache);
2611 kmem_cache_destroy(c->slab_cache);
2612 kmem_cache_destroy(c->slab_buffer);
2613 dm_io_client_destroy(c->dm_io);
2614 mutex_destroy(&c->lock);
2615 if (c->no_sleep)
2616 static_branch_dec(&no_sleep_enabled);
2617 kfree(c);
2618}
2619EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
2620
2621void dm_bufio_client_reset(struct dm_bufio_client *c)
2622{
2623 drop_buffers(c);
2624 flush_work(&c->shrink_work);
2625}
2626EXPORT_SYMBOL_GPL(dm_bufio_client_reset);
2627
2628void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
2629{
2630 c->start = start;
2631}
2632EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
2633
2634/*--------------------------------------------------------------*/
2635
2636static unsigned int get_max_age_hz(void)
2637{
2638 unsigned int max_age = READ_ONCE(dm_bufio_max_age);
2639
2640 if (max_age > UINT_MAX / HZ)
2641 max_age = UINT_MAX / HZ;
2642
2643 return max_age * HZ;
2644}
2645
2646static bool older_than(struct dm_buffer *b, unsigned long age_hz)
2647{
2648 return time_after_eq(jiffies, READ_ONCE(b->last_accessed) + age_hz);
2649}
2650
2651struct evict_params {
2652 gfp_t gfp;
2653 unsigned long age_hz;
2654
2655 /*
2656 * This gets updated with the largest last_accessed (ie. most
2657 * recently used) of the evicted buffers. It will not be reinitialised
2658 * by __evict_many(), so you can use it across multiple invocations.
2659 */
2660 unsigned long last_accessed;
2661};
2662
2663/*
2664 * We may not be able to evict this buffer if IO pending or the client
2665 * is still using it.
2666 *
2667 * And if GFP_NOFS is used, we must not do any I/O because we hold
2668 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
2669 * rerouted to different bufio client.
2670 */
2671static enum evict_result select_for_evict(struct dm_buffer *b, void *context)
2672{
2673 struct evict_params *params = context;
2674
2675 if (!(params->gfp & __GFP_FS) ||
2676 (static_branch_unlikely(&no_sleep_enabled) && b->c->no_sleep)) {
2677 if (test_bit_acquire(B_READING, &b->state) ||
2678 test_bit(B_WRITING, &b->state) ||
2679 test_bit(B_DIRTY, &b->state))
2680 return ER_DONT_EVICT;
2681 }
2682
2683 return older_than(b, params->age_hz) ? ER_EVICT : ER_STOP;
2684}
2685
2686static unsigned long __evict_many(struct dm_bufio_client *c,
2687 struct evict_params *params,
2688 int list_mode, unsigned long max_count)
2689{
2690 unsigned long count;
2691 unsigned long last_accessed;
2692 struct dm_buffer *b;
2693
2694 for (count = 0; count < max_count; count++) {
2695 b = cache_evict(&c->cache, list_mode, select_for_evict, params);
2696 if (!b)
2697 break;
2698
2699 last_accessed = READ_ONCE(b->last_accessed);
2700 if (time_after_eq(params->last_accessed, last_accessed))
2701 params->last_accessed = last_accessed;
2702
2703 __make_buffer_clean(b);
2704 __free_buffer_wake(b);
2705
2706 cond_resched();
2707 }
2708
2709 return count;
2710}
2711
2712static void evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
2713{
2714 struct evict_params params = {.gfp = 0, .age_hz = age_hz, .last_accessed = 0};
2715 unsigned long retain = get_retain_buffers(c);
2716 unsigned long count;
2717 LIST_HEAD(write_list);
2718
2719 dm_bufio_lock(c);
2720
2721 __check_watermark(c, &write_list);
2722 if (unlikely(!list_empty(&write_list))) {
2723 dm_bufio_unlock(c);
2724 __flush_write_list(&write_list);
2725 dm_bufio_lock(c);
2726 }
2727
2728 count = cache_total(&c->cache);
2729 if (count > retain)
2730 __evict_many(c, ¶ms, LIST_CLEAN, count - retain);
2731
2732 dm_bufio_unlock(c);
2733}
2734
2735static void cleanup_old_buffers(void)
2736{
2737 unsigned long max_age_hz = get_max_age_hz();
2738 struct dm_bufio_client *c;
2739
2740 mutex_lock(&dm_bufio_clients_lock);
2741
2742 __cache_size_refresh();
2743
2744 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2745 evict_old_buffers(c, max_age_hz);
2746
2747 mutex_unlock(&dm_bufio_clients_lock);
2748}
2749
2750static void work_fn(struct work_struct *w)
2751{
2752 cleanup_old_buffers();
2753
2754 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2755 DM_BUFIO_WORK_TIMER_SECS * HZ);
2756}
2757
2758/*--------------------------------------------------------------*/
2759
2760/*
2761 * Global cleanup tries to evict the oldest buffers from across _all_
2762 * the clients. It does this by repeatedly evicting a few buffers from
2763 * the client that holds the oldest buffer. It's approximate, but hopefully
2764 * good enough.
2765 */
2766static struct dm_bufio_client *__pop_client(void)
2767{
2768 struct list_head *h;
2769
2770 if (list_empty(&dm_bufio_all_clients))
2771 return NULL;
2772
2773 h = dm_bufio_all_clients.next;
2774 list_del(h);
2775 return container_of(h, struct dm_bufio_client, client_list);
2776}
2777
2778/*
2779 * Inserts the client in the global client list based on its
2780 * 'oldest_buffer' field.
2781 */
2782static void __insert_client(struct dm_bufio_client *new_client)
2783{
2784 struct dm_bufio_client *c;
2785 struct list_head *h = dm_bufio_all_clients.next;
2786
2787 while (h != &dm_bufio_all_clients) {
2788 c = container_of(h, struct dm_bufio_client, client_list);
2789 if (time_after_eq(c->oldest_buffer, new_client->oldest_buffer))
2790 break;
2791 h = h->next;
2792 }
2793
2794 list_add_tail(&new_client->client_list, h);
2795}
2796
2797static unsigned long __evict_a_few(unsigned long nr_buffers)
2798{
2799 unsigned long count;
2800 struct dm_bufio_client *c;
2801 struct evict_params params = {
2802 .gfp = GFP_KERNEL,
2803 .age_hz = 0,
2804 /* set to jiffies in case there are no buffers in this client */
2805 .last_accessed = jiffies
2806 };
2807
2808 c = __pop_client();
2809 if (!c)
2810 return 0;
2811
2812 dm_bufio_lock(c);
2813 count = __evict_many(c, ¶ms, LIST_CLEAN, nr_buffers);
2814 dm_bufio_unlock(c);
2815
2816 if (count)
2817 c->oldest_buffer = params.last_accessed;
2818 __insert_client(c);
2819
2820 return count;
2821}
2822
2823static void check_watermarks(void)
2824{
2825 LIST_HEAD(write_list);
2826 struct dm_bufio_client *c;
2827
2828 mutex_lock(&dm_bufio_clients_lock);
2829 list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
2830 dm_bufio_lock(c);
2831 __check_watermark(c, &write_list);
2832 dm_bufio_unlock(c);
2833 }
2834 mutex_unlock(&dm_bufio_clients_lock);
2835
2836 __flush_write_list(&write_list);
2837}
2838
2839static void evict_old(void)
2840{
2841 unsigned long threshold = dm_bufio_cache_size -
2842 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
2843
2844 mutex_lock(&dm_bufio_clients_lock);
2845 while (dm_bufio_current_allocated > threshold) {
2846 if (!__evict_a_few(64))
2847 break;
2848 cond_resched();
2849 }
2850 mutex_unlock(&dm_bufio_clients_lock);
2851}
2852
2853static void do_global_cleanup(struct work_struct *w)
2854{
2855 check_watermarks();
2856 evict_old();
2857}
2858
2859/*
2860 *--------------------------------------------------------------
2861 * Module setup
2862 *--------------------------------------------------------------
2863 */
2864
2865/*
2866 * This is called only once for the whole dm_bufio module.
2867 * It initializes memory limit.
2868 */
2869static int __init dm_bufio_init(void)
2870{
2871 __u64 mem;
2872
2873 dm_bufio_allocated_kmem_cache = 0;
2874 dm_bufio_allocated_get_free_pages = 0;
2875 dm_bufio_allocated_vmalloc = 0;
2876 dm_bufio_current_allocated = 0;
2877
2878 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2879 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2880
2881 if (mem > ULONG_MAX)
2882 mem = ULONG_MAX;
2883
2884#ifdef CONFIG_MMU
2885 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2886 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2887#endif
2888
2889 dm_bufio_default_cache_size = mem;
2890
2891 mutex_lock(&dm_bufio_clients_lock);
2892 __cache_size_refresh();
2893 mutex_unlock(&dm_bufio_clients_lock);
2894
2895 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2896 if (!dm_bufio_wq)
2897 return -ENOMEM;
2898
2899 INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2900 INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2901 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2902 DM_BUFIO_WORK_TIMER_SECS * HZ);
2903
2904 return 0;
2905}
2906
2907/*
2908 * This is called once when unloading the dm_bufio module.
2909 */
2910static void __exit dm_bufio_exit(void)
2911{
2912 int bug = 0;
2913
2914 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2915 destroy_workqueue(dm_bufio_wq);
2916
2917 if (dm_bufio_client_count) {
2918 DMCRIT("%s: dm_bufio_client_count leaked: %d",
2919 __func__, dm_bufio_client_count);
2920 bug = 1;
2921 }
2922
2923 if (dm_bufio_current_allocated) {
2924 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2925 __func__, dm_bufio_current_allocated);
2926 bug = 1;
2927 }
2928
2929 if (dm_bufio_allocated_get_free_pages) {
2930 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2931 __func__, dm_bufio_allocated_get_free_pages);
2932 bug = 1;
2933 }
2934
2935 if (dm_bufio_allocated_vmalloc) {
2936 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2937 __func__, dm_bufio_allocated_vmalloc);
2938 bug = 1;
2939 }
2940
2941 WARN_ON(bug); /* leaks are not worth crashing the system */
2942}
2943
2944module_init(dm_bufio_init)
2945module_exit(dm_bufio_exit)
2946
2947module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, 0644);
2948MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2949
2950module_param_named(max_age_seconds, dm_bufio_max_age, uint, 0644);
2951MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2952
2953module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, 0644);
2954MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2955
2956module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, 0644);
2957MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2958
2959module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, 0444);
2960MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2961
2962module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, 0444);
2963MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2964
2965module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, 0444);
2966MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2967
2968module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, 0444);
2969MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2970
2971MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2972MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2973MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2009-2011 Red Hat, Inc.
3 *
4 * Author: Mikulas Patocka <mpatocka@redhat.com>
5 *
6 * This file is released under the GPL.
7 */
8
9#include <linux/dm-bufio.h>
10
11#include <linux/device-mapper.h>
12#include <linux/dm-io.h>
13#include <linux/slab.h>
14#include <linux/sched/mm.h>
15#include <linux/jiffies.h>
16#include <linux/vmalloc.h>
17#include <linux/shrinker.h>
18#include <linux/module.h>
19#include <linux/rbtree.h>
20#include <linux/stacktrace.h>
21
22#define DM_MSG_PREFIX "bufio"
23
24/*
25 * Memory management policy:
26 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
27 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
28 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
29 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
30 * dirty buffers.
31 */
32#define DM_BUFIO_MIN_BUFFERS 8
33
34#define DM_BUFIO_MEMORY_PERCENT 2
35#define DM_BUFIO_VMALLOC_PERCENT 25
36#define DM_BUFIO_WRITEBACK_RATIO 3
37#define DM_BUFIO_LOW_WATERMARK_RATIO 16
38
39/*
40 * Check buffer ages in this interval (seconds)
41 */
42#define DM_BUFIO_WORK_TIMER_SECS 30
43
44/*
45 * Free buffers when they are older than this (seconds)
46 */
47#define DM_BUFIO_DEFAULT_AGE_SECS 300
48
49/*
50 * The nr of bytes of cached data to keep around.
51 */
52#define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
53
54/*
55 * Align buffer writes to this boundary.
56 * Tests show that SSDs have the highest IOPS when using 4k writes.
57 */
58#define DM_BUFIO_WRITE_ALIGN 4096
59
60/*
61 * dm_buffer->list_mode
62 */
63#define LIST_CLEAN 0
64#define LIST_DIRTY 1
65#define LIST_SIZE 2
66
67/*
68 * Linking of buffers:
69 * All buffers are linked to buffer_tree with their node field.
70 *
71 * Clean buffers that are not being written (B_WRITING not set)
72 * are linked to lru[LIST_CLEAN] with their lru_list field.
73 *
74 * Dirty and clean buffers that are being written are linked to
75 * lru[LIST_DIRTY] with their lru_list field. When the write
76 * finishes, the buffer cannot be relinked immediately (because we
77 * are in an interrupt context and relinking requires process
78 * context), so some clean-not-writing buffers can be held on
79 * dirty_lru too. They are later added to lru in the process
80 * context.
81 */
82struct dm_bufio_client {
83 struct mutex lock;
84
85 struct list_head lru[LIST_SIZE];
86 unsigned long n_buffers[LIST_SIZE];
87
88 struct block_device *bdev;
89 unsigned block_size;
90 s8 sectors_per_block_bits;
91 void (*alloc_callback)(struct dm_buffer *);
92 void (*write_callback)(struct dm_buffer *);
93
94 struct kmem_cache *slab_buffer;
95 struct kmem_cache *slab_cache;
96 struct dm_io_client *dm_io;
97
98 struct list_head reserved_buffers;
99 unsigned need_reserved_buffers;
100
101 unsigned minimum_buffers;
102
103 struct rb_root buffer_tree;
104 wait_queue_head_t free_buffer_wait;
105
106 sector_t start;
107
108 int async_write_error;
109
110 struct list_head client_list;
111 struct shrinker shrinker;
112};
113
114/*
115 * Buffer state bits.
116 */
117#define B_READING 0
118#define B_WRITING 1
119#define B_DIRTY 2
120
121/*
122 * Describes how the block was allocated:
123 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
124 * See the comment at alloc_buffer_data.
125 */
126enum data_mode {
127 DATA_MODE_SLAB = 0,
128 DATA_MODE_GET_FREE_PAGES = 1,
129 DATA_MODE_VMALLOC = 2,
130 DATA_MODE_LIMIT = 3
131};
132
133struct dm_buffer {
134 struct rb_node node;
135 struct list_head lru_list;
136 struct list_head global_list;
137 sector_t block;
138 void *data;
139 unsigned char data_mode; /* DATA_MODE_* */
140 unsigned char list_mode; /* LIST_* */
141 blk_status_t read_error;
142 blk_status_t write_error;
143 unsigned accessed;
144 unsigned hold_count;
145 unsigned long state;
146 unsigned long last_accessed;
147 unsigned dirty_start;
148 unsigned dirty_end;
149 unsigned write_start;
150 unsigned write_end;
151 struct dm_bufio_client *c;
152 struct list_head write_list;
153 void (*end_io)(struct dm_buffer *, blk_status_t);
154#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
155#define MAX_STACK 10
156 unsigned int stack_len;
157 unsigned long stack_entries[MAX_STACK];
158#endif
159};
160
161/*----------------------------------------------------------------*/
162
163#define dm_bufio_in_request() (!!current->bio_list)
164
165static void dm_bufio_lock(struct dm_bufio_client *c)
166{
167 mutex_lock_nested(&c->lock, dm_bufio_in_request());
168}
169
170static int dm_bufio_trylock(struct dm_bufio_client *c)
171{
172 return mutex_trylock(&c->lock);
173}
174
175static void dm_bufio_unlock(struct dm_bufio_client *c)
176{
177 mutex_unlock(&c->lock);
178}
179
180/*----------------------------------------------------------------*/
181
182/*
183 * Default cache size: available memory divided by the ratio.
184 */
185static unsigned long dm_bufio_default_cache_size;
186
187/*
188 * Total cache size set by the user.
189 */
190static unsigned long dm_bufio_cache_size;
191
192/*
193 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
194 * at any time. If it disagrees, the user has changed cache size.
195 */
196static unsigned long dm_bufio_cache_size_latch;
197
198static DEFINE_SPINLOCK(global_spinlock);
199
200static LIST_HEAD(global_queue);
201
202static unsigned long global_num = 0;
203
204/*
205 * Buffers are freed after this timeout
206 */
207static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
208static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
209
210static unsigned long dm_bufio_peak_allocated;
211static unsigned long dm_bufio_allocated_kmem_cache;
212static unsigned long dm_bufio_allocated_get_free_pages;
213static unsigned long dm_bufio_allocated_vmalloc;
214static unsigned long dm_bufio_current_allocated;
215
216/*----------------------------------------------------------------*/
217
218/*
219 * The current number of clients.
220 */
221static int dm_bufio_client_count;
222
223/*
224 * The list of all clients.
225 */
226static LIST_HEAD(dm_bufio_all_clients);
227
228/*
229 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
230 */
231static DEFINE_MUTEX(dm_bufio_clients_lock);
232
233static struct workqueue_struct *dm_bufio_wq;
234static struct delayed_work dm_bufio_cleanup_old_work;
235static struct work_struct dm_bufio_replacement_work;
236
237
238#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
239static void buffer_record_stack(struct dm_buffer *b)
240{
241 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
242}
243#endif
244
245/*----------------------------------------------------------------
246 * A red/black tree acts as an index for all the buffers.
247 *--------------------------------------------------------------*/
248static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
249{
250 struct rb_node *n = c->buffer_tree.rb_node;
251 struct dm_buffer *b;
252
253 while (n) {
254 b = container_of(n, struct dm_buffer, node);
255
256 if (b->block == block)
257 return b;
258
259 n = (b->block < block) ? n->rb_left : n->rb_right;
260 }
261
262 return NULL;
263}
264
265static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
266{
267 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
268 struct dm_buffer *found;
269
270 while (*new) {
271 found = container_of(*new, struct dm_buffer, node);
272
273 if (found->block == b->block) {
274 BUG_ON(found != b);
275 return;
276 }
277
278 parent = *new;
279 new = (found->block < b->block) ?
280 &((*new)->rb_left) : &((*new)->rb_right);
281 }
282
283 rb_link_node(&b->node, parent, new);
284 rb_insert_color(&b->node, &c->buffer_tree);
285}
286
287static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
288{
289 rb_erase(&b->node, &c->buffer_tree);
290}
291
292/*----------------------------------------------------------------*/
293
294static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
295{
296 unsigned char data_mode;
297 long diff;
298
299 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
300 &dm_bufio_allocated_kmem_cache,
301 &dm_bufio_allocated_get_free_pages,
302 &dm_bufio_allocated_vmalloc,
303 };
304
305 data_mode = b->data_mode;
306 diff = (long)b->c->block_size;
307 if (unlink)
308 diff = -diff;
309
310 spin_lock(&global_spinlock);
311
312 *class_ptr[data_mode] += diff;
313
314 dm_bufio_current_allocated += diff;
315
316 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
317 dm_bufio_peak_allocated = dm_bufio_current_allocated;
318
319 b->accessed = 1;
320
321 if (!unlink) {
322 list_add(&b->global_list, &global_queue);
323 global_num++;
324 if (dm_bufio_current_allocated > dm_bufio_cache_size)
325 queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
326 } else {
327 list_del(&b->global_list);
328 global_num--;
329 }
330
331 spin_unlock(&global_spinlock);
332}
333
334/*
335 * Change the number of clients and recalculate per-client limit.
336 */
337static void __cache_size_refresh(void)
338{
339 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
340 BUG_ON(dm_bufio_client_count < 0);
341
342 dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
343
344 /*
345 * Use default if set to 0 and report the actual cache size used.
346 */
347 if (!dm_bufio_cache_size_latch) {
348 (void)cmpxchg(&dm_bufio_cache_size, 0,
349 dm_bufio_default_cache_size);
350 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
351 }
352}
353
354/*
355 * Allocating buffer data.
356 *
357 * Small buffers are allocated with kmem_cache, to use space optimally.
358 *
359 * For large buffers, we choose between get_free_pages and vmalloc.
360 * Each has advantages and disadvantages.
361 *
362 * __get_free_pages can randomly fail if the memory is fragmented.
363 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
364 * as low as 128M) so using it for caching is not appropriate.
365 *
366 * If the allocation may fail we use __get_free_pages. Memory fragmentation
367 * won't have a fatal effect here, but it just causes flushes of some other
368 * buffers and more I/O will be performed. Don't use __get_free_pages if it
369 * always fails (i.e. order >= MAX_ORDER).
370 *
371 * If the allocation shouldn't fail we use __vmalloc. This is only for the
372 * initial reserve allocation, so there's no risk of wasting all vmalloc
373 * space.
374 */
375static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
376 unsigned char *data_mode)
377{
378 if (unlikely(c->slab_cache != NULL)) {
379 *data_mode = DATA_MODE_SLAB;
380 return kmem_cache_alloc(c->slab_cache, gfp_mask);
381 }
382
383 if (c->block_size <= KMALLOC_MAX_SIZE &&
384 gfp_mask & __GFP_NORETRY) {
385 *data_mode = DATA_MODE_GET_FREE_PAGES;
386 return (void *)__get_free_pages(gfp_mask,
387 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
388 }
389
390 *data_mode = DATA_MODE_VMALLOC;
391
392 /*
393 * __vmalloc allocates the data pages and auxiliary structures with
394 * gfp_flags that were specified, but pagetables are always allocated
395 * with GFP_KERNEL, no matter what was specified as gfp_mask.
396 *
397 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
398 * all allocations done by this process (including pagetables) are done
399 * as if GFP_NOIO was specified.
400 */
401 if (gfp_mask & __GFP_NORETRY) {
402 unsigned noio_flag = memalloc_noio_save();
403 void *ptr = __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
404
405 memalloc_noio_restore(noio_flag);
406 return ptr;
407 }
408
409 return __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
410}
411
412/*
413 * Free buffer's data.
414 */
415static void free_buffer_data(struct dm_bufio_client *c,
416 void *data, unsigned char data_mode)
417{
418 switch (data_mode) {
419 case DATA_MODE_SLAB:
420 kmem_cache_free(c->slab_cache, data);
421 break;
422
423 case DATA_MODE_GET_FREE_PAGES:
424 free_pages((unsigned long)data,
425 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
426 break;
427
428 case DATA_MODE_VMALLOC:
429 vfree(data);
430 break;
431
432 default:
433 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
434 data_mode);
435 BUG();
436 }
437}
438
439/*
440 * Allocate buffer and its data.
441 */
442static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
443{
444 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
445
446 if (!b)
447 return NULL;
448
449 b->c = c;
450
451 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
452 if (!b->data) {
453 kmem_cache_free(c->slab_buffer, b);
454 return NULL;
455 }
456
457#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
458 b->stack_len = 0;
459#endif
460 return b;
461}
462
463/*
464 * Free buffer and its data.
465 */
466static void free_buffer(struct dm_buffer *b)
467{
468 struct dm_bufio_client *c = b->c;
469
470 free_buffer_data(c, b->data, b->data_mode);
471 kmem_cache_free(c->slab_buffer, b);
472}
473
474/*
475 * Link buffer to the buffer tree and clean or dirty queue.
476 */
477static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
478{
479 struct dm_bufio_client *c = b->c;
480
481 c->n_buffers[dirty]++;
482 b->block = block;
483 b->list_mode = dirty;
484 list_add(&b->lru_list, &c->lru[dirty]);
485 __insert(b->c, b);
486 b->last_accessed = jiffies;
487
488 adjust_total_allocated(b, false);
489}
490
491/*
492 * Unlink buffer from the buffer tree and dirty or clean queue.
493 */
494static void __unlink_buffer(struct dm_buffer *b)
495{
496 struct dm_bufio_client *c = b->c;
497
498 BUG_ON(!c->n_buffers[b->list_mode]);
499
500 c->n_buffers[b->list_mode]--;
501 __remove(b->c, b);
502 list_del(&b->lru_list);
503
504 adjust_total_allocated(b, true);
505}
506
507/*
508 * Place the buffer to the head of dirty or clean LRU queue.
509 */
510static void __relink_lru(struct dm_buffer *b, int dirty)
511{
512 struct dm_bufio_client *c = b->c;
513
514 b->accessed = 1;
515
516 BUG_ON(!c->n_buffers[b->list_mode]);
517
518 c->n_buffers[b->list_mode]--;
519 c->n_buffers[dirty]++;
520 b->list_mode = dirty;
521 list_move(&b->lru_list, &c->lru[dirty]);
522 b->last_accessed = jiffies;
523}
524
525/*----------------------------------------------------------------
526 * Submit I/O on the buffer.
527 *
528 * Bio interface is faster but it has some problems:
529 * the vector list is limited (increasing this limit increases
530 * memory-consumption per buffer, so it is not viable);
531 *
532 * the memory must be direct-mapped, not vmalloced;
533 *
534 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
535 * it is not vmalloced, try using the bio interface.
536 *
537 * If the buffer is big, if it is vmalloced or if the underlying device
538 * rejects the bio because it is too large, use dm-io layer to do the I/O.
539 * The dm-io layer splits the I/O into multiple requests, avoiding the above
540 * shortcomings.
541 *--------------------------------------------------------------*/
542
543/*
544 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
545 * that the request was handled directly with bio interface.
546 */
547static void dmio_complete(unsigned long error, void *context)
548{
549 struct dm_buffer *b = context;
550
551 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
552}
553
554static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
555 unsigned n_sectors, unsigned offset)
556{
557 int r;
558 struct dm_io_request io_req = {
559 .bi_op = rw,
560 .bi_op_flags = 0,
561 .notify.fn = dmio_complete,
562 .notify.context = b,
563 .client = b->c->dm_io,
564 };
565 struct dm_io_region region = {
566 .bdev = b->c->bdev,
567 .sector = sector,
568 .count = n_sectors,
569 };
570
571 if (b->data_mode != DATA_MODE_VMALLOC) {
572 io_req.mem.type = DM_IO_KMEM;
573 io_req.mem.ptr.addr = (char *)b->data + offset;
574 } else {
575 io_req.mem.type = DM_IO_VMA;
576 io_req.mem.ptr.vma = (char *)b->data + offset;
577 }
578
579 r = dm_io(&io_req, 1, ®ion, NULL);
580 if (unlikely(r))
581 b->end_io(b, errno_to_blk_status(r));
582}
583
584static void bio_complete(struct bio *bio)
585{
586 struct dm_buffer *b = bio->bi_private;
587 blk_status_t status = bio->bi_status;
588 bio_put(bio);
589 b->end_io(b, status);
590}
591
592static void use_bio(struct dm_buffer *b, int rw, sector_t sector,
593 unsigned n_sectors, unsigned offset)
594{
595 struct bio *bio;
596 char *ptr;
597 unsigned vec_size, len;
598
599 vec_size = b->c->block_size >> PAGE_SHIFT;
600 if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
601 vec_size += 2;
602
603 bio = bio_kmalloc(GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN, vec_size);
604 if (!bio) {
605dmio:
606 use_dmio(b, rw, sector, n_sectors, offset);
607 return;
608 }
609
610 bio->bi_iter.bi_sector = sector;
611 bio_set_dev(bio, b->c->bdev);
612 bio_set_op_attrs(bio, rw, 0);
613 bio->bi_end_io = bio_complete;
614 bio->bi_private = b;
615
616 ptr = (char *)b->data + offset;
617 len = n_sectors << SECTOR_SHIFT;
618
619 do {
620 unsigned this_step = min((unsigned)(PAGE_SIZE - offset_in_page(ptr)), len);
621 if (!bio_add_page(bio, virt_to_page(ptr), this_step,
622 offset_in_page(ptr))) {
623 bio_put(bio);
624 goto dmio;
625 }
626
627 len -= this_step;
628 ptr += this_step;
629 } while (len > 0);
630
631 submit_bio(bio);
632}
633
634static void submit_io(struct dm_buffer *b, int rw, void (*end_io)(struct dm_buffer *, blk_status_t))
635{
636 unsigned n_sectors;
637 sector_t sector;
638 unsigned offset, end;
639
640 b->end_io = end_io;
641
642 if (likely(b->c->sectors_per_block_bits >= 0))
643 sector = b->block << b->c->sectors_per_block_bits;
644 else
645 sector = b->block * (b->c->block_size >> SECTOR_SHIFT);
646 sector += b->c->start;
647
648 if (rw != REQ_OP_WRITE) {
649 n_sectors = b->c->block_size >> SECTOR_SHIFT;
650 offset = 0;
651 } else {
652 if (b->c->write_callback)
653 b->c->write_callback(b);
654 offset = b->write_start;
655 end = b->write_end;
656 offset &= -DM_BUFIO_WRITE_ALIGN;
657 end += DM_BUFIO_WRITE_ALIGN - 1;
658 end &= -DM_BUFIO_WRITE_ALIGN;
659 if (unlikely(end > b->c->block_size))
660 end = b->c->block_size;
661
662 sector += offset >> SECTOR_SHIFT;
663 n_sectors = (end - offset) >> SECTOR_SHIFT;
664 }
665
666 if (b->data_mode != DATA_MODE_VMALLOC)
667 use_bio(b, rw, sector, n_sectors, offset);
668 else
669 use_dmio(b, rw, sector, n_sectors, offset);
670}
671
672/*----------------------------------------------------------------
673 * Writing dirty buffers
674 *--------------------------------------------------------------*/
675
676/*
677 * The endio routine for write.
678 *
679 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
680 * it.
681 */
682static void write_endio(struct dm_buffer *b, blk_status_t status)
683{
684 b->write_error = status;
685 if (unlikely(status)) {
686 struct dm_bufio_client *c = b->c;
687
688 (void)cmpxchg(&c->async_write_error, 0,
689 blk_status_to_errno(status));
690 }
691
692 BUG_ON(!test_bit(B_WRITING, &b->state));
693
694 smp_mb__before_atomic();
695 clear_bit(B_WRITING, &b->state);
696 smp_mb__after_atomic();
697
698 wake_up_bit(&b->state, B_WRITING);
699}
700
701/*
702 * Initiate a write on a dirty buffer, but don't wait for it.
703 *
704 * - If the buffer is not dirty, exit.
705 * - If there some previous write going on, wait for it to finish (we can't
706 * have two writes on the same buffer simultaneously).
707 * - Submit our write and don't wait on it. We set B_WRITING indicating
708 * that there is a write in progress.
709 */
710static void __write_dirty_buffer(struct dm_buffer *b,
711 struct list_head *write_list)
712{
713 if (!test_bit(B_DIRTY, &b->state))
714 return;
715
716 clear_bit(B_DIRTY, &b->state);
717 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
718
719 b->write_start = b->dirty_start;
720 b->write_end = b->dirty_end;
721
722 if (!write_list)
723 submit_io(b, REQ_OP_WRITE, write_endio);
724 else
725 list_add_tail(&b->write_list, write_list);
726}
727
728static void __flush_write_list(struct list_head *write_list)
729{
730 struct blk_plug plug;
731 blk_start_plug(&plug);
732 while (!list_empty(write_list)) {
733 struct dm_buffer *b =
734 list_entry(write_list->next, struct dm_buffer, write_list);
735 list_del(&b->write_list);
736 submit_io(b, REQ_OP_WRITE, write_endio);
737 cond_resched();
738 }
739 blk_finish_plug(&plug);
740}
741
742/*
743 * Wait until any activity on the buffer finishes. Possibly write the
744 * buffer if it is dirty. When this function finishes, there is no I/O
745 * running on the buffer and the buffer is not dirty.
746 */
747static void __make_buffer_clean(struct dm_buffer *b)
748{
749 BUG_ON(b->hold_count);
750
751 if (!b->state) /* fast case */
752 return;
753
754 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
755 __write_dirty_buffer(b, NULL);
756 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
757}
758
759/*
760 * Find some buffer that is not held by anybody, clean it, unlink it and
761 * return it.
762 */
763static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
764{
765 struct dm_buffer *b;
766
767 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
768 BUG_ON(test_bit(B_WRITING, &b->state));
769 BUG_ON(test_bit(B_DIRTY, &b->state));
770
771 if (!b->hold_count) {
772 __make_buffer_clean(b);
773 __unlink_buffer(b);
774 return b;
775 }
776 cond_resched();
777 }
778
779 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
780 BUG_ON(test_bit(B_READING, &b->state));
781
782 if (!b->hold_count) {
783 __make_buffer_clean(b);
784 __unlink_buffer(b);
785 return b;
786 }
787 cond_resched();
788 }
789
790 return NULL;
791}
792
793/*
794 * Wait until some other threads free some buffer or release hold count on
795 * some buffer.
796 *
797 * This function is entered with c->lock held, drops it and regains it
798 * before exiting.
799 */
800static void __wait_for_free_buffer(struct dm_bufio_client *c)
801{
802 DECLARE_WAITQUEUE(wait, current);
803
804 add_wait_queue(&c->free_buffer_wait, &wait);
805 set_current_state(TASK_UNINTERRUPTIBLE);
806 dm_bufio_unlock(c);
807
808 io_schedule();
809
810 remove_wait_queue(&c->free_buffer_wait, &wait);
811
812 dm_bufio_lock(c);
813}
814
815enum new_flag {
816 NF_FRESH = 0,
817 NF_READ = 1,
818 NF_GET = 2,
819 NF_PREFETCH = 3
820};
821
822/*
823 * Allocate a new buffer. If the allocation is not possible, wait until
824 * some other thread frees a buffer.
825 *
826 * May drop the lock and regain it.
827 */
828static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
829{
830 struct dm_buffer *b;
831 bool tried_noio_alloc = false;
832
833 /*
834 * dm-bufio is resistant to allocation failures (it just keeps
835 * one buffer reserved in cases all the allocations fail).
836 * So set flags to not try too hard:
837 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
838 * mutex and wait ourselves.
839 * __GFP_NORETRY: don't retry and rather return failure
840 * __GFP_NOMEMALLOC: don't use emergency reserves
841 * __GFP_NOWARN: don't print a warning in case of failure
842 *
843 * For debugging, if we set the cache size to 1, no new buffers will
844 * be allocated.
845 */
846 while (1) {
847 if (dm_bufio_cache_size_latch != 1) {
848 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
849 if (b)
850 return b;
851 }
852
853 if (nf == NF_PREFETCH)
854 return NULL;
855
856 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
857 dm_bufio_unlock(c);
858 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
859 dm_bufio_lock(c);
860 if (b)
861 return b;
862 tried_noio_alloc = true;
863 }
864
865 if (!list_empty(&c->reserved_buffers)) {
866 b = list_entry(c->reserved_buffers.next,
867 struct dm_buffer, lru_list);
868 list_del(&b->lru_list);
869 c->need_reserved_buffers++;
870
871 return b;
872 }
873
874 b = __get_unclaimed_buffer(c);
875 if (b)
876 return b;
877
878 __wait_for_free_buffer(c);
879 }
880}
881
882static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
883{
884 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
885
886 if (!b)
887 return NULL;
888
889 if (c->alloc_callback)
890 c->alloc_callback(b);
891
892 return b;
893}
894
895/*
896 * Free a buffer and wake other threads waiting for free buffers.
897 */
898static void __free_buffer_wake(struct dm_buffer *b)
899{
900 struct dm_bufio_client *c = b->c;
901
902 if (!c->need_reserved_buffers)
903 free_buffer(b);
904 else {
905 list_add(&b->lru_list, &c->reserved_buffers);
906 c->need_reserved_buffers--;
907 }
908
909 wake_up(&c->free_buffer_wait);
910}
911
912static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
913 struct list_head *write_list)
914{
915 struct dm_buffer *b, *tmp;
916
917 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
918 BUG_ON(test_bit(B_READING, &b->state));
919
920 if (!test_bit(B_DIRTY, &b->state) &&
921 !test_bit(B_WRITING, &b->state)) {
922 __relink_lru(b, LIST_CLEAN);
923 continue;
924 }
925
926 if (no_wait && test_bit(B_WRITING, &b->state))
927 return;
928
929 __write_dirty_buffer(b, write_list);
930 cond_resched();
931 }
932}
933
934/*
935 * Check if we're over watermark.
936 * If we are over threshold_buffers, start freeing buffers.
937 * If we're over "limit_buffers", block until we get under the limit.
938 */
939static void __check_watermark(struct dm_bufio_client *c,
940 struct list_head *write_list)
941{
942 if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
943 __write_dirty_buffers_async(c, 1, write_list);
944}
945
946/*----------------------------------------------------------------
947 * Getting a buffer
948 *--------------------------------------------------------------*/
949
950static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
951 enum new_flag nf, int *need_submit,
952 struct list_head *write_list)
953{
954 struct dm_buffer *b, *new_b = NULL;
955
956 *need_submit = 0;
957
958 b = __find(c, block);
959 if (b)
960 goto found_buffer;
961
962 if (nf == NF_GET)
963 return NULL;
964
965 new_b = __alloc_buffer_wait(c, nf);
966 if (!new_b)
967 return NULL;
968
969 /*
970 * We've had a period where the mutex was unlocked, so need to
971 * recheck the buffer tree.
972 */
973 b = __find(c, block);
974 if (b) {
975 __free_buffer_wake(new_b);
976 goto found_buffer;
977 }
978
979 __check_watermark(c, write_list);
980
981 b = new_b;
982 b->hold_count = 1;
983 b->read_error = 0;
984 b->write_error = 0;
985 __link_buffer(b, block, LIST_CLEAN);
986
987 if (nf == NF_FRESH) {
988 b->state = 0;
989 return b;
990 }
991
992 b->state = 1 << B_READING;
993 *need_submit = 1;
994
995 return b;
996
997found_buffer:
998 if (nf == NF_PREFETCH)
999 return NULL;
1000 /*
1001 * Note: it is essential that we don't wait for the buffer to be
1002 * read if dm_bufio_get function is used. Both dm_bufio_get and
1003 * dm_bufio_prefetch can be used in the driver request routine.
1004 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1005 * the same buffer, it would deadlock if we waited.
1006 */
1007 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1008 return NULL;
1009
1010 b->hold_count++;
1011 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1012 test_bit(B_WRITING, &b->state));
1013 return b;
1014}
1015
1016/*
1017 * The endio routine for reading: set the error, clear the bit and wake up
1018 * anyone waiting on the buffer.
1019 */
1020static void read_endio(struct dm_buffer *b, blk_status_t status)
1021{
1022 b->read_error = status;
1023
1024 BUG_ON(!test_bit(B_READING, &b->state));
1025
1026 smp_mb__before_atomic();
1027 clear_bit(B_READING, &b->state);
1028 smp_mb__after_atomic();
1029
1030 wake_up_bit(&b->state, B_READING);
1031}
1032
1033/*
1034 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1035 * functions is similar except that dm_bufio_new doesn't read the
1036 * buffer from the disk (assuming that the caller overwrites all the data
1037 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1038 */
1039static void *new_read(struct dm_bufio_client *c, sector_t block,
1040 enum new_flag nf, struct dm_buffer **bp)
1041{
1042 int need_submit;
1043 struct dm_buffer *b;
1044
1045 LIST_HEAD(write_list);
1046
1047 dm_bufio_lock(c);
1048 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1049#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1050 if (b && b->hold_count == 1)
1051 buffer_record_stack(b);
1052#endif
1053 dm_bufio_unlock(c);
1054
1055 __flush_write_list(&write_list);
1056
1057 if (!b)
1058 return NULL;
1059
1060 if (need_submit)
1061 submit_io(b, REQ_OP_READ, read_endio);
1062
1063 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1064
1065 if (b->read_error) {
1066 int error = blk_status_to_errno(b->read_error);
1067
1068 dm_bufio_release(b);
1069
1070 return ERR_PTR(error);
1071 }
1072
1073 *bp = b;
1074
1075 return b->data;
1076}
1077
1078void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1079 struct dm_buffer **bp)
1080{
1081 return new_read(c, block, NF_GET, bp);
1082}
1083EXPORT_SYMBOL_GPL(dm_bufio_get);
1084
1085void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1086 struct dm_buffer **bp)
1087{
1088 BUG_ON(dm_bufio_in_request());
1089
1090 return new_read(c, block, NF_READ, bp);
1091}
1092EXPORT_SYMBOL_GPL(dm_bufio_read);
1093
1094void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1095 struct dm_buffer **bp)
1096{
1097 BUG_ON(dm_bufio_in_request());
1098
1099 return new_read(c, block, NF_FRESH, bp);
1100}
1101EXPORT_SYMBOL_GPL(dm_bufio_new);
1102
1103void dm_bufio_prefetch(struct dm_bufio_client *c,
1104 sector_t block, unsigned n_blocks)
1105{
1106 struct blk_plug plug;
1107
1108 LIST_HEAD(write_list);
1109
1110 BUG_ON(dm_bufio_in_request());
1111
1112 blk_start_plug(&plug);
1113 dm_bufio_lock(c);
1114
1115 for (; n_blocks--; block++) {
1116 int need_submit;
1117 struct dm_buffer *b;
1118 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1119 &write_list);
1120 if (unlikely(!list_empty(&write_list))) {
1121 dm_bufio_unlock(c);
1122 blk_finish_plug(&plug);
1123 __flush_write_list(&write_list);
1124 blk_start_plug(&plug);
1125 dm_bufio_lock(c);
1126 }
1127 if (unlikely(b != NULL)) {
1128 dm_bufio_unlock(c);
1129
1130 if (need_submit)
1131 submit_io(b, REQ_OP_READ, read_endio);
1132 dm_bufio_release(b);
1133
1134 cond_resched();
1135
1136 if (!n_blocks)
1137 goto flush_plug;
1138 dm_bufio_lock(c);
1139 }
1140 }
1141
1142 dm_bufio_unlock(c);
1143
1144flush_plug:
1145 blk_finish_plug(&plug);
1146}
1147EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1148
1149void dm_bufio_release(struct dm_buffer *b)
1150{
1151 struct dm_bufio_client *c = b->c;
1152
1153 dm_bufio_lock(c);
1154
1155 BUG_ON(!b->hold_count);
1156
1157 b->hold_count--;
1158 if (!b->hold_count) {
1159 wake_up(&c->free_buffer_wait);
1160
1161 /*
1162 * If there were errors on the buffer, and the buffer is not
1163 * to be written, free the buffer. There is no point in caching
1164 * invalid buffer.
1165 */
1166 if ((b->read_error || b->write_error) &&
1167 !test_bit(B_READING, &b->state) &&
1168 !test_bit(B_WRITING, &b->state) &&
1169 !test_bit(B_DIRTY, &b->state)) {
1170 __unlink_buffer(b);
1171 __free_buffer_wake(b);
1172 }
1173 }
1174
1175 dm_bufio_unlock(c);
1176}
1177EXPORT_SYMBOL_GPL(dm_bufio_release);
1178
1179void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1180 unsigned start, unsigned end)
1181{
1182 struct dm_bufio_client *c = b->c;
1183
1184 BUG_ON(start >= end);
1185 BUG_ON(end > b->c->block_size);
1186
1187 dm_bufio_lock(c);
1188
1189 BUG_ON(test_bit(B_READING, &b->state));
1190
1191 if (!test_and_set_bit(B_DIRTY, &b->state)) {
1192 b->dirty_start = start;
1193 b->dirty_end = end;
1194 __relink_lru(b, LIST_DIRTY);
1195 } else {
1196 if (start < b->dirty_start)
1197 b->dirty_start = start;
1198 if (end > b->dirty_end)
1199 b->dirty_end = end;
1200 }
1201
1202 dm_bufio_unlock(c);
1203}
1204EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1205
1206void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1207{
1208 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1209}
1210EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1211
1212void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1213{
1214 LIST_HEAD(write_list);
1215
1216 BUG_ON(dm_bufio_in_request());
1217
1218 dm_bufio_lock(c);
1219 __write_dirty_buffers_async(c, 0, &write_list);
1220 dm_bufio_unlock(c);
1221 __flush_write_list(&write_list);
1222}
1223EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1224
1225/*
1226 * For performance, it is essential that the buffers are written asynchronously
1227 * and simultaneously (so that the block layer can merge the writes) and then
1228 * waited upon.
1229 *
1230 * Finally, we flush hardware disk cache.
1231 */
1232int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1233{
1234 int a, f;
1235 unsigned long buffers_processed = 0;
1236 struct dm_buffer *b, *tmp;
1237
1238 LIST_HEAD(write_list);
1239
1240 dm_bufio_lock(c);
1241 __write_dirty_buffers_async(c, 0, &write_list);
1242 dm_bufio_unlock(c);
1243 __flush_write_list(&write_list);
1244 dm_bufio_lock(c);
1245
1246again:
1247 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1248 int dropped_lock = 0;
1249
1250 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1251 buffers_processed++;
1252
1253 BUG_ON(test_bit(B_READING, &b->state));
1254
1255 if (test_bit(B_WRITING, &b->state)) {
1256 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1257 dropped_lock = 1;
1258 b->hold_count++;
1259 dm_bufio_unlock(c);
1260 wait_on_bit_io(&b->state, B_WRITING,
1261 TASK_UNINTERRUPTIBLE);
1262 dm_bufio_lock(c);
1263 b->hold_count--;
1264 } else
1265 wait_on_bit_io(&b->state, B_WRITING,
1266 TASK_UNINTERRUPTIBLE);
1267 }
1268
1269 if (!test_bit(B_DIRTY, &b->state) &&
1270 !test_bit(B_WRITING, &b->state))
1271 __relink_lru(b, LIST_CLEAN);
1272
1273 cond_resched();
1274
1275 /*
1276 * If we dropped the lock, the list is no longer consistent,
1277 * so we must restart the search.
1278 *
1279 * In the most common case, the buffer just processed is
1280 * relinked to the clean list, so we won't loop scanning the
1281 * same buffer again and again.
1282 *
1283 * This may livelock if there is another thread simultaneously
1284 * dirtying buffers, so we count the number of buffers walked
1285 * and if it exceeds the total number of buffers, it means that
1286 * someone is doing some writes simultaneously with us. In
1287 * this case, stop, dropping the lock.
1288 */
1289 if (dropped_lock)
1290 goto again;
1291 }
1292 wake_up(&c->free_buffer_wait);
1293 dm_bufio_unlock(c);
1294
1295 a = xchg(&c->async_write_error, 0);
1296 f = dm_bufio_issue_flush(c);
1297 if (a)
1298 return a;
1299
1300 return f;
1301}
1302EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1303
1304/*
1305 * Use dm-io to send an empty barrier to flush the device.
1306 */
1307int dm_bufio_issue_flush(struct dm_bufio_client *c)
1308{
1309 struct dm_io_request io_req = {
1310 .bi_op = REQ_OP_WRITE,
1311 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1312 .mem.type = DM_IO_KMEM,
1313 .mem.ptr.addr = NULL,
1314 .client = c->dm_io,
1315 };
1316 struct dm_io_region io_reg = {
1317 .bdev = c->bdev,
1318 .sector = 0,
1319 .count = 0,
1320 };
1321
1322 BUG_ON(dm_bufio_in_request());
1323
1324 return dm_io(&io_req, 1, &io_reg, NULL);
1325}
1326EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1327
1328/*
1329 * We first delete any other buffer that may be at that new location.
1330 *
1331 * Then, we write the buffer to the original location if it was dirty.
1332 *
1333 * Then, if we are the only one who is holding the buffer, relink the buffer
1334 * in the buffer tree for the new location.
1335 *
1336 * If there was someone else holding the buffer, we write it to the new
1337 * location but not relink it, because that other user needs to have the buffer
1338 * at the same place.
1339 */
1340void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1341{
1342 struct dm_bufio_client *c = b->c;
1343 struct dm_buffer *new;
1344
1345 BUG_ON(dm_bufio_in_request());
1346
1347 dm_bufio_lock(c);
1348
1349retry:
1350 new = __find(c, new_block);
1351 if (new) {
1352 if (new->hold_count) {
1353 __wait_for_free_buffer(c);
1354 goto retry;
1355 }
1356
1357 /*
1358 * FIXME: Is there any point waiting for a write that's going
1359 * to be overwritten in a bit?
1360 */
1361 __make_buffer_clean(new);
1362 __unlink_buffer(new);
1363 __free_buffer_wake(new);
1364 }
1365
1366 BUG_ON(!b->hold_count);
1367 BUG_ON(test_bit(B_READING, &b->state));
1368
1369 __write_dirty_buffer(b, NULL);
1370 if (b->hold_count == 1) {
1371 wait_on_bit_io(&b->state, B_WRITING,
1372 TASK_UNINTERRUPTIBLE);
1373 set_bit(B_DIRTY, &b->state);
1374 b->dirty_start = 0;
1375 b->dirty_end = c->block_size;
1376 __unlink_buffer(b);
1377 __link_buffer(b, new_block, LIST_DIRTY);
1378 } else {
1379 sector_t old_block;
1380 wait_on_bit_lock_io(&b->state, B_WRITING,
1381 TASK_UNINTERRUPTIBLE);
1382 /*
1383 * Relink buffer to "new_block" so that write_callback
1384 * sees "new_block" as a block number.
1385 * After the write, link the buffer back to old_block.
1386 * All this must be done in bufio lock, so that block number
1387 * change isn't visible to other threads.
1388 */
1389 old_block = b->block;
1390 __unlink_buffer(b);
1391 __link_buffer(b, new_block, b->list_mode);
1392 submit_io(b, REQ_OP_WRITE, write_endio);
1393 wait_on_bit_io(&b->state, B_WRITING,
1394 TASK_UNINTERRUPTIBLE);
1395 __unlink_buffer(b);
1396 __link_buffer(b, old_block, b->list_mode);
1397 }
1398
1399 dm_bufio_unlock(c);
1400 dm_bufio_release(b);
1401}
1402EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1403
1404/*
1405 * Free the given buffer.
1406 *
1407 * This is just a hint, if the buffer is in use or dirty, this function
1408 * does nothing.
1409 */
1410void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1411{
1412 struct dm_buffer *b;
1413
1414 dm_bufio_lock(c);
1415
1416 b = __find(c, block);
1417 if (b && likely(!b->hold_count) && likely(!b->state)) {
1418 __unlink_buffer(b);
1419 __free_buffer_wake(b);
1420 }
1421
1422 dm_bufio_unlock(c);
1423}
1424EXPORT_SYMBOL_GPL(dm_bufio_forget);
1425
1426void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1427{
1428 c->minimum_buffers = n;
1429}
1430EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1431
1432unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1433{
1434 return c->block_size;
1435}
1436EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1437
1438sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1439{
1440 sector_t s = i_size_read(c->bdev->bd_inode) >> SECTOR_SHIFT;
1441 if (likely(c->sectors_per_block_bits >= 0))
1442 s >>= c->sectors_per_block_bits;
1443 else
1444 sector_div(s, c->block_size >> SECTOR_SHIFT);
1445 return s;
1446}
1447EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1448
1449sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1450{
1451 return b->block;
1452}
1453EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1454
1455void *dm_bufio_get_block_data(struct dm_buffer *b)
1456{
1457 return b->data;
1458}
1459EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1460
1461void *dm_bufio_get_aux_data(struct dm_buffer *b)
1462{
1463 return b + 1;
1464}
1465EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1466
1467struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1468{
1469 return b->c;
1470}
1471EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1472
1473static void drop_buffers(struct dm_bufio_client *c)
1474{
1475 struct dm_buffer *b;
1476 int i;
1477 bool warned = false;
1478
1479 BUG_ON(dm_bufio_in_request());
1480
1481 /*
1482 * An optimization so that the buffers are not written one-by-one.
1483 */
1484 dm_bufio_write_dirty_buffers_async(c);
1485
1486 dm_bufio_lock(c);
1487
1488 while ((b = __get_unclaimed_buffer(c)))
1489 __free_buffer_wake(b);
1490
1491 for (i = 0; i < LIST_SIZE; i++)
1492 list_for_each_entry(b, &c->lru[i], lru_list) {
1493 WARN_ON(!warned);
1494 warned = true;
1495 DMERR("leaked buffer %llx, hold count %u, list %d",
1496 (unsigned long long)b->block, b->hold_count, i);
1497#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1498 stack_trace_print(b->stack_entries, b->stack_len, 1);
1499 /* mark unclaimed to avoid BUG_ON below */
1500 b->hold_count = 0;
1501#endif
1502 }
1503
1504#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1505 while ((b = __get_unclaimed_buffer(c)))
1506 __free_buffer_wake(b);
1507#endif
1508
1509 for (i = 0; i < LIST_SIZE; i++)
1510 BUG_ON(!list_empty(&c->lru[i]));
1511
1512 dm_bufio_unlock(c);
1513}
1514
1515/*
1516 * We may not be able to evict this buffer if IO pending or the client
1517 * is still using it. Caller is expected to know buffer is too old.
1518 *
1519 * And if GFP_NOFS is used, we must not do any I/O because we hold
1520 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1521 * rerouted to different bufio client.
1522 */
1523static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1524{
1525 if (!(gfp & __GFP_FS)) {
1526 if (test_bit(B_READING, &b->state) ||
1527 test_bit(B_WRITING, &b->state) ||
1528 test_bit(B_DIRTY, &b->state))
1529 return false;
1530 }
1531
1532 if (b->hold_count)
1533 return false;
1534
1535 __make_buffer_clean(b);
1536 __unlink_buffer(b);
1537 __free_buffer_wake(b);
1538
1539 return true;
1540}
1541
1542static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1543{
1544 unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1545 if (likely(c->sectors_per_block_bits >= 0))
1546 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1547 else
1548 retain_bytes /= c->block_size;
1549 return retain_bytes;
1550}
1551
1552static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1553 gfp_t gfp_mask)
1554{
1555 int l;
1556 struct dm_buffer *b, *tmp;
1557 unsigned long freed = 0;
1558 unsigned long count = c->n_buffers[LIST_CLEAN] +
1559 c->n_buffers[LIST_DIRTY];
1560 unsigned long retain_target = get_retain_buffers(c);
1561
1562 for (l = 0; l < LIST_SIZE; l++) {
1563 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1564 if (__try_evict_buffer(b, gfp_mask))
1565 freed++;
1566 if (!--nr_to_scan || ((count - freed) <= retain_target))
1567 return freed;
1568 cond_resched();
1569 }
1570 }
1571 return freed;
1572}
1573
1574static unsigned long
1575dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1576{
1577 struct dm_bufio_client *c;
1578 unsigned long freed;
1579
1580 c = container_of(shrink, struct dm_bufio_client, shrinker);
1581 if (sc->gfp_mask & __GFP_FS)
1582 dm_bufio_lock(c);
1583 else if (!dm_bufio_trylock(c))
1584 return SHRINK_STOP;
1585
1586 freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1587 dm_bufio_unlock(c);
1588 return freed;
1589}
1590
1591static unsigned long
1592dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1593{
1594 struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1595 unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1596 READ_ONCE(c->n_buffers[LIST_DIRTY]);
1597 unsigned long retain_target = get_retain_buffers(c);
1598
1599 return (count < retain_target) ? 0 : (count - retain_target);
1600}
1601
1602/*
1603 * Create the buffering interface
1604 */
1605struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1606 unsigned reserved_buffers, unsigned aux_size,
1607 void (*alloc_callback)(struct dm_buffer *),
1608 void (*write_callback)(struct dm_buffer *))
1609{
1610 int r;
1611 struct dm_bufio_client *c;
1612 unsigned i;
1613 char slab_name[27];
1614
1615 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1616 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1617 r = -EINVAL;
1618 goto bad_client;
1619 }
1620
1621 c = kzalloc(sizeof(*c), GFP_KERNEL);
1622 if (!c) {
1623 r = -ENOMEM;
1624 goto bad_client;
1625 }
1626 c->buffer_tree = RB_ROOT;
1627
1628 c->bdev = bdev;
1629 c->block_size = block_size;
1630 if (is_power_of_2(block_size))
1631 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1632 else
1633 c->sectors_per_block_bits = -1;
1634
1635 c->alloc_callback = alloc_callback;
1636 c->write_callback = write_callback;
1637
1638 for (i = 0; i < LIST_SIZE; i++) {
1639 INIT_LIST_HEAD(&c->lru[i]);
1640 c->n_buffers[i] = 0;
1641 }
1642
1643 mutex_init(&c->lock);
1644 INIT_LIST_HEAD(&c->reserved_buffers);
1645 c->need_reserved_buffers = reserved_buffers;
1646
1647 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1648
1649 init_waitqueue_head(&c->free_buffer_wait);
1650 c->async_write_error = 0;
1651
1652 c->dm_io = dm_io_client_create();
1653 if (IS_ERR(c->dm_io)) {
1654 r = PTR_ERR(c->dm_io);
1655 goto bad_dm_io;
1656 }
1657
1658 if (block_size <= KMALLOC_MAX_SIZE &&
1659 (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1660 unsigned align = min(1U << __ffs(block_size), (unsigned)PAGE_SIZE);
1661 snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1662 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1663 SLAB_RECLAIM_ACCOUNT, NULL);
1664 if (!c->slab_cache) {
1665 r = -ENOMEM;
1666 goto bad;
1667 }
1668 }
1669 if (aux_size)
1670 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1671 else
1672 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1673 c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1674 0, SLAB_RECLAIM_ACCOUNT, NULL);
1675 if (!c->slab_buffer) {
1676 r = -ENOMEM;
1677 goto bad;
1678 }
1679
1680 while (c->need_reserved_buffers) {
1681 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1682
1683 if (!b) {
1684 r = -ENOMEM;
1685 goto bad;
1686 }
1687 __free_buffer_wake(b);
1688 }
1689
1690 c->shrinker.count_objects = dm_bufio_shrink_count;
1691 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1692 c->shrinker.seeks = 1;
1693 c->shrinker.batch = 0;
1694 r = register_shrinker(&c->shrinker);
1695 if (r)
1696 goto bad;
1697
1698 mutex_lock(&dm_bufio_clients_lock);
1699 dm_bufio_client_count++;
1700 list_add(&c->client_list, &dm_bufio_all_clients);
1701 __cache_size_refresh();
1702 mutex_unlock(&dm_bufio_clients_lock);
1703
1704 return c;
1705
1706bad:
1707 while (!list_empty(&c->reserved_buffers)) {
1708 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1709 struct dm_buffer, lru_list);
1710 list_del(&b->lru_list);
1711 free_buffer(b);
1712 }
1713 kmem_cache_destroy(c->slab_cache);
1714 kmem_cache_destroy(c->slab_buffer);
1715 dm_io_client_destroy(c->dm_io);
1716bad_dm_io:
1717 mutex_destroy(&c->lock);
1718 kfree(c);
1719bad_client:
1720 return ERR_PTR(r);
1721}
1722EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1723
1724/*
1725 * Free the buffering interface.
1726 * It is required that there are no references on any buffers.
1727 */
1728void dm_bufio_client_destroy(struct dm_bufio_client *c)
1729{
1730 unsigned i;
1731
1732 drop_buffers(c);
1733
1734 unregister_shrinker(&c->shrinker);
1735
1736 mutex_lock(&dm_bufio_clients_lock);
1737
1738 list_del(&c->client_list);
1739 dm_bufio_client_count--;
1740 __cache_size_refresh();
1741
1742 mutex_unlock(&dm_bufio_clients_lock);
1743
1744 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1745 BUG_ON(c->need_reserved_buffers);
1746
1747 while (!list_empty(&c->reserved_buffers)) {
1748 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1749 struct dm_buffer, lru_list);
1750 list_del(&b->lru_list);
1751 free_buffer(b);
1752 }
1753
1754 for (i = 0; i < LIST_SIZE; i++)
1755 if (c->n_buffers[i])
1756 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1757
1758 for (i = 0; i < LIST_SIZE; i++)
1759 BUG_ON(c->n_buffers[i]);
1760
1761 kmem_cache_destroy(c->slab_cache);
1762 kmem_cache_destroy(c->slab_buffer);
1763 dm_io_client_destroy(c->dm_io);
1764 mutex_destroy(&c->lock);
1765 kfree(c);
1766}
1767EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1768
1769void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1770{
1771 c->start = start;
1772}
1773EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1774
1775static unsigned get_max_age_hz(void)
1776{
1777 unsigned max_age = READ_ONCE(dm_bufio_max_age);
1778
1779 if (max_age > UINT_MAX / HZ)
1780 max_age = UINT_MAX / HZ;
1781
1782 return max_age * HZ;
1783}
1784
1785static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1786{
1787 return time_after_eq(jiffies, b->last_accessed + age_hz);
1788}
1789
1790static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1791{
1792 struct dm_buffer *b, *tmp;
1793 unsigned long retain_target = get_retain_buffers(c);
1794 unsigned long count;
1795 LIST_HEAD(write_list);
1796
1797 dm_bufio_lock(c);
1798
1799 __check_watermark(c, &write_list);
1800 if (unlikely(!list_empty(&write_list))) {
1801 dm_bufio_unlock(c);
1802 __flush_write_list(&write_list);
1803 dm_bufio_lock(c);
1804 }
1805
1806 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1807 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1808 if (count <= retain_target)
1809 break;
1810
1811 if (!older_than(b, age_hz))
1812 break;
1813
1814 if (__try_evict_buffer(b, 0))
1815 count--;
1816
1817 cond_resched();
1818 }
1819
1820 dm_bufio_unlock(c);
1821}
1822
1823static void do_global_cleanup(struct work_struct *w)
1824{
1825 struct dm_bufio_client *locked_client = NULL;
1826 struct dm_bufio_client *current_client;
1827 struct dm_buffer *b;
1828 unsigned spinlock_hold_count;
1829 unsigned long threshold = dm_bufio_cache_size -
1830 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1831 unsigned long loops = global_num * 2;
1832
1833 mutex_lock(&dm_bufio_clients_lock);
1834
1835 while (1) {
1836 cond_resched();
1837
1838 spin_lock(&global_spinlock);
1839 if (unlikely(dm_bufio_current_allocated <= threshold))
1840 break;
1841
1842 spinlock_hold_count = 0;
1843get_next:
1844 if (!loops--)
1845 break;
1846 if (unlikely(list_empty(&global_queue)))
1847 break;
1848 b = list_entry(global_queue.prev, struct dm_buffer, global_list);
1849
1850 if (b->accessed) {
1851 b->accessed = 0;
1852 list_move(&b->global_list, &global_queue);
1853 if (likely(++spinlock_hold_count < 16))
1854 goto get_next;
1855 spin_unlock(&global_spinlock);
1856 continue;
1857 }
1858
1859 current_client = b->c;
1860 if (unlikely(current_client != locked_client)) {
1861 if (locked_client)
1862 dm_bufio_unlock(locked_client);
1863
1864 if (!dm_bufio_trylock(current_client)) {
1865 spin_unlock(&global_spinlock);
1866 dm_bufio_lock(current_client);
1867 locked_client = current_client;
1868 continue;
1869 }
1870
1871 locked_client = current_client;
1872 }
1873
1874 spin_unlock(&global_spinlock);
1875
1876 if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
1877 spin_lock(&global_spinlock);
1878 list_move(&b->global_list, &global_queue);
1879 spin_unlock(&global_spinlock);
1880 }
1881 }
1882
1883 spin_unlock(&global_spinlock);
1884
1885 if (locked_client)
1886 dm_bufio_unlock(locked_client);
1887
1888 mutex_unlock(&dm_bufio_clients_lock);
1889}
1890
1891static void cleanup_old_buffers(void)
1892{
1893 unsigned long max_age_hz = get_max_age_hz();
1894 struct dm_bufio_client *c;
1895
1896 mutex_lock(&dm_bufio_clients_lock);
1897
1898 __cache_size_refresh();
1899
1900 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1901 __evict_old_buffers(c, max_age_hz);
1902
1903 mutex_unlock(&dm_bufio_clients_lock);
1904}
1905
1906static void work_fn(struct work_struct *w)
1907{
1908 cleanup_old_buffers();
1909
1910 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
1911 DM_BUFIO_WORK_TIMER_SECS * HZ);
1912}
1913
1914/*----------------------------------------------------------------
1915 * Module setup
1916 *--------------------------------------------------------------*/
1917
1918/*
1919 * This is called only once for the whole dm_bufio module.
1920 * It initializes memory limit.
1921 */
1922static int __init dm_bufio_init(void)
1923{
1924 __u64 mem;
1925
1926 dm_bufio_allocated_kmem_cache = 0;
1927 dm_bufio_allocated_get_free_pages = 0;
1928 dm_bufio_allocated_vmalloc = 0;
1929 dm_bufio_current_allocated = 0;
1930
1931 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
1932 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
1933
1934 if (mem > ULONG_MAX)
1935 mem = ULONG_MAX;
1936
1937#ifdef CONFIG_MMU
1938 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
1939 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
1940#endif
1941
1942 dm_bufio_default_cache_size = mem;
1943
1944 mutex_lock(&dm_bufio_clients_lock);
1945 __cache_size_refresh();
1946 mutex_unlock(&dm_bufio_clients_lock);
1947
1948 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1949 if (!dm_bufio_wq)
1950 return -ENOMEM;
1951
1952 INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
1953 INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
1954 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
1955 DM_BUFIO_WORK_TIMER_SECS * HZ);
1956
1957 return 0;
1958}
1959
1960/*
1961 * This is called once when unloading the dm_bufio module.
1962 */
1963static void __exit dm_bufio_exit(void)
1964{
1965 int bug = 0;
1966
1967 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
1968 flush_workqueue(dm_bufio_wq);
1969 destroy_workqueue(dm_bufio_wq);
1970
1971 if (dm_bufio_client_count) {
1972 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1973 __func__, dm_bufio_client_count);
1974 bug = 1;
1975 }
1976
1977 if (dm_bufio_current_allocated) {
1978 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1979 __func__, dm_bufio_current_allocated);
1980 bug = 1;
1981 }
1982
1983 if (dm_bufio_allocated_get_free_pages) {
1984 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1985 __func__, dm_bufio_allocated_get_free_pages);
1986 bug = 1;
1987 }
1988
1989 if (dm_bufio_allocated_vmalloc) {
1990 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1991 __func__, dm_bufio_allocated_vmalloc);
1992 bug = 1;
1993 }
1994
1995 BUG_ON(bug);
1996}
1997
1998module_init(dm_bufio_init)
1999module_exit(dm_bufio_exit)
2000
2001module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2002MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2003
2004module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2005MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2006
2007module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2008MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2009
2010module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2011MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2012
2013module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2014MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2015
2016module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2017MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2018
2019module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2020MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2021
2022module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2023MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2024
2025MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2026MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2027MODULE_LICENSE("GPL");