projects/linux·blob master · refs

2940 lines · 69.9 KiB

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