Loading...
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 "dm-bufio.h"
10
11#include <linux/device-mapper.h>
12#include <linux/dm-io.h>
13#include <linux/slab.h>
14#include <linux/jiffies.h>
15#include <linux/vmalloc.h>
16#include <linux/shrinker.h>
17#include <linux/module.h>
18#include <linux/rbtree.h>
19#include <linux/stacktrace.h>
20
21#define DM_MSG_PREFIX "bufio"
22
23/*
24 * Memory management policy:
25 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
26 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
27 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
28 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
29 * dirty buffers.
30 */
31#define DM_BUFIO_MIN_BUFFERS 8
32
33#define DM_BUFIO_MEMORY_PERCENT 2
34#define DM_BUFIO_VMALLOC_PERCENT 25
35#define DM_BUFIO_WRITEBACK_PERCENT 75
36
37/*
38 * Check buffer ages in this interval (seconds)
39 */
40#define DM_BUFIO_WORK_TIMER_SECS 30
41
42/*
43 * Free buffers when they are older than this (seconds)
44 */
45#define DM_BUFIO_DEFAULT_AGE_SECS 300
46
47/*
48 * The nr of bytes of cached data to keep around.
49 */
50#define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
51
52/*
53 * The number of bvec entries that are embedded directly in the buffer.
54 * If the chunk size is larger, dm-io is used to do the io.
55 */
56#define DM_BUFIO_INLINE_VECS 16
57
58/*
59 * Don't try to use kmem_cache_alloc for blocks larger than this.
60 * For explanation, see alloc_buffer_data below.
61 */
62#define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT (PAGE_SIZE >> 1)
63#define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT (PAGE_SIZE << (MAX_ORDER - 1))
64
65/*
66 * dm_buffer->list_mode
67 */
68#define LIST_CLEAN 0
69#define LIST_DIRTY 1
70#define LIST_SIZE 2
71
72/*
73 * Linking of buffers:
74 * All buffers are linked to cache_hash with their hash_list field.
75 *
76 * Clean buffers that are not being written (B_WRITING not set)
77 * are linked to lru[LIST_CLEAN] with their lru_list field.
78 *
79 * Dirty and clean buffers that are being written are linked to
80 * lru[LIST_DIRTY] with their lru_list field. When the write
81 * finishes, the buffer cannot be relinked immediately (because we
82 * are in an interrupt context and relinking requires process
83 * context), so some clean-not-writing buffers can be held on
84 * dirty_lru too. They are later added to lru in the process
85 * context.
86 */
87struct dm_bufio_client {
88 struct mutex lock;
89
90 struct list_head lru[LIST_SIZE];
91 unsigned long n_buffers[LIST_SIZE];
92
93 struct block_device *bdev;
94 unsigned block_size;
95 unsigned char sectors_per_block_bits;
96 unsigned char pages_per_block_bits;
97 unsigned char blocks_per_page_bits;
98 unsigned aux_size;
99 void (*alloc_callback)(struct dm_buffer *);
100 void (*write_callback)(struct dm_buffer *);
101
102 struct dm_io_client *dm_io;
103
104 struct list_head reserved_buffers;
105 unsigned need_reserved_buffers;
106
107 unsigned minimum_buffers;
108
109 struct rb_root buffer_tree;
110 wait_queue_head_t free_buffer_wait;
111
112 int async_write_error;
113
114 struct list_head client_list;
115 struct shrinker shrinker;
116};
117
118/*
119 * Buffer state bits.
120 */
121#define B_READING 0
122#define B_WRITING 1
123#define B_DIRTY 2
124
125/*
126 * Describes how the block was allocated:
127 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
128 * See the comment at alloc_buffer_data.
129 */
130enum data_mode {
131 DATA_MODE_SLAB = 0,
132 DATA_MODE_GET_FREE_PAGES = 1,
133 DATA_MODE_VMALLOC = 2,
134 DATA_MODE_LIMIT = 3
135};
136
137struct dm_buffer {
138 struct rb_node node;
139 struct list_head lru_list;
140 sector_t block;
141 void *data;
142 enum data_mode data_mode;
143 unsigned char list_mode; /* LIST_* */
144 unsigned hold_count;
145 int read_error;
146 int write_error;
147 unsigned long state;
148 unsigned long last_accessed;
149 struct dm_bufio_client *c;
150 struct list_head write_list;
151 struct bio bio;
152 struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
153#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
154#define MAX_STACK 10
155 struct stack_trace stack_trace;
156 unsigned long stack_entries[MAX_STACK];
157#endif
158};
159
160/*----------------------------------------------------------------*/
161
162static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
163static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
164
165static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
166{
167 unsigned ret = c->blocks_per_page_bits - 1;
168
169 BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
170
171 return ret;
172}
173
174#define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)])
175#define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)])
176
177#define dm_bufio_in_request() (!!current->bio_list)
178
179static void dm_bufio_lock(struct dm_bufio_client *c)
180{
181 mutex_lock_nested(&c->lock, dm_bufio_in_request());
182}
183
184static int dm_bufio_trylock(struct dm_bufio_client *c)
185{
186 return mutex_trylock(&c->lock);
187}
188
189static void dm_bufio_unlock(struct dm_bufio_client *c)
190{
191 mutex_unlock(&c->lock);
192}
193
194/*
195 * FIXME Move to sched.h?
196 */
197#ifdef CONFIG_PREEMPT_VOLUNTARY
198# define dm_bufio_cond_resched() \
199do { \
200 if (unlikely(need_resched())) \
201 _cond_resched(); \
202} while (0)
203#else
204# define dm_bufio_cond_resched() do { } while (0)
205#endif
206
207/*----------------------------------------------------------------*/
208
209/*
210 * Default cache size: available memory divided by the ratio.
211 */
212static unsigned long dm_bufio_default_cache_size;
213
214/*
215 * Total cache size set by the user.
216 */
217static unsigned long dm_bufio_cache_size;
218
219/*
220 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
221 * at any time. If it disagrees, the user has changed cache size.
222 */
223static unsigned long dm_bufio_cache_size_latch;
224
225static DEFINE_SPINLOCK(param_spinlock);
226
227/*
228 * Buffers are freed after this timeout
229 */
230static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
231static unsigned dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
232
233static unsigned long dm_bufio_peak_allocated;
234static unsigned long dm_bufio_allocated_kmem_cache;
235static unsigned long dm_bufio_allocated_get_free_pages;
236static unsigned long dm_bufio_allocated_vmalloc;
237static unsigned long dm_bufio_current_allocated;
238
239/*----------------------------------------------------------------*/
240
241/*
242 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
243 */
244static unsigned long dm_bufio_cache_size_per_client;
245
246/*
247 * The current number of clients.
248 */
249static int dm_bufio_client_count;
250
251/*
252 * The list of all clients.
253 */
254static LIST_HEAD(dm_bufio_all_clients);
255
256/*
257 * This mutex protects dm_bufio_cache_size_latch,
258 * dm_bufio_cache_size_per_client and dm_bufio_client_count
259 */
260static DEFINE_MUTEX(dm_bufio_clients_lock);
261
262#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
263static void buffer_record_stack(struct dm_buffer *b)
264{
265 b->stack_trace.nr_entries = 0;
266 b->stack_trace.max_entries = MAX_STACK;
267 b->stack_trace.entries = b->stack_entries;
268 b->stack_trace.skip = 2;
269 save_stack_trace(&b->stack_trace);
270}
271#endif
272
273/*----------------------------------------------------------------
274 * A red/black tree acts as an index for all the buffers.
275 *--------------------------------------------------------------*/
276static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
277{
278 struct rb_node *n = c->buffer_tree.rb_node;
279 struct dm_buffer *b;
280
281 while (n) {
282 b = container_of(n, struct dm_buffer, node);
283
284 if (b->block == block)
285 return b;
286
287 n = (b->block < block) ? n->rb_left : n->rb_right;
288 }
289
290 return NULL;
291}
292
293static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
294{
295 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
296 struct dm_buffer *found;
297
298 while (*new) {
299 found = container_of(*new, struct dm_buffer, node);
300
301 if (found->block == b->block) {
302 BUG_ON(found != b);
303 return;
304 }
305
306 parent = *new;
307 new = (found->block < b->block) ?
308 &((*new)->rb_left) : &((*new)->rb_right);
309 }
310
311 rb_link_node(&b->node, parent, new);
312 rb_insert_color(&b->node, &c->buffer_tree);
313}
314
315static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
316{
317 rb_erase(&b->node, &c->buffer_tree);
318}
319
320/*----------------------------------------------------------------*/
321
322static void adjust_total_allocated(enum data_mode data_mode, long diff)
323{
324 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
325 &dm_bufio_allocated_kmem_cache,
326 &dm_bufio_allocated_get_free_pages,
327 &dm_bufio_allocated_vmalloc,
328 };
329
330 spin_lock(¶m_spinlock);
331
332 *class_ptr[data_mode] += diff;
333
334 dm_bufio_current_allocated += diff;
335
336 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
337 dm_bufio_peak_allocated = dm_bufio_current_allocated;
338
339 spin_unlock(¶m_spinlock);
340}
341
342/*
343 * Change the number of clients and recalculate per-client limit.
344 */
345static void __cache_size_refresh(void)
346{
347 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
348 BUG_ON(dm_bufio_client_count < 0);
349
350 dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
351
352 /*
353 * Use default if set to 0 and report the actual cache size used.
354 */
355 if (!dm_bufio_cache_size_latch) {
356 (void)cmpxchg(&dm_bufio_cache_size, 0,
357 dm_bufio_default_cache_size);
358 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
359 }
360
361 dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
362 (dm_bufio_client_count ? : 1);
363}
364
365/*
366 * Allocating buffer data.
367 *
368 * Small buffers are allocated with kmem_cache, to use space optimally.
369 *
370 * For large buffers, we choose between get_free_pages and vmalloc.
371 * Each has advantages and disadvantages.
372 *
373 * __get_free_pages can randomly fail if the memory is fragmented.
374 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
375 * as low as 128M) so using it for caching is not appropriate.
376 *
377 * If the allocation may fail we use __get_free_pages. Memory fragmentation
378 * won't have a fatal effect here, but it just causes flushes of some other
379 * buffers and more I/O will be performed. Don't use __get_free_pages if it
380 * always fails (i.e. order >= MAX_ORDER).
381 *
382 * If the allocation shouldn't fail we use __vmalloc. This is only for the
383 * initial reserve allocation, so there's no risk of wasting all vmalloc
384 * space.
385 */
386static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
387 enum data_mode *data_mode)
388{
389 unsigned noio_flag;
390 void *ptr;
391
392 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
393 *data_mode = DATA_MODE_SLAB;
394 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
395 }
396
397 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
398 gfp_mask & __GFP_NORETRY) {
399 *data_mode = DATA_MODE_GET_FREE_PAGES;
400 return (void *)__get_free_pages(gfp_mask,
401 c->pages_per_block_bits);
402 }
403
404 *data_mode = DATA_MODE_VMALLOC;
405
406 /*
407 * __vmalloc allocates the data pages and auxiliary structures with
408 * gfp_flags that were specified, but pagetables are always allocated
409 * with GFP_KERNEL, no matter what was specified as gfp_mask.
410 *
411 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
412 * all allocations done by this process (including pagetables) are done
413 * as if GFP_NOIO was specified.
414 */
415
416 if (gfp_mask & __GFP_NORETRY)
417 noio_flag = memalloc_noio_save();
418
419 ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL);
420
421 if (gfp_mask & __GFP_NORETRY)
422 memalloc_noio_restore(noio_flag);
423
424 return ptr;
425}
426
427/*
428 * Free buffer's data.
429 */
430static void free_buffer_data(struct dm_bufio_client *c,
431 void *data, enum data_mode data_mode)
432{
433 switch (data_mode) {
434 case DATA_MODE_SLAB:
435 kmem_cache_free(DM_BUFIO_CACHE(c), data);
436 break;
437
438 case DATA_MODE_GET_FREE_PAGES:
439 free_pages((unsigned long)data, c->pages_per_block_bits);
440 break;
441
442 case DATA_MODE_VMALLOC:
443 vfree(data);
444 break;
445
446 default:
447 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
448 data_mode);
449 BUG();
450 }
451}
452
453/*
454 * Allocate buffer and its data.
455 */
456static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
457{
458 struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
459 gfp_mask);
460
461 if (!b)
462 return NULL;
463
464 b->c = c;
465
466 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
467 if (!b->data) {
468 kfree(b);
469 return NULL;
470 }
471
472 adjust_total_allocated(b->data_mode, (long)c->block_size);
473
474#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
475 memset(&b->stack_trace, 0, sizeof(b->stack_trace));
476#endif
477 return b;
478}
479
480/*
481 * Free buffer and its data.
482 */
483static void free_buffer(struct dm_buffer *b)
484{
485 struct dm_bufio_client *c = b->c;
486
487 adjust_total_allocated(b->data_mode, -(long)c->block_size);
488
489 free_buffer_data(c, b->data, b->data_mode);
490 kfree(b);
491}
492
493/*
494 * Link buffer to the hash list and clean or dirty queue.
495 */
496static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
497{
498 struct dm_bufio_client *c = b->c;
499
500 c->n_buffers[dirty]++;
501 b->block = block;
502 b->list_mode = dirty;
503 list_add(&b->lru_list, &c->lru[dirty]);
504 __insert(b->c, b);
505 b->last_accessed = jiffies;
506}
507
508/*
509 * Unlink buffer from the hash list and dirty or clean queue.
510 */
511static void __unlink_buffer(struct dm_buffer *b)
512{
513 struct dm_bufio_client *c = b->c;
514
515 BUG_ON(!c->n_buffers[b->list_mode]);
516
517 c->n_buffers[b->list_mode]--;
518 __remove(b->c, b);
519 list_del(&b->lru_list);
520}
521
522/*
523 * Place the buffer to the head of dirty or clean LRU queue.
524 */
525static void __relink_lru(struct dm_buffer *b, int dirty)
526{
527 struct dm_bufio_client *c = b->c;
528
529 BUG_ON(!c->n_buffers[b->list_mode]);
530
531 c->n_buffers[b->list_mode]--;
532 c->n_buffers[dirty]++;
533 b->list_mode = dirty;
534 list_move(&b->lru_list, &c->lru[dirty]);
535 b->last_accessed = jiffies;
536}
537
538/*----------------------------------------------------------------
539 * Submit I/O on the buffer.
540 *
541 * Bio interface is faster but it has some problems:
542 * the vector list is limited (increasing this limit increases
543 * memory-consumption per buffer, so it is not viable);
544 *
545 * the memory must be direct-mapped, not vmalloced;
546 *
547 * the I/O driver can reject requests spuriously if it thinks that
548 * the requests are too big for the device or if they cross a
549 * controller-defined memory boundary.
550 *
551 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
552 * it is not vmalloced, try using the bio interface.
553 *
554 * If the buffer is big, if it is vmalloced or if the underlying device
555 * rejects the bio because it is too large, use dm-io layer to do the I/O.
556 * The dm-io layer splits the I/O into multiple requests, avoiding the above
557 * shortcomings.
558 *--------------------------------------------------------------*/
559
560/*
561 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
562 * that the request was handled directly with bio interface.
563 */
564static void dmio_complete(unsigned long error, void *context)
565{
566 struct dm_buffer *b = context;
567
568 b->bio.bi_error = error ? -EIO : 0;
569 b->bio.bi_end_io(&b->bio);
570}
571
572static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
573 bio_end_io_t *end_io)
574{
575 int r;
576 struct dm_io_request io_req = {
577 .bi_rw = rw,
578 .notify.fn = dmio_complete,
579 .notify.context = b,
580 .client = b->c->dm_io,
581 };
582 struct dm_io_region region = {
583 .bdev = b->c->bdev,
584 .sector = block << b->c->sectors_per_block_bits,
585 .count = b->c->block_size >> SECTOR_SHIFT,
586 };
587
588 if (b->data_mode != DATA_MODE_VMALLOC) {
589 io_req.mem.type = DM_IO_KMEM;
590 io_req.mem.ptr.addr = b->data;
591 } else {
592 io_req.mem.type = DM_IO_VMA;
593 io_req.mem.ptr.vma = b->data;
594 }
595
596 b->bio.bi_end_io = end_io;
597
598 r = dm_io(&io_req, 1, ®ion, NULL);
599 if (r) {
600 b->bio.bi_error = r;
601 end_io(&b->bio);
602 }
603}
604
605static void inline_endio(struct bio *bio)
606{
607 bio_end_io_t *end_fn = bio->bi_private;
608 int error = bio->bi_error;
609
610 /*
611 * Reset the bio to free any attached resources
612 * (e.g. bio integrity profiles).
613 */
614 bio_reset(bio);
615
616 bio->bi_error = error;
617 end_fn(bio);
618}
619
620static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
621 bio_end_io_t *end_io)
622{
623 char *ptr;
624 int len;
625
626 bio_init(&b->bio);
627 b->bio.bi_io_vec = b->bio_vec;
628 b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
629 b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
630 b->bio.bi_bdev = b->c->bdev;
631 b->bio.bi_end_io = inline_endio;
632 /*
633 * Use of .bi_private isn't a problem here because
634 * the dm_buffer's inline bio is local to bufio.
635 */
636 b->bio.bi_private = end_io;
637
638 /*
639 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
640 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
641 */
642 ptr = b->data;
643 len = b->c->block_size;
644
645 if (len >= PAGE_SIZE)
646 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
647 else
648 BUG_ON((unsigned long)ptr & (len - 1));
649
650 do {
651 if (!bio_add_page(&b->bio, virt_to_page(ptr),
652 len < PAGE_SIZE ? len : PAGE_SIZE,
653 offset_in_page(ptr))) {
654 BUG_ON(b->c->block_size <= PAGE_SIZE);
655 use_dmio(b, rw, block, end_io);
656 return;
657 }
658
659 len -= PAGE_SIZE;
660 ptr += PAGE_SIZE;
661 } while (len > 0);
662
663 submit_bio(rw, &b->bio);
664}
665
666static void submit_io(struct dm_buffer *b, int rw, sector_t block,
667 bio_end_io_t *end_io)
668{
669 if (rw == WRITE && b->c->write_callback)
670 b->c->write_callback(b);
671
672 if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
673 b->data_mode != DATA_MODE_VMALLOC)
674 use_inline_bio(b, rw, block, end_io);
675 else
676 use_dmio(b, rw, block, end_io);
677}
678
679/*----------------------------------------------------------------
680 * Writing dirty buffers
681 *--------------------------------------------------------------*/
682
683/*
684 * The endio routine for write.
685 *
686 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
687 * it.
688 */
689static void write_endio(struct bio *bio)
690{
691 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
692
693 b->write_error = bio->bi_error;
694 if (unlikely(bio->bi_error)) {
695 struct dm_bufio_client *c = b->c;
696 int error = bio->bi_error;
697 (void)cmpxchg(&c->async_write_error, 0, error);
698 }
699
700 BUG_ON(!test_bit(B_WRITING, &b->state));
701
702 smp_mb__before_atomic();
703 clear_bit(B_WRITING, &b->state);
704 smp_mb__after_atomic();
705
706 wake_up_bit(&b->state, B_WRITING);
707}
708
709/*
710 * Initiate a write on a dirty buffer, but don't wait for it.
711 *
712 * - If the buffer is not dirty, exit.
713 * - If there some previous write going on, wait for it to finish (we can't
714 * have two writes on the same buffer simultaneously).
715 * - Submit our write and don't wait on it. We set B_WRITING indicating
716 * that there is a write in progress.
717 */
718static void __write_dirty_buffer(struct dm_buffer *b,
719 struct list_head *write_list)
720{
721 if (!test_bit(B_DIRTY, &b->state))
722 return;
723
724 clear_bit(B_DIRTY, &b->state);
725 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
726
727 if (!write_list)
728 submit_io(b, WRITE, b->block, write_endio);
729 else
730 list_add_tail(&b->write_list, write_list);
731}
732
733static void __flush_write_list(struct list_head *write_list)
734{
735 struct blk_plug plug;
736 blk_start_plug(&plug);
737 while (!list_empty(write_list)) {
738 struct dm_buffer *b =
739 list_entry(write_list->next, struct dm_buffer, write_list);
740 list_del(&b->write_list);
741 submit_io(b, WRITE, b->block, write_endio);
742 dm_bufio_cond_resched();
743 }
744 blk_finish_plug(&plug);
745}
746
747/*
748 * Wait until any activity on the buffer finishes. Possibly write the
749 * buffer if it is dirty. When this function finishes, there is no I/O
750 * running on the buffer and the buffer is not dirty.
751 */
752static void __make_buffer_clean(struct dm_buffer *b)
753{
754 BUG_ON(b->hold_count);
755
756 if (!b->state) /* fast case */
757 return;
758
759 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
760 __write_dirty_buffer(b, NULL);
761 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
762}
763
764/*
765 * Find some buffer that is not held by anybody, clean it, unlink it and
766 * return it.
767 */
768static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
769{
770 struct dm_buffer *b;
771
772 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
773 BUG_ON(test_bit(B_WRITING, &b->state));
774 BUG_ON(test_bit(B_DIRTY, &b->state));
775
776 if (!b->hold_count) {
777 __make_buffer_clean(b);
778 __unlink_buffer(b);
779 return b;
780 }
781 dm_bufio_cond_resched();
782 }
783
784 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
785 BUG_ON(test_bit(B_READING, &b->state));
786
787 if (!b->hold_count) {
788 __make_buffer_clean(b);
789 __unlink_buffer(b);
790 return b;
791 }
792 dm_bufio_cond_resched();
793 }
794
795 return NULL;
796}
797
798/*
799 * Wait until some other threads free some buffer or release hold count on
800 * some buffer.
801 *
802 * This function is entered with c->lock held, drops it and regains it
803 * before exiting.
804 */
805static void __wait_for_free_buffer(struct dm_bufio_client *c)
806{
807 DECLARE_WAITQUEUE(wait, current);
808
809 add_wait_queue(&c->free_buffer_wait, &wait);
810 set_task_state(current, TASK_UNINTERRUPTIBLE);
811 dm_bufio_unlock(c);
812
813 io_schedule();
814
815 remove_wait_queue(&c->free_buffer_wait, &wait);
816
817 dm_bufio_lock(c);
818}
819
820enum new_flag {
821 NF_FRESH = 0,
822 NF_READ = 1,
823 NF_GET = 2,
824 NF_PREFETCH = 3
825};
826
827/*
828 * Allocate a new buffer. If the allocation is not possible, wait until
829 * some other thread frees a buffer.
830 *
831 * May drop the lock and regain it.
832 */
833static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
834{
835 struct dm_buffer *b;
836
837 /*
838 * dm-bufio is resistant to allocation failures (it just keeps
839 * one buffer reserved in cases all the allocations fail).
840 * So set flags to not try too hard:
841 * GFP_NOIO: don't recurse into the I/O layer
842 * __GFP_NORETRY: don't retry and rather return failure
843 * __GFP_NOMEMALLOC: don't use emergency reserves
844 * __GFP_NOWARN: don't print a warning in case of failure
845 *
846 * For debugging, if we set the cache size to 1, no new buffers will
847 * be allocated.
848 */
849 while (1) {
850 if (dm_bufio_cache_size_latch != 1) {
851 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
852 if (b)
853 return b;
854 }
855
856 if (nf == NF_PREFETCH)
857 return NULL;
858
859 if (!list_empty(&c->reserved_buffers)) {
860 b = list_entry(c->reserved_buffers.next,
861 struct dm_buffer, lru_list);
862 list_del(&b->lru_list);
863 c->need_reserved_buffers++;
864
865 return b;
866 }
867
868 b = __get_unclaimed_buffer(c);
869 if (b)
870 return b;
871
872 __wait_for_free_buffer(c);
873 }
874}
875
876static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
877{
878 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
879
880 if (!b)
881 return NULL;
882
883 if (c->alloc_callback)
884 c->alloc_callback(b);
885
886 return b;
887}
888
889/*
890 * Free a buffer and wake other threads waiting for free buffers.
891 */
892static void __free_buffer_wake(struct dm_buffer *b)
893{
894 struct dm_bufio_client *c = b->c;
895
896 if (!c->need_reserved_buffers)
897 free_buffer(b);
898 else {
899 list_add(&b->lru_list, &c->reserved_buffers);
900 c->need_reserved_buffers--;
901 }
902
903 wake_up(&c->free_buffer_wait);
904}
905
906static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
907 struct list_head *write_list)
908{
909 struct dm_buffer *b, *tmp;
910
911 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
912 BUG_ON(test_bit(B_READING, &b->state));
913
914 if (!test_bit(B_DIRTY, &b->state) &&
915 !test_bit(B_WRITING, &b->state)) {
916 __relink_lru(b, LIST_CLEAN);
917 continue;
918 }
919
920 if (no_wait && test_bit(B_WRITING, &b->state))
921 return;
922
923 __write_dirty_buffer(b, write_list);
924 dm_bufio_cond_resched();
925 }
926}
927
928/*
929 * Get writeback threshold and buffer limit for a given client.
930 */
931static void __get_memory_limit(struct dm_bufio_client *c,
932 unsigned long *threshold_buffers,
933 unsigned long *limit_buffers)
934{
935 unsigned long buffers;
936
937 if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
938 mutex_lock(&dm_bufio_clients_lock);
939 __cache_size_refresh();
940 mutex_unlock(&dm_bufio_clients_lock);
941 }
942
943 buffers = dm_bufio_cache_size_per_client >>
944 (c->sectors_per_block_bits + SECTOR_SHIFT);
945
946 if (buffers < c->minimum_buffers)
947 buffers = c->minimum_buffers;
948
949 *limit_buffers = buffers;
950 *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
951}
952
953/*
954 * Check if we're over watermark.
955 * If we are over threshold_buffers, start freeing buffers.
956 * If we're over "limit_buffers", block until we get under the limit.
957 */
958static void __check_watermark(struct dm_bufio_client *c,
959 struct list_head *write_list)
960{
961 unsigned long threshold_buffers, limit_buffers;
962
963 __get_memory_limit(c, &threshold_buffers, &limit_buffers);
964
965 while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
966 limit_buffers) {
967
968 struct dm_buffer *b = __get_unclaimed_buffer(c);
969
970 if (!b)
971 return;
972
973 __free_buffer_wake(b);
974 dm_bufio_cond_resched();
975 }
976
977 if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
978 __write_dirty_buffers_async(c, 1, write_list);
979}
980
981/*----------------------------------------------------------------
982 * Getting a buffer
983 *--------------------------------------------------------------*/
984
985static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
986 enum new_flag nf, int *need_submit,
987 struct list_head *write_list)
988{
989 struct dm_buffer *b, *new_b = NULL;
990
991 *need_submit = 0;
992
993 b = __find(c, block);
994 if (b)
995 goto found_buffer;
996
997 if (nf == NF_GET)
998 return NULL;
999
1000 new_b = __alloc_buffer_wait(c, nf);
1001 if (!new_b)
1002 return NULL;
1003
1004 /*
1005 * We've had a period where the mutex was unlocked, so need to
1006 * recheck the hash table.
1007 */
1008 b = __find(c, block);
1009 if (b) {
1010 __free_buffer_wake(new_b);
1011 goto found_buffer;
1012 }
1013
1014 __check_watermark(c, write_list);
1015
1016 b = new_b;
1017 b->hold_count = 1;
1018 b->read_error = 0;
1019 b->write_error = 0;
1020 __link_buffer(b, block, LIST_CLEAN);
1021
1022 if (nf == NF_FRESH) {
1023 b->state = 0;
1024 return b;
1025 }
1026
1027 b->state = 1 << B_READING;
1028 *need_submit = 1;
1029
1030 return b;
1031
1032found_buffer:
1033 if (nf == NF_PREFETCH)
1034 return NULL;
1035 /*
1036 * Note: it is essential that we don't wait for the buffer to be
1037 * read if dm_bufio_get function is used. Both dm_bufio_get and
1038 * dm_bufio_prefetch can be used in the driver request routine.
1039 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1040 * the same buffer, it would deadlock if we waited.
1041 */
1042 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1043 return NULL;
1044
1045 b->hold_count++;
1046 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1047 test_bit(B_WRITING, &b->state));
1048 return b;
1049}
1050
1051/*
1052 * The endio routine for reading: set the error, clear the bit and wake up
1053 * anyone waiting on the buffer.
1054 */
1055static void read_endio(struct bio *bio)
1056{
1057 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
1058
1059 b->read_error = bio->bi_error;
1060
1061 BUG_ON(!test_bit(B_READING, &b->state));
1062
1063 smp_mb__before_atomic();
1064 clear_bit(B_READING, &b->state);
1065 smp_mb__after_atomic();
1066
1067 wake_up_bit(&b->state, B_READING);
1068}
1069
1070/*
1071 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1072 * functions is similar except that dm_bufio_new doesn't read the
1073 * buffer from the disk (assuming that the caller overwrites all the data
1074 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1075 */
1076static void *new_read(struct dm_bufio_client *c, sector_t block,
1077 enum new_flag nf, struct dm_buffer **bp)
1078{
1079 int need_submit;
1080 struct dm_buffer *b;
1081
1082 LIST_HEAD(write_list);
1083
1084 dm_bufio_lock(c);
1085 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1086#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1087 if (b && b->hold_count == 1)
1088 buffer_record_stack(b);
1089#endif
1090 dm_bufio_unlock(c);
1091
1092 __flush_write_list(&write_list);
1093
1094 if (!b)
1095 return NULL;
1096
1097 if (need_submit)
1098 submit_io(b, READ, b->block, read_endio);
1099
1100 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1101
1102 if (b->read_error) {
1103 int error = b->read_error;
1104
1105 dm_bufio_release(b);
1106
1107 return ERR_PTR(error);
1108 }
1109
1110 *bp = b;
1111
1112 return b->data;
1113}
1114
1115void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1116 struct dm_buffer **bp)
1117{
1118 return new_read(c, block, NF_GET, bp);
1119}
1120EXPORT_SYMBOL_GPL(dm_bufio_get);
1121
1122void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1123 struct dm_buffer **bp)
1124{
1125 BUG_ON(dm_bufio_in_request());
1126
1127 return new_read(c, block, NF_READ, bp);
1128}
1129EXPORT_SYMBOL_GPL(dm_bufio_read);
1130
1131void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1132 struct dm_buffer **bp)
1133{
1134 BUG_ON(dm_bufio_in_request());
1135
1136 return new_read(c, block, NF_FRESH, bp);
1137}
1138EXPORT_SYMBOL_GPL(dm_bufio_new);
1139
1140void dm_bufio_prefetch(struct dm_bufio_client *c,
1141 sector_t block, unsigned n_blocks)
1142{
1143 struct blk_plug plug;
1144
1145 LIST_HEAD(write_list);
1146
1147 BUG_ON(dm_bufio_in_request());
1148
1149 blk_start_plug(&plug);
1150 dm_bufio_lock(c);
1151
1152 for (; n_blocks--; block++) {
1153 int need_submit;
1154 struct dm_buffer *b;
1155 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1156 &write_list);
1157 if (unlikely(!list_empty(&write_list))) {
1158 dm_bufio_unlock(c);
1159 blk_finish_plug(&plug);
1160 __flush_write_list(&write_list);
1161 blk_start_plug(&plug);
1162 dm_bufio_lock(c);
1163 }
1164 if (unlikely(b != NULL)) {
1165 dm_bufio_unlock(c);
1166
1167 if (need_submit)
1168 submit_io(b, READ, b->block, read_endio);
1169 dm_bufio_release(b);
1170
1171 dm_bufio_cond_resched();
1172
1173 if (!n_blocks)
1174 goto flush_plug;
1175 dm_bufio_lock(c);
1176 }
1177 }
1178
1179 dm_bufio_unlock(c);
1180
1181flush_plug:
1182 blk_finish_plug(&plug);
1183}
1184EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1185
1186void dm_bufio_release(struct dm_buffer *b)
1187{
1188 struct dm_bufio_client *c = b->c;
1189
1190 dm_bufio_lock(c);
1191
1192 BUG_ON(!b->hold_count);
1193
1194 b->hold_count--;
1195 if (!b->hold_count) {
1196 wake_up(&c->free_buffer_wait);
1197
1198 /*
1199 * If there were errors on the buffer, and the buffer is not
1200 * to be written, free the buffer. There is no point in caching
1201 * invalid buffer.
1202 */
1203 if ((b->read_error || b->write_error) &&
1204 !test_bit(B_READING, &b->state) &&
1205 !test_bit(B_WRITING, &b->state) &&
1206 !test_bit(B_DIRTY, &b->state)) {
1207 __unlink_buffer(b);
1208 __free_buffer_wake(b);
1209 }
1210 }
1211
1212 dm_bufio_unlock(c);
1213}
1214EXPORT_SYMBOL_GPL(dm_bufio_release);
1215
1216void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1217{
1218 struct dm_bufio_client *c = b->c;
1219
1220 dm_bufio_lock(c);
1221
1222 BUG_ON(test_bit(B_READING, &b->state));
1223
1224 if (!test_and_set_bit(B_DIRTY, &b->state))
1225 __relink_lru(b, LIST_DIRTY);
1226
1227 dm_bufio_unlock(c);
1228}
1229EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1230
1231void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1232{
1233 LIST_HEAD(write_list);
1234
1235 BUG_ON(dm_bufio_in_request());
1236
1237 dm_bufio_lock(c);
1238 __write_dirty_buffers_async(c, 0, &write_list);
1239 dm_bufio_unlock(c);
1240 __flush_write_list(&write_list);
1241}
1242EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1243
1244/*
1245 * For performance, it is essential that the buffers are written asynchronously
1246 * and simultaneously (so that the block layer can merge the writes) and then
1247 * waited upon.
1248 *
1249 * Finally, we flush hardware disk cache.
1250 */
1251int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1252{
1253 int a, f;
1254 unsigned long buffers_processed = 0;
1255 struct dm_buffer *b, *tmp;
1256
1257 LIST_HEAD(write_list);
1258
1259 dm_bufio_lock(c);
1260 __write_dirty_buffers_async(c, 0, &write_list);
1261 dm_bufio_unlock(c);
1262 __flush_write_list(&write_list);
1263 dm_bufio_lock(c);
1264
1265again:
1266 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1267 int dropped_lock = 0;
1268
1269 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1270 buffers_processed++;
1271
1272 BUG_ON(test_bit(B_READING, &b->state));
1273
1274 if (test_bit(B_WRITING, &b->state)) {
1275 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1276 dropped_lock = 1;
1277 b->hold_count++;
1278 dm_bufio_unlock(c);
1279 wait_on_bit_io(&b->state, B_WRITING,
1280 TASK_UNINTERRUPTIBLE);
1281 dm_bufio_lock(c);
1282 b->hold_count--;
1283 } else
1284 wait_on_bit_io(&b->state, B_WRITING,
1285 TASK_UNINTERRUPTIBLE);
1286 }
1287
1288 if (!test_bit(B_DIRTY, &b->state) &&
1289 !test_bit(B_WRITING, &b->state))
1290 __relink_lru(b, LIST_CLEAN);
1291
1292 dm_bufio_cond_resched();
1293
1294 /*
1295 * If we dropped the lock, the list is no longer consistent,
1296 * so we must restart the search.
1297 *
1298 * In the most common case, the buffer just processed is
1299 * relinked to the clean list, so we won't loop scanning the
1300 * same buffer again and again.
1301 *
1302 * This may livelock if there is another thread simultaneously
1303 * dirtying buffers, so we count the number of buffers walked
1304 * and if it exceeds the total number of buffers, it means that
1305 * someone is doing some writes simultaneously with us. In
1306 * this case, stop, dropping the lock.
1307 */
1308 if (dropped_lock)
1309 goto again;
1310 }
1311 wake_up(&c->free_buffer_wait);
1312 dm_bufio_unlock(c);
1313
1314 a = xchg(&c->async_write_error, 0);
1315 f = dm_bufio_issue_flush(c);
1316 if (a)
1317 return a;
1318
1319 return f;
1320}
1321EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1322
1323/*
1324 * Use dm-io to send and empty barrier flush the device.
1325 */
1326int dm_bufio_issue_flush(struct dm_bufio_client *c)
1327{
1328 struct dm_io_request io_req = {
1329 .bi_rw = WRITE_FLUSH,
1330 .mem.type = DM_IO_KMEM,
1331 .mem.ptr.addr = NULL,
1332 .client = c->dm_io,
1333 };
1334 struct dm_io_region io_reg = {
1335 .bdev = c->bdev,
1336 .sector = 0,
1337 .count = 0,
1338 };
1339
1340 BUG_ON(dm_bufio_in_request());
1341
1342 return dm_io(&io_req, 1, &io_reg, NULL);
1343}
1344EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1345
1346/*
1347 * We first delete any other buffer that may be at that new location.
1348 *
1349 * Then, we write the buffer to the original location if it was dirty.
1350 *
1351 * Then, if we are the only one who is holding the buffer, relink the buffer
1352 * in the hash queue for the new location.
1353 *
1354 * If there was someone else holding the buffer, we write it to the new
1355 * location but not relink it, because that other user needs to have the buffer
1356 * at the same place.
1357 */
1358void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1359{
1360 struct dm_bufio_client *c = b->c;
1361 struct dm_buffer *new;
1362
1363 BUG_ON(dm_bufio_in_request());
1364
1365 dm_bufio_lock(c);
1366
1367retry:
1368 new = __find(c, new_block);
1369 if (new) {
1370 if (new->hold_count) {
1371 __wait_for_free_buffer(c);
1372 goto retry;
1373 }
1374
1375 /*
1376 * FIXME: Is there any point waiting for a write that's going
1377 * to be overwritten in a bit?
1378 */
1379 __make_buffer_clean(new);
1380 __unlink_buffer(new);
1381 __free_buffer_wake(new);
1382 }
1383
1384 BUG_ON(!b->hold_count);
1385 BUG_ON(test_bit(B_READING, &b->state));
1386
1387 __write_dirty_buffer(b, NULL);
1388 if (b->hold_count == 1) {
1389 wait_on_bit_io(&b->state, B_WRITING,
1390 TASK_UNINTERRUPTIBLE);
1391 set_bit(B_DIRTY, &b->state);
1392 __unlink_buffer(b);
1393 __link_buffer(b, new_block, LIST_DIRTY);
1394 } else {
1395 sector_t old_block;
1396 wait_on_bit_lock_io(&b->state, B_WRITING,
1397 TASK_UNINTERRUPTIBLE);
1398 /*
1399 * Relink buffer to "new_block" so that write_callback
1400 * sees "new_block" as a block number.
1401 * After the write, link the buffer back to old_block.
1402 * All this must be done in bufio lock, so that block number
1403 * change isn't visible to other threads.
1404 */
1405 old_block = b->block;
1406 __unlink_buffer(b);
1407 __link_buffer(b, new_block, b->list_mode);
1408 submit_io(b, WRITE, new_block, write_endio);
1409 wait_on_bit_io(&b->state, B_WRITING,
1410 TASK_UNINTERRUPTIBLE);
1411 __unlink_buffer(b);
1412 __link_buffer(b, old_block, b->list_mode);
1413 }
1414
1415 dm_bufio_unlock(c);
1416 dm_bufio_release(b);
1417}
1418EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1419
1420/*
1421 * Free the given buffer.
1422 *
1423 * This is just a hint, if the buffer is in use or dirty, this function
1424 * does nothing.
1425 */
1426void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1427{
1428 struct dm_buffer *b;
1429
1430 dm_bufio_lock(c);
1431
1432 b = __find(c, block);
1433 if (b && likely(!b->hold_count) && likely(!b->state)) {
1434 __unlink_buffer(b);
1435 __free_buffer_wake(b);
1436 }
1437
1438 dm_bufio_unlock(c);
1439}
1440EXPORT_SYMBOL(dm_bufio_forget);
1441
1442void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1443{
1444 c->minimum_buffers = n;
1445}
1446EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1447
1448unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1449{
1450 return c->block_size;
1451}
1452EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1453
1454sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1455{
1456 return i_size_read(c->bdev->bd_inode) >>
1457 (SECTOR_SHIFT + c->sectors_per_block_bits);
1458}
1459EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1460
1461sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1462{
1463 return b->block;
1464}
1465EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1466
1467void *dm_bufio_get_block_data(struct dm_buffer *b)
1468{
1469 return b->data;
1470}
1471EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1472
1473void *dm_bufio_get_aux_data(struct dm_buffer *b)
1474{
1475 return b + 1;
1476}
1477EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1478
1479struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1480{
1481 return b->c;
1482}
1483EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1484
1485static void drop_buffers(struct dm_bufio_client *c)
1486{
1487 struct dm_buffer *b;
1488 int i;
1489 bool warned = false;
1490
1491 BUG_ON(dm_bufio_in_request());
1492
1493 /*
1494 * An optimization so that the buffers are not written one-by-one.
1495 */
1496 dm_bufio_write_dirty_buffers_async(c);
1497
1498 dm_bufio_lock(c);
1499
1500 while ((b = __get_unclaimed_buffer(c)))
1501 __free_buffer_wake(b);
1502
1503 for (i = 0; i < LIST_SIZE; i++)
1504 list_for_each_entry(b, &c->lru[i], lru_list) {
1505 WARN_ON(!warned);
1506 warned = true;
1507 DMERR("leaked buffer %llx, hold count %u, list %d",
1508 (unsigned long long)b->block, b->hold_count, i);
1509#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1510 print_stack_trace(&b->stack_trace, 1);
1511 b->hold_count = 0; /* mark unclaimed to avoid BUG_ON below */
1512#endif
1513 }
1514
1515#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1516 while ((b = __get_unclaimed_buffer(c)))
1517 __free_buffer_wake(b);
1518#endif
1519
1520 for (i = 0; i < LIST_SIZE; i++)
1521 BUG_ON(!list_empty(&c->lru[i]));
1522
1523 dm_bufio_unlock(c);
1524}
1525
1526/*
1527 * We may not be able to evict this buffer if IO pending or the client
1528 * is still using it. Caller is expected to know buffer is too old.
1529 *
1530 * And if GFP_NOFS is used, we must not do any I/O because we hold
1531 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1532 * rerouted to different bufio client.
1533 */
1534static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1535{
1536 if (!(gfp & __GFP_FS)) {
1537 if (test_bit(B_READING, &b->state) ||
1538 test_bit(B_WRITING, &b->state) ||
1539 test_bit(B_DIRTY, &b->state))
1540 return false;
1541 }
1542
1543 if (b->hold_count)
1544 return false;
1545
1546 __make_buffer_clean(b);
1547 __unlink_buffer(b);
1548 __free_buffer_wake(b);
1549
1550 return true;
1551}
1552
1553static unsigned get_retain_buffers(struct dm_bufio_client *c)
1554{
1555 unsigned retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes);
1556 return retain_bytes / c->block_size;
1557}
1558
1559static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1560 gfp_t gfp_mask)
1561{
1562 int l;
1563 struct dm_buffer *b, *tmp;
1564 unsigned long freed = 0;
1565 unsigned long count = nr_to_scan;
1566 unsigned retain_target = get_retain_buffers(c);
1567
1568 for (l = 0; l < LIST_SIZE; l++) {
1569 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1570 if (__try_evict_buffer(b, gfp_mask))
1571 freed++;
1572 if (!--nr_to_scan || ((count - freed) <= retain_target))
1573 return freed;
1574 dm_bufio_cond_resched();
1575 }
1576 }
1577 return freed;
1578}
1579
1580static unsigned long
1581dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1582{
1583 struct dm_bufio_client *c;
1584 unsigned long freed;
1585
1586 c = container_of(shrink, struct dm_bufio_client, shrinker);
1587 if (sc->gfp_mask & __GFP_FS)
1588 dm_bufio_lock(c);
1589 else if (!dm_bufio_trylock(c))
1590 return SHRINK_STOP;
1591
1592 freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1593 dm_bufio_unlock(c);
1594 return freed;
1595}
1596
1597static unsigned long
1598dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1599{
1600 struct dm_bufio_client *c;
1601 unsigned long count;
1602
1603 c = container_of(shrink, struct dm_bufio_client, shrinker);
1604 if (sc->gfp_mask & __GFP_FS)
1605 dm_bufio_lock(c);
1606 else if (!dm_bufio_trylock(c))
1607 return 0;
1608
1609 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1610 dm_bufio_unlock(c);
1611 return count;
1612}
1613
1614/*
1615 * Create the buffering interface
1616 */
1617struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1618 unsigned reserved_buffers, unsigned aux_size,
1619 void (*alloc_callback)(struct dm_buffer *),
1620 void (*write_callback)(struct dm_buffer *))
1621{
1622 int r;
1623 struct dm_bufio_client *c;
1624 unsigned i;
1625
1626 BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1627 (block_size & (block_size - 1)));
1628
1629 c = kzalloc(sizeof(*c), GFP_KERNEL);
1630 if (!c) {
1631 r = -ENOMEM;
1632 goto bad_client;
1633 }
1634 c->buffer_tree = RB_ROOT;
1635
1636 c->bdev = bdev;
1637 c->block_size = block_size;
1638 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1639 c->pages_per_block_bits = (__ffs(block_size) >= PAGE_SHIFT) ?
1640 __ffs(block_size) - PAGE_SHIFT : 0;
1641 c->blocks_per_page_bits = (__ffs(block_size) < PAGE_SHIFT ?
1642 PAGE_SHIFT - __ffs(block_size) : 0);
1643
1644 c->aux_size = aux_size;
1645 c->alloc_callback = alloc_callback;
1646 c->write_callback = write_callback;
1647
1648 for (i = 0; i < LIST_SIZE; i++) {
1649 INIT_LIST_HEAD(&c->lru[i]);
1650 c->n_buffers[i] = 0;
1651 }
1652
1653 mutex_init(&c->lock);
1654 INIT_LIST_HEAD(&c->reserved_buffers);
1655 c->need_reserved_buffers = reserved_buffers;
1656
1657 c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1658
1659 init_waitqueue_head(&c->free_buffer_wait);
1660 c->async_write_error = 0;
1661
1662 c->dm_io = dm_io_client_create();
1663 if (IS_ERR(c->dm_io)) {
1664 r = PTR_ERR(c->dm_io);
1665 goto bad_dm_io;
1666 }
1667
1668 mutex_lock(&dm_bufio_clients_lock);
1669 if (c->blocks_per_page_bits) {
1670 if (!DM_BUFIO_CACHE_NAME(c)) {
1671 DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1672 if (!DM_BUFIO_CACHE_NAME(c)) {
1673 r = -ENOMEM;
1674 mutex_unlock(&dm_bufio_clients_lock);
1675 goto bad_cache;
1676 }
1677 }
1678
1679 if (!DM_BUFIO_CACHE(c)) {
1680 DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1681 c->block_size,
1682 c->block_size, 0, NULL);
1683 if (!DM_BUFIO_CACHE(c)) {
1684 r = -ENOMEM;
1685 mutex_unlock(&dm_bufio_clients_lock);
1686 goto bad_cache;
1687 }
1688 }
1689 }
1690 mutex_unlock(&dm_bufio_clients_lock);
1691
1692 while (c->need_reserved_buffers) {
1693 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1694
1695 if (!b) {
1696 r = -ENOMEM;
1697 goto bad_buffer;
1698 }
1699 __free_buffer_wake(b);
1700 }
1701
1702 mutex_lock(&dm_bufio_clients_lock);
1703 dm_bufio_client_count++;
1704 list_add(&c->client_list, &dm_bufio_all_clients);
1705 __cache_size_refresh();
1706 mutex_unlock(&dm_bufio_clients_lock);
1707
1708 c->shrinker.count_objects = dm_bufio_shrink_count;
1709 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1710 c->shrinker.seeks = 1;
1711 c->shrinker.batch = 0;
1712 register_shrinker(&c->shrinker);
1713
1714 return c;
1715
1716bad_buffer:
1717bad_cache:
1718 while (!list_empty(&c->reserved_buffers)) {
1719 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1720 struct dm_buffer, lru_list);
1721 list_del(&b->lru_list);
1722 free_buffer(b);
1723 }
1724 dm_io_client_destroy(c->dm_io);
1725bad_dm_io:
1726 kfree(c);
1727bad_client:
1728 return ERR_PTR(r);
1729}
1730EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1731
1732/*
1733 * Free the buffering interface.
1734 * It is required that there are no references on any buffers.
1735 */
1736void dm_bufio_client_destroy(struct dm_bufio_client *c)
1737{
1738 unsigned i;
1739
1740 drop_buffers(c);
1741
1742 unregister_shrinker(&c->shrinker);
1743
1744 mutex_lock(&dm_bufio_clients_lock);
1745
1746 list_del(&c->client_list);
1747 dm_bufio_client_count--;
1748 __cache_size_refresh();
1749
1750 mutex_unlock(&dm_bufio_clients_lock);
1751
1752 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1753 BUG_ON(c->need_reserved_buffers);
1754
1755 while (!list_empty(&c->reserved_buffers)) {
1756 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1757 struct dm_buffer, lru_list);
1758 list_del(&b->lru_list);
1759 free_buffer(b);
1760 }
1761
1762 for (i = 0; i < LIST_SIZE; i++)
1763 if (c->n_buffers[i])
1764 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1765
1766 for (i = 0; i < LIST_SIZE; i++)
1767 BUG_ON(c->n_buffers[i]);
1768
1769 dm_io_client_destroy(c->dm_io);
1770 kfree(c);
1771}
1772EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1773
1774static unsigned get_max_age_hz(void)
1775{
1776 unsigned max_age = ACCESS_ONCE(dm_bufio_max_age);
1777
1778 if (max_age > UINT_MAX / HZ)
1779 max_age = UINT_MAX / HZ;
1780
1781 return max_age * HZ;
1782}
1783
1784static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1785{
1786 return time_after_eq(jiffies, b->last_accessed + age_hz);
1787}
1788
1789static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1790{
1791 struct dm_buffer *b, *tmp;
1792 unsigned retain_target = get_retain_buffers(c);
1793 unsigned count;
1794
1795 dm_bufio_lock(c);
1796
1797 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1798 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1799 if (count <= retain_target)
1800 break;
1801
1802 if (!older_than(b, age_hz))
1803 break;
1804
1805 if (__try_evict_buffer(b, 0))
1806 count--;
1807
1808 dm_bufio_cond_resched();
1809 }
1810
1811 dm_bufio_unlock(c);
1812}
1813
1814static void cleanup_old_buffers(void)
1815{
1816 unsigned long max_age_hz = get_max_age_hz();
1817 struct dm_bufio_client *c;
1818
1819 mutex_lock(&dm_bufio_clients_lock);
1820
1821 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1822 __evict_old_buffers(c, max_age_hz);
1823
1824 mutex_unlock(&dm_bufio_clients_lock);
1825}
1826
1827static struct workqueue_struct *dm_bufio_wq;
1828static struct delayed_work dm_bufio_work;
1829
1830static void work_fn(struct work_struct *w)
1831{
1832 cleanup_old_buffers();
1833
1834 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1835 DM_BUFIO_WORK_TIMER_SECS * HZ);
1836}
1837
1838/*----------------------------------------------------------------
1839 * Module setup
1840 *--------------------------------------------------------------*/
1841
1842/*
1843 * This is called only once for the whole dm_bufio module.
1844 * It initializes memory limit.
1845 */
1846static int __init dm_bufio_init(void)
1847{
1848 __u64 mem;
1849
1850 dm_bufio_allocated_kmem_cache = 0;
1851 dm_bufio_allocated_get_free_pages = 0;
1852 dm_bufio_allocated_vmalloc = 0;
1853 dm_bufio_current_allocated = 0;
1854
1855 memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1856 memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1857
1858 mem = (__u64)((totalram_pages - totalhigh_pages) *
1859 DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1860
1861 if (mem > ULONG_MAX)
1862 mem = ULONG_MAX;
1863
1864#ifdef CONFIG_MMU
1865 /*
1866 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1867 * in fs/proc/internal.h
1868 */
1869 if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1870 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1871#endif
1872
1873 dm_bufio_default_cache_size = mem;
1874
1875 mutex_lock(&dm_bufio_clients_lock);
1876 __cache_size_refresh();
1877 mutex_unlock(&dm_bufio_clients_lock);
1878
1879 dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1880 if (!dm_bufio_wq)
1881 return -ENOMEM;
1882
1883 INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1884 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1885 DM_BUFIO_WORK_TIMER_SECS * HZ);
1886
1887 return 0;
1888}
1889
1890/*
1891 * This is called once when unloading the dm_bufio module.
1892 */
1893static void __exit dm_bufio_exit(void)
1894{
1895 int bug = 0;
1896 int i;
1897
1898 cancel_delayed_work_sync(&dm_bufio_work);
1899 destroy_workqueue(dm_bufio_wq);
1900
1901 for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++)
1902 kmem_cache_destroy(dm_bufio_caches[i]);
1903
1904 for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1905 kfree(dm_bufio_cache_names[i]);
1906
1907 if (dm_bufio_client_count) {
1908 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1909 __func__, dm_bufio_client_count);
1910 bug = 1;
1911 }
1912
1913 if (dm_bufio_current_allocated) {
1914 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1915 __func__, dm_bufio_current_allocated);
1916 bug = 1;
1917 }
1918
1919 if (dm_bufio_allocated_get_free_pages) {
1920 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1921 __func__, dm_bufio_allocated_get_free_pages);
1922 bug = 1;
1923 }
1924
1925 if (dm_bufio_allocated_vmalloc) {
1926 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1927 __func__, dm_bufio_allocated_vmalloc);
1928 bug = 1;
1929 }
1930
1931 BUG_ON(bug);
1932}
1933
1934module_init(dm_bufio_init)
1935module_exit(dm_bufio_exit)
1936
1937module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1938MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1939
1940module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1941MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1942
1943module_param_named(retain_bytes, dm_bufio_retain_bytes, uint, S_IRUGO | S_IWUSR);
1944MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
1945
1946module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1947MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1948
1949module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1950MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1951
1952module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1953MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1954
1955module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1956MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1957
1958module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1959MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1960
1961MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1962MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1963MODULE_LICENSE("GPL");
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_KMALLOC = 1,
322 DATA_MODE_GET_FREE_PAGES = 2,
323 DATA_MODE_VMALLOC = 3,
324 DATA_MODE_LIMIT = 4
325};
326
327struct dm_buffer {
328 /* protected by the locks in dm_buffer_cache */
329 struct rb_node node;
330
331 /* immutable, so don't need protecting */
332 sector_t block;
333 void *data;
334 unsigned char data_mode; /* DATA_MODE_* */
335
336 /*
337 * These two fields are used in isolation, so do not need
338 * a surrounding lock.
339 */
340 atomic_t hold_count;
341 unsigned long last_accessed;
342
343 /*
344 * Everything else is protected by the mutex in
345 * dm_bufio_client
346 */
347 unsigned long state;
348 struct lru_entry lru;
349 unsigned char list_mode; /* LIST_* */
350 blk_status_t read_error;
351 blk_status_t write_error;
352 unsigned int dirty_start;
353 unsigned int dirty_end;
354 unsigned int write_start;
355 unsigned int write_end;
356 struct list_head write_list;
357 struct dm_bufio_client *c;
358 void (*end_io)(struct dm_buffer *b, blk_status_t bs);
359#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
360#define MAX_STACK 10
361 unsigned int stack_len;
362 unsigned long stack_entries[MAX_STACK];
363#endif
364};
365
366/*--------------------------------------------------------------*/
367
368/*
369 * The buffer cache manages buffers, particularly:
370 * - inc/dec of holder count
371 * - setting the last_accessed field
372 * - maintains clean/dirty state along with lru
373 * - selecting buffers that match predicates
374 *
375 * It does *not* handle:
376 * - allocation/freeing of buffers.
377 * - IO
378 * - Eviction or cache sizing.
379 *
380 * cache_get() and cache_put() are threadsafe, you do not need to
381 * protect these calls with a surrounding mutex. All the other
382 * methods are not threadsafe; they do use locking primitives, but
383 * only enough to ensure get/put are threadsafe.
384 */
385
386struct buffer_tree {
387 union {
388 struct rw_semaphore lock;
389 rwlock_t spinlock;
390 } u;
391 struct rb_root root;
392} ____cacheline_aligned_in_smp;
393
394struct dm_buffer_cache {
395 struct lru lru[LIST_SIZE];
396 /*
397 * We spread entries across multiple trees to reduce contention
398 * on the locks.
399 */
400 unsigned int num_locks;
401 bool no_sleep;
402 struct buffer_tree trees[];
403};
404
405static DEFINE_STATIC_KEY_FALSE(no_sleep_enabled);
406
407static inline unsigned int cache_index(sector_t block, unsigned int num_locks)
408{
409 return dm_hash_locks_index(block, num_locks);
410}
411
412static inline void cache_read_lock(struct dm_buffer_cache *bc, sector_t block)
413{
414 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
415 read_lock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
416 else
417 down_read(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
418}
419
420static inline void cache_read_unlock(struct dm_buffer_cache *bc, sector_t block)
421{
422 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
423 read_unlock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
424 else
425 up_read(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
426}
427
428static inline void cache_write_lock(struct dm_buffer_cache *bc, sector_t block)
429{
430 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
431 write_lock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
432 else
433 down_write(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
434}
435
436static inline void cache_write_unlock(struct dm_buffer_cache *bc, sector_t block)
437{
438 if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
439 write_unlock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
440 else
441 up_write(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
442}
443
444/*
445 * Sometimes we want to repeatedly get and drop locks as part of an iteration.
446 * This struct helps avoid redundant drop and gets of the same lock.
447 */
448struct lock_history {
449 struct dm_buffer_cache *cache;
450 bool write;
451 unsigned int previous;
452 unsigned int no_previous;
453};
454
455static void lh_init(struct lock_history *lh, struct dm_buffer_cache *cache, bool write)
456{
457 lh->cache = cache;
458 lh->write = write;
459 lh->no_previous = cache->num_locks;
460 lh->previous = lh->no_previous;
461}
462
463static void __lh_lock(struct lock_history *lh, unsigned int index)
464{
465 if (lh->write) {
466 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
467 write_lock_bh(&lh->cache->trees[index].u.spinlock);
468 else
469 down_write(&lh->cache->trees[index].u.lock);
470 } else {
471 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
472 read_lock_bh(&lh->cache->trees[index].u.spinlock);
473 else
474 down_read(&lh->cache->trees[index].u.lock);
475 }
476}
477
478static void __lh_unlock(struct lock_history *lh, unsigned int index)
479{
480 if (lh->write) {
481 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
482 write_unlock_bh(&lh->cache->trees[index].u.spinlock);
483 else
484 up_write(&lh->cache->trees[index].u.lock);
485 } else {
486 if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
487 read_unlock_bh(&lh->cache->trees[index].u.spinlock);
488 else
489 up_read(&lh->cache->trees[index].u.lock);
490 }
491}
492
493/*
494 * Make sure you call this since it will unlock the final lock.
495 */
496static void lh_exit(struct lock_history *lh)
497{
498 if (lh->previous != lh->no_previous) {
499 __lh_unlock(lh, lh->previous);
500 lh->previous = lh->no_previous;
501 }
502}
503
504/*
505 * Named 'next' because there is no corresponding
506 * 'up/unlock' call since it's done automatically.
507 */
508static void lh_next(struct lock_history *lh, sector_t b)
509{
510 unsigned int index = cache_index(b, lh->no_previous); /* no_previous is num_locks */
511
512 if (lh->previous != lh->no_previous) {
513 if (lh->previous != index) {
514 __lh_unlock(lh, lh->previous);
515 __lh_lock(lh, index);
516 lh->previous = index;
517 }
518 } else {
519 __lh_lock(lh, index);
520 lh->previous = index;
521 }
522}
523
524static inline struct dm_buffer *le_to_buffer(struct lru_entry *le)
525{
526 return container_of(le, struct dm_buffer, lru);
527}
528
529static struct dm_buffer *list_to_buffer(struct list_head *l)
530{
531 struct lru_entry *le = list_entry(l, struct lru_entry, list);
532
533 return le_to_buffer(le);
534}
535
536static void cache_init(struct dm_buffer_cache *bc, unsigned int num_locks, bool no_sleep)
537{
538 unsigned int i;
539
540 bc->num_locks = num_locks;
541 bc->no_sleep = no_sleep;
542
543 for (i = 0; i < bc->num_locks; i++) {
544 if (no_sleep)
545 rwlock_init(&bc->trees[i].u.spinlock);
546 else
547 init_rwsem(&bc->trees[i].u.lock);
548 bc->trees[i].root = RB_ROOT;
549 }
550
551 lru_init(&bc->lru[LIST_CLEAN]);
552 lru_init(&bc->lru[LIST_DIRTY]);
553}
554
555static void cache_destroy(struct dm_buffer_cache *bc)
556{
557 unsigned int i;
558
559 for (i = 0; i < bc->num_locks; i++)
560 WARN_ON_ONCE(!RB_EMPTY_ROOT(&bc->trees[i].root));
561
562 lru_destroy(&bc->lru[LIST_CLEAN]);
563 lru_destroy(&bc->lru[LIST_DIRTY]);
564}
565
566/*--------------*/
567
568/*
569 * not threadsafe, or racey depending how you look at it
570 */
571static inline unsigned long cache_count(struct dm_buffer_cache *bc, int list_mode)
572{
573 return bc->lru[list_mode].count;
574}
575
576static inline unsigned long cache_total(struct dm_buffer_cache *bc)
577{
578 return cache_count(bc, LIST_CLEAN) + cache_count(bc, LIST_DIRTY);
579}
580
581/*--------------*/
582
583/*
584 * Gets a specific buffer, indexed by block.
585 * If the buffer is found then its holder count will be incremented and
586 * lru_reference will be called.
587 *
588 * threadsafe
589 */
590static struct dm_buffer *__cache_get(const struct rb_root *root, sector_t block)
591{
592 struct rb_node *n = root->rb_node;
593 struct dm_buffer *b;
594
595 while (n) {
596 b = container_of(n, struct dm_buffer, node);
597
598 if (b->block == block)
599 return b;
600
601 n = block < b->block ? n->rb_left : n->rb_right;
602 }
603
604 return NULL;
605}
606
607static void __cache_inc_buffer(struct dm_buffer *b)
608{
609 atomic_inc(&b->hold_count);
610 WRITE_ONCE(b->last_accessed, jiffies);
611}
612
613static struct dm_buffer *cache_get(struct dm_buffer_cache *bc, sector_t block)
614{
615 struct dm_buffer *b;
616
617 cache_read_lock(bc, block);
618 b = __cache_get(&bc->trees[cache_index(block, bc->num_locks)].root, block);
619 if (b) {
620 lru_reference(&b->lru);
621 __cache_inc_buffer(b);
622 }
623 cache_read_unlock(bc, block);
624
625 return b;
626}
627
628/*--------------*/
629
630/*
631 * Returns true if the hold count hits zero.
632 * threadsafe
633 */
634static bool cache_put(struct dm_buffer_cache *bc, struct dm_buffer *b)
635{
636 bool r;
637
638 cache_read_lock(bc, b->block);
639 BUG_ON(!atomic_read(&b->hold_count));
640 r = atomic_dec_and_test(&b->hold_count);
641 cache_read_unlock(bc, b->block);
642
643 return r;
644}
645
646/*--------------*/
647
648typedef enum evict_result (*b_predicate)(struct dm_buffer *, void *);
649
650/*
651 * Evicts a buffer based on a predicate. The oldest buffer that
652 * matches the predicate will be selected. In addition to the
653 * predicate the hold_count of the selected buffer will be zero.
654 */
655struct evict_wrapper {
656 struct lock_history *lh;
657 b_predicate pred;
658 void *context;
659};
660
661/*
662 * Wraps the buffer predicate turning it into an lru predicate. Adds
663 * extra test for hold_count.
664 */
665static enum evict_result __evict_pred(struct lru_entry *le, void *context)
666{
667 struct evict_wrapper *w = context;
668 struct dm_buffer *b = le_to_buffer(le);
669
670 lh_next(w->lh, b->block);
671
672 if (atomic_read(&b->hold_count))
673 return ER_DONT_EVICT;
674
675 return w->pred(b, w->context);
676}
677
678static struct dm_buffer *__cache_evict(struct dm_buffer_cache *bc, int list_mode,
679 b_predicate pred, void *context,
680 struct lock_history *lh)
681{
682 struct evict_wrapper w = {.lh = lh, .pred = pred, .context = context};
683 struct lru_entry *le;
684 struct dm_buffer *b;
685
686 le = lru_evict(&bc->lru[list_mode], __evict_pred, &w, bc->no_sleep);
687 if (!le)
688 return NULL;
689
690 b = le_to_buffer(le);
691 /* __evict_pred will have locked the appropriate tree. */
692 rb_erase(&b->node, &bc->trees[cache_index(b->block, bc->num_locks)].root);
693
694 return b;
695}
696
697static struct dm_buffer *cache_evict(struct dm_buffer_cache *bc, int list_mode,
698 b_predicate pred, void *context)
699{
700 struct dm_buffer *b;
701 struct lock_history lh;
702
703 lh_init(&lh, bc, true);
704 b = __cache_evict(bc, list_mode, pred, context, &lh);
705 lh_exit(&lh);
706
707 return b;
708}
709
710/*--------------*/
711
712/*
713 * Mark a buffer as clean or dirty. Not threadsafe.
714 */
715static void cache_mark(struct dm_buffer_cache *bc, struct dm_buffer *b, int list_mode)
716{
717 cache_write_lock(bc, b->block);
718 if (list_mode != b->list_mode) {
719 lru_remove(&bc->lru[b->list_mode], &b->lru);
720 b->list_mode = list_mode;
721 lru_insert(&bc->lru[b->list_mode], &b->lru);
722 }
723 cache_write_unlock(bc, b->block);
724}
725
726/*--------------*/
727
728/*
729 * Runs through the lru associated with 'old_mode', if the predicate matches then
730 * it moves them to 'new_mode'. Not threadsafe.
731 */
732static void __cache_mark_many(struct dm_buffer_cache *bc, int old_mode, int new_mode,
733 b_predicate pred, void *context, struct lock_history *lh)
734{
735 struct lru_entry *le;
736 struct dm_buffer *b;
737 struct evict_wrapper w = {.lh = lh, .pred = pred, .context = context};
738
739 while (true) {
740 le = lru_evict(&bc->lru[old_mode], __evict_pred, &w, bc->no_sleep);
741 if (!le)
742 break;
743
744 b = le_to_buffer(le);
745 b->list_mode = new_mode;
746 lru_insert(&bc->lru[b->list_mode], &b->lru);
747 }
748}
749
750static void cache_mark_many(struct dm_buffer_cache *bc, int old_mode, int new_mode,
751 b_predicate pred, void *context)
752{
753 struct lock_history lh;
754
755 lh_init(&lh, bc, true);
756 __cache_mark_many(bc, old_mode, new_mode, pred, context, &lh);
757 lh_exit(&lh);
758}
759
760/*--------------*/
761
762/*
763 * Iterates through all clean or dirty entries calling a function for each
764 * entry. The callback may terminate the iteration early. Not threadsafe.
765 */
766
767/*
768 * Iterator functions should return one of these actions to indicate
769 * how the iteration should proceed.
770 */
771enum it_action {
772 IT_NEXT,
773 IT_COMPLETE,
774};
775
776typedef enum it_action (*iter_fn)(struct dm_buffer *b, void *context);
777
778static void __cache_iterate(struct dm_buffer_cache *bc, int list_mode,
779 iter_fn fn, void *context, struct lock_history *lh)
780{
781 struct lru *lru = &bc->lru[list_mode];
782 struct lru_entry *le, *first;
783
784 if (!lru->cursor)
785 return;
786
787 first = le = to_le(lru->cursor);
788 do {
789 struct dm_buffer *b = le_to_buffer(le);
790
791 lh_next(lh, b->block);
792
793 switch (fn(b, context)) {
794 case IT_NEXT:
795 break;
796
797 case IT_COMPLETE:
798 return;
799 }
800 cond_resched();
801
802 le = to_le(le->list.next);
803 } while (le != first);
804}
805
806static void cache_iterate(struct dm_buffer_cache *bc, int list_mode,
807 iter_fn fn, void *context)
808{
809 struct lock_history lh;
810
811 lh_init(&lh, bc, false);
812 __cache_iterate(bc, list_mode, fn, context, &lh);
813 lh_exit(&lh);
814}
815
816/*--------------*/
817
818/*
819 * Passes ownership of the buffer to the cache. Returns false if the
820 * buffer was already present (in which case ownership does not pass).
821 * eg, a race with another thread.
822 *
823 * Holder count should be 1 on insertion.
824 *
825 * Not threadsafe.
826 */
827static bool __cache_insert(struct rb_root *root, struct dm_buffer *b)
828{
829 struct rb_node **new = &root->rb_node, *parent = NULL;
830 struct dm_buffer *found;
831
832 while (*new) {
833 found = container_of(*new, struct dm_buffer, node);
834
835 if (found->block == b->block)
836 return false;
837
838 parent = *new;
839 new = b->block < found->block ?
840 &found->node.rb_left : &found->node.rb_right;
841 }
842
843 rb_link_node(&b->node, parent, new);
844 rb_insert_color(&b->node, root);
845
846 return true;
847}
848
849static bool cache_insert(struct dm_buffer_cache *bc, struct dm_buffer *b)
850{
851 bool r;
852
853 if (WARN_ON_ONCE(b->list_mode >= LIST_SIZE))
854 return false;
855
856 cache_write_lock(bc, b->block);
857 BUG_ON(atomic_read(&b->hold_count) != 1);
858 r = __cache_insert(&bc->trees[cache_index(b->block, bc->num_locks)].root, b);
859 if (r)
860 lru_insert(&bc->lru[b->list_mode], &b->lru);
861 cache_write_unlock(bc, b->block);
862
863 return r;
864}
865
866/*--------------*/
867
868/*
869 * Removes buffer from cache, ownership of the buffer passes back to the caller.
870 * Fails if the hold_count is not one (ie. the caller holds the only reference).
871 *
872 * Not threadsafe.
873 */
874static bool cache_remove(struct dm_buffer_cache *bc, struct dm_buffer *b)
875{
876 bool r;
877
878 cache_write_lock(bc, b->block);
879
880 if (atomic_read(&b->hold_count) != 1) {
881 r = false;
882 } else {
883 r = true;
884 rb_erase(&b->node, &bc->trees[cache_index(b->block, bc->num_locks)].root);
885 lru_remove(&bc->lru[b->list_mode], &b->lru);
886 }
887
888 cache_write_unlock(bc, b->block);
889
890 return r;
891}
892
893/*--------------*/
894
895typedef void (*b_release)(struct dm_buffer *);
896
897static struct dm_buffer *__find_next(struct rb_root *root, sector_t block)
898{
899 struct rb_node *n = root->rb_node;
900 struct dm_buffer *b;
901 struct dm_buffer *best = NULL;
902
903 while (n) {
904 b = container_of(n, struct dm_buffer, node);
905
906 if (b->block == block)
907 return b;
908
909 if (block <= b->block) {
910 n = n->rb_left;
911 best = b;
912 } else {
913 n = n->rb_right;
914 }
915 }
916
917 return best;
918}
919
920static void __remove_range(struct dm_buffer_cache *bc,
921 struct rb_root *root,
922 sector_t begin, sector_t end,
923 b_predicate pred, b_release release)
924{
925 struct dm_buffer *b;
926
927 while (true) {
928 cond_resched();
929
930 b = __find_next(root, begin);
931 if (!b || (b->block >= end))
932 break;
933
934 begin = b->block + 1;
935
936 if (atomic_read(&b->hold_count))
937 continue;
938
939 if (pred(b, NULL) == ER_EVICT) {
940 rb_erase(&b->node, root);
941 lru_remove(&bc->lru[b->list_mode], &b->lru);
942 release(b);
943 }
944 }
945}
946
947static void cache_remove_range(struct dm_buffer_cache *bc,
948 sector_t begin, sector_t end,
949 b_predicate pred, b_release release)
950{
951 unsigned int i;
952
953 BUG_ON(bc->no_sleep);
954 for (i = 0; i < bc->num_locks; i++) {
955 down_write(&bc->trees[i].u.lock);
956 __remove_range(bc, &bc->trees[i].root, begin, end, pred, release);
957 up_write(&bc->trees[i].u.lock);
958 }
959}
960
961/*----------------------------------------------------------------*/
962
963/*
964 * Linking of buffers:
965 * All buffers are linked to buffer_cache with their node field.
966 *
967 * Clean buffers that are not being written (B_WRITING not set)
968 * are linked to lru[LIST_CLEAN] with their lru_list field.
969 *
970 * Dirty and clean buffers that are being written are linked to
971 * lru[LIST_DIRTY] with their lru_list field. When the write
972 * finishes, the buffer cannot be relinked immediately (because we
973 * are in an interrupt context and relinking requires process
974 * context), so some clean-not-writing buffers can be held on
975 * dirty_lru too. They are later added to lru in the process
976 * context.
977 */
978struct dm_bufio_client {
979 struct block_device *bdev;
980 unsigned int block_size;
981 s8 sectors_per_block_bits;
982
983 bool no_sleep;
984 struct mutex lock;
985 spinlock_t spinlock;
986
987 int async_write_error;
988
989 void (*alloc_callback)(struct dm_buffer *buf);
990 void (*write_callback)(struct dm_buffer *buf);
991 struct kmem_cache *slab_buffer;
992 struct kmem_cache *slab_cache;
993 struct dm_io_client *dm_io;
994
995 struct list_head reserved_buffers;
996 unsigned int need_reserved_buffers;
997
998 unsigned int minimum_buffers;
999
1000 sector_t start;
1001
1002 struct shrinker *shrinker;
1003 struct work_struct shrink_work;
1004 atomic_long_t need_shrink;
1005
1006 wait_queue_head_t free_buffer_wait;
1007
1008 struct list_head client_list;
1009
1010 /*
1011 * Used by global_cleanup to sort the clients list.
1012 */
1013 unsigned long oldest_buffer;
1014
1015 struct dm_buffer_cache cache; /* must be last member */
1016};
1017
1018/*----------------------------------------------------------------*/
1019
1020#define dm_bufio_in_request() (!!current->bio_list)
1021
1022static void dm_bufio_lock(struct dm_bufio_client *c)
1023{
1024 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1025 spin_lock_bh(&c->spinlock);
1026 else
1027 mutex_lock_nested(&c->lock, dm_bufio_in_request());
1028}
1029
1030static void dm_bufio_unlock(struct dm_bufio_client *c)
1031{
1032 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1033 spin_unlock_bh(&c->spinlock);
1034 else
1035 mutex_unlock(&c->lock);
1036}
1037
1038/*----------------------------------------------------------------*/
1039
1040/*
1041 * Default cache size: available memory divided by the ratio.
1042 */
1043static unsigned long dm_bufio_default_cache_size;
1044
1045/*
1046 * Total cache size set by the user.
1047 */
1048static unsigned long dm_bufio_cache_size;
1049
1050/*
1051 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
1052 * at any time. If it disagrees, the user has changed cache size.
1053 */
1054static unsigned long dm_bufio_cache_size_latch;
1055
1056static DEFINE_SPINLOCK(global_spinlock);
1057
1058/*
1059 * Buffers are freed after this timeout
1060 */
1061static unsigned int dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
1062static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
1063
1064static unsigned long dm_bufio_peak_allocated;
1065static unsigned long dm_bufio_allocated_kmem_cache;
1066static unsigned long dm_bufio_allocated_kmalloc;
1067static unsigned long dm_bufio_allocated_get_free_pages;
1068static unsigned long dm_bufio_allocated_vmalloc;
1069static unsigned long dm_bufio_current_allocated;
1070
1071/*----------------------------------------------------------------*/
1072
1073/*
1074 * The current number of clients.
1075 */
1076static int dm_bufio_client_count;
1077
1078/*
1079 * The list of all clients.
1080 */
1081static LIST_HEAD(dm_bufio_all_clients);
1082
1083/*
1084 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
1085 */
1086static DEFINE_MUTEX(dm_bufio_clients_lock);
1087
1088static struct workqueue_struct *dm_bufio_wq;
1089static struct delayed_work dm_bufio_cleanup_old_work;
1090static struct work_struct dm_bufio_replacement_work;
1091
1092
1093#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1094static void buffer_record_stack(struct dm_buffer *b)
1095{
1096 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
1097}
1098#endif
1099
1100/*----------------------------------------------------------------*/
1101
1102static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
1103{
1104 unsigned char data_mode;
1105 long diff;
1106
1107 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
1108 &dm_bufio_allocated_kmem_cache,
1109 &dm_bufio_allocated_kmalloc,
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 (unlikely(c->block_size < PAGE_SIZE)) {
1188 *data_mode = DATA_MODE_KMALLOC;
1189 return kmalloc(c->block_size, gfp_mask | __GFP_RECLAIMABLE);
1190 }
1191
1192 if (c->block_size <= KMALLOC_MAX_SIZE &&
1193 gfp_mask & __GFP_NORETRY) {
1194 *data_mode = DATA_MODE_GET_FREE_PAGES;
1195 return (void *)__get_free_pages(gfp_mask,
1196 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
1197 }
1198
1199 *data_mode = DATA_MODE_VMALLOC;
1200
1201 return __vmalloc(c->block_size, gfp_mask);
1202}
1203
1204/*
1205 * Free buffer's data.
1206 */
1207static void free_buffer_data(struct dm_bufio_client *c,
1208 void *data, unsigned char data_mode)
1209{
1210 switch (data_mode) {
1211 case DATA_MODE_SLAB:
1212 kmem_cache_free(c->slab_cache, data);
1213 break;
1214
1215 case DATA_MODE_KMALLOC:
1216 kfree(data);
1217 break;
1218
1219 case DATA_MODE_GET_FREE_PAGES:
1220 free_pages((unsigned long)data,
1221 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
1222 break;
1223
1224 case DATA_MODE_VMALLOC:
1225 vfree(data);
1226 break;
1227
1228 default:
1229 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
1230 data_mode);
1231 BUG();
1232 }
1233}
1234
1235/*
1236 * Allocate buffer and its data.
1237 */
1238static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
1239{
1240 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
1241
1242 if (!b)
1243 return NULL;
1244
1245 b->c = c;
1246
1247 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
1248 if (!b->data) {
1249 kmem_cache_free(c->slab_buffer, b);
1250 return NULL;
1251 }
1252 adjust_total_allocated(b, false);
1253
1254#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1255 b->stack_len = 0;
1256#endif
1257 return b;
1258}
1259
1260/*
1261 * Free buffer and its data.
1262 */
1263static void free_buffer(struct dm_buffer *b)
1264{
1265 struct dm_bufio_client *c = b->c;
1266
1267 adjust_total_allocated(b, true);
1268 free_buffer_data(c, b->data, b->data_mode);
1269 kmem_cache_free(c->slab_buffer, b);
1270}
1271
1272/*
1273 *--------------------------------------------------------------------------
1274 * Submit I/O on the buffer.
1275 *
1276 * Bio interface is faster but it has some problems:
1277 * the vector list is limited (increasing this limit increases
1278 * memory-consumption per buffer, so it is not viable);
1279 *
1280 * the memory must be direct-mapped, not vmalloced;
1281 *
1282 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
1283 * it is not vmalloced, try using the bio interface.
1284 *
1285 * If the buffer is big, if it is vmalloced or if the underlying device
1286 * rejects the bio because it is too large, use dm-io layer to do the I/O.
1287 * The dm-io layer splits the I/O into multiple requests, avoiding the above
1288 * shortcomings.
1289 *--------------------------------------------------------------------------
1290 */
1291
1292/*
1293 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
1294 * that the request was handled directly with bio interface.
1295 */
1296static void dmio_complete(unsigned long error, void *context)
1297{
1298 struct dm_buffer *b = context;
1299
1300 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
1301}
1302
1303static void use_dmio(struct dm_buffer *b, enum req_op op, sector_t sector,
1304 unsigned int n_sectors, unsigned int offset,
1305 unsigned short ioprio)
1306{
1307 int r;
1308 struct dm_io_request io_req = {
1309 .bi_opf = op,
1310 .notify.fn = dmio_complete,
1311 .notify.context = b,
1312 .client = b->c->dm_io,
1313 };
1314 struct dm_io_region region = {
1315 .bdev = b->c->bdev,
1316 .sector = sector,
1317 .count = n_sectors,
1318 };
1319
1320 if (b->data_mode != DATA_MODE_VMALLOC) {
1321 io_req.mem.type = DM_IO_KMEM;
1322 io_req.mem.ptr.addr = (char *)b->data + offset;
1323 } else {
1324 io_req.mem.type = DM_IO_VMA;
1325 io_req.mem.ptr.vma = (char *)b->data + offset;
1326 }
1327
1328 r = dm_io(&io_req, 1, ®ion, NULL, ioprio);
1329 if (unlikely(r))
1330 b->end_io(b, errno_to_blk_status(r));
1331}
1332
1333static void bio_complete(struct bio *bio)
1334{
1335 struct dm_buffer *b = bio->bi_private;
1336 blk_status_t status = bio->bi_status;
1337
1338 bio_uninit(bio);
1339 kfree(bio);
1340 b->end_io(b, status);
1341}
1342
1343static void use_bio(struct dm_buffer *b, enum req_op op, sector_t sector,
1344 unsigned int n_sectors, unsigned int offset,
1345 unsigned short ioprio)
1346{
1347 struct bio *bio;
1348 char *ptr;
1349 unsigned int len;
1350
1351 bio = bio_kmalloc(1, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN);
1352 if (!bio) {
1353 use_dmio(b, op, sector, n_sectors, offset, ioprio);
1354 return;
1355 }
1356 bio_init(bio, b->c->bdev, bio->bi_inline_vecs, 1, op);
1357 bio->bi_iter.bi_sector = sector;
1358 bio->bi_end_io = bio_complete;
1359 bio->bi_private = b;
1360 bio->bi_ioprio = ioprio;
1361
1362 ptr = (char *)b->data + offset;
1363 len = n_sectors << SECTOR_SHIFT;
1364
1365 __bio_add_page(bio, virt_to_page(ptr), len, offset_in_page(ptr));
1366
1367 submit_bio(bio);
1368}
1369
1370static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
1371{
1372 sector_t sector;
1373
1374 if (likely(c->sectors_per_block_bits >= 0))
1375 sector = block << c->sectors_per_block_bits;
1376 else
1377 sector = block * (c->block_size >> SECTOR_SHIFT);
1378 sector += c->start;
1379
1380 return sector;
1381}
1382
1383static void submit_io(struct dm_buffer *b, enum req_op op, unsigned short ioprio,
1384 void (*end_io)(struct dm_buffer *, blk_status_t))
1385{
1386 unsigned int n_sectors;
1387 sector_t sector;
1388 unsigned int offset, end;
1389
1390 b->end_io = end_io;
1391
1392 sector = block_to_sector(b->c, b->block);
1393
1394 if (op != REQ_OP_WRITE) {
1395 n_sectors = b->c->block_size >> SECTOR_SHIFT;
1396 offset = 0;
1397 } else {
1398 if (b->c->write_callback)
1399 b->c->write_callback(b);
1400 offset = b->write_start;
1401 end = b->write_end;
1402 offset &= -DM_BUFIO_WRITE_ALIGN;
1403 end += DM_BUFIO_WRITE_ALIGN - 1;
1404 end &= -DM_BUFIO_WRITE_ALIGN;
1405 if (unlikely(end > b->c->block_size))
1406 end = b->c->block_size;
1407
1408 sector += offset >> SECTOR_SHIFT;
1409 n_sectors = (end - offset) >> SECTOR_SHIFT;
1410 }
1411
1412 if (b->data_mode != DATA_MODE_VMALLOC)
1413 use_bio(b, op, sector, n_sectors, offset, ioprio);
1414 else
1415 use_dmio(b, op, sector, n_sectors, offset, ioprio);
1416}
1417
1418/*
1419 *--------------------------------------------------------------
1420 * Writing dirty buffers
1421 *--------------------------------------------------------------
1422 */
1423
1424/*
1425 * The endio routine for write.
1426 *
1427 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
1428 * it.
1429 */
1430static void write_endio(struct dm_buffer *b, blk_status_t status)
1431{
1432 b->write_error = status;
1433 if (unlikely(status)) {
1434 struct dm_bufio_client *c = b->c;
1435
1436 (void)cmpxchg(&c->async_write_error, 0,
1437 blk_status_to_errno(status));
1438 }
1439
1440 BUG_ON(!test_bit(B_WRITING, &b->state));
1441
1442 smp_mb__before_atomic();
1443 clear_bit(B_WRITING, &b->state);
1444 smp_mb__after_atomic();
1445
1446 wake_up_bit(&b->state, B_WRITING);
1447}
1448
1449/*
1450 * Initiate a write on a dirty buffer, but don't wait for it.
1451 *
1452 * - If the buffer is not dirty, exit.
1453 * - If there some previous write going on, wait for it to finish (we can't
1454 * have two writes on the same buffer simultaneously).
1455 * - Submit our write and don't wait on it. We set B_WRITING indicating
1456 * that there is a write in progress.
1457 */
1458static void __write_dirty_buffer(struct dm_buffer *b,
1459 struct list_head *write_list)
1460{
1461 if (!test_bit(B_DIRTY, &b->state))
1462 return;
1463
1464 clear_bit(B_DIRTY, &b->state);
1465 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
1466
1467 b->write_start = b->dirty_start;
1468 b->write_end = b->dirty_end;
1469
1470 if (!write_list)
1471 submit_io(b, REQ_OP_WRITE, IOPRIO_DEFAULT, write_endio);
1472 else
1473 list_add_tail(&b->write_list, write_list);
1474}
1475
1476static void __flush_write_list(struct list_head *write_list)
1477{
1478 struct blk_plug plug;
1479
1480 blk_start_plug(&plug);
1481 while (!list_empty(write_list)) {
1482 struct dm_buffer *b =
1483 list_entry(write_list->next, struct dm_buffer, write_list);
1484 list_del(&b->write_list);
1485 submit_io(b, REQ_OP_WRITE, IOPRIO_DEFAULT, write_endio);
1486 cond_resched();
1487 }
1488 blk_finish_plug(&plug);
1489}
1490
1491/*
1492 * Wait until any activity on the buffer finishes. Possibly write the
1493 * buffer if it is dirty. When this function finishes, there is no I/O
1494 * running on the buffer and the buffer is not dirty.
1495 */
1496static void __make_buffer_clean(struct dm_buffer *b)
1497{
1498 BUG_ON(atomic_read(&b->hold_count));
1499
1500 /* smp_load_acquire() pairs with read_endio()'s smp_mb__before_atomic() */
1501 if (!smp_load_acquire(&b->state)) /* fast case */
1502 return;
1503
1504 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1505 __write_dirty_buffer(b, NULL);
1506 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
1507}
1508
1509static enum evict_result is_clean(struct dm_buffer *b, void *context)
1510{
1511 struct dm_bufio_client *c = context;
1512
1513 /* These should never happen */
1514 if (WARN_ON_ONCE(test_bit(B_WRITING, &b->state)))
1515 return ER_DONT_EVICT;
1516 if (WARN_ON_ONCE(test_bit(B_DIRTY, &b->state)))
1517 return ER_DONT_EVICT;
1518 if (WARN_ON_ONCE(b->list_mode != LIST_CLEAN))
1519 return ER_DONT_EVICT;
1520
1521 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep &&
1522 unlikely(test_bit(B_READING, &b->state)))
1523 return ER_DONT_EVICT;
1524
1525 return ER_EVICT;
1526}
1527
1528static enum evict_result is_dirty(struct dm_buffer *b, void *context)
1529{
1530 /* These should never happen */
1531 if (WARN_ON_ONCE(test_bit(B_READING, &b->state)))
1532 return ER_DONT_EVICT;
1533 if (WARN_ON_ONCE(b->list_mode != LIST_DIRTY))
1534 return ER_DONT_EVICT;
1535
1536 return ER_EVICT;
1537}
1538
1539/*
1540 * Find some buffer that is not held by anybody, clean it, unlink it and
1541 * return it.
1542 */
1543static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
1544{
1545 struct dm_buffer *b;
1546
1547 b = cache_evict(&c->cache, LIST_CLEAN, is_clean, c);
1548 if (b) {
1549 /* this also waits for pending reads */
1550 __make_buffer_clean(b);
1551 return b;
1552 }
1553
1554 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1555 return NULL;
1556
1557 b = cache_evict(&c->cache, LIST_DIRTY, is_dirty, NULL);
1558 if (b) {
1559 __make_buffer_clean(b);
1560 return b;
1561 }
1562
1563 return NULL;
1564}
1565
1566/*
1567 * Wait until some other threads free some buffer or release hold count on
1568 * some buffer.
1569 *
1570 * This function is entered with c->lock held, drops it and regains it
1571 * before exiting.
1572 */
1573static void __wait_for_free_buffer(struct dm_bufio_client *c)
1574{
1575 DECLARE_WAITQUEUE(wait, current);
1576
1577 add_wait_queue(&c->free_buffer_wait, &wait);
1578 set_current_state(TASK_UNINTERRUPTIBLE);
1579 dm_bufio_unlock(c);
1580
1581 /*
1582 * It's possible to miss a wake up event since we don't always
1583 * hold c->lock when wake_up is called. So we have a timeout here,
1584 * just in case.
1585 */
1586 io_schedule_timeout(5 * HZ);
1587
1588 remove_wait_queue(&c->free_buffer_wait, &wait);
1589
1590 dm_bufio_lock(c);
1591}
1592
1593enum new_flag {
1594 NF_FRESH = 0,
1595 NF_READ = 1,
1596 NF_GET = 2,
1597 NF_PREFETCH = 3
1598};
1599
1600/*
1601 * Allocate a new buffer. If the allocation is not possible, wait until
1602 * some other thread frees a buffer.
1603 *
1604 * May drop the lock and regain it.
1605 */
1606static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
1607{
1608 struct dm_buffer *b;
1609 bool tried_noio_alloc = false;
1610
1611 /*
1612 * dm-bufio is resistant to allocation failures (it just keeps
1613 * one buffer reserved in cases all the allocations fail).
1614 * So set flags to not try too hard:
1615 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
1616 * mutex and wait ourselves.
1617 * __GFP_NORETRY: don't retry and rather return failure
1618 * __GFP_NOMEMALLOC: don't use emergency reserves
1619 * __GFP_NOWARN: don't print a warning in case of failure
1620 *
1621 * For debugging, if we set the cache size to 1, no new buffers will
1622 * be allocated.
1623 */
1624 while (1) {
1625 if (dm_bufio_cache_size_latch != 1) {
1626 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
1627 if (b)
1628 return b;
1629 }
1630
1631 if (nf == NF_PREFETCH)
1632 return NULL;
1633
1634 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
1635 dm_bufio_unlock(c);
1636 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
1637 dm_bufio_lock(c);
1638 if (b)
1639 return b;
1640 tried_noio_alloc = true;
1641 }
1642
1643 if (!list_empty(&c->reserved_buffers)) {
1644 b = list_to_buffer(c->reserved_buffers.next);
1645 list_del(&b->lru.list);
1646 c->need_reserved_buffers++;
1647
1648 return b;
1649 }
1650
1651 b = __get_unclaimed_buffer(c);
1652 if (b)
1653 return b;
1654
1655 __wait_for_free_buffer(c);
1656 }
1657}
1658
1659static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
1660{
1661 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
1662
1663 if (!b)
1664 return NULL;
1665
1666 if (c->alloc_callback)
1667 c->alloc_callback(b);
1668
1669 return b;
1670}
1671
1672/*
1673 * Free a buffer and wake other threads waiting for free buffers.
1674 */
1675static void __free_buffer_wake(struct dm_buffer *b)
1676{
1677 struct dm_bufio_client *c = b->c;
1678
1679 b->block = -1;
1680 if (!c->need_reserved_buffers)
1681 free_buffer(b);
1682 else {
1683 list_add(&b->lru.list, &c->reserved_buffers);
1684 c->need_reserved_buffers--;
1685 }
1686
1687 /*
1688 * We hold the bufio lock here, so no one can add entries to the
1689 * wait queue anyway.
1690 */
1691 if (unlikely(waitqueue_active(&c->free_buffer_wait)))
1692 wake_up(&c->free_buffer_wait);
1693}
1694
1695static enum evict_result cleaned(struct dm_buffer *b, void *context)
1696{
1697 if (WARN_ON_ONCE(test_bit(B_READING, &b->state)))
1698 return ER_DONT_EVICT; /* should never happen */
1699
1700 if (test_bit(B_DIRTY, &b->state) || test_bit(B_WRITING, &b->state))
1701 return ER_DONT_EVICT;
1702 else
1703 return ER_EVICT;
1704}
1705
1706static void __move_clean_buffers(struct dm_bufio_client *c)
1707{
1708 cache_mark_many(&c->cache, LIST_DIRTY, LIST_CLEAN, cleaned, NULL);
1709}
1710
1711struct write_context {
1712 int no_wait;
1713 struct list_head *write_list;
1714};
1715
1716static enum it_action write_one(struct dm_buffer *b, void *context)
1717{
1718 struct write_context *wc = context;
1719
1720 if (wc->no_wait && test_bit(B_WRITING, &b->state))
1721 return IT_COMPLETE;
1722
1723 __write_dirty_buffer(b, wc->write_list);
1724 return IT_NEXT;
1725}
1726
1727static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
1728 struct list_head *write_list)
1729{
1730 struct write_context wc = {.no_wait = no_wait, .write_list = write_list};
1731
1732 __move_clean_buffers(c);
1733 cache_iterate(&c->cache, LIST_DIRTY, write_one, &wc);
1734}
1735
1736/*
1737 * Check if we're over watermark.
1738 * If we are over threshold_buffers, start freeing buffers.
1739 * If we're over "limit_buffers", block until we get under the limit.
1740 */
1741static void __check_watermark(struct dm_bufio_client *c,
1742 struct list_head *write_list)
1743{
1744 if (cache_count(&c->cache, LIST_DIRTY) >
1745 cache_count(&c->cache, LIST_CLEAN) * DM_BUFIO_WRITEBACK_RATIO)
1746 __write_dirty_buffers_async(c, 1, write_list);
1747}
1748
1749/*
1750 *--------------------------------------------------------------
1751 * Getting a buffer
1752 *--------------------------------------------------------------
1753 */
1754
1755static void cache_put_and_wake(struct dm_bufio_client *c, struct dm_buffer *b)
1756{
1757 /*
1758 * Relying on waitqueue_active() is racey, but we sleep
1759 * with schedule_timeout anyway.
1760 */
1761 if (cache_put(&c->cache, b) &&
1762 unlikely(waitqueue_active(&c->free_buffer_wait)))
1763 wake_up(&c->free_buffer_wait);
1764}
1765
1766/*
1767 * This assumes you have already checked the cache to see if the buffer
1768 * is already present (it will recheck after dropping the lock for allocation).
1769 */
1770static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
1771 enum new_flag nf, int *need_submit,
1772 struct list_head *write_list)
1773{
1774 struct dm_buffer *b, *new_b = NULL;
1775
1776 *need_submit = 0;
1777
1778 /* This can't be called with NF_GET */
1779 if (WARN_ON_ONCE(nf == NF_GET))
1780 return NULL;
1781
1782 new_b = __alloc_buffer_wait(c, nf);
1783 if (!new_b)
1784 return NULL;
1785
1786 /*
1787 * We've had a period where the mutex was unlocked, so need to
1788 * recheck the buffer tree.
1789 */
1790 b = cache_get(&c->cache, block);
1791 if (b) {
1792 __free_buffer_wake(new_b);
1793 goto found_buffer;
1794 }
1795
1796 __check_watermark(c, write_list);
1797
1798 b = new_b;
1799 atomic_set(&b->hold_count, 1);
1800 WRITE_ONCE(b->last_accessed, jiffies);
1801 b->block = block;
1802 b->read_error = 0;
1803 b->write_error = 0;
1804 b->list_mode = LIST_CLEAN;
1805
1806 if (nf == NF_FRESH)
1807 b->state = 0;
1808 else {
1809 b->state = 1 << B_READING;
1810 *need_submit = 1;
1811 }
1812
1813 /*
1814 * We mustn't insert into the cache until the B_READING state
1815 * is set. Otherwise another thread could get it and use
1816 * it before it had been read.
1817 */
1818 cache_insert(&c->cache, b);
1819
1820 return b;
1821
1822found_buffer:
1823 if (nf == NF_PREFETCH) {
1824 cache_put_and_wake(c, b);
1825 return NULL;
1826 }
1827
1828 /*
1829 * Note: it is essential that we don't wait for the buffer to be
1830 * read if dm_bufio_get function is used. Both dm_bufio_get and
1831 * dm_bufio_prefetch can be used in the driver request routine.
1832 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1833 * the same buffer, it would deadlock if we waited.
1834 */
1835 if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state))) {
1836 cache_put_and_wake(c, b);
1837 return NULL;
1838 }
1839
1840 return b;
1841}
1842
1843/*
1844 * The endio routine for reading: set the error, clear the bit and wake up
1845 * anyone waiting on the buffer.
1846 */
1847static void read_endio(struct dm_buffer *b, blk_status_t status)
1848{
1849 b->read_error = status;
1850
1851 BUG_ON(!test_bit(B_READING, &b->state));
1852
1853 smp_mb__before_atomic();
1854 clear_bit(B_READING, &b->state);
1855 smp_mb__after_atomic();
1856
1857 wake_up_bit(&b->state, B_READING);
1858}
1859
1860/*
1861 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1862 * functions is similar except that dm_bufio_new doesn't read the
1863 * buffer from the disk (assuming that the caller overwrites all the data
1864 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1865 */
1866static void *new_read(struct dm_bufio_client *c, sector_t block,
1867 enum new_flag nf, struct dm_buffer **bp,
1868 unsigned short ioprio)
1869{
1870 int need_submit = 0;
1871 struct dm_buffer *b;
1872
1873 LIST_HEAD(write_list);
1874
1875 *bp = NULL;
1876
1877 /*
1878 * Fast path, hopefully the block is already in the cache. No need
1879 * to get the client lock for this.
1880 */
1881 b = cache_get(&c->cache, block);
1882 if (b) {
1883 if (nf == NF_PREFETCH) {
1884 cache_put_and_wake(c, b);
1885 return NULL;
1886 }
1887
1888 /*
1889 * Note: it is essential that we don't wait for the buffer to be
1890 * read if dm_bufio_get function is used. Both dm_bufio_get and
1891 * dm_bufio_prefetch can be used in the driver request routine.
1892 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1893 * the same buffer, it would deadlock if we waited.
1894 */
1895 if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state))) {
1896 cache_put_and_wake(c, b);
1897 return NULL;
1898 }
1899 }
1900
1901 if (!b) {
1902 if (nf == NF_GET)
1903 return NULL;
1904
1905 dm_bufio_lock(c);
1906 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1907 dm_bufio_unlock(c);
1908 }
1909
1910#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1911 if (b && (atomic_read(&b->hold_count) == 1))
1912 buffer_record_stack(b);
1913#endif
1914
1915 __flush_write_list(&write_list);
1916
1917 if (!b)
1918 return NULL;
1919
1920 if (need_submit)
1921 submit_io(b, REQ_OP_READ, ioprio, read_endio);
1922
1923 if (nf != NF_GET) /* we already tested this condition above */
1924 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1925
1926 if (b->read_error) {
1927 int error = blk_status_to_errno(b->read_error);
1928
1929 dm_bufio_release(b);
1930
1931 return ERR_PTR(error);
1932 }
1933
1934 *bp = b;
1935
1936 return b->data;
1937}
1938
1939void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1940 struct dm_buffer **bp)
1941{
1942 return new_read(c, block, NF_GET, bp, IOPRIO_DEFAULT);
1943}
1944EXPORT_SYMBOL_GPL(dm_bufio_get);
1945
1946static void *__dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1947 struct dm_buffer **bp, unsigned short ioprio)
1948{
1949 if (WARN_ON_ONCE(dm_bufio_in_request()))
1950 return ERR_PTR(-EINVAL);
1951
1952 return new_read(c, block, NF_READ, bp, ioprio);
1953}
1954
1955void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1956 struct dm_buffer **bp)
1957{
1958 return __dm_bufio_read(c, block, bp, IOPRIO_DEFAULT);
1959}
1960EXPORT_SYMBOL_GPL(dm_bufio_read);
1961
1962void *dm_bufio_read_with_ioprio(struct dm_bufio_client *c, sector_t block,
1963 struct dm_buffer **bp, unsigned short ioprio)
1964{
1965 return __dm_bufio_read(c, block, bp, ioprio);
1966}
1967EXPORT_SYMBOL_GPL(dm_bufio_read_with_ioprio);
1968
1969void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1970 struct dm_buffer **bp)
1971{
1972 if (WARN_ON_ONCE(dm_bufio_in_request()))
1973 return ERR_PTR(-EINVAL);
1974
1975 return new_read(c, block, NF_FRESH, bp, IOPRIO_DEFAULT);
1976}
1977EXPORT_SYMBOL_GPL(dm_bufio_new);
1978
1979static void __dm_bufio_prefetch(struct dm_bufio_client *c,
1980 sector_t block, unsigned int n_blocks,
1981 unsigned short ioprio)
1982{
1983 struct blk_plug plug;
1984
1985 LIST_HEAD(write_list);
1986
1987 if (WARN_ON_ONCE(dm_bufio_in_request()))
1988 return; /* should never happen */
1989
1990 blk_start_plug(&plug);
1991
1992 for (; n_blocks--; block++) {
1993 int need_submit;
1994 struct dm_buffer *b;
1995
1996 b = cache_get(&c->cache, block);
1997 if (b) {
1998 /* already in cache */
1999 cache_put_and_wake(c, b);
2000 continue;
2001 }
2002
2003 dm_bufio_lock(c);
2004 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
2005 &write_list);
2006 if (unlikely(!list_empty(&write_list))) {
2007 dm_bufio_unlock(c);
2008 blk_finish_plug(&plug);
2009 __flush_write_list(&write_list);
2010 blk_start_plug(&plug);
2011 dm_bufio_lock(c);
2012 }
2013 if (unlikely(b != NULL)) {
2014 dm_bufio_unlock(c);
2015
2016 if (need_submit)
2017 submit_io(b, REQ_OP_READ, ioprio, read_endio);
2018 dm_bufio_release(b);
2019
2020 cond_resched();
2021
2022 if (!n_blocks)
2023 goto flush_plug;
2024 dm_bufio_lock(c);
2025 }
2026 dm_bufio_unlock(c);
2027 }
2028
2029flush_plug:
2030 blk_finish_plug(&plug);
2031}
2032
2033void dm_bufio_prefetch(struct dm_bufio_client *c, sector_t block, unsigned int n_blocks)
2034{
2035 return __dm_bufio_prefetch(c, block, n_blocks, IOPRIO_DEFAULT);
2036}
2037EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
2038
2039void dm_bufio_prefetch_with_ioprio(struct dm_bufio_client *c, sector_t block,
2040 unsigned int n_blocks, unsigned short ioprio)
2041{
2042 return __dm_bufio_prefetch(c, block, n_blocks, ioprio);
2043}
2044EXPORT_SYMBOL_GPL(dm_bufio_prefetch_with_ioprio);
2045
2046void dm_bufio_release(struct dm_buffer *b)
2047{
2048 struct dm_bufio_client *c = b->c;
2049
2050 /*
2051 * If there were errors on the buffer, and the buffer is not
2052 * to be written, free the buffer. There is no point in caching
2053 * invalid buffer.
2054 */
2055 if ((b->read_error || b->write_error) &&
2056 !test_bit_acquire(B_READING, &b->state) &&
2057 !test_bit(B_WRITING, &b->state) &&
2058 !test_bit(B_DIRTY, &b->state)) {
2059 dm_bufio_lock(c);
2060
2061 /* cache remove can fail if there are other holders */
2062 if (cache_remove(&c->cache, b)) {
2063 __free_buffer_wake(b);
2064 dm_bufio_unlock(c);
2065 return;
2066 }
2067
2068 dm_bufio_unlock(c);
2069 }
2070
2071 cache_put_and_wake(c, b);
2072}
2073EXPORT_SYMBOL_GPL(dm_bufio_release);
2074
2075void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
2076 unsigned int start, unsigned int end)
2077{
2078 struct dm_bufio_client *c = b->c;
2079
2080 BUG_ON(start >= end);
2081 BUG_ON(end > b->c->block_size);
2082
2083 dm_bufio_lock(c);
2084
2085 BUG_ON(test_bit(B_READING, &b->state));
2086
2087 if (!test_and_set_bit(B_DIRTY, &b->state)) {
2088 b->dirty_start = start;
2089 b->dirty_end = end;
2090 cache_mark(&c->cache, b, LIST_DIRTY);
2091 } else {
2092 if (start < b->dirty_start)
2093 b->dirty_start = start;
2094 if (end > b->dirty_end)
2095 b->dirty_end = end;
2096 }
2097
2098 dm_bufio_unlock(c);
2099}
2100EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
2101
2102void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
2103{
2104 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
2105}
2106EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
2107
2108void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
2109{
2110 LIST_HEAD(write_list);
2111
2112 if (WARN_ON_ONCE(dm_bufio_in_request()))
2113 return; /* should never happen */
2114
2115 dm_bufio_lock(c);
2116 __write_dirty_buffers_async(c, 0, &write_list);
2117 dm_bufio_unlock(c);
2118 __flush_write_list(&write_list);
2119}
2120EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
2121
2122/*
2123 * For performance, it is essential that the buffers are written asynchronously
2124 * and simultaneously (so that the block layer can merge the writes) and then
2125 * waited upon.
2126 *
2127 * Finally, we flush hardware disk cache.
2128 */
2129static bool is_writing(struct lru_entry *e, void *context)
2130{
2131 struct dm_buffer *b = le_to_buffer(e);
2132
2133 return test_bit(B_WRITING, &b->state);
2134}
2135
2136int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
2137{
2138 int a, f;
2139 unsigned long nr_buffers;
2140 struct lru_entry *e;
2141 struct lru_iter it;
2142
2143 LIST_HEAD(write_list);
2144
2145 dm_bufio_lock(c);
2146 __write_dirty_buffers_async(c, 0, &write_list);
2147 dm_bufio_unlock(c);
2148 __flush_write_list(&write_list);
2149 dm_bufio_lock(c);
2150
2151 nr_buffers = cache_count(&c->cache, LIST_DIRTY);
2152 lru_iter_begin(&c->cache.lru[LIST_DIRTY], &it);
2153 while ((e = lru_iter_next(&it, is_writing, c))) {
2154 struct dm_buffer *b = le_to_buffer(e);
2155 __cache_inc_buffer(b);
2156
2157 BUG_ON(test_bit(B_READING, &b->state));
2158
2159 if (nr_buffers) {
2160 nr_buffers--;
2161 dm_bufio_unlock(c);
2162 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
2163 dm_bufio_lock(c);
2164 } else {
2165 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
2166 }
2167
2168 if (!test_bit(B_DIRTY, &b->state) && !test_bit(B_WRITING, &b->state))
2169 cache_mark(&c->cache, b, LIST_CLEAN);
2170
2171 cache_put_and_wake(c, b);
2172
2173 cond_resched();
2174 }
2175 lru_iter_end(&it);
2176
2177 wake_up(&c->free_buffer_wait);
2178 dm_bufio_unlock(c);
2179
2180 a = xchg(&c->async_write_error, 0);
2181 f = dm_bufio_issue_flush(c);
2182 if (a)
2183 return a;
2184
2185 return f;
2186}
2187EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
2188
2189/*
2190 * Use dm-io to send an empty barrier to flush the device.
2191 */
2192int dm_bufio_issue_flush(struct dm_bufio_client *c)
2193{
2194 struct dm_io_request io_req = {
2195 .bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
2196 .mem.type = DM_IO_KMEM,
2197 .mem.ptr.addr = NULL,
2198 .client = c->dm_io,
2199 };
2200 struct dm_io_region io_reg = {
2201 .bdev = c->bdev,
2202 .sector = 0,
2203 .count = 0,
2204 };
2205
2206 if (WARN_ON_ONCE(dm_bufio_in_request()))
2207 return -EINVAL;
2208
2209 return dm_io(&io_req, 1, &io_reg, NULL, IOPRIO_DEFAULT);
2210}
2211EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
2212
2213/*
2214 * Use dm-io to send a discard request to flush the device.
2215 */
2216int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
2217{
2218 struct dm_io_request io_req = {
2219 .bi_opf = REQ_OP_DISCARD | REQ_SYNC,
2220 .mem.type = DM_IO_KMEM,
2221 .mem.ptr.addr = NULL,
2222 .client = c->dm_io,
2223 };
2224 struct dm_io_region io_reg = {
2225 .bdev = c->bdev,
2226 .sector = block_to_sector(c, block),
2227 .count = block_to_sector(c, count),
2228 };
2229
2230 if (WARN_ON_ONCE(dm_bufio_in_request()))
2231 return -EINVAL; /* discards are optional */
2232
2233 return dm_io(&io_req, 1, &io_reg, NULL, IOPRIO_DEFAULT);
2234}
2235EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
2236
2237static bool forget_buffer(struct dm_bufio_client *c, sector_t block)
2238{
2239 struct dm_buffer *b;
2240
2241 b = cache_get(&c->cache, block);
2242 if (b) {
2243 if (likely(!smp_load_acquire(&b->state))) {
2244 if (cache_remove(&c->cache, b))
2245 __free_buffer_wake(b);
2246 else
2247 cache_put_and_wake(c, b);
2248 } else {
2249 cache_put_and_wake(c, b);
2250 }
2251 }
2252
2253 return b ? true : false;
2254}
2255
2256/*
2257 * Free the given buffer.
2258 *
2259 * This is just a hint, if the buffer is in use or dirty, this function
2260 * does nothing.
2261 */
2262void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
2263{
2264 dm_bufio_lock(c);
2265 forget_buffer(c, block);
2266 dm_bufio_unlock(c);
2267}
2268EXPORT_SYMBOL_GPL(dm_bufio_forget);
2269
2270static enum evict_result idle(struct dm_buffer *b, void *context)
2271{
2272 return b->state ? ER_DONT_EVICT : ER_EVICT;
2273}
2274
2275void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
2276{
2277 dm_bufio_lock(c);
2278 cache_remove_range(&c->cache, block, block + n_blocks, idle, __free_buffer_wake);
2279 dm_bufio_unlock(c);
2280}
2281EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
2282
2283void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned int n)
2284{
2285 c->minimum_buffers = n;
2286}
2287EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
2288
2289unsigned int dm_bufio_get_block_size(struct dm_bufio_client *c)
2290{
2291 return c->block_size;
2292}
2293EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
2294
2295sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
2296{
2297 sector_t s = bdev_nr_sectors(c->bdev);
2298
2299 if (s >= c->start)
2300 s -= c->start;
2301 else
2302 s = 0;
2303 if (likely(c->sectors_per_block_bits >= 0))
2304 s >>= c->sectors_per_block_bits;
2305 else
2306 sector_div(s, c->block_size >> SECTOR_SHIFT);
2307 return s;
2308}
2309EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
2310
2311struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
2312{
2313 return c->dm_io;
2314}
2315EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
2316
2317sector_t dm_bufio_get_block_number(struct dm_buffer *b)
2318{
2319 return b->block;
2320}
2321EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
2322
2323void *dm_bufio_get_block_data(struct dm_buffer *b)
2324{
2325 return b->data;
2326}
2327EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
2328
2329void *dm_bufio_get_aux_data(struct dm_buffer *b)
2330{
2331 return b + 1;
2332}
2333EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
2334
2335struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
2336{
2337 return b->c;
2338}
2339EXPORT_SYMBOL_GPL(dm_bufio_get_client);
2340
2341static enum it_action warn_leak(struct dm_buffer *b, void *context)
2342{
2343 bool *warned = context;
2344
2345 WARN_ON(!(*warned));
2346 *warned = true;
2347 DMERR("leaked buffer %llx, hold count %u, list %d",
2348 (unsigned long long)b->block, atomic_read(&b->hold_count), b->list_mode);
2349#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
2350 stack_trace_print(b->stack_entries, b->stack_len, 1);
2351 /* mark unclaimed to avoid WARN_ON at end of drop_buffers() */
2352 atomic_set(&b->hold_count, 0);
2353#endif
2354 return IT_NEXT;
2355}
2356
2357static void drop_buffers(struct dm_bufio_client *c)
2358{
2359 int i;
2360 struct dm_buffer *b;
2361
2362 if (WARN_ON(dm_bufio_in_request()))
2363 return; /* should never happen */
2364
2365 /*
2366 * An optimization so that the buffers are not written one-by-one.
2367 */
2368 dm_bufio_write_dirty_buffers_async(c);
2369
2370 dm_bufio_lock(c);
2371
2372 while ((b = __get_unclaimed_buffer(c)))
2373 __free_buffer_wake(b);
2374
2375 for (i = 0; i < LIST_SIZE; i++) {
2376 bool warned = false;
2377
2378 cache_iterate(&c->cache, i, warn_leak, &warned);
2379 }
2380
2381#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
2382 while ((b = __get_unclaimed_buffer(c)))
2383 __free_buffer_wake(b);
2384#endif
2385
2386 for (i = 0; i < LIST_SIZE; i++)
2387 WARN_ON(cache_count(&c->cache, i));
2388
2389 dm_bufio_unlock(c);
2390}
2391
2392static unsigned long get_retain_buffers(struct dm_bufio_client *c)
2393{
2394 unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
2395
2396 if (likely(c->sectors_per_block_bits >= 0))
2397 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
2398 else
2399 retain_bytes /= c->block_size;
2400
2401 return retain_bytes;
2402}
2403
2404static void __scan(struct dm_bufio_client *c)
2405{
2406 int l;
2407 struct dm_buffer *b;
2408 unsigned long freed = 0;
2409 unsigned long retain_target = get_retain_buffers(c);
2410 unsigned long count = cache_total(&c->cache);
2411
2412 for (l = 0; l < LIST_SIZE; l++) {
2413 while (true) {
2414 if (count - freed <= retain_target)
2415 atomic_long_set(&c->need_shrink, 0);
2416 if (!atomic_long_read(&c->need_shrink))
2417 break;
2418
2419 b = cache_evict(&c->cache, l,
2420 l == LIST_CLEAN ? is_clean : is_dirty, c);
2421 if (!b)
2422 break;
2423
2424 __make_buffer_clean(b);
2425 __free_buffer_wake(b);
2426
2427 atomic_long_dec(&c->need_shrink);
2428 freed++;
2429 cond_resched();
2430 }
2431 }
2432}
2433
2434static void shrink_work(struct work_struct *w)
2435{
2436 struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
2437
2438 dm_bufio_lock(c);
2439 __scan(c);
2440 dm_bufio_unlock(c);
2441}
2442
2443static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
2444{
2445 struct dm_bufio_client *c;
2446
2447 c = shrink->private_data;
2448 atomic_long_add(sc->nr_to_scan, &c->need_shrink);
2449 queue_work(dm_bufio_wq, &c->shrink_work);
2450
2451 return sc->nr_to_scan;
2452}
2453
2454static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
2455{
2456 struct dm_bufio_client *c = shrink->private_data;
2457 unsigned long count = cache_total(&c->cache);
2458 unsigned long retain_target = get_retain_buffers(c);
2459 unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
2460
2461 if (unlikely(count < retain_target))
2462 count = 0;
2463 else
2464 count -= retain_target;
2465
2466 if (unlikely(count < queued_for_cleanup))
2467 count = 0;
2468 else
2469 count -= queued_for_cleanup;
2470
2471 return count;
2472}
2473
2474/*
2475 * Create the buffering interface
2476 */
2477struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned int block_size,
2478 unsigned int reserved_buffers, unsigned int aux_size,
2479 void (*alloc_callback)(struct dm_buffer *),
2480 void (*write_callback)(struct dm_buffer *),
2481 unsigned int flags)
2482{
2483 int r;
2484 unsigned int num_locks;
2485 struct dm_bufio_client *c;
2486 char slab_name[64];
2487 static atomic_t seqno = ATOMIC_INIT(0);
2488
2489 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
2490 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
2491 r = -EINVAL;
2492 goto bad_client;
2493 }
2494
2495 num_locks = dm_num_hash_locks();
2496 c = kzalloc(sizeof(*c) + (num_locks * sizeof(struct buffer_tree)), GFP_KERNEL);
2497 if (!c) {
2498 r = -ENOMEM;
2499 goto bad_client;
2500 }
2501 cache_init(&c->cache, num_locks, (flags & DM_BUFIO_CLIENT_NO_SLEEP) != 0);
2502
2503 c->bdev = bdev;
2504 c->block_size = block_size;
2505 if (is_power_of_2(block_size))
2506 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
2507 else
2508 c->sectors_per_block_bits = -1;
2509
2510 c->alloc_callback = alloc_callback;
2511 c->write_callback = write_callback;
2512
2513 if (flags & DM_BUFIO_CLIENT_NO_SLEEP) {
2514 c->no_sleep = true;
2515 static_branch_inc(&no_sleep_enabled);
2516 }
2517
2518 mutex_init(&c->lock);
2519 spin_lock_init(&c->spinlock);
2520 INIT_LIST_HEAD(&c->reserved_buffers);
2521 c->need_reserved_buffers = reserved_buffers;
2522
2523 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
2524
2525 init_waitqueue_head(&c->free_buffer_wait);
2526 c->async_write_error = 0;
2527
2528 c->dm_io = dm_io_client_create();
2529 if (IS_ERR(c->dm_io)) {
2530 r = PTR_ERR(c->dm_io);
2531 goto bad_dm_io;
2532 }
2533
2534 if (block_size <= KMALLOC_MAX_SIZE && !is_power_of_2(block_size)) {
2535 unsigned int align = min(1U << __ffs(block_size), (unsigned int)PAGE_SIZE);
2536
2537 snprintf(slab_name, sizeof(slab_name), "dm_bufio_cache-%u-%u",
2538 block_size, atomic_inc_return(&seqno));
2539 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
2540 SLAB_RECLAIM_ACCOUNT, NULL);
2541 if (!c->slab_cache) {
2542 r = -ENOMEM;
2543 goto bad;
2544 }
2545 }
2546 if (aux_size)
2547 snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer-%u-%u",
2548 aux_size, atomic_inc_return(&seqno));
2549 else
2550 snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer-%u",
2551 atomic_inc_return(&seqno));
2552 c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
2553 0, SLAB_RECLAIM_ACCOUNT, NULL);
2554 if (!c->slab_buffer) {
2555 r = -ENOMEM;
2556 goto bad;
2557 }
2558
2559 while (c->need_reserved_buffers) {
2560 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
2561
2562 if (!b) {
2563 r = -ENOMEM;
2564 goto bad;
2565 }
2566 __free_buffer_wake(b);
2567 }
2568
2569 INIT_WORK(&c->shrink_work, shrink_work);
2570 atomic_long_set(&c->need_shrink, 0);
2571
2572 c->shrinker = shrinker_alloc(0, "dm-bufio:(%u:%u)",
2573 MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev));
2574 if (!c->shrinker) {
2575 r = -ENOMEM;
2576 goto bad;
2577 }
2578
2579 c->shrinker->count_objects = dm_bufio_shrink_count;
2580 c->shrinker->scan_objects = dm_bufio_shrink_scan;
2581 c->shrinker->seeks = 1;
2582 c->shrinker->batch = 0;
2583 c->shrinker->private_data = c;
2584
2585 shrinker_register(c->shrinker);
2586
2587 mutex_lock(&dm_bufio_clients_lock);
2588 dm_bufio_client_count++;
2589 list_add(&c->client_list, &dm_bufio_all_clients);
2590 __cache_size_refresh();
2591 mutex_unlock(&dm_bufio_clients_lock);
2592
2593 return c;
2594
2595bad:
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 kmem_cache_destroy(c->slab_cache);
2603 kmem_cache_destroy(c->slab_buffer);
2604 dm_io_client_destroy(c->dm_io);
2605bad_dm_io:
2606 mutex_destroy(&c->lock);
2607 if (c->no_sleep)
2608 static_branch_dec(&no_sleep_enabled);
2609 kfree(c);
2610bad_client:
2611 return ERR_PTR(r);
2612}
2613EXPORT_SYMBOL_GPL(dm_bufio_client_create);
2614
2615/*
2616 * Free the buffering interface.
2617 * It is required that there are no references on any buffers.
2618 */
2619void dm_bufio_client_destroy(struct dm_bufio_client *c)
2620{
2621 unsigned int i;
2622
2623 drop_buffers(c);
2624
2625 shrinker_free(c->shrinker);
2626 flush_work(&c->shrink_work);
2627
2628 mutex_lock(&dm_bufio_clients_lock);
2629
2630 list_del(&c->client_list);
2631 dm_bufio_client_count--;
2632 __cache_size_refresh();
2633
2634 mutex_unlock(&dm_bufio_clients_lock);
2635
2636 WARN_ON(c->need_reserved_buffers);
2637
2638 while (!list_empty(&c->reserved_buffers)) {
2639 struct dm_buffer *b = list_to_buffer(c->reserved_buffers.next);
2640
2641 list_del(&b->lru.list);
2642 free_buffer(b);
2643 }
2644
2645 for (i = 0; i < LIST_SIZE; i++)
2646 if (cache_count(&c->cache, i))
2647 DMERR("leaked buffer count %d: %lu", i, cache_count(&c->cache, i));
2648
2649 for (i = 0; i < LIST_SIZE; i++)
2650 WARN_ON(cache_count(&c->cache, i));
2651
2652 cache_destroy(&c->cache);
2653 kmem_cache_destroy(c->slab_cache);
2654 kmem_cache_destroy(c->slab_buffer);
2655 dm_io_client_destroy(c->dm_io);
2656 mutex_destroy(&c->lock);
2657 if (c->no_sleep)
2658 static_branch_dec(&no_sleep_enabled);
2659 kfree(c);
2660}
2661EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
2662
2663void dm_bufio_client_reset(struct dm_bufio_client *c)
2664{
2665 drop_buffers(c);
2666 flush_work(&c->shrink_work);
2667}
2668EXPORT_SYMBOL_GPL(dm_bufio_client_reset);
2669
2670void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
2671{
2672 c->start = start;
2673}
2674EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
2675
2676/*--------------------------------------------------------------*/
2677
2678static unsigned int get_max_age_hz(void)
2679{
2680 unsigned int max_age = READ_ONCE(dm_bufio_max_age);
2681
2682 if (max_age > UINT_MAX / HZ)
2683 max_age = UINT_MAX / HZ;
2684
2685 return max_age * HZ;
2686}
2687
2688static bool older_than(struct dm_buffer *b, unsigned long age_hz)
2689{
2690 return time_after_eq(jiffies, READ_ONCE(b->last_accessed) + age_hz);
2691}
2692
2693struct evict_params {
2694 gfp_t gfp;
2695 unsigned long age_hz;
2696
2697 /*
2698 * This gets updated with the largest last_accessed (ie. most
2699 * recently used) of the evicted buffers. It will not be reinitialised
2700 * by __evict_many(), so you can use it across multiple invocations.
2701 */
2702 unsigned long last_accessed;
2703};
2704
2705/*
2706 * We may not be able to evict this buffer if IO pending or the client
2707 * is still using it.
2708 *
2709 * And if GFP_NOFS is used, we must not do any I/O because we hold
2710 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
2711 * rerouted to different bufio client.
2712 */
2713static enum evict_result select_for_evict(struct dm_buffer *b, void *context)
2714{
2715 struct evict_params *params = context;
2716
2717 if (!(params->gfp & __GFP_FS) ||
2718 (static_branch_unlikely(&no_sleep_enabled) && b->c->no_sleep)) {
2719 if (test_bit_acquire(B_READING, &b->state) ||
2720 test_bit(B_WRITING, &b->state) ||
2721 test_bit(B_DIRTY, &b->state))
2722 return ER_DONT_EVICT;
2723 }
2724
2725 return older_than(b, params->age_hz) ? ER_EVICT : ER_STOP;
2726}
2727
2728static unsigned long __evict_many(struct dm_bufio_client *c,
2729 struct evict_params *params,
2730 int list_mode, unsigned long max_count)
2731{
2732 unsigned long count;
2733 unsigned long last_accessed;
2734 struct dm_buffer *b;
2735
2736 for (count = 0; count < max_count; count++) {
2737 b = cache_evict(&c->cache, list_mode, select_for_evict, params);
2738 if (!b)
2739 break;
2740
2741 last_accessed = READ_ONCE(b->last_accessed);
2742 if (time_after_eq(params->last_accessed, last_accessed))
2743 params->last_accessed = last_accessed;
2744
2745 __make_buffer_clean(b);
2746 __free_buffer_wake(b);
2747
2748 cond_resched();
2749 }
2750
2751 return count;
2752}
2753
2754static void evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
2755{
2756 struct evict_params params = {.gfp = 0, .age_hz = age_hz, .last_accessed = 0};
2757 unsigned long retain = get_retain_buffers(c);
2758 unsigned long count;
2759 LIST_HEAD(write_list);
2760
2761 dm_bufio_lock(c);
2762
2763 __check_watermark(c, &write_list);
2764 if (unlikely(!list_empty(&write_list))) {
2765 dm_bufio_unlock(c);
2766 __flush_write_list(&write_list);
2767 dm_bufio_lock(c);
2768 }
2769
2770 count = cache_total(&c->cache);
2771 if (count > retain)
2772 __evict_many(c, ¶ms, LIST_CLEAN, count - retain);
2773
2774 dm_bufio_unlock(c);
2775}
2776
2777static void cleanup_old_buffers(void)
2778{
2779 unsigned long max_age_hz = get_max_age_hz();
2780 struct dm_bufio_client *c;
2781
2782 mutex_lock(&dm_bufio_clients_lock);
2783
2784 __cache_size_refresh();
2785
2786 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2787 evict_old_buffers(c, max_age_hz);
2788
2789 mutex_unlock(&dm_bufio_clients_lock);
2790}
2791
2792static void work_fn(struct work_struct *w)
2793{
2794 cleanup_old_buffers();
2795
2796 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2797 DM_BUFIO_WORK_TIMER_SECS * HZ);
2798}
2799
2800/*--------------------------------------------------------------*/
2801
2802/*
2803 * Global cleanup tries to evict the oldest buffers from across _all_
2804 * the clients. It does this by repeatedly evicting a few buffers from
2805 * the client that holds the oldest buffer. It's approximate, but hopefully
2806 * good enough.
2807 */
2808static struct dm_bufio_client *__pop_client(void)
2809{
2810 struct list_head *h;
2811
2812 if (list_empty(&dm_bufio_all_clients))
2813 return NULL;
2814
2815 h = dm_bufio_all_clients.next;
2816 list_del(h);
2817 return container_of(h, struct dm_bufio_client, client_list);
2818}
2819
2820/*
2821 * Inserts the client in the global client list based on its
2822 * 'oldest_buffer' field.
2823 */
2824static void __insert_client(struct dm_bufio_client *new_client)
2825{
2826 struct dm_bufio_client *c;
2827 struct list_head *h = dm_bufio_all_clients.next;
2828
2829 while (h != &dm_bufio_all_clients) {
2830 c = container_of(h, struct dm_bufio_client, client_list);
2831 if (time_after_eq(c->oldest_buffer, new_client->oldest_buffer))
2832 break;
2833 h = h->next;
2834 }
2835
2836 list_add_tail(&new_client->client_list, h);
2837}
2838
2839static unsigned long __evict_a_few(unsigned long nr_buffers)
2840{
2841 unsigned long count;
2842 struct dm_bufio_client *c;
2843 struct evict_params params = {
2844 .gfp = GFP_KERNEL,
2845 .age_hz = 0,
2846 /* set to jiffies in case there are no buffers in this client */
2847 .last_accessed = jiffies
2848 };
2849
2850 c = __pop_client();
2851 if (!c)
2852 return 0;
2853
2854 dm_bufio_lock(c);
2855 count = __evict_many(c, ¶ms, LIST_CLEAN, nr_buffers);
2856 dm_bufio_unlock(c);
2857
2858 if (count)
2859 c->oldest_buffer = params.last_accessed;
2860 __insert_client(c);
2861
2862 return count;
2863}
2864
2865static void check_watermarks(void)
2866{
2867 LIST_HEAD(write_list);
2868 struct dm_bufio_client *c;
2869
2870 mutex_lock(&dm_bufio_clients_lock);
2871 list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
2872 dm_bufio_lock(c);
2873 __check_watermark(c, &write_list);
2874 dm_bufio_unlock(c);
2875 }
2876 mutex_unlock(&dm_bufio_clients_lock);
2877
2878 __flush_write_list(&write_list);
2879}
2880
2881static void evict_old(void)
2882{
2883 unsigned long threshold = dm_bufio_cache_size -
2884 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
2885
2886 mutex_lock(&dm_bufio_clients_lock);
2887 while (dm_bufio_current_allocated > threshold) {
2888 if (!__evict_a_few(64))
2889 break;
2890 cond_resched();
2891 }
2892 mutex_unlock(&dm_bufio_clients_lock);
2893}
2894
2895static void do_global_cleanup(struct work_struct *w)
2896{
2897 check_watermarks();
2898 evict_old();
2899}
2900
2901/*
2902 *--------------------------------------------------------------
2903 * Module setup
2904 *--------------------------------------------------------------
2905 */
2906
2907/*
2908 * This is called only once for the whole dm_bufio module.
2909 * It initializes memory limit.
2910 */
2911static int __init dm_bufio_init(void)
2912{
2913 __u64 mem;
2914
2915 dm_bufio_allocated_kmem_cache = 0;
2916 dm_bufio_allocated_kmalloc = 0;
2917 dm_bufio_allocated_get_free_pages = 0;
2918 dm_bufio_allocated_vmalloc = 0;
2919 dm_bufio_current_allocated = 0;
2920
2921 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2922 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2923
2924 if (mem > ULONG_MAX)
2925 mem = ULONG_MAX;
2926
2927#ifdef CONFIG_MMU
2928 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2929 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2930#endif
2931
2932 dm_bufio_default_cache_size = mem;
2933
2934 mutex_lock(&dm_bufio_clients_lock);
2935 __cache_size_refresh();
2936 mutex_unlock(&dm_bufio_clients_lock);
2937
2938 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2939 if (!dm_bufio_wq)
2940 return -ENOMEM;
2941
2942 INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2943 INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2944 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2945 DM_BUFIO_WORK_TIMER_SECS * HZ);
2946
2947 return 0;
2948}
2949
2950/*
2951 * This is called once when unloading the dm_bufio module.
2952 */
2953static void __exit dm_bufio_exit(void)
2954{
2955 int bug = 0;
2956
2957 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2958 destroy_workqueue(dm_bufio_wq);
2959
2960 if (dm_bufio_client_count) {
2961 DMCRIT("%s: dm_bufio_client_count leaked: %d",
2962 __func__, dm_bufio_client_count);
2963 bug = 1;
2964 }
2965
2966 if (dm_bufio_current_allocated) {
2967 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2968 __func__, dm_bufio_current_allocated);
2969 bug = 1;
2970 }
2971
2972 if (dm_bufio_allocated_get_free_pages) {
2973 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2974 __func__, dm_bufio_allocated_get_free_pages);
2975 bug = 1;
2976 }
2977
2978 if (dm_bufio_allocated_vmalloc) {
2979 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2980 __func__, dm_bufio_allocated_vmalloc);
2981 bug = 1;
2982 }
2983
2984 WARN_ON(bug); /* leaks are not worth crashing the system */
2985}
2986
2987module_init(dm_bufio_init)
2988module_exit(dm_bufio_exit)
2989
2990module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, 0644);
2991MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2992
2993module_param_named(max_age_seconds, dm_bufio_max_age, uint, 0644);
2994MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2995
2996module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, 0644);
2997MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2998
2999module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, 0644);
3000MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
3001
3002module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, 0444);
3003MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
3004
3005module_param_named(allocated_kmalloc_bytes, dm_bufio_allocated_kmalloc, ulong, 0444);
3006MODULE_PARM_DESC(allocated_kmalloc_bytes, "Memory allocated with kmalloc_alloc");
3007
3008module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, 0444);
3009MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
3010
3011module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, 0444);
3012MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
3013
3014module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, 0444);
3015MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
3016
3017MODULE_AUTHOR("Mikulas Patocka <dm-devel@lists.linux.dev>");
3018MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
3019MODULE_LICENSE("GPL");