aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/md
diff options
context:
space:
mode:
authorMikulas Patocka <mpatocka@redhat.com>2011-10-31 16:19:09 -0400
committerAlasdair G Kergon <agk@redhat.com>2011-10-31 16:19:09 -0400
commit95d402f057f2e208e4631893f6cd4a59c7c05e41 (patch)
tree4edb6c01f62689617ee3098104c135ef2c43f0e2 /drivers/md
parent3cf2e4ba74ca1bf5d8ad26cd18592f02b32964f5 (diff)
dm: add bufio
The dm-bufio interface allows you to do cached I/O on devices, holding recently-read blocks in memory and performing delayed writes. We don't use buffer cache or page cache already present in the kernel, because: * we need to handle block sizes larger than a page * we can't allocate memory to perform reads or we'd have deadlocks Currently, when a cache is required, we limit its size to a fraction of available memory. Usage can be viewed and changed in /sys/module/dm_bufio/parameters/ . The first user is thin provisioning, but more dm users are planned. Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Diffstat (limited to 'drivers/md')
-rw-r--r--drivers/md/Kconfig8
-rw-r--r--drivers/md/Makefile1
-rw-r--r--drivers/md/dm-bufio.c1699
-rw-r--r--drivers/md/dm-bufio.h112
4 files changed, 1820 insertions, 0 deletions
diff --git a/drivers/md/Kconfig b/drivers/md/Kconfig
index f75a66e7d312..b1a921497043 100644
--- a/drivers/md/Kconfig
+++ b/drivers/md/Kconfig
@@ -208,6 +208,14 @@ config DM_DEBUG
208 208
209 If unsure, say N. 209 If unsure, say N.
210 210
211config DM_BUFIO
212 tristate
213 depends on BLK_DEV_DM && EXPERIMENTAL
214 ---help---
215 This interface allows you to do buffered I/O on a device and acts
216 as a cache, holding recently-read blocks in memory and performing
217 delayed writes.
218
211config DM_CRYPT 219config DM_CRYPT
212 tristate "Crypt target support" 220 tristate "Crypt target support"
213 depends on BLK_DEV_DM 221 depends on BLK_DEV_DM
diff --git a/drivers/md/Makefile b/drivers/md/Makefile
index 448838b1f92a..56661c4272f2 100644
--- a/drivers/md/Makefile
+++ b/drivers/md/Makefile
@@ -27,6 +27,7 @@ obj-$(CONFIG_MD_MULTIPATH) += multipath.o
27obj-$(CONFIG_MD_FAULTY) += faulty.o 27obj-$(CONFIG_MD_FAULTY) += faulty.o
28obj-$(CONFIG_BLK_DEV_MD) += md-mod.o 28obj-$(CONFIG_BLK_DEV_MD) += md-mod.o
29obj-$(CONFIG_BLK_DEV_DM) += dm-mod.o 29obj-$(CONFIG_BLK_DEV_DM) += dm-mod.o
30obj-$(CONFIG_DM_BUFIO) += dm-bufio.o
30obj-$(CONFIG_DM_CRYPT) += dm-crypt.o 31obj-$(CONFIG_DM_CRYPT) += dm-crypt.o
31obj-$(CONFIG_DM_DELAY) += dm-delay.o 32obj-$(CONFIG_DM_DELAY) += dm-delay.o
32obj-$(CONFIG_DM_FLAKEY) += dm-flakey.o 33obj-$(CONFIG_DM_FLAKEY) += dm-flakey.o
diff --git a/drivers/md/dm-bufio.c b/drivers/md/dm-bufio.c
new file mode 100644
index 000000000000..cb246667dd52
--- /dev/null
+++ b/drivers/md/dm-bufio.c
@@ -0,0 +1,1699 @@
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/vmalloc.h>
15#include <linux/version.h>
16#include <linux/shrinker.h>
17
18#define DM_MSG_PREFIX "bufio"
19
20/*
21 * Memory management policy:
22 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
23 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
24 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
25 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
26 * dirty buffers.
27 */
28#define DM_BUFIO_MIN_BUFFERS 8
29
30#define DM_BUFIO_MEMORY_PERCENT 2
31#define DM_BUFIO_VMALLOC_PERCENT 25
32#define DM_BUFIO_WRITEBACK_PERCENT 75
33
34/*
35 * Check buffer ages in this interval (seconds)
36 */
37#define DM_BUFIO_WORK_TIMER_SECS 10
38
39/*
40 * Free buffers when they are older than this (seconds)
41 */
42#define DM_BUFIO_DEFAULT_AGE_SECS 60
43
44/*
45 * The number of bvec entries that are embedded directly in the buffer.
46 * If the chunk size is larger, dm-io is used to do the io.
47 */
48#define DM_BUFIO_INLINE_VECS 16
49
50/*
51 * Buffer hash
52 */
53#define DM_BUFIO_HASH_BITS 20
54#define DM_BUFIO_HASH(block) \
55 ((((block) >> DM_BUFIO_HASH_BITS) ^ (block)) & \
56 ((1 << DM_BUFIO_HASH_BITS) - 1))
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 struct hlist_head *cache_hash;
108 wait_queue_head_t free_buffer_wait;
109
110 int async_write_error;
111
112 struct list_head client_list;
113 struct shrinker shrinker;
114};
115
116/*
117 * Buffer state bits.
118 */
119#define B_READING 0
120#define B_WRITING 1
121#define B_DIRTY 2
122
123/*
124 * Describes how the block was allocated:
125 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
126 * See the comment at alloc_buffer_data.
127 */
128enum data_mode {
129 DATA_MODE_SLAB = 0,
130 DATA_MODE_GET_FREE_PAGES = 1,
131 DATA_MODE_VMALLOC = 2,
132 DATA_MODE_LIMIT = 3
133};
134
135struct dm_buffer {
136 struct hlist_node hash_list;
137 struct list_head lru_list;
138 sector_t block;
139 void *data;
140 enum data_mode data_mode;
141 unsigned char list_mode; /* LIST_* */
142 unsigned hold_count;
143 int read_error;
144 int write_error;
145 unsigned long state;
146 unsigned long last_accessed;
147 struct dm_bufio_client *c;
148 struct bio bio;
149 struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
150};
151
152/*----------------------------------------------------------------*/
153
154static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
155static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
156
157static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
158{
159 unsigned ret = c->blocks_per_page_bits - 1;
160
161 BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
162
163 return ret;
164}
165
166#define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)])
167#define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)])
168
169#define dm_bufio_in_request() (!!current->bio_list)
170
171static void dm_bufio_lock(struct dm_bufio_client *c)
172{
173 mutex_lock_nested(&c->lock, dm_bufio_in_request());
174}
175
176static int dm_bufio_trylock(struct dm_bufio_client *c)
177{
178 return mutex_trylock(&c->lock);
179}
180
181static void dm_bufio_unlock(struct dm_bufio_client *c)
182{
183 mutex_unlock(&c->lock);
184}
185
186/*
187 * FIXME Move to sched.h?
188 */
189#ifdef CONFIG_PREEMPT_VOLUNTARY
190# define dm_bufio_cond_resched() \
191do { \
192 if (unlikely(need_resched())) \
193 _cond_resched(); \
194} while (0)
195#else
196# define dm_bufio_cond_resched() do { } while (0)
197#endif
198
199/*----------------------------------------------------------------*/
200
201/*
202 * Default cache size: available memory divided by the ratio.
203 */
204static unsigned long dm_bufio_default_cache_size;
205
206/*
207 * Total cache size set by the user.
208 */
209static unsigned long dm_bufio_cache_size;
210
211/*
212 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
213 * at any time. If it disagrees, the user has changed cache size.
214 */
215static unsigned long dm_bufio_cache_size_latch;
216
217static DEFINE_SPINLOCK(param_spinlock);
218
219/*
220 * Buffers are freed after this timeout
221 */
222static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
223
224static unsigned long dm_bufio_peak_allocated;
225static unsigned long dm_bufio_allocated_kmem_cache;
226static unsigned long dm_bufio_allocated_get_free_pages;
227static unsigned long dm_bufio_allocated_vmalloc;
228static unsigned long dm_bufio_current_allocated;
229
230/*----------------------------------------------------------------*/
231
232/*
233 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
234 */
235static unsigned long dm_bufio_cache_size_per_client;
236
237/*
238 * The current number of clients.
239 */
240static int dm_bufio_client_count;
241
242/*
243 * The list of all clients.
244 */
245static LIST_HEAD(dm_bufio_all_clients);
246
247/*
248 * This mutex protects dm_bufio_cache_size_latch,
249 * dm_bufio_cache_size_per_client and dm_bufio_client_count
250 */
251static DEFINE_MUTEX(dm_bufio_clients_lock);
252
253/*----------------------------------------------------------------*/
254
255static void adjust_total_allocated(enum data_mode data_mode, long diff)
256{
257 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
258 &dm_bufio_allocated_kmem_cache,
259 &dm_bufio_allocated_get_free_pages,
260 &dm_bufio_allocated_vmalloc,
261 };
262
263 spin_lock(&param_spinlock);
264
265 *class_ptr[data_mode] += diff;
266
267 dm_bufio_current_allocated += diff;
268
269 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
270 dm_bufio_peak_allocated = dm_bufio_current_allocated;
271
272 spin_unlock(&param_spinlock);
273}
274
275/*
276 * Change the number of clients and recalculate per-client limit.
277 */
278static void __cache_size_refresh(void)
279{
280 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
281 BUG_ON(dm_bufio_client_count < 0);
282
283 dm_bufio_cache_size_latch = dm_bufio_cache_size;
284
285 barrier();
286
287 /*
288 * Use default if set to 0 and report the actual cache size used.
289 */
290 if (!dm_bufio_cache_size_latch) {
291 (void)cmpxchg(&dm_bufio_cache_size, 0,
292 dm_bufio_default_cache_size);
293 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
294 }
295
296 dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
297 (dm_bufio_client_count ? : 1);
298}
299
300/*
301 * Allocating buffer data.
302 *
303 * Small buffers are allocated with kmem_cache, to use space optimally.
304 *
305 * For large buffers, we choose between get_free_pages and vmalloc.
306 * Each has advantages and disadvantages.
307 *
308 * __get_free_pages can randomly fail if the memory is fragmented.
309 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
310 * as low as 128M) so using it for caching is not appropriate.
311 *
312 * If the allocation may fail we use __get_free_pages. Memory fragmentation
313 * won't have a fatal effect here, but it just causes flushes of some other
314 * buffers and more I/O will be performed. Don't use __get_free_pages if it
315 * always fails (i.e. order >= MAX_ORDER).
316 *
317 * If the allocation shouldn't fail we use __vmalloc. This is only for the
318 * initial reserve allocation, so there's no risk of wasting all vmalloc
319 * space.
320 */
321static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
322 enum data_mode *data_mode)
323{
324 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
325 *data_mode = DATA_MODE_SLAB;
326 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
327 }
328
329 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
330 gfp_mask & __GFP_NORETRY) {
331 *data_mode = DATA_MODE_GET_FREE_PAGES;
332 return (void *)__get_free_pages(gfp_mask,
333 c->pages_per_block_bits);
334 }
335
336 *data_mode = DATA_MODE_VMALLOC;
337 return __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
338}
339
340/*
341 * Free buffer's data.
342 */
343static void free_buffer_data(struct dm_bufio_client *c,
344 void *data, enum data_mode data_mode)
345{
346 switch (data_mode) {
347 case DATA_MODE_SLAB:
348 kmem_cache_free(DM_BUFIO_CACHE(c), data);
349 break;
350
351 case DATA_MODE_GET_FREE_PAGES:
352 free_pages((unsigned long)data, c->pages_per_block_bits);
353 break;
354
355 case DATA_MODE_VMALLOC:
356 vfree(data);
357 break;
358
359 default:
360 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
361 data_mode);
362 BUG();
363 }
364}
365
366/*
367 * Allocate buffer and its data.
368 */
369static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
370{
371 struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
372 gfp_mask);
373
374 if (!b)
375 return NULL;
376
377 b->c = c;
378
379 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
380 if (!b->data) {
381 kfree(b);
382 return NULL;
383 }
384
385 adjust_total_allocated(b->data_mode, (long)c->block_size);
386
387 return b;
388}
389
390/*
391 * Free buffer and its data.
392 */
393static void free_buffer(struct dm_buffer *b)
394{
395 struct dm_bufio_client *c = b->c;
396
397 adjust_total_allocated(b->data_mode, -(long)c->block_size);
398
399 free_buffer_data(c, b->data, b->data_mode);
400 kfree(b);
401}
402
403/*
404 * Link buffer to the hash list and clean or dirty queue.
405 */
406static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
407{
408 struct dm_bufio_client *c = b->c;
409
410 c->n_buffers[dirty]++;
411 b->block = block;
412 b->list_mode = dirty;
413 list_add(&b->lru_list, &c->lru[dirty]);
414 hlist_add_head(&b->hash_list, &c->cache_hash[DM_BUFIO_HASH(block)]);
415 b->last_accessed = jiffies;
416}
417
418/*
419 * Unlink buffer from the hash list and dirty or clean queue.
420 */
421static void __unlink_buffer(struct dm_buffer *b)
422{
423 struct dm_bufio_client *c = b->c;
424
425 BUG_ON(!c->n_buffers[b->list_mode]);
426
427 c->n_buffers[b->list_mode]--;
428 hlist_del(&b->hash_list);
429 list_del(&b->lru_list);
430}
431
432/*
433 * Place the buffer to the head of dirty or clean LRU queue.
434 */
435static void __relink_lru(struct dm_buffer *b, int dirty)
436{
437 struct dm_bufio_client *c = b->c;
438
439 BUG_ON(!c->n_buffers[b->list_mode]);
440
441 c->n_buffers[b->list_mode]--;
442 c->n_buffers[dirty]++;
443 b->list_mode = dirty;
444 list_del(&b->lru_list);
445 list_add(&b->lru_list, &c->lru[dirty]);
446}
447
448/*----------------------------------------------------------------
449 * Submit I/O on the buffer.
450 *
451 * Bio interface is faster but it has some problems:
452 * the vector list is limited (increasing this limit increases
453 * memory-consumption per buffer, so it is not viable);
454 *
455 * the memory must be direct-mapped, not vmalloced;
456 *
457 * the I/O driver can reject requests spuriously if it thinks that
458 * the requests are too big for the device or if they cross a
459 * controller-defined memory boundary.
460 *
461 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
462 * it is not vmalloced, try using the bio interface.
463 *
464 * If the buffer is big, if it is vmalloced or if the underlying device
465 * rejects the bio because it is too large, use dm-io layer to do the I/O.
466 * The dm-io layer splits the I/O into multiple requests, avoiding the above
467 * shortcomings.
468 *--------------------------------------------------------------*/
469
470/*
471 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
472 * that the request was handled directly with bio interface.
473 */
474static void dmio_complete(unsigned long error, void *context)
475{
476 struct dm_buffer *b = context;
477
478 b->bio.bi_end_io(&b->bio, error ? -EIO : 0);
479}
480
481static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
482 bio_end_io_t *end_io)
483{
484 int r;
485 struct dm_io_request io_req = {
486 .bi_rw = rw,
487 .notify.fn = dmio_complete,
488 .notify.context = b,
489 .client = b->c->dm_io,
490 };
491 struct dm_io_region region = {
492 .bdev = b->c->bdev,
493 .sector = block << b->c->sectors_per_block_bits,
494 .count = b->c->block_size >> SECTOR_SHIFT,
495 };
496
497 if (b->data_mode != DATA_MODE_VMALLOC) {
498 io_req.mem.type = DM_IO_KMEM;
499 io_req.mem.ptr.addr = b->data;
500 } else {
501 io_req.mem.type = DM_IO_VMA;
502 io_req.mem.ptr.vma = b->data;
503 }
504
505 b->bio.bi_end_io = end_io;
506
507 r = dm_io(&io_req, 1, &region, NULL);
508 if (r)
509 end_io(&b->bio, r);
510}
511
512static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
513 bio_end_io_t *end_io)
514{
515 char *ptr;
516 int len;
517
518 bio_init(&b->bio);
519 b->bio.bi_io_vec = b->bio_vec;
520 b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
521 b->bio.bi_sector = block << b->c->sectors_per_block_bits;
522 b->bio.bi_bdev = b->c->bdev;
523 b->bio.bi_end_io = end_io;
524
525 /*
526 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
527 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
528 */
529 ptr = b->data;
530 len = b->c->block_size;
531
532 if (len >= PAGE_SIZE)
533 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
534 else
535 BUG_ON((unsigned long)ptr & (len - 1));
536
537 do {
538 if (!bio_add_page(&b->bio, virt_to_page(ptr),
539 len < PAGE_SIZE ? len : PAGE_SIZE,
540 virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
541 BUG_ON(b->c->block_size <= PAGE_SIZE);
542 use_dmio(b, rw, block, end_io);
543 return;
544 }
545
546 len -= PAGE_SIZE;
547 ptr += PAGE_SIZE;
548 } while (len > 0);
549
550 submit_bio(rw, &b->bio);
551}
552
553static void submit_io(struct dm_buffer *b, int rw, sector_t block,
554 bio_end_io_t *end_io)
555{
556 if (rw == WRITE && b->c->write_callback)
557 b->c->write_callback(b);
558
559 if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
560 b->data_mode != DATA_MODE_VMALLOC)
561 use_inline_bio(b, rw, block, end_io);
562 else
563 use_dmio(b, rw, block, end_io);
564}
565
566/*----------------------------------------------------------------
567 * Writing dirty buffers
568 *--------------------------------------------------------------*/
569
570/*
571 * The endio routine for write.
572 *
573 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
574 * it.
575 */
576static void write_endio(struct bio *bio, int error)
577{
578 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
579
580 b->write_error = error;
581 if (error) {
582 struct dm_bufio_client *c = b->c;
583 (void)cmpxchg(&c->async_write_error, 0, error);
584 }
585
586 BUG_ON(!test_bit(B_WRITING, &b->state));
587
588 smp_mb__before_clear_bit();
589 clear_bit(B_WRITING, &b->state);
590 smp_mb__after_clear_bit();
591
592 wake_up_bit(&b->state, B_WRITING);
593}
594
595/*
596 * This function is called when wait_on_bit is actually waiting.
597 */
598static int do_io_schedule(void *word)
599{
600 io_schedule();
601
602 return 0;
603}
604
605/*
606 * Initiate a write on a dirty buffer, but don't wait for it.
607 *
608 * - If the buffer is not dirty, exit.
609 * - If there some previous write going on, wait for it to finish (we can't
610 * have two writes on the same buffer simultaneously).
611 * - Submit our write and don't wait on it. We set B_WRITING indicating
612 * that there is a write in progress.
613 */
614static void __write_dirty_buffer(struct dm_buffer *b)
615{
616 if (!test_bit(B_DIRTY, &b->state))
617 return;
618
619 clear_bit(B_DIRTY, &b->state);
620 wait_on_bit_lock(&b->state, B_WRITING,
621 do_io_schedule, TASK_UNINTERRUPTIBLE);
622
623 submit_io(b, WRITE, b->block, write_endio);
624}
625
626/*
627 * Wait until any activity on the buffer finishes. Possibly write the
628 * buffer if it is dirty. When this function finishes, there is no I/O
629 * running on the buffer and the buffer is not dirty.
630 */
631static void __make_buffer_clean(struct dm_buffer *b)
632{
633 BUG_ON(b->hold_count);
634
635 if (!b->state) /* fast case */
636 return;
637
638 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
639 __write_dirty_buffer(b);
640 wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
641}
642
643/*
644 * Find some buffer that is not held by anybody, clean it, unlink it and
645 * return it.
646 */
647static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
648{
649 struct dm_buffer *b;
650
651 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
652 BUG_ON(test_bit(B_WRITING, &b->state));
653 BUG_ON(test_bit(B_DIRTY, &b->state));
654
655 if (!b->hold_count) {
656 __make_buffer_clean(b);
657 __unlink_buffer(b);
658 return b;
659 }
660 dm_bufio_cond_resched();
661 }
662
663 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
664 BUG_ON(test_bit(B_READING, &b->state));
665
666 if (!b->hold_count) {
667 __make_buffer_clean(b);
668 __unlink_buffer(b);
669 return b;
670 }
671 dm_bufio_cond_resched();
672 }
673
674 return NULL;
675}
676
677/*
678 * Wait until some other threads free some buffer or release hold count on
679 * some buffer.
680 *
681 * This function is entered with c->lock held, drops it and regains it
682 * before exiting.
683 */
684static void __wait_for_free_buffer(struct dm_bufio_client *c)
685{
686 DECLARE_WAITQUEUE(wait, current);
687
688 add_wait_queue(&c->free_buffer_wait, &wait);
689 set_task_state(current, TASK_UNINTERRUPTIBLE);
690 dm_bufio_unlock(c);
691
692 io_schedule();
693
694 set_task_state(current, TASK_RUNNING);
695 remove_wait_queue(&c->free_buffer_wait, &wait);
696
697 dm_bufio_lock(c);
698}
699
700/*
701 * Allocate a new buffer. If the allocation is not possible, wait until
702 * some other thread frees a buffer.
703 *
704 * May drop the lock and regain it.
705 */
706static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c)
707{
708 struct dm_buffer *b;
709
710 /*
711 * dm-bufio is resistant to allocation failures (it just keeps
712 * one buffer reserved in cases all the allocations fail).
713 * So set flags to not try too hard:
714 * GFP_NOIO: don't recurse into the I/O layer
715 * __GFP_NORETRY: don't retry and rather return failure
716 * __GFP_NOMEMALLOC: don't use emergency reserves
717 * __GFP_NOWARN: don't print a warning in case of failure
718 *
719 * For debugging, if we set the cache size to 1, no new buffers will
720 * be allocated.
721 */
722 while (1) {
723 if (dm_bufio_cache_size_latch != 1) {
724 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
725 if (b)
726 return b;
727 }
728
729 if (!list_empty(&c->reserved_buffers)) {
730 b = list_entry(c->reserved_buffers.next,
731 struct dm_buffer, lru_list);
732 list_del(&b->lru_list);
733 c->need_reserved_buffers++;
734
735 return b;
736 }
737
738 b = __get_unclaimed_buffer(c);
739 if (b)
740 return b;
741
742 __wait_for_free_buffer(c);
743 }
744}
745
746static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c)
747{
748 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c);
749
750 if (c->alloc_callback)
751 c->alloc_callback(b);
752
753 return b;
754}
755
756/*
757 * Free a buffer and wake other threads waiting for free buffers.
758 */
759static void __free_buffer_wake(struct dm_buffer *b)
760{
761 struct dm_bufio_client *c = b->c;
762
763 if (!c->need_reserved_buffers)
764 free_buffer(b);
765 else {
766 list_add(&b->lru_list, &c->reserved_buffers);
767 c->need_reserved_buffers--;
768 }
769
770 wake_up(&c->free_buffer_wait);
771}
772
773static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait)
774{
775 struct dm_buffer *b, *tmp;
776
777 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
778 BUG_ON(test_bit(B_READING, &b->state));
779
780 if (!test_bit(B_DIRTY, &b->state) &&
781 !test_bit(B_WRITING, &b->state)) {
782 __relink_lru(b, LIST_CLEAN);
783 continue;
784 }
785
786 if (no_wait && test_bit(B_WRITING, &b->state))
787 return;
788
789 __write_dirty_buffer(b);
790 dm_bufio_cond_resched();
791 }
792}
793
794/*
795 * Get writeback threshold and buffer limit for a given client.
796 */
797static void __get_memory_limit(struct dm_bufio_client *c,
798 unsigned long *threshold_buffers,
799 unsigned long *limit_buffers)
800{
801 unsigned long buffers;
802
803 if (dm_bufio_cache_size != dm_bufio_cache_size_latch) {
804 mutex_lock(&dm_bufio_clients_lock);
805 __cache_size_refresh();
806 mutex_unlock(&dm_bufio_clients_lock);
807 }
808
809 buffers = dm_bufio_cache_size_per_client >>
810 (c->sectors_per_block_bits + SECTOR_SHIFT);
811
812 if (buffers < DM_BUFIO_MIN_BUFFERS)
813 buffers = DM_BUFIO_MIN_BUFFERS;
814
815 *limit_buffers = buffers;
816 *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
817}
818
819/*
820 * Check if we're over watermark.
821 * If we are over threshold_buffers, start freeing buffers.
822 * If we're over "limit_buffers", block until we get under the limit.
823 */
824static void __check_watermark(struct dm_bufio_client *c)
825{
826 unsigned long threshold_buffers, limit_buffers;
827
828 __get_memory_limit(c, &threshold_buffers, &limit_buffers);
829
830 while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
831 limit_buffers) {
832
833 struct dm_buffer *b = __get_unclaimed_buffer(c);
834
835 if (!b)
836 return;
837
838 __free_buffer_wake(b);
839 dm_bufio_cond_resched();
840 }
841
842 if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
843 __write_dirty_buffers_async(c, 1);
844}
845
846/*
847 * Find a buffer in the hash.
848 */
849static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
850{
851 struct dm_buffer *b;
852 struct hlist_node *hn;
853
854 hlist_for_each_entry(b, hn, &c->cache_hash[DM_BUFIO_HASH(block)],
855 hash_list) {
856 dm_bufio_cond_resched();
857 if (b->block == block)
858 return b;
859 }
860
861 return NULL;
862}
863
864/*----------------------------------------------------------------
865 * Getting a buffer
866 *--------------------------------------------------------------*/
867
868enum new_flag {
869 NF_FRESH = 0,
870 NF_READ = 1,
871 NF_GET = 2
872};
873
874static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
875 enum new_flag nf, struct dm_buffer **bp,
876 int *need_submit)
877{
878 struct dm_buffer *b, *new_b = NULL;
879
880 *need_submit = 0;
881
882 b = __find(c, block);
883 if (b) {
884 b->hold_count++;
885 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
886 test_bit(B_WRITING, &b->state));
887 return b;
888 }
889
890 if (nf == NF_GET)
891 return NULL;
892
893 new_b = __alloc_buffer_wait(c);
894
895 /*
896 * We've had a period where the mutex was unlocked, so need to
897 * recheck the hash table.
898 */
899 b = __find(c, block);
900 if (b) {
901 __free_buffer_wake(new_b);
902 b->hold_count++;
903 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
904 test_bit(B_WRITING, &b->state));
905 return b;
906 }
907
908 __check_watermark(c);
909
910 b = new_b;
911 b->hold_count = 1;
912 b->read_error = 0;
913 b->write_error = 0;
914 __link_buffer(b, block, LIST_CLEAN);
915
916 if (nf == NF_FRESH) {
917 b->state = 0;
918 return b;
919 }
920
921 b->state = 1 << B_READING;
922 *need_submit = 1;
923
924 return b;
925}
926
927/*
928 * The endio routine for reading: set the error, clear the bit and wake up
929 * anyone waiting on the buffer.
930 */
931static void read_endio(struct bio *bio, int error)
932{
933 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
934
935 b->read_error = error;
936
937 BUG_ON(!test_bit(B_READING, &b->state));
938
939 smp_mb__before_clear_bit();
940 clear_bit(B_READING, &b->state);
941 smp_mb__after_clear_bit();
942
943 wake_up_bit(&b->state, B_READING);
944}
945
946/*
947 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
948 * functions is similar except that dm_bufio_new doesn't read the
949 * buffer from the disk (assuming that the caller overwrites all the data
950 * and uses dm_bufio_mark_buffer_dirty to write new data back).
951 */
952static void *new_read(struct dm_bufio_client *c, sector_t block,
953 enum new_flag nf, struct dm_buffer **bp)
954{
955 int need_submit;
956 struct dm_buffer *b;
957
958 dm_bufio_lock(c);
959 b = __bufio_new(c, block, nf, bp, &need_submit);
960 dm_bufio_unlock(c);
961
962 if (!b || IS_ERR(b))
963 return b;
964
965 if (need_submit)
966 submit_io(b, READ, b->block, read_endio);
967
968 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
969
970 if (b->read_error) {
971 int error = b->read_error;
972
973 dm_bufio_release(b);
974
975 return ERR_PTR(error);
976 }
977
978 *bp = b;
979
980 return b->data;
981}
982
983void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
984 struct dm_buffer **bp)
985{
986 return new_read(c, block, NF_GET, bp);
987}
988EXPORT_SYMBOL_GPL(dm_bufio_get);
989
990void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
991 struct dm_buffer **bp)
992{
993 BUG_ON(dm_bufio_in_request());
994
995 return new_read(c, block, NF_READ, bp);
996}
997EXPORT_SYMBOL_GPL(dm_bufio_read);
998
999void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1000 struct dm_buffer **bp)
1001{
1002 BUG_ON(dm_bufio_in_request());
1003
1004 return new_read(c, block, NF_FRESH, bp);
1005}
1006EXPORT_SYMBOL_GPL(dm_bufio_new);
1007
1008void dm_bufio_release(struct dm_buffer *b)
1009{
1010 struct dm_bufio_client *c = b->c;
1011
1012 dm_bufio_lock(c);
1013
1014 BUG_ON(test_bit(B_READING, &b->state));
1015 BUG_ON(!b->hold_count);
1016
1017 b->hold_count--;
1018 if (!b->hold_count) {
1019 wake_up(&c->free_buffer_wait);
1020
1021 /*
1022 * If there were errors on the buffer, and the buffer is not
1023 * to be written, free the buffer. There is no point in caching
1024 * invalid buffer.
1025 */
1026 if ((b->read_error || b->write_error) &&
1027 !test_bit(B_WRITING, &b->state) &&
1028 !test_bit(B_DIRTY, &b->state)) {
1029 __unlink_buffer(b);
1030 __free_buffer_wake(b);
1031 }
1032 }
1033
1034 dm_bufio_unlock(c);
1035}
1036EXPORT_SYMBOL_GPL(dm_bufio_release);
1037
1038void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1039{
1040 struct dm_bufio_client *c = b->c;
1041
1042 dm_bufio_lock(c);
1043
1044 if (!test_and_set_bit(B_DIRTY, &b->state))
1045 __relink_lru(b, LIST_DIRTY);
1046
1047 dm_bufio_unlock(c);
1048}
1049EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1050
1051void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1052{
1053 BUG_ON(dm_bufio_in_request());
1054
1055 dm_bufio_lock(c);
1056 __write_dirty_buffers_async(c, 0);
1057 dm_bufio_unlock(c);
1058}
1059EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1060
1061/*
1062 * For performance, it is essential that the buffers are written asynchronously
1063 * and simultaneously (so that the block layer can merge the writes) and then
1064 * waited upon.
1065 *
1066 * Finally, we flush hardware disk cache.
1067 */
1068int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1069{
1070 int a, f;
1071 unsigned long buffers_processed = 0;
1072 struct dm_buffer *b, *tmp;
1073
1074 dm_bufio_lock(c);
1075 __write_dirty_buffers_async(c, 0);
1076
1077again:
1078 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1079 int dropped_lock = 0;
1080
1081 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1082 buffers_processed++;
1083
1084 BUG_ON(test_bit(B_READING, &b->state));
1085
1086 if (test_bit(B_WRITING, &b->state)) {
1087 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1088 dropped_lock = 1;
1089 b->hold_count++;
1090 dm_bufio_unlock(c);
1091 wait_on_bit(&b->state, B_WRITING,
1092 do_io_schedule,
1093 TASK_UNINTERRUPTIBLE);
1094 dm_bufio_lock(c);
1095 b->hold_count--;
1096 } else
1097 wait_on_bit(&b->state, B_WRITING,
1098 do_io_schedule,
1099 TASK_UNINTERRUPTIBLE);
1100 }
1101
1102 if (!test_bit(B_DIRTY, &b->state) &&
1103 !test_bit(B_WRITING, &b->state))
1104 __relink_lru(b, LIST_CLEAN);
1105
1106 dm_bufio_cond_resched();
1107
1108 /*
1109 * If we dropped the lock, the list is no longer consistent,
1110 * so we must restart the search.
1111 *
1112 * In the most common case, the buffer just processed is
1113 * relinked to the clean list, so we won't loop scanning the
1114 * same buffer again and again.
1115 *
1116 * This may livelock if there is another thread simultaneously
1117 * dirtying buffers, so we count the number of buffers walked
1118 * and if it exceeds the total number of buffers, it means that
1119 * someone is doing some writes simultaneously with us. In
1120 * this case, stop, dropping the lock.
1121 */
1122 if (dropped_lock)
1123 goto again;
1124 }
1125 wake_up(&c->free_buffer_wait);
1126 dm_bufio_unlock(c);
1127
1128 a = xchg(&c->async_write_error, 0);
1129 f = dm_bufio_issue_flush(c);
1130 if (a)
1131 return a;
1132
1133 return f;
1134}
1135EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1136
1137/*
1138 * Use dm-io to send and empty barrier flush the device.
1139 */
1140int dm_bufio_issue_flush(struct dm_bufio_client *c)
1141{
1142 struct dm_io_request io_req = {
1143 .bi_rw = REQ_FLUSH,
1144 .mem.type = DM_IO_KMEM,
1145 .mem.ptr.addr = NULL,
1146 .client = c->dm_io,
1147 };
1148 struct dm_io_region io_reg = {
1149 .bdev = c->bdev,
1150 .sector = 0,
1151 .count = 0,
1152 };
1153
1154 BUG_ON(dm_bufio_in_request());
1155
1156 return dm_io(&io_req, 1, &io_reg, NULL);
1157}
1158EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1159
1160/*
1161 * We first delete any other buffer that may be at that new location.
1162 *
1163 * Then, we write the buffer to the original location if it was dirty.
1164 *
1165 * Then, if we are the only one who is holding the buffer, relink the buffer
1166 * in the hash queue for the new location.
1167 *
1168 * If there was someone else holding the buffer, we write it to the new
1169 * location but not relink it, because that other user needs to have the buffer
1170 * at the same place.
1171 */
1172void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1173{
1174 struct dm_bufio_client *c = b->c;
1175 struct dm_buffer *new;
1176
1177 BUG_ON(dm_bufio_in_request());
1178
1179 dm_bufio_lock(c);
1180
1181retry:
1182 new = __find(c, new_block);
1183 if (new) {
1184 if (new->hold_count) {
1185 __wait_for_free_buffer(c);
1186 goto retry;
1187 }
1188
1189 /*
1190 * FIXME: Is there any point waiting for a write that's going
1191 * to be overwritten in a bit?
1192 */
1193 __make_buffer_clean(new);
1194 __unlink_buffer(new);
1195 __free_buffer_wake(new);
1196 }
1197
1198 BUG_ON(!b->hold_count);
1199 BUG_ON(test_bit(B_READING, &b->state));
1200
1201 __write_dirty_buffer(b);
1202 if (b->hold_count == 1) {
1203 wait_on_bit(&b->state, B_WRITING,
1204 do_io_schedule, TASK_UNINTERRUPTIBLE);
1205 set_bit(B_DIRTY, &b->state);
1206 __unlink_buffer(b);
1207 __link_buffer(b, new_block, LIST_DIRTY);
1208 } else {
1209 sector_t old_block;
1210 wait_on_bit_lock(&b->state, B_WRITING,
1211 do_io_schedule, TASK_UNINTERRUPTIBLE);
1212 /*
1213 * Relink buffer to "new_block" so that write_callback
1214 * sees "new_block" as a block number.
1215 * After the write, link the buffer back to old_block.
1216 * All this must be done in bufio lock, so that block number
1217 * change isn't visible to other threads.
1218 */
1219 old_block = b->block;
1220 __unlink_buffer(b);
1221 __link_buffer(b, new_block, b->list_mode);
1222 submit_io(b, WRITE, new_block, write_endio);
1223 wait_on_bit(&b->state, B_WRITING,
1224 do_io_schedule, TASK_UNINTERRUPTIBLE);
1225 __unlink_buffer(b);
1226 __link_buffer(b, old_block, b->list_mode);
1227 }
1228
1229 dm_bufio_unlock(c);
1230 dm_bufio_release(b);
1231}
1232EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1233
1234unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1235{
1236 return c->block_size;
1237}
1238EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1239
1240sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1241{
1242 return i_size_read(c->bdev->bd_inode) >>
1243 (SECTOR_SHIFT + c->sectors_per_block_bits);
1244}
1245EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1246
1247sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1248{
1249 return b->block;
1250}
1251EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1252
1253void *dm_bufio_get_block_data(struct dm_buffer *b)
1254{
1255 return b->data;
1256}
1257EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1258
1259void *dm_bufio_get_aux_data(struct dm_buffer *b)
1260{
1261 return b + 1;
1262}
1263EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1264
1265struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1266{
1267 return b->c;
1268}
1269EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1270
1271static void drop_buffers(struct dm_bufio_client *c)
1272{
1273 struct dm_buffer *b;
1274 int i;
1275
1276 BUG_ON(dm_bufio_in_request());
1277
1278 /*
1279 * An optimization so that the buffers are not written one-by-one.
1280 */
1281 dm_bufio_write_dirty_buffers_async(c);
1282
1283 dm_bufio_lock(c);
1284
1285 while ((b = __get_unclaimed_buffer(c)))
1286 __free_buffer_wake(b);
1287
1288 for (i = 0; i < LIST_SIZE; i++)
1289 list_for_each_entry(b, &c->lru[i], lru_list)
1290 DMERR("leaked buffer %llx, hold count %u, list %d",
1291 (unsigned long long)b->block, b->hold_count, i);
1292
1293 for (i = 0; i < LIST_SIZE; i++)
1294 BUG_ON(!list_empty(&c->lru[i]));
1295
1296 dm_bufio_unlock(c);
1297}
1298
1299/*
1300 * Test if the buffer is unused and too old, and commit it.
1301 * At if noio is set, we must not do any I/O because we hold
1302 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets rerouted to
1303 * different bufio client.
1304 */
1305static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1306 unsigned long max_jiffies)
1307{
1308 if (jiffies - b->last_accessed < max_jiffies)
1309 return 1;
1310
1311 if (!(gfp & __GFP_IO)) {
1312 if (test_bit(B_READING, &b->state) ||
1313 test_bit(B_WRITING, &b->state) ||
1314 test_bit(B_DIRTY, &b->state))
1315 return 1;
1316 }
1317
1318 if (b->hold_count)
1319 return 1;
1320
1321 __make_buffer_clean(b);
1322 __unlink_buffer(b);
1323 __free_buffer_wake(b);
1324
1325 return 0;
1326}
1327
1328static void __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1329 struct shrink_control *sc)
1330{
1331 int l;
1332 struct dm_buffer *b, *tmp;
1333
1334 for (l = 0; l < LIST_SIZE; l++) {
1335 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list)
1336 if (!__cleanup_old_buffer(b, sc->gfp_mask, 0) &&
1337 !--nr_to_scan)
1338 return;
1339 dm_bufio_cond_resched();
1340 }
1341}
1342
1343static int shrink(struct shrinker *shrinker, struct shrink_control *sc)
1344{
1345 struct dm_bufio_client *c =
1346 container_of(shrinker, struct dm_bufio_client, shrinker);
1347 unsigned long r;
1348 unsigned long nr_to_scan = sc->nr_to_scan;
1349
1350 if (sc->gfp_mask & __GFP_IO)
1351 dm_bufio_lock(c);
1352 else if (!dm_bufio_trylock(c))
1353 return !nr_to_scan ? 0 : -1;
1354
1355 if (nr_to_scan)
1356 __scan(c, nr_to_scan, sc);
1357
1358 r = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1359 if (r > INT_MAX)
1360 r = INT_MAX;
1361
1362 dm_bufio_unlock(c);
1363
1364 return r;
1365}
1366
1367/*
1368 * Create the buffering interface
1369 */
1370struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1371 unsigned reserved_buffers, unsigned aux_size,
1372 void (*alloc_callback)(struct dm_buffer *),
1373 void (*write_callback)(struct dm_buffer *))
1374{
1375 int r;
1376 struct dm_bufio_client *c;
1377 unsigned i;
1378
1379 BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1380 (block_size & (block_size - 1)));
1381
1382 c = kmalloc(sizeof(*c), GFP_KERNEL);
1383 if (!c) {
1384 r = -ENOMEM;
1385 goto bad_client;
1386 }
1387 c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1388 if (!c->cache_hash) {
1389 r = -ENOMEM;
1390 goto bad_hash;
1391 }
1392
1393 c->bdev = bdev;
1394 c->block_size = block_size;
1395 c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1396 c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1397 ffs(block_size) - 1 - PAGE_SHIFT : 0;
1398 c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1399 PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1400
1401 c->aux_size = aux_size;
1402 c->alloc_callback = alloc_callback;
1403 c->write_callback = write_callback;
1404
1405 for (i = 0; i < LIST_SIZE; i++) {
1406 INIT_LIST_HEAD(&c->lru[i]);
1407 c->n_buffers[i] = 0;
1408 }
1409
1410 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1411 INIT_HLIST_HEAD(&c->cache_hash[i]);
1412
1413 mutex_init(&c->lock);
1414 INIT_LIST_HEAD(&c->reserved_buffers);
1415 c->need_reserved_buffers = reserved_buffers;
1416
1417 init_waitqueue_head(&c->free_buffer_wait);
1418 c->async_write_error = 0;
1419
1420 c->dm_io = dm_io_client_create();
1421 if (IS_ERR(c->dm_io)) {
1422 r = PTR_ERR(c->dm_io);
1423 goto bad_dm_io;
1424 }
1425
1426 mutex_lock(&dm_bufio_clients_lock);
1427 if (c->blocks_per_page_bits) {
1428 if (!DM_BUFIO_CACHE_NAME(c)) {
1429 DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1430 if (!DM_BUFIO_CACHE_NAME(c)) {
1431 r = -ENOMEM;
1432 mutex_unlock(&dm_bufio_clients_lock);
1433 goto bad_cache;
1434 }
1435 }
1436
1437 if (!DM_BUFIO_CACHE(c)) {
1438 DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1439 c->block_size,
1440 c->block_size, 0, NULL);
1441 if (!DM_BUFIO_CACHE(c)) {
1442 r = -ENOMEM;
1443 mutex_unlock(&dm_bufio_clients_lock);
1444 goto bad_cache;
1445 }
1446 }
1447 }
1448 mutex_unlock(&dm_bufio_clients_lock);
1449
1450 while (c->need_reserved_buffers) {
1451 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1452
1453 if (!b) {
1454 r = -ENOMEM;
1455 goto bad_buffer;
1456 }
1457 __free_buffer_wake(b);
1458 }
1459
1460 mutex_lock(&dm_bufio_clients_lock);
1461 dm_bufio_client_count++;
1462 list_add(&c->client_list, &dm_bufio_all_clients);
1463 __cache_size_refresh();
1464 mutex_unlock(&dm_bufio_clients_lock);
1465
1466 c->shrinker.shrink = shrink;
1467 c->shrinker.seeks = 1;
1468 c->shrinker.batch = 0;
1469 register_shrinker(&c->shrinker);
1470
1471 return c;
1472
1473bad_buffer:
1474bad_cache:
1475 while (!list_empty(&c->reserved_buffers)) {
1476 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1477 struct dm_buffer, lru_list);
1478 list_del(&b->lru_list);
1479 free_buffer(b);
1480 }
1481 dm_io_client_destroy(c->dm_io);
1482bad_dm_io:
1483 vfree(c->cache_hash);
1484bad_hash:
1485 kfree(c);
1486bad_client:
1487 return ERR_PTR(r);
1488}
1489EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1490
1491/*
1492 * Free the buffering interface.
1493 * It is required that there are no references on any buffers.
1494 */
1495void dm_bufio_client_destroy(struct dm_bufio_client *c)
1496{
1497 unsigned i;
1498
1499 drop_buffers(c);
1500
1501 unregister_shrinker(&c->shrinker);
1502
1503 mutex_lock(&dm_bufio_clients_lock);
1504
1505 list_del(&c->client_list);
1506 dm_bufio_client_count--;
1507 __cache_size_refresh();
1508
1509 mutex_unlock(&dm_bufio_clients_lock);
1510
1511 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1512 BUG_ON(!hlist_empty(&c->cache_hash[i]));
1513
1514 BUG_ON(c->need_reserved_buffers);
1515
1516 while (!list_empty(&c->reserved_buffers)) {
1517 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1518 struct dm_buffer, lru_list);
1519 list_del(&b->lru_list);
1520 free_buffer(b);
1521 }
1522
1523 for (i = 0; i < LIST_SIZE; i++)
1524 if (c->n_buffers[i])
1525 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1526
1527 for (i = 0; i < LIST_SIZE; i++)
1528 BUG_ON(c->n_buffers[i]);
1529
1530 dm_io_client_destroy(c->dm_io);
1531 vfree(c->cache_hash);
1532 kfree(c);
1533}
1534EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1535
1536static void cleanup_old_buffers(void)
1537{
1538 unsigned long max_age = dm_bufio_max_age;
1539 struct dm_bufio_client *c;
1540
1541 barrier();
1542
1543 if (max_age > ULONG_MAX / HZ)
1544 max_age = ULONG_MAX / HZ;
1545
1546 mutex_lock(&dm_bufio_clients_lock);
1547 list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1548 if (!dm_bufio_trylock(c))
1549 continue;
1550
1551 while (!list_empty(&c->lru[LIST_CLEAN])) {
1552 struct dm_buffer *b;
1553 b = list_entry(c->lru[LIST_CLEAN].prev,
1554 struct dm_buffer, lru_list);
1555 if (__cleanup_old_buffer(b, 0, max_age * HZ))
1556 break;
1557 dm_bufio_cond_resched();
1558 }
1559
1560 dm_bufio_unlock(c);
1561 dm_bufio_cond_resched();
1562 }
1563 mutex_unlock(&dm_bufio_clients_lock);
1564}
1565
1566static struct workqueue_struct *dm_bufio_wq;
1567static struct delayed_work dm_bufio_work;
1568
1569static void work_fn(struct work_struct *w)
1570{
1571 cleanup_old_buffers();
1572
1573 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1574 DM_BUFIO_WORK_TIMER_SECS * HZ);
1575}
1576
1577/*----------------------------------------------------------------
1578 * Module setup
1579 *--------------------------------------------------------------*/
1580
1581/*
1582 * This is called only once for the whole dm_bufio module.
1583 * It initializes memory limit.
1584 */
1585static int __init dm_bufio_init(void)
1586{
1587 __u64 mem;
1588
1589 memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1590 memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1591
1592 mem = (__u64)((totalram_pages - totalhigh_pages) *
1593 DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1594
1595 if (mem > ULONG_MAX)
1596 mem = ULONG_MAX;
1597
1598#ifdef CONFIG_MMU
1599 /*
1600 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1601 * in fs/proc/internal.h
1602 */
1603 if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1604 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1605#endif
1606
1607 dm_bufio_default_cache_size = mem;
1608
1609 mutex_lock(&dm_bufio_clients_lock);
1610 __cache_size_refresh();
1611 mutex_unlock(&dm_bufio_clients_lock);
1612
1613 dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1614 if (!dm_bufio_wq)
1615 return -ENOMEM;
1616
1617 INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1618 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1619 DM_BUFIO_WORK_TIMER_SECS * HZ);
1620
1621 return 0;
1622}
1623
1624/*
1625 * This is called once when unloading the dm_bufio module.
1626 */
1627static void __exit dm_bufio_exit(void)
1628{
1629 int bug = 0;
1630 int i;
1631
1632 cancel_delayed_work_sync(&dm_bufio_work);
1633 destroy_workqueue(dm_bufio_wq);
1634
1635 for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1636 struct kmem_cache *kc = dm_bufio_caches[i];
1637
1638 if (kc)
1639 kmem_cache_destroy(kc);
1640 }
1641
1642 for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1643 kfree(dm_bufio_cache_names[i]);
1644
1645 if (dm_bufio_client_count) {
1646 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1647 __func__, dm_bufio_client_count);
1648 bug = 1;
1649 }
1650
1651 if (dm_bufio_current_allocated) {
1652 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1653 __func__, dm_bufio_current_allocated);
1654 bug = 1;
1655 }
1656
1657 if (dm_bufio_allocated_get_free_pages) {
1658 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1659 __func__, dm_bufio_allocated_get_free_pages);
1660 bug = 1;
1661 }
1662
1663 if (dm_bufio_allocated_vmalloc) {
1664 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1665 __func__, dm_bufio_allocated_vmalloc);
1666 bug = 1;
1667 }
1668
1669 if (bug)
1670 BUG();
1671}
1672
1673module_init(dm_bufio_init)
1674module_exit(dm_bufio_exit)
1675
1676module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1677MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1678
1679module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1680MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1681
1682module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1683MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1684
1685module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1686MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1687
1688module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1689MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1690
1691module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1692MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1693
1694module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1695MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1696
1697MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1698MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1699MODULE_LICENSE("GPL");
diff --git a/drivers/md/dm-bufio.h b/drivers/md/dm-bufio.h
new file mode 100644
index 000000000000..5c4c3a04e381
--- /dev/null
+++ b/drivers/md/dm-bufio.h
@@ -0,0 +1,112 @@
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#ifndef DM_BUFIO_H
10#define DM_BUFIO_H
11
12#include <linux/blkdev.h>
13#include <linux/types.h>
14
15/*----------------------------------------------------------------*/
16
17struct dm_bufio_client;
18struct dm_buffer;
19
20/*
21 * Create a buffered IO cache on a given device
22 */
23struct dm_bufio_client *
24dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
25 unsigned reserved_buffers, unsigned aux_size,
26 void (*alloc_callback)(struct dm_buffer *),
27 void (*write_callback)(struct dm_buffer *));
28
29/*
30 * Release a buffered IO cache.
31 */
32void dm_bufio_client_destroy(struct dm_bufio_client *c);
33
34/*
35 * WARNING: to avoid deadlocks, these conditions are observed:
36 *
37 * - At most one thread can hold at most "reserved_buffers" simultaneously.
38 * - Each other threads can hold at most one buffer.
39 * - Threads which call only dm_bufio_get can hold unlimited number of
40 * buffers.
41 */
42
43/*
44 * Read a given block from disk. Returns pointer to data. Returns a
45 * pointer to dm_buffer that can be used to release the buffer or to make
46 * it dirty.
47 */
48void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
49 struct dm_buffer **bp);
50
51/*
52 * Like dm_bufio_read, but return buffer from cache, don't read
53 * it. If the buffer is not in the cache, return NULL.
54 */
55void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
56 struct dm_buffer **bp);
57
58/*
59 * Like dm_bufio_read, but don't read anything from the disk. It is
60 * expected that the caller initializes the buffer and marks it dirty.
61 */
62void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
63 struct dm_buffer **bp);
64
65/*
66 * Release a reference obtained with dm_bufio_{read,get,new}. The data
67 * pointer and dm_buffer pointer is no longer valid after this call.
68 */
69void dm_bufio_release(struct dm_buffer *b);
70
71/*
72 * Mark a buffer dirty. It should be called after the buffer is modified.
73 *
74 * In case of memory pressure, the buffer may be written after
75 * dm_bufio_mark_buffer_dirty, but before dm_bufio_write_dirty_buffers. So
76 * dm_bufio_write_dirty_buffers guarantees that the buffer is on-disk but
77 * the actual writing may occur earlier.
78 */
79void dm_bufio_mark_buffer_dirty(struct dm_buffer *b);
80
81/*
82 * Initiate writing of dirty buffers, without waiting for completion.
83 */
84void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c);
85
86/*
87 * Write all dirty buffers. Guarantees that all dirty buffers created prior
88 * to this call are on disk when this call exits.
89 */
90int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c);
91
92/*
93 * Send an empty write barrier to the device to flush hardware disk cache.
94 */
95int dm_bufio_issue_flush(struct dm_bufio_client *c);
96
97/*
98 * Like dm_bufio_release but also move the buffer to the new
99 * block. dm_bufio_write_dirty_buffers is needed to commit the new block.
100 */
101void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block);
102
103unsigned dm_bufio_get_block_size(struct dm_bufio_client *c);
104sector_t dm_bufio_get_device_size(struct dm_bufio_client *c);
105sector_t dm_bufio_get_block_number(struct dm_buffer *b);
106void *dm_bufio_get_block_data(struct dm_buffer *b);
107void *dm_bufio_get_aux_data(struct dm_buffer *b);
108struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b);
109
110/*----------------------------------------------------------------*/
111
112#endif
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-26 18:28:32 -0500