Commit | Line | Data |
---|---|---|
95d402f0 MP |
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> | |
95d402f0 | 15 | #include <linux/shrinker.h> |
6f66263f | 16 | #include <linux/module.h> |
4e420c45 | 17 | #include <linux/rbtree.h> |
95d402f0 MP |
18 | |
19 | #define DM_MSG_PREFIX "bufio" | |
20 | ||
21 | /* | |
22 | * Memory management policy: | |
23 | * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory | |
24 | * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower). | |
25 | * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers. | |
26 | * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT | |
27 | * dirty buffers. | |
28 | */ | |
29 | #define DM_BUFIO_MIN_BUFFERS 8 | |
30 | ||
31 | #define DM_BUFIO_MEMORY_PERCENT 2 | |
32 | #define DM_BUFIO_VMALLOC_PERCENT 25 | |
33 | #define DM_BUFIO_WRITEBACK_PERCENT 75 | |
34 | ||
35 | /* | |
36 | * Check buffer ages in this interval (seconds) | |
37 | */ | |
33096a78 | 38 | #define DM_BUFIO_WORK_TIMER_SECS 30 |
95d402f0 MP |
39 | |
40 | /* | |
41 | * Free buffers when they are older than this (seconds) | |
42 | */ | |
33096a78 | 43 | #define DM_BUFIO_DEFAULT_AGE_SECS 300 |
95d402f0 MP |
44 | |
45 | /* | |
33096a78 | 46 | * The nr of bytes of cached data to keep around. |
95d402f0 | 47 | */ |
33096a78 | 48 | #define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024) |
95d402f0 MP |
49 | |
50 | /* | |
95d402f0 MP |
51 | * The number of bvec entries that are embedded directly in the buffer. |
52 | * If the chunk size is larger, dm-io is used to do the io. | |
95d402f0 | 53 | */ |
95d402f0 | 54 | #define DM_BUFIO_INLINE_VECS 16 |
95d402f0 MP |
55 | |
56 | /* | |
57 | * Don't try to use kmem_cache_alloc for blocks larger than this. | |
58 | * For explanation, see alloc_buffer_data below. | |
59 | */ | |
60 | #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT (PAGE_SIZE >> 1) | |
61 | #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT (PAGE_SIZE << (MAX_ORDER - 1)) | |
62 | ||
63 | /* | |
64 | * dm_buffer->list_mode | |
65 | */ | |
66 | #define LIST_CLEAN 0 | |
67 | #define LIST_DIRTY 1 | |
68 | #define LIST_SIZE 2 | |
69 | ||
70 | /* | |
71 | * Linking of buffers: | |
72 | * All buffers are linked to cache_hash with their hash_list field. | |
73 | * | |
74 | * Clean buffers that are not being written (B_WRITING not set) | |
75 | * are linked to lru[LIST_CLEAN] with their lru_list field. | |
76 | * | |
77 | * Dirty and clean buffers that are being written are linked to | |
78 | * lru[LIST_DIRTY] with their lru_list field. When the write | |
79 | * finishes, the buffer cannot be relinked immediately (because we | |
80 | * are in an interrupt context and relinking requires process | |
81 | * context), so some clean-not-writing buffers can be held on | |
82 | * dirty_lru too. They are later added to lru in the process | |
83 | * context. | |
84 | */ | |
85 | struct dm_bufio_client { | |
86 | struct mutex lock; | |
87 | ||
88 | struct list_head lru[LIST_SIZE]; | |
89 | unsigned long n_buffers[LIST_SIZE]; | |
90 | ||
91 | struct block_device *bdev; | |
92 | unsigned block_size; | |
93 | unsigned char sectors_per_block_bits; | |
94 | unsigned char pages_per_block_bits; | |
95 | unsigned char blocks_per_page_bits; | |
96 | unsigned aux_size; | |
97 | void (*alloc_callback)(struct dm_buffer *); | |
98 | void (*write_callback)(struct dm_buffer *); | |
99 | ||
100 | struct dm_io_client *dm_io; | |
101 | ||
102 | struct list_head reserved_buffers; | |
103 | unsigned need_reserved_buffers; | |
104 | ||
55b082e6 MP |
105 | unsigned minimum_buffers; |
106 | ||
4e420c45 | 107 | struct rb_root buffer_tree; |
95d402f0 MP |
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 | */ | |
128 | enum 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 | ||
135 | struct dm_buffer { | |
4e420c45 | 136 | struct rb_node node; |
95d402f0 MP |
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; | |
2480945c | 148 | struct list_head write_list; |
95d402f0 MP |
149 | struct bio bio; |
150 | struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS]; | |
151 | }; | |
152 | ||
153 | /*----------------------------------------------------------------*/ | |
154 | ||
155 | static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT]; | |
156 | static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT]; | |
157 | ||
158 | static inline int dm_bufio_cache_index(struct dm_bufio_client *c) | |
159 | { | |
160 | unsigned ret = c->blocks_per_page_bits - 1; | |
161 | ||
162 | BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches)); | |
163 | ||
164 | return ret; | |
165 | } | |
166 | ||
167 | #define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)]) | |
168 | #define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)]) | |
169 | ||
170 | #define dm_bufio_in_request() (!!current->bio_list) | |
171 | ||
172 | static void dm_bufio_lock(struct dm_bufio_client *c) | |
173 | { | |
174 | mutex_lock_nested(&c->lock, dm_bufio_in_request()); | |
175 | } | |
176 | ||
177 | static int dm_bufio_trylock(struct dm_bufio_client *c) | |
178 | { | |
179 | return mutex_trylock(&c->lock); | |
180 | } | |
181 | ||
182 | static void dm_bufio_unlock(struct dm_bufio_client *c) | |
183 | { | |
184 | mutex_unlock(&c->lock); | |
185 | } | |
186 | ||
187 | /* | |
188 | * FIXME Move to sched.h? | |
189 | */ | |
190 | #ifdef CONFIG_PREEMPT_VOLUNTARY | |
191 | # define dm_bufio_cond_resched() \ | |
192 | do { \ | |
193 | if (unlikely(need_resched())) \ | |
194 | _cond_resched(); \ | |
195 | } while (0) | |
196 | #else | |
197 | # define dm_bufio_cond_resched() do { } while (0) | |
198 | #endif | |
199 | ||
200 | /*----------------------------------------------------------------*/ | |
201 | ||
202 | /* | |
203 | * Default cache size: available memory divided by the ratio. | |
204 | */ | |
205 | static unsigned long dm_bufio_default_cache_size; | |
206 | ||
207 | /* | |
208 | * Total cache size set by the user. | |
209 | */ | |
210 | static unsigned long dm_bufio_cache_size; | |
211 | ||
212 | /* | |
213 | * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change | |
214 | * at any time. If it disagrees, the user has changed cache size. | |
215 | */ | |
216 | static unsigned long dm_bufio_cache_size_latch; | |
217 | ||
218 | static DEFINE_SPINLOCK(param_spinlock); | |
219 | ||
220 | /* | |
221 | * Buffers are freed after this timeout | |
222 | */ | |
223 | static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS; | |
33096a78 | 224 | static unsigned dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES; |
95d402f0 MP |
225 | |
226 | static unsigned long dm_bufio_peak_allocated; | |
227 | static unsigned long dm_bufio_allocated_kmem_cache; | |
228 | static unsigned long dm_bufio_allocated_get_free_pages; | |
229 | static unsigned long dm_bufio_allocated_vmalloc; | |
230 | static unsigned long dm_bufio_current_allocated; | |
231 | ||
232 | /*----------------------------------------------------------------*/ | |
233 | ||
234 | /* | |
235 | * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count | |
236 | */ | |
237 | static unsigned long dm_bufio_cache_size_per_client; | |
238 | ||
239 | /* | |
240 | * The current number of clients. | |
241 | */ | |
242 | static int dm_bufio_client_count; | |
243 | ||
244 | /* | |
245 | * The list of all clients. | |
246 | */ | |
247 | static LIST_HEAD(dm_bufio_all_clients); | |
248 | ||
249 | /* | |
250 | * This mutex protects dm_bufio_cache_size_latch, | |
251 | * dm_bufio_cache_size_per_client and dm_bufio_client_count | |
252 | */ | |
253 | static DEFINE_MUTEX(dm_bufio_clients_lock); | |
254 | ||
4e420c45 JT |
255 | /*---------------------------------------------------------------- |
256 | * A red/black tree acts as an index for all the buffers. | |
257 | *--------------------------------------------------------------*/ | |
258 | static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block) | |
259 | { | |
260 | struct rb_node *n = c->buffer_tree.rb_node; | |
261 | struct dm_buffer *b; | |
262 | ||
263 | while (n) { | |
264 | b = container_of(n, struct dm_buffer, node); | |
265 | ||
266 | if (b->block == block) | |
267 | return b; | |
268 | ||
269 | n = (b->block < block) ? n->rb_left : n->rb_right; | |
270 | } | |
271 | ||
272 | return NULL; | |
273 | } | |
274 | ||
275 | static void __insert(struct dm_bufio_client *c, struct dm_buffer *b) | |
276 | { | |
277 | struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL; | |
278 | struct dm_buffer *found; | |
279 | ||
280 | while (*new) { | |
281 | found = container_of(*new, struct dm_buffer, node); | |
282 | ||
283 | if (found->block == b->block) { | |
284 | BUG_ON(found != b); | |
285 | return; | |
286 | } | |
287 | ||
288 | parent = *new; | |
289 | new = (found->block < b->block) ? | |
290 | &((*new)->rb_left) : &((*new)->rb_right); | |
291 | } | |
292 | ||
293 | rb_link_node(&b->node, parent, new); | |
294 | rb_insert_color(&b->node, &c->buffer_tree); | |
295 | } | |
296 | ||
297 | static void __remove(struct dm_bufio_client *c, struct dm_buffer *b) | |
298 | { | |
299 | rb_erase(&b->node, &c->buffer_tree); | |
300 | } | |
301 | ||
95d402f0 MP |
302 | /*----------------------------------------------------------------*/ |
303 | ||
304 | static void adjust_total_allocated(enum data_mode data_mode, long diff) | |
305 | { | |
306 | static unsigned long * const class_ptr[DATA_MODE_LIMIT] = { | |
307 | &dm_bufio_allocated_kmem_cache, | |
308 | &dm_bufio_allocated_get_free_pages, | |
309 | &dm_bufio_allocated_vmalloc, | |
310 | }; | |
311 | ||
312 | spin_lock(¶m_spinlock); | |
313 | ||
314 | *class_ptr[data_mode] += diff; | |
315 | ||
316 | dm_bufio_current_allocated += diff; | |
317 | ||
318 | if (dm_bufio_current_allocated > dm_bufio_peak_allocated) | |
319 | dm_bufio_peak_allocated = dm_bufio_current_allocated; | |
320 | ||
321 | spin_unlock(¶m_spinlock); | |
322 | } | |
323 | ||
324 | /* | |
325 | * Change the number of clients and recalculate per-client limit. | |
326 | */ | |
327 | static void __cache_size_refresh(void) | |
328 | { | |
329 | BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock)); | |
330 | BUG_ON(dm_bufio_client_count < 0); | |
331 | ||
fe5fe906 | 332 | dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size); |
95d402f0 MP |
333 | |
334 | /* | |
335 | * Use default if set to 0 and report the actual cache size used. | |
336 | */ | |
337 | if (!dm_bufio_cache_size_latch) { | |
338 | (void)cmpxchg(&dm_bufio_cache_size, 0, | |
339 | dm_bufio_default_cache_size); | |
340 | dm_bufio_cache_size_latch = dm_bufio_default_cache_size; | |
341 | } | |
342 | ||
343 | dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch / | |
344 | (dm_bufio_client_count ? : 1); | |
345 | } | |
346 | ||
347 | /* | |
348 | * Allocating buffer data. | |
349 | * | |
350 | * Small buffers are allocated with kmem_cache, to use space optimally. | |
351 | * | |
352 | * For large buffers, we choose between get_free_pages and vmalloc. | |
353 | * Each has advantages and disadvantages. | |
354 | * | |
355 | * __get_free_pages can randomly fail if the memory is fragmented. | |
356 | * __vmalloc won't randomly fail, but vmalloc space is limited (it may be | |
357 | * as low as 128M) so using it for caching is not appropriate. | |
358 | * | |
359 | * If the allocation may fail we use __get_free_pages. Memory fragmentation | |
360 | * won't have a fatal effect here, but it just causes flushes of some other | |
361 | * buffers and more I/O will be performed. Don't use __get_free_pages if it | |
362 | * always fails (i.e. order >= MAX_ORDER). | |
363 | * | |
364 | * If the allocation shouldn't fail we use __vmalloc. This is only for the | |
365 | * initial reserve allocation, so there's no risk of wasting all vmalloc | |
366 | * space. | |
367 | */ | |
368 | static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask, | |
369 | enum data_mode *data_mode) | |
370 | { | |
502624bd MP |
371 | unsigned noio_flag; |
372 | void *ptr; | |
373 | ||
95d402f0 MP |
374 | if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) { |
375 | *data_mode = DATA_MODE_SLAB; | |
376 | return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask); | |
377 | } | |
378 | ||
379 | if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT && | |
380 | gfp_mask & __GFP_NORETRY) { | |
381 | *data_mode = DATA_MODE_GET_FREE_PAGES; | |
382 | return (void *)__get_free_pages(gfp_mask, | |
383 | c->pages_per_block_bits); | |
384 | } | |
385 | ||
386 | *data_mode = DATA_MODE_VMALLOC; | |
502624bd MP |
387 | |
388 | /* | |
389 | * __vmalloc allocates the data pages and auxiliary structures with | |
390 | * gfp_flags that were specified, but pagetables are always allocated | |
391 | * with GFP_KERNEL, no matter what was specified as gfp_mask. | |
392 | * | |
393 | * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that | |
394 | * all allocations done by this process (including pagetables) are done | |
395 | * as if GFP_NOIO was specified. | |
396 | */ | |
397 | ||
398 | if (gfp_mask & __GFP_NORETRY) | |
399 | noio_flag = memalloc_noio_save(); | |
400 | ||
220cd058 | 401 | ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL); |
502624bd MP |
402 | |
403 | if (gfp_mask & __GFP_NORETRY) | |
404 | memalloc_noio_restore(noio_flag); | |
405 | ||
406 | return ptr; | |
95d402f0 MP |
407 | } |
408 | ||
409 | /* | |
410 | * Free buffer's data. | |
411 | */ | |
412 | static void free_buffer_data(struct dm_bufio_client *c, | |
413 | void *data, enum data_mode data_mode) | |
414 | { | |
415 | switch (data_mode) { | |
416 | case DATA_MODE_SLAB: | |
417 | kmem_cache_free(DM_BUFIO_CACHE(c), data); | |
418 | break; | |
419 | ||
420 | case DATA_MODE_GET_FREE_PAGES: | |
421 | free_pages((unsigned long)data, c->pages_per_block_bits); | |
422 | break; | |
423 | ||
424 | case DATA_MODE_VMALLOC: | |
425 | vfree(data); | |
426 | break; | |
427 | ||
428 | default: | |
429 | DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d", | |
430 | data_mode); | |
431 | BUG(); | |
432 | } | |
433 | } | |
434 | ||
435 | /* | |
436 | * Allocate buffer and its data. | |
437 | */ | |
438 | static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask) | |
439 | { | |
440 | struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size, | |
441 | gfp_mask); | |
442 | ||
443 | if (!b) | |
444 | return NULL; | |
445 | ||
446 | b->c = c; | |
447 | ||
448 | b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode); | |
449 | if (!b->data) { | |
450 | kfree(b); | |
451 | return NULL; | |
452 | } | |
453 | ||
454 | adjust_total_allocated(b->data_mode, (long)c->block_size); | |
455 | ||
456 | return b; | |
457 | } | |
458 | ||
459 | /* | |
460 | * Free buffer and its data. | |
461 | */ | |
462 | static void free_buffer(struct dm_buffer *b) | |
463 | { | |
464 | struct dm_bufio_client *c = b->c; | |
465 | ||
466 | adjust_total_allocated(b->data_mode, -(long)c->block_size); | |
467 | ||
468 | free_buffer_data(c, b->data, b->data_mode); | |
469 | kfree(b); | |
470 | } | |
471 | ||
472 | /* | |
473 | * Link buffer to the hash list and clean or dirty queue. | |
474 | */ | |
475 | static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty) | |
476 | { | |
477 | struct dm_bufio_client *c = b->c; | |
478 | ||
479 | c->n_buffers[dirty]++; | |
480 | b->block = block; | |
481 | b->list_mode = dirty; | |
482 | list_add(&b->lru_list, &c->lru[dirty]); | |
4e420c45 | 483 | __insert(b->c, b); |
95d402f0 MP |
484 | b->last_accessed = jiffies; |
485 | } | |
486 | ||
487 | /* | |
488 | * Unlink buffer from the hash list and dirty or clean queue. | |
489 | */ | |
490 | static void __unlink_buffer(struct dm_buffer *b) | |
491 | { | |
492 | struct dm_bufio_client *c = b->c; | |
493 | ||
494 | BUG_ON(!c->n_buffers[b->list_mode]); | |
495 | ||
496 | c->n_buffers[b->list_mode]--; | |
4e420c45 | 497 | __remove(b->c, b); |
95d402f0 MP |
498 | list_del(&b->lru_list); |
499 | } | |
500 | ||
501 | /* | |
502 | * Place the buffer to the head of dirty or clean LRU queue. | |
503 | */ | |
504 | static void __relink_lru(struct dm_buffer *b, int dirty) | |
505 | { | |
506 | struct dm_bufio_client *c = b->c; | |
507 | ||
508 | BUG_ON(!c->n_buffers[b->list_mode]); | |
509 | ||
510 | c->n_buffers[b->list_mode]--; | |
511 | c->n_buffers[dirty]++; | |
512 | b->list_mode = dirty; | |
54499afb | 513 | list_move(&b->lru_list, &c->lru[dirty]); |
eb76faf5 | 514 | b->last_accessed = jiffies; |
95d402f0 MP |
515 | } |
516 | ||
517 | /*---------------------------------------------------------------- | |
518 | * Submit I/O on the buffer. | |
519 | * | |
520 | * Bio interface is faster but it has some problems: | |
521 | * the vector list is limited (increasing this limit increases | |
522 | * memory-consumption per buffer, so it is not viable); | |
523 | * | |
524 | * the memory must be direct-mapped, not vmalloced; | |
525 | * | |
526 | * the I/O driver can reject requests spuriously if it thinks that | |
527 | * the requests are too big for the device or if they cross a | |
528 | * controller-defined memory boundary. | |
529 | * | |
530 | * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and | |
531 | * it is not vmalloced, try using the bio interface. | |
532 | * | |
533 | * If the buffer is big, if it is vmalloced or if the underlying device | |
534 | * rejects the bio because it is too large, use dm-io layer to do the I/O. | |
535 | * The dm-io layer splits the I/O into multiple requests, avoiding the above | |
536 | * shortcomings. | |
537 | *--------------------------------------------------------------*/ | |
538 | ||
539 | /* | |
540 | * dm-io completion routine. It just calls b->bio.bi_end_io, pretending | |
541 | * that the request was handled directly with bio interface. | |
542 | */ | |
543 | static void dmio_complete(unsigned long error, void *context) | |
544 | { | |
545 | struct dm_buffer *b = context; | |
546 | ||
547 | b->bio.bi_end_io(&b->bio, error ? -EIO : 0); | |
548 | } | |
549 | ||
550 | static void use_dmio(struct dm_buffer *b, int rw, sector_t block, | |
551 | bio_end_io_t *end_io) | |
552 | { | |
553 | int r; | |
554 | struct dm_io_request io_req = { | |
555 | .bi_rw = rw, | |
556 | .notify.fn = dmio_complete, | |
557 | .notify.context = b, | |
558 | .client = b->c->dm_io, | |
559 | }; | |
560 | struct dm_io_region region = { | |
561 | .bdev = b->c->bdev, | |
562 | .sector = block << b->c->sectors_per_block_bits, | |
563 | .count = b->c->block_size >> SECTOR_SHIFT, | |
564 | }; | |
565 | ||
566 | if (b->data_mode != DATA_MODE_VMALLOC) { | |
567 | io_req.mem.type = DM_IO_KMEM; | |
568 | io_req.mem.ptr.addr = b->data; | |
569 | } else { | |
570 | io_req.mem.type = DM_IO_VMA; | |
571 | io_req.mem.ptr.vma = b->data; | |
572 | } | |
573 | ||
574 | b->bio.bi_end_io = end_io; | |
575 | ||
576 | r = dm_io(&io_req, 1, ®ion, NULL); | |
577 | if (r) | |
578 | end_io(&b->bio, r); | |
579 | } | |
580 | ||
445559cd DW |
581 | static void inline_endio(struct bio *bio, int error) |
582 | { | |
583 | bio_end_io_t *end_fn = bio->bi_private; | |
584 | ||
585 | /* | |
586 | * Reset the bio to free any attached resources | |
587 | * (e.g. bio integrity profiles). | |
588 | */ | |
589 | bio_reset(bio); | |
590 | ||
591 | end_fn(bio, error); | |
592 | } | |
593 | ||
95d402f0 MP |
594 | static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block, |
595 | bio_end_io_t *end_io) | |
596 | { | |
597 | char *ptr; | |
598 | int len; | |
599 | ||
600 | bio_init(&b->bio); | |
601 | b->bio.bi_io_vec = b->bio_vec; | |
602 | b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS; | |
4f024f37 | 603 | b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits; |
95d402f0 | 604 | b->bio.bi_bdev = b->c->bdev; |
445559cd DW |
605 | b->bio.bi_end_io = inline_endio; |
606 | /* | |
607 | * Use of .bi_private isn't a problem here because | |
608 | * the dm_buffer's inline bio is local to bufio. | |
609 | */ | |
610 | b->bio.bi_private = end_io; | |
95d402f0 MP |
611 | |
612 | /* | |
613 | * We assume that if len >= PAGE_SIZE ptr is page-aligned. | |
614 | * If len < PAGE_SIZE the buffer doesn't cross page boundary. | |
615 | */ | |
616 | ptr = b->data; | |
617 | len = b->c->block_size; | |
618 | ||
619 | if (len >= PAGE_SIZE) | |
620 | BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1)); | |
621 | else | |
622 | BUG_ON((unsigned long)ptr & (len - 1)); | |
623 | ||
624 | do { | |
625 | if (!bio_add_page(&b->bio, virt_to_page(ptr), | |
626 | len < PAGE_SIZE ? len : PAGE_SIZE, | |
627 | virt_to_phys(ptr) & (PAGE_SIZE - 1))) { | |
628 | BUG_ON(b->c->block_size <= PAGE_SIZE); | |
629 | use_dmio(b, rw, block, end_io); | |
630 | return; | |
631 | } | |
632 | ||
633 | len -= PAGE_SIZE; | |
634 | ptr += PAGE_SIZE; | |
635 | } while (len > 0); | |
636 | ||
637 | submit_bio(rw, &b->bio); | |
638 | } | |
639 | ||
640 | static void submit_io(struct dm_buffer *b, int rw, sector_t block, | |
641 | bio_end_io_t *end_io) | |
642 | { | |
643 | if (rw == WRITE && b->c->write_callback) | |
644 | b->c->write_callback(b); | |
645 | ||
646 | if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE && | |
647 | b->data_mode != DATA_MODE_VMALLOC) | |
648 | use_inline_bio(b, rw, block, end_io); | |
649 | else | |
650 | use_dmio(b, rw, block, end_io); | |
651 | } | |
652 | ||
653 | /*---------------------------------------------------------------- | |
654 | * Writing dirty buffers | |
655 | *--------------------------------------------------------------*/ | |
656 | ||
657 | /* | |
658 | * The endio routine for write. | |
659 | * | |
660 | * Set the error, clear B_WRITING bit and wake anyone who was waiting on | |
661 | * it. | |
662 | */ | |
663 | static void write_endio(struct bio *bio, int error) | |
664 | { | |
665 | struct dm_buffer *b = container_of(bio, struct dm_buffer, bio); | |
666 | ||
667 | b->write_error = error; | |
a66cc28f | 668 | if (unlikely(error)) { |
95d402f0 MP |
669 | struct dm_bufio_client *c = b->c; |
670 | (void)cmpxchg(&c->async_write_error, 0, error); | |
671 | } | |
672 | ||
673 | BUG_ON(!test_bit(B_WRITING, &b->state)); | |
674 | ||
4e857c58 | 675 | smp_mb__before_atomic(); |
95d402f0 | 676 | clear_bit(B_WRITING, &b->state); |
4e857c58 | 677 | smp_mb__after_atomic(); |
95d402f0 MP |
678 | |
679 | wake_up_bit(&b->state, B_WRITING); | |
680 | } | |
681 | ||
95d402f0 MP |
682 | /* |
683 | * Initiate a write on a dirty buffer, but don't wait for it. | |
684 | * | |
685 | * - If the buffer is not dirty, exit. | |
686 | * - If there some previous write going on, wait for it to finish (we can't | |
687 | * have two writes on the same buffer simultaneously). | |
688 | * - Submit our write and don't wait on it. We set B_WRITING indicating | |
689 | * that there is a write in progress. | |
690 | */ | |
2480945c MP |
691 | static void __write_dirty_buffer(struct dm_buffer *b, |
692 | struct list_head *write_list) | |
95d402f0 MP |
693 | { |
694 | if (!test_bit(B_DIRTY, &b->state)) | |
695 | return; | |
696 | ||
697 | clear_bit(B_DIRTY, &b->state); | |
74316201 | 698 | wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE); |
95d402f0 | 699 | |
2480945c MP |
700 | if (!write_list) |
701 | submit_io(b, WRITE, b->block, write_endio); | |
702 | else | |
703 | list_add_tail(&b->write_list, write_list); | |
704 | } | |
705 | ||
706 | static void __flush_write_list(struct list_head *write_list) | |
707 | { | |
708 | struct blk_plug plug; | |
709 | blk_start_plug(&plug); | |
710 | while (!list_empty(write_list)) { | |
711 | struct dm_buffer *b = | |
712 | list_entry(write_list->next, struct dm_buffer, write_list); | |
713 | list_del(&b->write_list); | |
714 | submit_io(b, WRITE, b->block, write_endio); | |
715 | dm_bufio_cond_resched(); | |
716 | } | |
717 | blk_finish_plug(&plug); | |
95d402f0 MP |
718 | } |
719 | ||
720 | /* | |
721 | * Wait until any activity on the buffer finishes. Possibly write the | |
722 | * buffer if it is dirty. When this function finishes, there is no I/O | |
723 | * running on the buffer and the buffer is not dirty. | |
724 | */ | |
725 | static void __make_buffer_clean(struct dm_buffer *b) | |
726 | { | |
727 | BUG_ON(b->hold_count); | |
728 | ||
729 | if (!b->state) /* fast case */ | |
730 | return; | |
731 | ||
74316201 | 732 | wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE); |
2480945c | 733 | __write_dirty_buffer(b, NULL); |
74316201 | 734 | wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE); |
95d402f0 MP |
735 | } |
736 | ||
737 | /* | |
738 | * Find some buffer that is not held by anybody, clean it, unlink it and | |
739 | * return it. | |
740 | */ | |
741 | static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c) | |
742 | { | |
743 | struct dm_buffer *b; | |
744 | ||
745 | list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) { | |
746 | BUG_ON(test_bit(B_WRITING, &b->state)); | |
747 | BUG_ON(test_bit(B_DIRTY, &b->state)); | |
748 | ||
749 | if (!b->hold_count) { | |
750 | __make_buffer_clean(b); | |
751 | __unlink_buffer(b); | |
752 | return b; | |
753 | } | |
754 | dm_bufio_cond_resched(); | |
755 | } | |
756 | ||
757 | list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) { | |
758 | BUG_ON(test_bit(B_READING, &b->state)); | |
759 | ||
760 | if (!b->hold_count) { | |
761 | __make_buffer_clean(b); | |
762 | __unlink_buffer(b); | |
763 | return b; | |
764 | } | |
765 | dm_bufio_cond_resched(); | |
766 | } | |
767 | ||
768 | return NULL; | |
769 | } | |
770 | ||
771 | /* | |
772 | * Wait until some other threads free some buffer or release hold count on | |
773 | * some buffer. | |
774 | * | |
775 | * This function is entered with c->lock held, drops it and regains it | |
776 | * before exiting. | |
777 | */ | |
778 | static void __wait_for_free_buffer(struct dm_bufio_client *c) | |
779 | { | |
780 | DECLARE_WAITQUEUE(wait, current); | |
781 | ||
782 | add_wait_queue(&c->free_buffer_wait, &wait); | |
783 | set_task_state(current, TASK_UNINTERRUPTIBLE); | |
784 | dm_bufio_unlock(c); | |
785 | ||
786 | io_schedule(); | |
787 | ||
95d402f0 MP |
788 | remove_wait_queue(&c->free_buffer_wait, &wait); |
789 | ||
790 | dm_bufio_lock(c); | |
791 | } | |
792 | ||
a66cc28f MP |
793 | enum new_flag { |
794 | NF_FRESH = 0, | |
795 | NF_READ = 1, | |
796 | NF_GET = 2, | |
797 | NF_PREFETCH = 3 | |
798 | }; | |
799 | ||
95d402f0 MP |
800 | /* |
801 | * Allocate a new buffer. If the allocation is not possible, wait until | |
802 | * some other thread frees a buffer. | |
803 | * | |
804 | * May drop the lock and regain it. | |
805 | */ | |
a66cc28f | 806 | static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf) |
95d402f0 MP |
807 | { |
808 | struct dm_buffer *b; | |
809 | ||
810 | /* | |
811 | * dm-bufio is resistant to allocation failures (it just keeps | |
812 | * one buffer reserved in cases all the allocations fail). | |
813 | * So set flags to not try too hard: | |
814 | * GFP_NOIO: don't recurse into the I/O layer | |
815 | * __GFP_NORETRY: don't retry and rather return failure | |
816 | * __GFP_NOMEMALLOC: don't use emergency reserves | |
817 | * __GFP_NOWARN: don't print a warning in case of failure | |
818 | * | |
819 | * For debugging, if we set the cache size to 1, no new buffers will | |
820 | * be allocated. | |
821 | */ | |
822 | while (1) { | |
823 | if (dm_bufio_cache_size_latch != 1) { | |
824 | b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); | |
825 | if (b) | |
826 | return b; | |
827 | } | |
828 | ||
a66cc28f MP |
829 | if (nf == NF_PREFETCH) |
830 | return NULL; | |
831 | ||
95d402f0 MP |
832 | if (!list_empty(&c->reserved_buffers)) { |
833 | b = list_entry(c->reserved_buffers.next, | |
834 | struct dm_buffer, lru_list); | |
835 | list_del(&b->lru_list); | |
836 | c->need_reserved_buffers++; | |
837 | ||
838 | return b; | |
839 | } | |
840 | ||
841 | b = __get_unclaimed_buffer(c); | |
842 | if (b) | |
843 | return b; | |
844 | ||
845 | __wait_for_free_buffer(c); | |
846 | } | |
847 | } | |
848 | ||
a66cc28f | 849 | static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf) |
95d402f0 | 850 | { |
a66cc28f MP |
851 | struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf); |
852 | ||
853 | if (!b) | |
854 | return NULL; | |
95d402f0 MP |
855 | |
856 | if (c->alloc_callback) | |
857 | c->alloc_callback(b); | |
858 | ||
859 | return b; | |
860 | } | |
861 | ||
862 | /* | |
863 | * Free a buffer and wake other threads waiting for free buffers. | |
864 | */ | |
865 | static void __free_buffer_wake(struct dm_buffer *b) | |
866 | { | |
867 | struct dm_bufio_client *c = b->c; | |
868 | ||
869 | if (!c->need_reserved_buffers) | |
870 | free_buffer(b); | |
871 | else { | |
872 | list_add(&b->lru_list, &c->reserved_buffers); | |
873 | c->need_reserved_buffers--; | |
874 | } | |
875 | ||
876 | wake_up(&c->free_buffer_wait); | |
877 | } | |
878 | ||
2480945c MP |
879 | static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait, |
880 | struct list_head *write_list) | |
95d402f0 MP |
881 | { |
882 | struct dm_buffer *b, *tmp; | |
883 | ||
884 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) { | |
885 | BUG_ON(test_bit(B_READING, &b->state)); | |
886 | ||
887 | if (!test_bit(B_DIRTY, &b->state) && | |
888 | !test_bit(B_WRITING, &b->state)) { | |
889 | __relink_lru(b, LIST_CLEAN); | |
890 | continue; | |
891 | } | |
892 | ||
893 | if (no_wait && test_bit(B_WRITING, &b->state)) | |
894 | return; | |
895 | ||
2480945c | 896 | __write_dirty_buffer(b, write_list); |
95d402f0 MP |
897 | dm_bufio_cond_resched(); |
898 | } | |
899 | } | |
900 | ||
901 | /* | |
902 | * Get writeback threshold and buffer limit for a given client. | |
903 | */ | |
904 | static void __get_memory_limit(struct dm_bufio_client *c, | |
905 | unsigned long *threshold_buffers, | |
906 | unsigned long *limit_buffers) | |
907 | { | |
908 | unsigned long buffers; | |
909 | ||
fe5fe906 | 910 | if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) { |
95d402f0 MP |
911 | mutex_lock(&dm_bufio_clients_lock); |
912 | __cache_size_refresh(); | |
913 | mutex_unlock(&dm_bufio_clients_lock); | |
914 | } | |
915 | ||
916 | buffers = dm_bufio_cache_size_per_client >> | |
917 | (c->sectors_per_block_bits + SECTOR_SHIFT); | |
918 | ||
55b082e6 MP |
919 | if (buffers < c->minimum_buffers) |
920 | buffers = c->minimum_buffers; | |
95d402f0 MP |
921 | |
922 | *limit_buffers = buffers; | |
923 | *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100; | |
924 | } | |
925 | ||
926 | /* | |
927 | * Check if we're over watermark. | |
928 | * If we are over threshold_buffers, start freeing buffers. | |
929 | * If we're over "limit_buffers", block until we get under the limit. | |
930 | */ | |
2480945c MP |
931 | static void __check_watermark(struct dm_bufio_client *c, |
932 | struct list_head *write_list) | |
95d402f0 MP |
933 | { |
934 | unsigned long threshold_buffers, limit_buffers; | |
935 | ||
936 | __get_memory_limit(c, &threshold_buffers, &limit_buffers); | |
937 | ||
938 | while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] > | |
939 | limit_buffers) { | |
940 | ||
941 | struct dm_buffer *b = __get_unclaimed_buffer(c); | |
942 | ||
943 | if (!b) | |
944 | return; | |
945 | ||
946 | __free_buffer_wake(b); | |
947 | dm_bufio_cond_resched(); | |
948 | } | |
949 | ||
950 | if (c->n_buffers[LIST_DIRTY] > threshold_buffers) | |
2480945c | 951 | __write_dirty_buffers_async(c, 1, write_list); |
95d402f0 MP |
952 | } |
953 | ||
95d402f0 MP |
954 | /*---------------------------------------------------------------- |
955 | * Getting a buffer | |
956 | *--------------------------------------------------------------*/ | |
957 | ||
95d402f0 | 958 | static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block, |
2480945c MP |
959 | enum new_flag nf, int *need_submit, |
960 | struct list_head *write_list) | |
95d402f0 MP |
961 | { |
962 | struct dm_buffer *b, *new_b = NULL; | |
963 | ||
964 | *need_submit = 0; | |
965 | ||
966 | b = __find(c, block); | |
a66cc28f MP |
967 | if (b) |
968 | goto found_buffer; | |
95d402f0 MP |
969 | |
970 | if (nf == NF_GET) | |
971 | return NULL; | |
972 | ||
a66cc28f MP |
973 | new_b = __alloc_buffer_wait(c, nf); |
974 | if (!new_b) | |
975 | return NULL; | |
95d402f0 MP |
976 | |
977 | /* | |
978 | * We've had a period where the mutex was unlocked, so need to | |
979 | * recheck the hash table. | |
980 | */ | |
981 | b = __find(c, block); | |
982 | if (b) { | |
983 | __free_buffer_wake(new_b); | |
a66cc28f | 984 | goto found_buffer; |
95d402f0 MP |
985 | } |
986 | ||
2480945c | 987 | __check_watermark(c, write_list); |
95d402f0 MP |
988 | |
989 | b = new_b; | |
990 | b->hold_count = 1; | |
991 | b->read_error = 0; | |
992 | b->write_error = 0; | |
993 | __link_buffer(b, block, LIST_CLEAN); | |
994 | ||
995 | if (nf == NF_FRESH) { | |
996 | b->state = 0; | |
997 | return b; | |
998 | } | |
999 | ||
1000 | b->state = 1 << B_READING; | |
1001 | *need_submit = 1; | |
1002 | ||
1003 | return b; | |
a66cc28f MP |
1004 | |
1005 | found_buffer: | |
1006 | if (nf == NF_PREFETCH) | |
1007 | return NULL; | |
1008 | /* | |
1009 | * Note: it is essential that we don't wait for the buffer to be | |
1010 | * read if dm_bufio_get function is used. Both dm_bufio_get and | |
1011 | * dm_bufio_prefetch can be used in the driver request routine. | |
1012 | * If the user called both dm_bufio_prefetch and dm_bufio_get on | |
1013 | * the same buffer, it would deadlock if we waited. | |
1014 | */ | |
1015 | if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state))) | |
1016 | return NULL; | |
1017 | ||
1018 | b->hold_count++; | |
1019 | __relink_lru(b, test_bit(B_DIRTY, &b->state) || | |
1020 | test_bit(B_WRITING, &b->state)); | |
1021 | return b; | |
95d402f0 MP |
1022 | } |
1023 | ||
1024 | /* | |
1025 | * The endio routine for reading: set the error, clear the bit and wake up | |
1026 | * anyone waiting on the buffer. | |
1027 | */ | |
1028 | static void read_endio(struct bio *bio, int error) | |
1029 | { | |
1030 | struct dm_buffer *b = container_of(bio, struct dm_buffer, bio); | |
1031 | ||
1032 | b->read_error = error; | |
1033 | ||
1034 | BUG_ON(!test_bit(B_READING, &b->state)); | |
1035 | ||
4e857c58 | 1036 | smp_mb__before_atomic(); |
95d402f0 | 1037 | clear_bit(B_READING, &b->state); |
4e857c58 | 1038 | smp_mb__after_atomic(); |
95d402f0 MP |
1039 | |
1040 | wake_up_bit(&b->state, B_READING); | |
1041 | } | |
1042 | ||
1043 | /* | |
1044 | * A common routine for dm_bufio_new and dm_bufio_read. Operation of these | |
1045 | * functions is similar except that dm_bufio_new doesn't read the | |
1046 | * buffer from the disk (assuming that the caller overwrites all the data | |
1047 | * and uses dm_bufio_mark_buffer_dirty to write new data back). | |
1048 | */ | |
1049 | static void *new_read(struct dm_bufio_client *c, sector_t block, | |
1050 | enum new_flag nf, struct dm_buffer **bp) | |
1051 | { | |
1052 | int need_submit; | |
1053 | struct dm_buffer *b; | |
1054 | ||
2480945c MP |
1055 | LIST_HEAD(write_list); |
1056 | ||
95d402f0 | 1057 | dm_bufio_lock(c); |
2480945c | 1058 | b = __bufio_new(c, block, nf, &need_submit, &write_list); |
95d402f0 MP |
1059 | dm_bufio_unlock(c); |
1060 | ||
2480945c MP |
1061 | __flush_write_list(&write_list); |
1062 | ||
a66cc28f | 1063 | if (!b) |
95d402f0 MP |
1064 | return b; |
1065 | ||
1066 | if (need_submit) | |
1067 | submit_io(b, READ, b->block, read_endio); | |
1068 | ||
74316201 | 1069 | wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE); |
95d402f0 MP |
1070 | |
1071 | if (b->read_error) { | |
1072 | int error = b->read_error; | |
1073 | ||
1074 | dm_bufio_release(b); | |
1075 | ||
1076 | return ERR_PTR(error); | |
1077 | } | |
1078 | ||
1079 | *bp = b; | |
1080 | ||
1081 | return b->data; | |
1082 | } | |
1083 | ||
1084 | void *dm_bufio_get(struct dm_bufio_client *c, sector_t block, | |
1085 | struct dm_buffer **bp) | |
1086 | { | |
1087 | return new_read(c, block, NF_GET, bp); | |
1088 | } | |
1089 | EXPORT_SYMBOL_GPL(dm_bufio_get); | |
1090 | ||
1091 | void *dm_bufio_read(struct dm_bufio_client *c, sector_t block, | |
1092 | struct dm_buffer **bp) | |
1093 | { | |
1094 | BUG_ON(dm_bufio_in_request()); | |
1095 | ||
1096 | return new_read(c, block, NF_READ, bp); | |
1097 | } | |
1098 | EXPORT_SYMBOL_GPL(dm_bufio_read); | |
1099 | ||
1100 | void *dm_bufio_new(struct dm_bufio_client *c, sector_t block, | |
1101 | struct dm_buffer **bp) | |
1102 | { | |
1103 | BUG_ON(dm_bufio_in_request()); | |
1104 | ||
1105 | return new_read(c, block, NF_FRESH, bp); | |
1106 | } | |
1107 | EXPORT_SYMBOL_GPL(dm_bufio_new); | |
1108 | ||
a66cc28f MP |
1109 | void dm_bufio_prefetch(struct dm_bufio_client *c, |
1110 | sector_t block, unsigned n_blocks) | |
1111 | { | |
1112 | struct blk_plug plug; | |
1113 | ||
2480945c MP |
1114 | LIST_HEAD(write_list); |
1115 | ||
3b6b7813 MP |
1116 | BUG_ON(dm_bufio_in_request()); |
1117 | ||
a66cc28f MP |
1118 | blk_start_plug(&plug); |
1119 | dm_bufio_lock(c); | |
1120 | ||
1121 | for (; n_blocks--; block++) { | |
1122 | int need_submit; | |
1123 | struct dm_buffer *b; | |
2480945c MP |
1124 | b = __bufio_new(c, block, NF_PREFETCH, &need_submit, |
1125 | &write_list); | |
1126 | if (unlikely(!list_empty(&write_list))) { | |
1127 | dm_bufio_unlock(c); | |
1128 | blk_finish_plug(&plug); | |
1129 | __flush_write_list(&write_list); | |
1130 | blk_start_plug(&plug); | |
1131 | dm_bufio_lock(c); | |
1132 | } | |
a66cc28f MP |
1133 | if (unlikely(b != NULL)) { |
1134 | dm_bufio_unlock(c); | |
1135 | ||
1136 | if (need_submit) | |
1137 | submit_io(b, READ, b->block, read_endio); | |
1138 | dm_bufio_release(b); | |
1139 | ||
1140 | dm_bufio_cond_resched(); | |
1141 | ||
1142 | if (!n_blocks) | |
1143 | goto flush_plug; | |
1144 | dm_bufio_lock(c); | |
1145 | } | |
a66cc28f MP |
1146 | } |
1147 | ||
1148 | dm_bufio_unlock(c); | |
1149 | ||
1150 | flush_plug: | |
1151 | blk_finish_plug(&plug); | |
1152 | } | |
1153 | EXPORT_SYMBOL_GPL(dm_bufio_prefetch); | |
1154 | ||
95d402f0 MP |
1155 | void dm_bufio_release(struct dm_buffer *b) |
1156 | { | |
1157 | struct dm_bufio_client *c = b->c; | |
1158 | ||
1159 | dm_bufio_lock(c); | |
1160 | ||
95d402f0 MP |
1161 | BUG_ON(!b->hold_count); |
1162 | ||
1163 | b->hold_count--; | |
1164 | if (!b->hold_count) { | |
1165 | wake_up(&c->free_buffer_wait); | |
1166 | ||
1167 | /* | |
1168 | * If there were errors on the buffer, and the buffer is not | |
1169 | * to be written, free the buffer. There is no point in caching | |
1170 | * invalid buffer. | |
1171 | */ | |
1172 | if ((b->read_error || b->write_error) && | |
a66cc28f | 1173 | !test_bit(B_READING, &b->state) && |
95d402f0 MP |
1174 | !test_bit(B_WRITING, &b->state) && |
1175 | !test_bit(B_DIRTY, &b->state)) { | |
1176 | __unlink_buffer(b); | |
1177 | __free_buffer_wake(b); | |
1178 | } | |
1179 | } | |
1180 | ||
1181 | dm_bufio_unlock(c); | |
1182 | } | |
1183 | EXPORT_SYMBOL_GPL(dm_bufio_release); | |
1184 | ||
1185 | void dm_bufio_mark_buffer_dirty(struct dm_buffer *b) | |
1186 | { | |
1187 | struct dm_bufio_client *c = b->c; | |
1188 | ||
1189 | dm_bufio_lock(c); | |
1190 | ||
a66cc28f MP |
1191 | BUG_ON(test_bit(B_READING, &b->state)); |
1192 | ||
95d402f0 MP |
1193 | if (!test_and_set_bit(B_DIRTY, &b->state)) |
1194 | __relink_lru(b, LIST_DIRTY); | |
1195 | ||
1196 | dm_bufio_unlock(c); | |
1197 | } | |
1198 | EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty); | |
1199 | ||
1200 | void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c) | |
1201 | { | |
2480945c MP |
1202 | LIST_HEAD(write_list); |
1203 | ||
95d402f0 MP |
1204 | BUG_ON(dm_bufio_in_request()); |
1205 | ||
1206 | dm_bufio_lock(c); | |
2480945c | 1207 | __write_dirty_buffers_async(c, 0, &write_list); |
95d402f0 | 1208 | dm_bufio_unlock(c); |
2480945c | 1209 | __flush_write_list(&write_list); |
95d402f0 MP |
1210 | } |
1211 | EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async); | |
1212 | ||
1213 | /* | |
1214 | * For performance, it is essential that the buffers are written asynchronously | |
1215 | * and simultaneously (so that the block layer can merge the writes) and then | |
1216 | * waited upon. | |
1217 | * | |
1218 | * Finally, we flush hardware disk cache. | |
1219 | */ | |
1220 | int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c) | |
1221 | { | |
1222 | int a, f; | |
1223 | unsigned long buffers_processed = 0; | |
1224 | struct dm_buffer *b, *tmp; | |
1225 | ||
2480945c MP |
1226 | LIST_HEAD(write_list); |
1227 | ||
1228 | dm_bufio_lock(c); | |
1229 | __write_dirty_buffers_async(c, 0, &write_list); | |
1230 | dm_bufio_unlock(c); | |
1231 | __flush_write_list(&write_list); | |
95d402f0 | 1232 | dm_bufio_lock(c); |
95d402f0 MP |
1233 | |
1234 | again: | |
1235 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) { | |
1236 | int dropped_lock = 0; | |
1237 | ||
1238 | if (buffers_processed < c->n_buffers[LIST_DIRTY]) | |
1239 | buffers_processed++; | |
1240 | ||
1241 | BUG_ON(test_bit(B_READING, &b->state)); | |
1242 | ||
1243 | if (test_bit(B_WRITING, &b->state)) { | |
1244 | if (buffers_processed < c->n_buffers[LIST_DIRTY]) { | |
1245 | dropped_lock = 1; | |
1246 | b->hold_count++; | |
1247 | dm_bufio_unlock(c); | |
74316201 N |
1248 | wait_on_bit_io(&b->state, B_WRITING, |
1249 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1250 | dm_bufio_lock(c); |
1251 | b->hold_count--; | |
1252 | } else | |
74316201 N |
1253 | wait_on_bit_io(&b->state, B_WRITING, |
1254 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1255 | } |
1256 | ||
1257 | if (!test_bit(B_DIRTY, &b->state) && | |
1258 | !test_bit(B_WRITING, &b->state)) | |
1259 | __relink_lru(b, LIST_CLEAN); | |
1260 | ||
1261 | dm_bufio_cond_resched(); | |
1262 | ||
1263 | /* | |
1264 | * If we dropped the lock, the list is no longer consistent, | |
1265 | * so we must restart the search. | |
1266 | * | |
1267 | * In the most common case, the buffer just processed is | |
1268 | * relinked to the clean list, so we won't loop scanning the | |
1269 | * same buffer again and again. | |
1270 | * | |
1271 | * This may livelock if there is another thread simultaneously | |
1272 | * dirtying buffers, so we count the number of buffers walked | |
1273 | * and if it exceeds the total number of buffers, it means that | |
1274 | * someone is doing some writes simultaneously with us. In | |
1275 | * this case, stop, dropping the lock. | |
1276 | */ | |
1277 | if (dropped_lock) | |
1278 | goto again; | |
1279 | } | |
1280 | wake_up(&c->free_buffer_wait); | |
1281 | dm_bufio_unlock(c); | |
1282 | ||
1283 | a = xchg(&c->async_write_error, 0); | |
1284 | f = dm_bufio_issue_flush(c); | |
1285 | if (a) | |
1286 | return a; | |
1287 | ||
1288 | return f; | |
1289 | } | |
1290 | EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers); | |
1291 | ||
1292 | /* | |
1293 | * Use dm-io to send and empty barrier flush the device. | |
1294 | */ | |
1295 | int dm_bufio_issue_flush(struct dm_bufio_client *c) | |
1296 | { | |
1297 | struct dm_io_request io_req = { | |
3daec3b4 | 1298 | .bi_rw = WRITE_FLUSH, |
95d402f0 MP |
1299 | .mem.type = DM_IO_KMEM, |
1300 | .mem.ptr.addr = NULL, | |
1301 | .client = c->dm_io, | |
1302 | }; | |
1303 | struct dm_io_region io_reg = { | |
1304 | .bdev = c->bdev, | |
1305 | .sector = 0, | |
1306 | .count = 0, | |
1307 | }; | |
1308 | ||
1309 | BUG_ON(dm_bufio_in_request()); | |
1310 | ||
1311 | return dm_io(&io_req, 1, &io_reg, NULL); | |
1312 | } | |
1313 | EXPORT_SYMBOL_GPL(dm_bufio_issue_flush); | |
1314 | ||
1315 | /* | |
1316 | * We first delete any other buffer that may be at that new location. | |
1317 | * | |
1318 | * Then, we write the buffer to the original location if it was dirty. | |
1319 | * | |
1320 | * Then, if we are the only one who is holding the buffer, relink the buffer | |
1321 | * in the hash queue for the new location. | |
1322 | * | |
1323 | * If there was someone else holding the buffer, we write it to the new | |
1324 | * location but not relink it, because that other user needs to have the buffer | |
1325 | * at the same place. | |
1326 | */ | |
1327 | void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block) | |
1328 | { | |
1329 | struct dm_bufio_client *c = b->c; | |
1330 | struct dm_buffer *new; | |
1331 | ||
1332 | BUG_ON(dm_bufio_in_request()); | |
1333 | ||
1334 | dm_bufio_lock(c); | |
1335 | ||
1336 | retry: | |
1337 | new = __find(c, new_block); | |
1338 | if (new) { | |
1339 | if (new->hold_count) { | |
1340 | __wait_for_free_buffer(c); | |
1341 | goto retry; | |
1342 | } | |
1343 | ||
1344 | /* | |
1345 | * FIXME: Is there any point waiting for a write that's going | |
1346 | * to be overwritten in a bit? | |
1347 | */ | |
1348 | __make_buffer_clean(new); | |
1349 | __unlink_buffer(new); | |
1350 | __free_buffer_wake(new); | |
1351 | } | |
1352 | ||
1353 | BUG_ON(!b->hold_count); | |
1354 | BUG_ON(test_bit(B_READING, &b->state)); | |
1355 | ||
2480945c | 1356 | __write_dirty_buffer(b, NULL); |
95d402f0 | 1357 | if (b->hold_count == 1) { |
74316201 N |
1358 | wait_on_bit_io(&b->state, B_WRITING, |
1359 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1360 | set_bit(B_DIRTY, &b->state); |
1361 | __unlink_buffer(b); | |
1362 | __link_buffer(b, new_block, LIST_DIRTY); | |
1363 | } else { | |
1364 | sector_t old_block; | |
74316201 N |
1365 | wait_on_bit_lock_io(&b->state, B_WRITING, |
1366 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1367 | /* |
1368 | * Relink buffer to "new_block" so that write_callback | |
1369 | * sees "new_block" as a block number. | |
1370 | * After the write, link the buffer back to old_block. | |
1371 | * All this must be done in bufio lock, so that block number | |
1372 | * change isn't visible to other threads. | |
1373 | */ | |
1374 | old_block = b->block; | |
1375 | __unlink_buffer(b); | |
1376 | __link_buffer(b, new_block, b->list_mode); | |
1377 | submit_io(b, WRITE, new_block, write_endio); | |
74316201 N |
1378 | wait_on_bit_io(&b->state, B_WRITING, |
1379 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1380 | __unlink_buffer(b); |
1381 | __link_buffer(b, old_block, b->list_mode); | |
1382 | } | |
1383 | ||
1384 | dm_bufio_unlock(c); | |
1385 | dm_bufio_release(b); | |
1386 | } | |
1387 | EXPORT_SYMBOL_GPL(dm_bufio_release_move); | |
1388 | ||
55494bf2 MP |
1389 | /* |
1390 | * Free the given buffer. | |
1391 | * | |
1392 | * This is just a hint, if the buffer is in use or dirty, this function | |
1393 | * does nothing. | |
1394 | */ | |
1395 | void dm_bufio_forget(struct dm_bufio_client *c, sector_t block) | |
1396 | { | |
1397 | struct dm_buffer *b; | |
1398 | ||
1399 | dm_bufio_lock(c); | |
1400 | ||
1401 | b = __find(c, block); | |
1402 | if (b && likely(!b->hold_count) && likely(!b->state)) { | |
1403 | __unlink_buffer(b); | |
1404 | __free_buffer_wake(b); | |
1405 | } | |
1406 | ||
1407 | dm_bufio_unlock(c); | |
1408 | } | |
1409 | EXPORT_SYMBOL(dm_bufio_forget); | |
1410 | ||
55b082e6 MP |
1411 | void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n) |
1412 | { | |
1413 | c->minimum_buffers = n; | |
1414 | } | |
1415 | EXPORT_SYMBOL(dm_bufio_set_minimum_buffers); | |
1416 | ||
95d402f0 MP |
1417 | unsigned dm_bufio_get_block_size(struct dm_bufio_client *c) |
1418 | { | |
1419 | return c->block_size; | |
1420 | } | |
1421 | EXPORT_SYMBOL_GPL(dm_bufio_get_block_size); | |
1422 | ||
1423 | sector_t dm_bufio_get_device_size(struct dm_bufio_client *c) | |
1424 | { | |
1425 | return i_size_read(c->bdev->bd_inode) >> | |
1426 | (SECTOR_SHIFT + c->sectors_per_block_bits); | |
1427 | } | |
1428 | EXPORT_SYMBOL_GPL(dm_bufio_get_device_size); | |
1429 | ||
1430 | sector_t dm_bufio_get_block_number(struct dm_buffer *b) | |
1431 | { | |
1432 | return b->block; | |
1433 | } | |
1434 | EXPORT_SYMBOL_GPL(dm_bufio_get_block_number); | |
1435 | ||
1436 | void *dm_bufio_get_block_data(struct dm_buffer *b) | |
1437 | { | |
1438 | return b->data; | |
1439 | } | |
1440 | EXPORT_SYMBOL_GPL(dm_bufio_get_block_data); | |
1441 | ||
1442 | void *dm_bufio_get_aux_data(struct dm_buffer *b) | |
1443 | { | |
1444 | return b + 1; | |
1445 | } | |
1446 | EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data); | |
1447 | ||
1448 | struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b) | |
1449 | { | |
1450 | return b->c; | |
1451 | } | |
1452 | EXPORT_SYMBOL_GPL(dm_bufio_get_client); | |
1453 | ||
1454 | static void drop_buffers(struct dm_bufio_client *c) | |
1455 | { | |
1456 | struct dm_buffer *b; | |
1457 | int i; | |
1458 | ||
1459 | BUG_ON(dm_bufio_in_request()); | |
1460 | ||
1461 | /* | |
1462 | * An optimization so that the buffers are not written one-by-one. | |
1463 | */ | |
1464 | dm_bufio_write_dirty_buffers_async(c); | |
1465 | ||
1466 | dm_bufio_lock(c); | |
1467 | ||
1468 | while ((b = __get_unclaimed_buffer(c))) | |
1469 | __free_buffer_wake(b); | |
1470 | ||
1471 | for (i = 0; i < LIST_SIZE; i++) | |
1472 | list_for_each_entry(b, &c->lru[i], lru_list) | |
1473 | DMERR("leaked buffer %llx, hold count %u, list %d", | |
1474 | (unsigned long long)b->block, b->hold_count, i); | |
1475 | ||
1476 | for (i = 0; i < LIST_SIZE; i++) | |
1477 | BUG_ON(!list_empty(&c->lru[i])); | |
1478 | ||
1479 | dm_bufio_unlock(c); | |
1480 | } | |
1481 | ||
1482 | /* | |
33096a78 JT |
1483 | * We may not be able to evict this buffer if IO pending or the client |
1484 | * is still using it. Caller is expected to know buffer is too old. | |
1485 | * | |
9d28eb12 MP |
1486 | * And if GFP_NOFS is used, we must not do any I/O because we hold |
1487 | * dm_bufio_clients_lock and we would risk deadlock if the I/O gets | |
1488 | * rerouted to different bufio client. | |
95d402f0 | 1489 | */ |
33096a78 | 1490 | static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp) |
95d402f0 | 1491 | { |
9d28eb12 | 1492 | if (!(gfp & __GFP_FS)) { |
95d402f0 MP |
1493 | if (test_bit(B_READING, &b->state) || |
1494 | test_bit(B_WRITING, &b->state) || | |
1495 | test_bit(B_DIRTY, &b->state)) | |
33096a78 | 1496 | return false; |
95d402f0 MP |
1497 | } |
1498 | ||
1499 | if (b->hold_count) | |
33096a78 | 1500 | return false; |
95d402f0 MP |
1501 | |
1502 | __make_buffer_clean(b); | |
1503 | __unlink_buffer(b); | |
1504 | __free_buffer_wake(b); | |
1505 | ||
33096a78 | 1506 | return true; |
95d402f0 MP |
1507 | } |
1508 | ||
33096a78 JT |
1509 | static unsigned get_retain_buffers(struct dm_bufio_client *c) |
1510 | { | |
1511 | unsigned retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes); | |
1512 | return retain_bytes / c->block_size; | |
1513 | } | |
1514 | ||
1515 | static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan, | |
1516 | gfp_t gfp_mask) | |
95d402f0 MP |
1517 | { |
1518 | int l; | |
1519 | struct dm_buffer *b, *tmp; | |
33096a78 JT |
1520 | unsigned long freed = 0; |
1521 | unsigned long count = nr_to_scan; | |
1522 | unsigned retain_target = get_retain_buffers(c); | |
95d402f0 MP |
1523 | |
1524 | for (l = 0; l < LIST_SIZE; l++) { | |
7dc19d5a | 1525 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) { |
33096a78 JT |
1526 | if (__try_evict_buffer(b, gfp_mask)) |
1527 | freed++; | |
1528 | if (!--nr_to_scan || ((count - freed) <= retain_target)) | |
0e825862 MP |
1529 | return freed; |
1530 | dm_bufio_cond_resched(); | |
7dc19d5a | 1531 | } |
95d402f0 | 1532 | } |
7dc19d5a | 1533 | return freed; |
95d402f0 MP |
1534 | } |
1535 | ||
7dc19d5a DC |
1536 | static unsigned long |
1537 | dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) | |
95d402f0 | 1538 | { |
7dc19d5a DC |
1539 | struct dm_bufio_client *c; |
1540 | unsigned long freed; | |
95d402f0 | 1541 | |
7dc19d5a | 1542 | c = container_of(shrink, struct dm_bufio_client, shrinker); |
9d28eb12 | 1543 | if (sc->gfp_mask & __GFP_FS) |
95d402f0 MP |
1544 | dm_bufio_lock(c); |
1545 | else if (!dm_bufio_trylock(c)) | |
7dc19d5a | 1546 | return SHRINK_STOP; |
95d402f0 | 1547 | |
7dc19d5a DC |
1548 | freed = __scan(c, sc->nr_to_scan, sc->gfp_mask); |
1549 | dm_bufio_unlock(c); | |
1550 | return freed; | |
1551 | } | |
95d402f0 | 1552 | |
7dc19d5a DC |
1553 | static unsigned long |
1554 | dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc) | |
1555 | { | |
1556 | struct dm_bufio_client *c; | |
1557 | unsigned long count; | |
95d402f0 | 1558 | |
7dc19d5a | 1559 | c = container_of(shrink, struct dm_bufio_client, shrinker); |
9d28eb12 | 1560 | if (sc->gfp_mask & __GFP_FS) |
7dc19d5a DC |
1561 | dm_bufio_lock(c); |
1562 | else if (!dm_bufio_trylock(c)) | |
1563 | return 0; | |
95d402f0 | 1564 | |
7dc19d5a DC |
1565 | count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY]; |
1566 | dm_bufio_unlock(c); | |
1567 | return count; | |
95d402f0 MP |
1568 | } |
1569 | ||
1570 | /* | |
1571 | * Create the buffering interface | |
1572 | */ | |
1573 | struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size, | |
1574 | unsigned reserved_buffers, unsigned aux_size, | |
1575 | void (*alloc_callback)(struct dm_buffer *), | |
1576 | void (*write_callback)(struct dm_buffer *)) | |
1577 | { | |
1578 | int r; | |
1579 | struct dm_bufio_client *c; | |
1580 | unsigned i; | |
1581 | ||
1582 | BUG_ON(block_size < 1 << SECTOR_SHIFT || | |
1583 | (block_size & (block_size - 1))); | |
1584 | ||
d8c712ea | 1585 | c = kzalloc(sizeof(*c), GFP_KERNEL); |
95d402f0 MP |
1586 | if (!c) { |
1587 | r = -ENOMEM; | |
1588 | goto bad_client; | |
1589 | } | |
4e420c45 | 1590 | c->buffer_tree = RB_ROOT; |
95d402f0 MP |
1591 | |
1592 | c->bdev = bdev; | |
1593 | c->block_size = block_size; | |
1594 | c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT; | |
1595 | c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ? | |
1596 | ffs(block_size) - 1 - PAGE_SHIFT : 0; | |
1597 | c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ? | |
1598 | PAGE_SHIFT - (ffs(block_size) - 1) : 0); | |
1599 | ||
1600 | c->aux_size = aux_size; | |
1601 | c->alloc_callback = alloc_callback; | |
1602 | c->write_callback = write_callback; | |
1603 | ||
1604 | for (i = 0; i < LIST_SIZE; i++) { | |
1605 | INIT_LIST_HEAD(&c->lru[i]); | |
1606 | c->n_buffers[i] = 0; | |
1607 | } | |
1608 | ||
95d402f0 MP |
1609 | mutex_init(&c->lock); |
1610 | INIT_LIST_HEAD(&c->reserved_buffers); | |
1611 | c->need_reserved_buffers = reserved_buffers; | |
1612 | ||
55b082e6 MP |
1613 | c->minimum_buffers = DM_BUFIO_MIN_BUFFERS; |
1614 | ||
95d402f0 MP |
1615 | init_waitqueue_head(&c->free_buffer_wait); |
1616 | c->async_write_error = 0; | |
1617 | ||
1618 | c->dm_io = dm_io_client_create(); | |
1619 | if (IS_ERR(c->dm_io)) { | |
1620 | r = PTR_ERR(c->dm_io); | |
1621 | goto bad_dm_io; | |
1622 | } | |
1623 | ||
1624 | mutex_lock(&dm_bufio_clients_lock); | |
1625 | if (c->blocks_per_page_bits) { | |
1626 | if (!DM_BUFIO_CACHE_NAME(c)) { | |
1627 | DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size); | |
1628 | if (!DM_BUFIO_CACHE_NAME(c)) { | |
1629 | r = -ENOMEM; | |
1630 | mutex_unlock(&dm_bufio_clients_lock); | |
1631 | goto bad_cache; | |
1632 | } | |
1633 | } | |
1634 | ||
1635 | if (!DM_BUFIO_CACHE(c)) { | |
1636 | DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c), | |
1637 | c->block_size, | |
1638 | c->block_size, 0, NULL); | |
1639 | if (!DM_BUFIO_CACHE(c)) { | |
1640 | r = -ENOMEM; | |
1641 | mutex_unlock(&dm_bufio_clients_lock); | |
1642 | goto bad_cache; | |
1643 | } | |
1644 | } | |
1645 | } | |
1646 | mutex_unlock(&dm_bufio_clients_lock); | |
1647 | ||
1648 | while (c->need_reserved_buffers) { | |
1649 | struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL); | |
1650 | ||
1651 | if (!b) { | |
1652 | r = -ENOMEM; | |
1653 | goto bad_buffer; | |
1654 | } | |
1655 | __free_buffer_wake(b); | |
1656 | } | |
1657 | ||
1658 | mutex_lock(&dm_bufio_clients_lock); | |
1659 | dm_bufio_client_count++; | |
1660 | list_add(&c->client_list, &dm_bufio_all_clients); | |
1661 | __cache_size_refresh(); | |
1662 | mutex_unlock(&dm_bufio_clients_lock); | |
1663 | ||
7dc19d5a DC |
1664 | c->shrinker.count_objects = dm_bufio_shrink_count; |
1665 | c->shrinker.scan_objects = dm_bufio_shrink_scan; | |
95d402f0 MP |
1666 | c->shrinker.seeks = 1; |
1667 | c->shrinker.batch = 0; | |
1668 | register_shrinker(&c->shrinker); | |
1669 | ||
1670 | return c; | |
1671 | ||
1672 | bad_buffer: | |
1673 | bad_cache: | |
1674 | while (!list_empty(&c->reserved_buffers)) { | |
1675 | struct dm_buffer *b = list_entry(c->reserved_buffers.next, | |
1676 | struct dm_buffer, lru_list); | |
1677 | list_del(&b->lru_list); | |
1678 | free_buffer(b); | |
1679 | } | |
1680 | dm_io_client_destroy(c->dm_io); | |
1681 | bad_dm_io: | |
95d402f0 MP |
1682 | kfree(c); |
1683 | bad_client: | |
1684 | return ERR_PTR(r); | |
1685 | } | |
1686 | EXPORT_SYMBOL_GPL(dm_bufio_client_create); | |
1687 | ||
1688 | /* | |
1689 | * Free the buffering interface. | |
1690 | * It is required that there are no references on any buffers. | |
1691 | */ | |
1692 | void dm_bufio_client_destroy(struct dm_bufio_client *c) | |
1693 | { | |
1694 | unsigned i; | |
1695 | ||
1696 | drop_buffers(c); | |
1697 | ||
1698 | unregister_shrinker(&c->shrinker); | |
1699 | ||
1700 | mutex_lock(&dm_bufio_clients_lock); | |
1701 | ||
1702 | list_del(&c->client_list); | |
1703 | dm_bufio_client_count--; | |
1704 | __cache_size_refresh(); | |
1705 | ||
1706 | mutex_unlock(&dm_bufio_clients_lock); | |
1707 | ||
4e420c45 | 1708 | BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree)); |
95d402f0 MP |
1709 | BUG_ON(c->need_reserved_buffers); |
1710 | ||
1711 | while (!list_empty(&c->reserved_buffers)) { | |
1712 | struct dm_buffer *b = list_entry(c->reserved_buffers.next, | |
1713 | struct dm_buffer, lru_list); | |
1714 | list_del(&b->lru_list); | |
1715 | free_buffer(b); | |
1716 | } | |
1717 | ||
1718 | for (i = 0; i < LIST_SIZE; i++) | |
1719 | if (c->n_buffers[i]) | |
1720 | DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]); | |
1721 | ||
1722 | for (i = 0; i < LIST_SIZE; i++) | |
1723 | BUG_ON(c->n_buffers[i]); | |
1724 | ||
1725 | dm_io_client_destroy(c->dm_io); | |
95d402f0 MP |
1726 | kfree(c); |
1727 | } | |
1728 | EXPORT_SYMBOL_GPL(dm_bufio_client_destroy); | |
1729 | ||
33096a78 | 1730 | static unsigned get_max_age_hz(void) |
95d402f0 | 1731 | { |
33096a78 | 1732 | unsigned max_age = ACCESS_ONCE(dm_bufio_max_age); |
95d402f0 | 1733 | |
33096a78 JT |
1734 | if (max_age > UINT_MAX / HZ) |
1735 | max_age = UINT_MAX / HZ; | |
95d402f0 | 1736 | |
33096a78 JT |
1737 | return max_age * HZ; |
1738 | } | |
95d402f0 | 1739 | |
33096a78 JT |
1740 | static bool older_than(struct dm_buffer *b, unsigned long age_hz) |
1741 | { | |
1742 | return (jiffies - b->last_accessed) >= age_hz; | |
1743 | } | |
1744 | ||
1745 | static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz) | |
1746 | { | |
1747 | struct dm_buffer *b, *tmp; | |
1748 | unsigned retain_target = get_retain_buffers(c); | |
1749 | unsigned count; | |
1750 | ||
1751 | dm_bufio_lock(c); | |
1752 | ||
1753 | count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY]; | |
1754 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) { | |
1755 | if (count <= retain_target) | |
1756 | break; | |
1757 | ||
1758 | if (!older_than(b, age_hz)) | |
1759 | break; | |
1760 | ||
1761 | if (__try_evict_buffer(b, 0)) | |
1762 | count--; | |
95d402f0 | 1763 | |
95d402f0 MP |
1764 | dm_bufio_cond_resched(); |
1765 | } | |
33096a78 JT |
1766 | |
1767 | dm_bufio_unlock(c); | |
1768 | } | |
1769 | ||
1770 | static void cleanup_old_buffers(void) | |
1771 | { | |
1772 | unsigned long max_age_hz = get_max_age_hz(); | |
1773 | struct dm_bufio_client *c; | |
1774 | ||
1775 | mutex_lock(&dm_bufio_clients_lock); | |
1776 | ||
1777 | list_for_each_entry(c, &dm_bufio_all_clients, client_list) | |
1778 | __evict_old_buffers(c, max_age_hz); | |
1779 | ||
95d402f0 MP |
1780 | mutex_unlock(&dm_bufio_clients_lock); |
1781 | } | |
1782 | ||
1783 | static struct workqueue_struct *dm_bufio_wq; | |
1784 | static struct delayed_work dm_bufio_work; | |
1785 | ||
1786 | static void work_fn(struct work_struct *w) | |
1787 | { | |
1788 | cleanup_old_buffers(); | |
1789 | ||
1790 | queue_delayed_work(dm_bufio_wq, &dm_bufio_work, | |
1791 | DM_BUFIO_WORK_TIMER_SECS * HZ); | |
1792 | } | |
1793 | ||
1794 | /*---------------------------------------------------------------- | |
1795 | * Module setup | |
1796 | *--------------------------------------------------------------*/ | |
1797 | ||
1798 | /* | |
1799 | * This is called only once for the whole dm_bufio module. | |
1800 | * It initializes memory limit. | |
1801 | */ | |
1802 | static int __init dm_bufio_init(void) | |
1803 | { | |
1804 | __u64 mem; | |
1805 | ||
4cb57ab4 MP |
1806 | dm_bufio_allocated_kmem_cache = 0; |
1807 | dm_bufio_allocated_get_free_pages = 0; | |
1808 | dm_bufio_allocated_vmalloc = 0; | |
1809 | dm_bufio_current_allocated = 0; | |
1810 | ||
95d402f0 MP |
1811 | memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches); |
1812 | memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names); | |
1813 | ||
1814 | mem = (__u64)((totalram_pages - totalhigh_pages) * | |
1815 | DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT; | |
1816 | ||
1817 | if (mem > ULONG_MAX) | |
1818 | mem = ULONG_MAX; | |
1819 | ||
1820 | #ifdef CONFIG_MMU | |
1821 | /* | |
1822 | * Get the size of vmalloc space the same way as VMALLOC_TOTAL | |
1823 | * in fs/proc/internal.h | |
1824 | */ | |
1825 | if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100) | |
1826 | mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100; | |
1827 | #endif | |
1828 | ||
1829 | dm_bufio_default_cache_size = mem; | |
1830 | ||
1831 | mutex_lock(&dm_bufio_clients_lock); | |
1832 | __cache_size_refresh(); | |
1833 | mutex_unlock(&dm_bufio_clients_lock); | |
1834 | ||
1835 | dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache"); | |
1836 | if (!dm_bufio_wq) | |
1837 | return -ENOMEM; | |
1838 | ||
1839 | INIT_DELAYED_WORK(&dm_bufio_work, work_fn); | |
1840 | queue_delayed_work(dm_bufio_wq, &dm_bufio_work, | |
1841 | DM_BUFIO_WORK_TIMER_SECS * HZ); | |
1842 | ||
1843 | return 0; | |
1844 | } | |
1845 | ||
1846 | /* | |
1847 | * This is called once when unloading the dm_bufio module. | |
1848 | */ | |
1849 | static void __exit dm_bufio_exit(void) | |
1850 | { | |
1851 | int bug = 0; | |
1852 | int i; | |
1853 | ||
1854 | cancel_delayed_work_sync(&dm_bufio_work); | |
1855 | destroy_workqueue(dm_bufio_wq); | |
1856 | ||
1857 | for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) { | |
1858 | struct kmem_cache *kc = dm_bufio_caches[i]; | |
1859 | ||
1860 | if (kc) | |
1861 | kmem_cache_destroy(kc); | |
1862 | } | |
1863 | ||
1864 | for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++) | |
1865 | kfree(dm_bufio_cache_names[i]); | |
1866 | ||
1867 | if (dm_bufio_client_count) { | |
1868 | DMCRIT("%s: dm_bufio_client_count leaked: %d", | |
1869 | __func__, dm_bufio_client_count); | |
1870 | bug = 1; | |
1871 | } | |
1872 | ||
1873 | if (dm_bufio_current_allocated) { | |
1874 | DMCRIT("%s: dm_bufio_current_allocated leaked: %lu", | |
1875 | __func__, dm_bufio_current_allocated); | |
1876 | bug = 1; | |
1877 | } | |
1878 | ||
1879 | if (dm_bufio_allocated_get_free_pages) { | |
1880 | DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu", | |
1881 | __func__, dm_bufio_allocated_get_free_pages); | |
1882 | bug = 1; | |
1883 | } | |
1884 | ||
1885 | if (dm_bufio_allocated_vmalloc) { | |
1886 | DMCRIT("%s: dm_bufio_vmalloc leaked: %lu", | |
1887 | __func__, dm_bufio_allocated_vmalloc); | |
1888 | bug = 1; | |
1889 | } | |
1890 | ||
1891 | if (bug) | |
1892 | BUG(); | |
1893 | } | |
1894 | ||
1895 | module_init(dm_bufio_init) | |
1896 | module_exit(dm_bufio_exit) | |
1897 | ||
1898 | module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR); | |
1899 | MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache"); | |
1900 | ||
1901 | module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR); | |
1902 | MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds"); | |
33096a78 JT |
1903 | |
1904 | module_param_named(retain_bytes, dm_bufio_retain_bytes, uint, S_IRUGO | S_IWUSR); | |
1905 | MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory"); | |
95d402f0 MP |
1906 | |
1907 | module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR); | |
1908 | MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory"); | |
1909 | ||
1910 | module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO); | |
1911 | MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc"); | |
1912 | ||
1913 | module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO); | |
1914 | MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages"); | |
1915 | ||
1916 | module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO); | |
1917 | MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc"); | |
1918 | ||
1919 | module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO); | |
1920 | MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache"); | |
1921 | ||
1922 | MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>"); | |
1923 | MODULE_DESCRIPTION(DM_NAME " buffered I/O library"); | |
1924 | MODULE_LICENSE("GPL"); |