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> | |
f495339c | 14 | #include <linux/jiffies.h> |
95d402f0 | 15 | #include <linux/vmalloc.h> |
95d402f0 | 16 | #include <linux/shrinker.h> |
6f66263f | 17 | #include <linux/module.h> |
4e420c45 | 18 | #include <linux/rbtree.h> |
95d402f0 MP |
19 | |
20 | #define DM_MSG_PREFIX "bufio" | |
21 | ||
22 | /* | |
23 | * Memory management policy: | |
24 | * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory | |
25 | * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower). | |
26 | * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers. | |
27 | * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT | |
28 | * dirty buffers. | |
29 | */ | |
30 | #define DM_BUFIO_MIN_BUFFERS 8 | |
31 | ||
32 | #define DM_BUFIO_MEMORY_PERCENT 2 | |
33 | #define DM_BUFIO_VMALLOC_PERCENT 25 | |
34 | #define DM_BUFIO_WRITEBACK_PERCENT 75 | |
35 | ||
36 | /* | |
37 | * Check buffer ages in this interval (seconds) | |
38 | */ | |
33096a78 | 39 | #define DM_BUFIO_WORK_TIMER_SECS 30 |
95d402f0 MP |
40 | |
41 | /* | |
42 | * Free buffers when they are older than this (seconds) | |
43 | */ | |
33096a78 | 44 | #define DM_BUFIO_DEFAULT_AGE_SECS 300 |
95d402f0 MP |
45 | |
46 | /* | |
33096a78 | 47 | * The nr of bytes of cached data to keep around. |
95d402f0 | 48 | */ |
33096a78 | 49 | #define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024) |
95d402f0 MP |
50 | |
51 | /* | |
95d402f0 MP |
52 | * The number of bvec entries that are embedded directly in the buffer. |
53 | * If the chunk size is larger, dm-io is used to do the io. | |
95d402f0 | 54 | */ |
95d402f0 | 55 | #define DM_BUFIO_INLINE_VECS 16 |
95d402f0 MP |
56 | |
57 | /* | |
58 | * Don't try to use kmem_cache_alloc for blocks larger than this. | |
59 | * For explanation, see alloc_buffer_data below. | |
60 | */ | |
61 | #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT (PAGE_SIZE >> 1) | |
62 | #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT (PAGE_SIZE << (MAX_ORDER - 1)) | |
63 | ||
64 | /* | |
65 | * dm_buffer->list_mode | |
66 | */ | |
67 | #define LIST_CLEAN 0 | |
68 | #define LIST_DIRTY 1 | |
69 | #define LIST_SIZE 2 | |
70 | ||
71 | /* | |
72 | * Linking of buffers: | |
73 | * All buffers are linked to cache_hash with their hash_list field. | |
74 | * | |
75 | * Clean buffers that are not being written (B_WRITING not set) | |
76 | * are linked to lru[LIST_CLEAN] with their lru_list field. | |
77 | * | |
78 | * Dirty and clean buffers that are being written are linked to | |
79 | * lru[LIST_DIRTY] with their lru_list field. When the write | |
80 | * finishes, the buffer cannot be relinked immediately (because we | |
81 | * are in an interrupt context and relinking requires process | |
82 | * context), so some clean-not-writing buffers can be held on | |
83 | * dirty_lru too. They are later added to lru in the process | |
84 | * context. | |
85 | */ | |
86 | struct dm_bufio_client { | |
87 | struct mutex lock; | |
88 | ||
89 | struct list_head lru[LIST_SIZE]; | |
90 | unsigned long n_buffers[LIST_SIZE]; | |
91 | ||
92 | struct block_device *bdev; | |
93 | unsigned block_size; | |
94 | unsigned char sectors_per_block_bits; | |
95 | unsigned char pages_per_block_bits; | |
96 | unsigned char blocks_per_page_bits; | |
97 | unsigned aux_size; | |
98 | void (*alloc_callback)(struct dm_buffer *); | |
99 | void (*write_callback)(struct dm_buffer *); | |
100 | ||
101 | struct dm_io_client *dm_io; | |
102 | ||
103 | struct list_head reserved_buffers; | |
104 | unsigned need_reserved_buffers; | |
105 | ||
55b082e6 MP |
106 | unsigned minimum_buffers; |
107 | ||
4e420c45 | 108 | struct rb_root buffer_tree; |
95d402f0 MP |
109 | wait_queue_head_t free_buffer_wait; |
110 | ||
111 | int async_write_error; | |
112 | ||
113 | struct list_head client_list; | |
114 | struct shrinker shrinker; | |
115 | }; | |
116 | ||
117 | /* | |
118 | * Buffer state bits. | |
119 | */ | |
120 | #define B_READING 0 | |
121 | #define B_WRITING 1 | |
122 | #define B_DIRTY 2 | |
123 | ||
124 | /* | |
125 | * Describes how the block was allocated: | |
126 | * kmem_cache_alloc(), __get_free_pages() or vmalloc(). | |
127 | * See the comment at alloc_buffer_data. | |
128 | */ | |
129 | enum data_mode { | |
130 | DATA_MODE_SLAB = 0, | |
131 | DATA_MODE_GET_FREE_PAGES = 1, | |
132 | DATA_MODE_VMALLOC = 2, | |
133 | DATA_MODE_LIMIT = 3 | |
134 | }; | |
135 | ||
136 | struct dm_buffer { | |
4e420c45 | 137 | struct rb_node node; |
95d402f0 MP |
138 | struct list_head lru_list; |
139 | sector_t block; | |
140 | void *data; | |
141 | enum data_mode data_mode; | |
142 | unsigned char list_mode; /* LIST_* */ | |
143 | unsigned hold_count; | |
144 | int read_error; | |
145 | int write_error; | |
146 | unsigned long state; | |
147 | unsigned long last_accessed; | |
148 | struct dm_bufio_client *c; | |
2480945c | 149 | struct list_head write_list; |
95d402f0 MP |
150 | struct bio bio; |
151 | struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS]; | |
152 | }; | |
153 | ||
154 | /*----------------------------------------------------------------*/ | |
155 | ||
156 | static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT]; | |
157 | static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT]; | |
158 | ||
159 | static inline int dm_bufio_cache_index(struct dm_bufio_client *c) | |
160 | { | |
161 | unsigned ret = c->blocks_per_page_bits - 1; | |
162 | ||
163 | BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches)); | |
164 | ||
165 | return ret; | |
166 | } | |
167 | ||
168 | #define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)]) | |
169 | #define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)]) | |
170 | ||
171 | #define dm_bufio_in_request() (!!current->bio_list) | |
172 | ||
173 | static void dm_bufio_lock(struct dm_bufio_client *c) | |
174 | { | |
175 | mutex_lock_nested(&c->lock, dm_bufio_in_request()); | |
176 | } | |
177 | ||
178 | static int dm_bufio_trylock(struct dm_bufio_client *c) | |
179 | { | |
180 | return mutex_trylock(&c->lock); | |
181 | } | |
182 | ||
183 | static void dm_bufio_unlock(struct dm_bufio_client *c) | |
184 | { | |
185 | mutex_unlock(&c->lock); | |
186 | } | |
187 | ||
188 | /* | |
189 | * FIXME Move to sched.h? | |
190 | */ | |
191 | #ifdef CONFIG_PREEMPT_VOLUNTARY | |
192 | # define dm_bufio_cond_resched() \ | |
193 | do { \ | |
194 | if (unlikely(need_resched())) \ | |
195 | _cond_resched(); \ | |
196 | } while (0) | |
197 | #else | |
198 | # define dm_bufio_cond_resched() do { } while (0) | |
199 | #endif | |
200 | ||
201 | /*----------------------------------------------------------------*/ | |
202 | ||
203 | /* | |
204 | * Default cache size: available memory divided by the ratio. | |
205 | */ | |
206 | static unsigned long dm_bufio_default_cache_size; | |
207 | ||
208 | /* | |
209 | * Total cache size set by the user. | |
210 | */ | |
211 | static unsigned long dm_bufio_cache_size; | |
212 | ||
213 | /* | |
214 | * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change | |
215 | * at any time. If it disagrees, the user has changed cache size. | |
216 | */ | |
217 | static unsigned long dm_bufio_cache_size_latch; | |
218 | ||
219 | static DEFINE_SPINLOCK(param_spinlock); | |
220 | ||
221 | /* | |
222 | * Buffers are freed after this timeout | |
223 | */ | |
224 | static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS; | |
33096a78 | 225 | static unsigned dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES; |
95d402f0 MP |
226 | |
227 | static unsigned long dm_bufio_peak_allocated; | |
228 | static unsigned long dm_bufio_allocated_kmem_cache; | |
229 | static unsigned long dm_bufio_allocated_get_free_pages; | |
230 | static unsigned long dm_bufio_allocated_vmalloc; | |
231 | static unsigned long dm_bufio_current_allocated; | |
232 | ||
233 | /*----------------------------------------------------------------*/ | |
234 | ||
235 | /* | |
236 | * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count | |
237 | */ | |
238 | static unsigned long dm_bufio_cache_size_per_client; | |
239 | ||
240 | /* | |
241 | * The current number of clients. | |
242 | */ | |
243 | static int dm_bufio_client_count; | |
244 | ||
245 | /* | |
246 | * The list of all clients. | |
247 | */ | |
248 | static LIST_HEAD(dm_bufio_all_clients); | |
249 | ||
250 | /* | |
251 | * This mutex protects dm_bufio_cache_size_latch, | |
252 | * dm_bufio_cache_size_per_client and dm_bufio_client_count | |
253 | */ | |
254 | static DEFINE_MUTEX(dm_bufio_clients_lock); | |
255 | ||
4e420c45 JT |
256 | /*---------------------------------------------------------------- |
257 | * A red/black tree acts as an index for all the buffers. | |
258 | *--------------------------------------------------------------*/ | |
259 | static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block) | |
260 | { | |
261 | struct rb_node *n = c->buffer_tree.rb_node; | |
262 | struct dm_buffer *b; | |
263 | ||
264 | while (n) { | |
265 | b = container_of(n, struct dm_buffer, node); | |
266 | ||
267 | if (b->block == block) | |
268 | return b; | |
269 | ||
270 | n = (b->block < block) ? n->rb_left : n->rb_right; | |
271 | } | |
272 | ||
273 | return NULL; | |
274 | } | |
275 | ||
276 | static void __insert(struct dm_bufio_client *c, struct dm_buffer *b) | |
277 | { | |
278 | struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL; | |
279 | struct dm_buffer *found; | |
280 | ||
281 | while (*new) { | |
282 | found = container_of(*new, struct dm_buffer, node); | |
283 | ||
284 | if (found->block == b->block) { | |
285 | BUG_ON(found != b); | |
286 | return; | |
287 | } | |
288 | ||
289 | parent = *new; | |
290 | new = (found->block < b->block) ? | |
291 | &((*new)->rb_left) : &((*new)->rb_right); | |
292 | } | |
293 | ||
294 | rb_link_node(&b->node, parent, new); | |
295 | rb_insert_color(&b->node, &c->buffer_tree); | |
296 | } | |
297 | ||
298 | static void __remove(struct dm_bufio_client *c, struct dm_buffer *b) | |
299 | { | |
300 | rb_erase(&b->node, &c->buffer_tree); | |
301 | } | |
302 | ||
95d402f0 MP |
303 | /*----------------------------------------------------------------*/ |
304 | ||
305 | static void adjust_total_allocated(enum data_mode data_mode, long diff) | |
306 | { | |
307 | static unsigned long * const class_ptr[DATA_MODE_LIMIT] = { | |
308 | &dm_bufio_allocated_kmem_cache, | |
309 | &dm_bufio_allocated_get_free_pages, | |
310 | &dm_bufio_allocated_vmalloc, | |
311 | }; | |
312 | ||
313 | spin_lock(¶m_spinlock); | |
314 | ||
315 | *class_ptr[data_mode] += diff; | |
316 | ||
317 | dm_bufio_current_allocated += diff; | |
318 | ||
319 | if (dm_bufio_current_allocated > dm_bufio_peak_allocated) | |
320 | dm_bufio_peak_allocated = dm_bufio_current_allocated; | |
321 | ||
322 | spin_unlock(¶m_spinlock); | |
323 | } | |
324 | ||
325 | /* | |
326 | * Change the number of clients and recalculate per-client limit. | |
327 | */ | |
328 | static void __cache_size_refresh(void) | |
329 | { | |
330 | BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock)); | |
331 | BUG_ON(dm_bufio_client_count < 0); | |
332 | ||
fe5fe906 | 333 | dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size); |
95d402f0 MP |
334 | |
335 | /* | |
336 | * Use default if set to 0 and report the actual cache size used. | |
337 | */ | |
338 | if (!dm_bufio_cache_size_latch) { | |
339 | (void)cmpxchg(&dm_bufio_cache_size, 0, | |
340 | dm_bufio_default_cache_size); | |
341 | dm_bufio_cache_size_latch = dm_bufio_default_cache_size; | |
342 | } | |
343 | ||
344 | dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch / | |
345 | (dm_bufio_client_count ? : 1); | |
346 | } | |
347 | ||
348 | /* | |
349 | * Allocating buffer data. | |
350 | * | |
351 | * Small buffers are allocated with kmem_cache, to use space optimally. | |
352 | * | |
353 | * For large buffers, we choose between get_free_pages and vmalloc. | |
354 | * Each has advantages and disadvantages. | |
355 | * | |
356 | * __get_free_pages can randomly fail if the memory is fragmented. | |
357 | * __vmalloc won't randomly fail, but vmalloc space is limited (it may be | |
358 | * as low as 128M) so using it for caching is not appropriate. | |
359 | * | |
360 | * If the allocation may fail we use __get_free_pages. Memory fragmentation | |
361 | * won't have a fatal effect here, but it just causes flushes of some other | |
362 | * buffers and more I/O will be performed. Don't use __get_free_pages if it | |
363 | * always fails (i.e. order >= MAX_ORDER). | |
364 | * | |
365 | * If the allocation shouldn't fail we use __vmalloc. This is only for the | |
366 | * initial reserve allocation, so there's no risk of wasting all vmalloc | |
367 | * space. | |
368 | */ | |
369 | static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask, | |
370 | enum data_mode *data_mode) | |
371 | { | |
502624bd MP |
372 | unsigned noio_flag; |
373 | void *ptr; | |
374 | ||
95d402f0 MP |
375 | if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) { |
376 | *data_mode = DATA_MODE_SLAB; | |
377 | return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask); | |
378 | } | |
379 | ||
380 | if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT && | |
381 | gfp_mask & __GFP_NORETRY) { | |
382 | *data_mode = DATA_MODE_GET_FREE_PAGES; | |
383 | return (void *)__get_free_pages(gfp_mask, | |
384 | c->pages_per_block_bits); | |
385 | } | |
386 | ||
387 | *data_mode = DATA_MODE_VMALLOC; | |
502624bd MP |
388 | |
389 | /* | |
390 | * __vmalloc allocates the data pages and auxiliary structures with | |
391 | * gfp_flags that were specified, but pagetables are always allocated | |
392 | * with GFP_KERNEL, no matter what was specified as gfp_mask. | |
393 | * | |
394 | * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that | |
395 | * all allocations done by this process (including pagetables) are done | |
396 | * as if GFP_NOIO was specified. | |
397 | */ | |
398 | ||
399 | if (gfp_mask & __GFP_NORETRY) | |
400 | noio_flag = memalloc_noio_save(); | |
401 | ||
220cd058 | 402 | ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL); |
502624bd MP |
403 | |
404 | if (gfp_mask & __GFP_NORETRY) | |
405 | memalloc_noio_restore(noio_flag); | |
406 | ||
407 | return ptr; | |
95d402f0 MP |
408 | } |
409 | ||
410 | /* | |
411 | * Free buffer's data. | |
412 | */ | |
413 | static void free_buffer_data(struct dm_bufio_client *c, | |
414 | void *data, enum data_mode data_mode) | |
415 | { | |
416 | switch (data_mode) { | |
417 | case DATA_MODE_SLAB: | |
418 | kmem_cache_free(DM_BUFIO_CACHE(c), data); | |
419 | break; | |
420 | ||
421 | case DATA_MODE_GET_FREE_PAGES: | |
422 | free_pages((unsigned long)data, c->pages_per_block_bits); | |
423 | break; | |
424 | ||
425 | case DATA_MODE_VMALLOC: | |
426 | vfree(data); | |
427 | break; | |
428 | ||
429 | default: | |
430 | DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d", | |
431 | data_mode); | |
432 | BUG(); | |
433 | } | |
434 | } | |
435 | ||
436 | /* | |
437 | * Allocate buffer and its data. | |
438 | */ | |
439 | static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask) | |
440 | { | |
441 | struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size, | |
442 | gfp_mask); | |
443 | ||
444 | if (!b) | |
445 | return NULL; | |
446 | ||
447 | b->c = c; | |
448 | ||
449 | b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode); | |
450 | if (!b->data) { | |
451 | kfree(b); | |
452 | return NULL; | |
453 | } | |
454 | ||
455 | adjust_total_allocated(b->data_mode, (long)c->block_size); | |
456 | ||
457 | return b; | |
458 | } | |
459 | ||
460 | /* | |
461 | * Free buffer and its data. | |
462 | */ | |
463 | static void free_buffer(struct dm_buffer *b) | |
464 | { | |
465 | struct dm_bufio_client *c = b->c; | |
466 | ||
467 | adjust_total_allocated(b->data_mode, -(long)c->block_size); | |
468 | ||
469 | free_buffer_data(c, b->data, b->data_mode); | |
470 | kfree(b); | |
471 | } | |
472 | ||
473 | /* | |
474 | * Link buffer to the hash list and clean or dirty queue. | |
475 | */ | |
476 | static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty) | |
477 | { | |
478 | struct dm_bufio_client *c = b->c; | |
479 | ||
480 | c->n_buffers[dirty]++; | |
481 | b->block = block; | |
482 | b->list_mode = dirty; | |
483 | list_add(&b->lru_list, &c->lru[dirty]); | |
4e420c45 | 484 | __insert(b->c, b); |
95d402f0 MP |
485 | b->last_accessed = jiffies; |
486 | } | |
487 | ||
488 | /* | |
489 | * Unlink buffer from the hash list and dirty or clean queue. | |
490 | */ | |
491 | static void __unlink_buffer(struct dm_buffer *b) | |
492 | { | |
493 | struct dm_bufio_client *c = b->c; | |
494 | ||
495 | BUG_ON(!c->n_buffers[b->list_mode]); | |
496 | ||
497 | c->n_buffers[b->list_mode]--; | |
4e420c45 | 498 | __remove(b->c, b); |
95d402f0 MP |
499 | list_del(&b->lru_list); |
500 | } | |
501 | ||
502 | /* | |
503 | * Place the buffer to the head of dirty or clean LRU queue. | |
504 | */ | |
505 | static void __relink_lru(struct dm_buffer *b, int dirty) | |
506 | { | |
507 | struct dm_bufio_client *c = b->c; | |
508 | ||
509 | BUG_ON(!c->n_buffers[b->list_mode]); | |
510 | ||
511 | c->n_buffers[b->list_mode]--; | |
512 | c->n_buffers[dirty]++; | |
513 | b->list_mode = dirty; | |
54499afb | 514 | list_move(&b->lru_list, &c->lru[dirty]); |
eb76faf5 | 515 | b->last_accessed = jiffies; |
95d402f0 MP |
516 | } |
517 | ||
518 | /*---------------------------------------------------------------- | |
519 | * Submit I/O on the buffer. | |
520 | * | |
521 | * Bio interface is faster but it has some problems: | |
522 | * the vector list is limited (increasing this limit increases | |
523 | * memory-consumption per buffer, so it is not viable); | |
524 | * | |
525 | * the memory must be direct-mapped, not vmalloced; | |
526 | * | |
527 | * the I/O driver can reject requests spuriously if it thinks that | |
528 | * the requests are too big for the device or if they cross a | |
529 | * controller-defined memory boundary. | |
530 | * | |
531 | * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and | |
532 | * it is not vmalloced, try using the bio interface. | |
533 | * | |
534 | * If the buffer is big, if it is vmalloced or if the underlying device | |
535 | * rejects the bio because it is too large, use dm-io layer to do the I/O. | |
536 | * The dm-io layer splits the I/O into multiple requests, avoiding the above | |
537 | * shortcomings. | |
538 | *--------------------------------------------------------------*/ | |
539 | ||
540 | /* | |
541 | * dm-io completion routine. It just calls b->bio.bi_end_io, pretending | |
542 | * that the request was handled directly with bio interface. | |
543 | */ | |
544 | static void dmio_complete(unsigned long error, void *context) | |
545 | { | |
546 | struct dm_buffer *b = context; | |
547 | ||
548 | b->bio.bi_end_io(&b->bio, error ? -EIO : 0); | |
549 | } | |
550 | ||
551 | static void use_dmio(struct dm_buffer *b, int rw, sector_t block, | |
552 | bio_end_io_t *end_io) | |
553 | { | |
554 | int r; | |
555 | struct dm_io_request io_req = { | |
556 | .bi_rw = rw, | |
557 | .notify.fn = dmio_complete, | |
558 | .notify.context = b, | |
559 | .client = b->c->dm_io, | |
560 | }; | |
561 | struct dm_io_region region = { | |
562 | .bdev = b->c->bdev, | |
563 | .sector = block << b->c->sectors_per_block_bits, | |
564 | .count = b->c->block_size >> SECTOR_SHIFT, | |
565 | }; | |
566 | ||
567 | if (b->data_mode != DATA_MODE_VMALLOC) { | |
568 | io_req.mem.type = DM_IO_KMEM; | |
569 | io_req.mem.ptr.addr = b->data; | |
570 | } else { | |
571 | io_req.mem.type = DM_IO_VMA; | |
572 | io_req.mem.ptr.vma = b->data; | |
573 | } | |
574 | ||
575 | b->bio.bi_end_io = end_io; | |
576 | ||
577 | r = dm_io(&io_req, 1, ®ion, NULL); | |
578 | if (r) | |
579 | end_io(&b->bio, r); | |
580 | } | |
581 | ||
445559cd DW |
582 | static void inline_endio(struct bio *bio, int error) |
583 | { | |
584 | bio_end_io_t *end_fn = bio->bi_private; | |
585 | ||
586 | /* | |
587 | * Reset the bio to free any attached resources | |
588 | * (e.g. bio integrity profiles). | |
589 | */ | |
590 | bio_reset(bio); | |
591 | ||
592 | end_fn(bio, error); | |
593 | } | |
594 | ||
95d402f0 MP |
595 | static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block, |
596 | bio_end_io_t *end_io) | |
597 | { | |
598 | char *ptr; | |
599 | int len; | |
600 | ||
601 | bio_init(&b->bio); | |
602 | b->bio.bi_io_vec = b->bio_vec; | |
603 | b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS; | |
4f024f37 | 604 | b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits; |
95d402f0 | 605 | b->bio.bi_bdev = b->c->bdev; |
445559cd DW |
606 | b->bio.bi_end_io = inline_endio; |
607 | /* | |
608 | * Use of .bi_private isn't a problem here because | |
609 | * the dm_buffer's inline bio is local to bufio. | |
610 | */ | |
611 | b->bio.bi_private = end_io; | |
95d402f0 MP |
612 | |
613 | /* | |
614 | * We assume that if len >= PAGE_SIZE ptr is page-aligned. | |
615 | * If len < PAGE_SIZE the buffer doesn't cross page boundary. | |
616 | */ | |
617 | ptr = b->data; | |
618 | len = b->c->block_size; | |
619 | ||
620 | if (len >= PAGE_SIZE) | |
621 | BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1)); | |
622 | else | |
623 | BUG_ON((unsigned long)ptr & (len - 1)); | |
624 | ||
625 | do { | |
626 | if (!bio_add_page(&b->bio, virt_to_page(ptr), | |
627 | len < PAGE_SIZE ? len : PAGE_SIZE, | |
628 | virt_to_phys(ptr) & (PAGE_SIZE - 1))) { | |
629 | BUG_ON(b->c->block_size <= PAGE_SIZE); | |
630 | use_dmio(b, rw, block, end_io); | |
631 | return; | |
632 | } | |
633 | ||
634 | len -= PAGE_SIZE; | |
635 | ptr += PAGE_SIZE; | |
636 | } while (len > 0); | |
637 | ||
638 | submit_bio(rw, &b->bio); | |
639 | } | |
640 | ||
641 | static void submit_io(struct dm_buffer *b, int rw, sector_t block, | |
642 | bio_end_io_t *end_io) | |
643 | { | |
644 | if (rw == WRITE && b->c->write_callback) | |
645 | b->c->write_callback(b); | |
646 | ||
647 | if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE && | |
648 | b->data_mode != DATA_MODE_VMALLOC) | |
649 | use_inline_bio(b, rw, block, end_io); | |
650 | else | |
651 | use_dmio(b, rw, block, end_io); | |
652 | } | |
653 | ||
654 | /*---------------------------------------------------------------- | |
655 | * Writing dirty buffers | |
656 | *--------------------------------------------------------------*/ | |
657 | ||
658 | /* | |
659 | * The endio routine for write. | |
660 | * | |
661 | * Set the error, clear B_WRITING bit and wake anyone who was waiting on | |
662 | * it. | |
663 | */ | |
664 | static void write_endio(struct bio *bio, int error) | |
665 | { | |
666 | struct dm_buffer *b = container_of(bio, struct dm_buffer, bio); | |
667 | ||
668 | b->write_error = error; | |
a66cc28f | 669 | if (unlikely(error)) { |
95d402f0 MP |
670 | struct dm_bufio_client *c = b->c; |
671 | (void)cmpxchg(&c->async_write_error, 0, error); | |
672 | } | |
673 | ||
674 | BUG_ON(!test_bit(B_WRITING, &b->state)); | |
675 | ||
4e857c58 | 676 | smp_mb__before_atomic(); |
95d402f0 | 677 | clear_bit(B_WRITING, &b->state); |
4e857c58 | 678 | smp_mb__after_atomic(); |
95d402f0 MP |
679 | |
680 | wake_up_bit(&b->state, B_WRITING); | |
681 | } | |
682 | ||
95d402f0 MP |
683 | /* |
684 | * Initiate a write on a dirty buffer, but don't wait for it. | |
685 | * | |
686 | * - If the buffer is not dirty, exit. | |
687 | * - If there some previous write going on, wait for it to finish (we can't | |
688 | * have two writes on the same buffer simultaneously). | |
689 | * - Submit our write and don't wait on it. We set B_WRITING indicating | |
690 | * that there is a write in progress. | |
691 | */ | |
2480945c MP |
692 | static void __write_dirty_buffer(struct dm_buffer *b, |
693 | struct list_head *write_list) | |
95d402f0 MP |
694 | { |
695 | if (!test_bit(B_DIRTY, &b->state)) | |
696 | return; | |
697 | ||
698 | clear_bit(B_DIRTY, &b->state); | |
74316201 | 699 | wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE); |
95d402f0 | 700 | |
2480945c MP |
701 | if (!write_list) |
702 | submit_io(b, WRITE, b->block, write_endio); | |
703 | else | |
704 | list_add_tail(&b->write_list, write_list); | |
705 | } | |
706 | ||
707 | static void __flush_write_list(struct list_head *write_list) | |
708 | { | |
709 | struct blk_plug plug; | |
710 | blk_start_plug(&plug); | |
711 | while (!list_empty(write_list)) { | |
712 | struct dm_buffer *b = | |
713 | list_entry(write_list->next, struct dm_buffer, write_list); | |
714 | list_del(&b->write_list); | |
715 | submit_io(b, WRITE, b->block, write_endio); | |
716 | dm_bufio_cond_resched(); | |
717 | } | |
718 | blk_finish_plug(&plug); | |
95d402f0 MP |
719 | } |
720 | ||
721 | /* | |
722 | * Wait until any activity on the buffer finishes. Possibly write the | |
723 | * buffer if it is dirty. When this function finishes, there is no I/O | |
724 | * running on the buffer and the buffer is not dirty. | |
725 | */ | |
726 | static void __make_buffer_clean(struct dm_buffer *b) | |
727 | { | |
728 | BUG_ON(b->hold_count); | |
729 | ||
730 | if (!b->state) /* fast case */ | |
731 | return; | |
732 | ||
74316201 | 733 | wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE); |
2480945c | 734 | __write_dirty_buffer(b, NULL); |
74316201 | 735 | wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE); |
95d402f0 MP |
736 | } |
737 | ||
738 | /* | |
739 | * Find some buffer that is not held by anybody, clean it, unlink it and | |
740 | * return it. | |
741 | */ | |
742 | static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c) | |
743 | { | |
744 | struct dm_buffer *b; | |
745 | ||
746 | list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) { | |
747 | BUG_ON(test_bit(B_WRITING, &b->state)); | |
748 | BUG_ON(test_bit(B_DIRTY, &b->state)); | |
749 | ||
750 | if (!b->hold_count) { | |
751 | __make_buffer_clean(b); | |
752 | __unlink_buffer(b); | |
753 | return b; | |
754 | } | |
755 | dm_bufio_cond_resched(); | |
756 | } | |
757 | ||
758 | list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) { | |
759 | BUG_ON(test_bit(B_READING, &b->state)); | |
760 | ||
761 | if (!b->hold_count) { | |
762 | __make_buffer_clean(b); | |
763 | __unlink_buffer(b); | |
764 | return b; | |
765 | } | |
766 | dm_bufio_cond_resched(); | |
767 | } | |
768 | ||
769 | return NULL; | |
770 | } | |
771 | ||
772 | /* | |
773 | * Wait until some other threads free some buffer or release hold count on | |
774 | * some buffer. | |
775 | * | |
776 | * This function is entered with c->lock held, drops it and regains it | |
777 | * before exiting. | |
778 | */ | |
779 | static void __wait_for_free_buffer(struct dm_bufio_client *c) | |
780 | { | |
781 | DECLARE_WAITQUEUE(wait, current); | |
782 | ||
783 | add_wait_queue(&c->free_buffer_wait, &wait); | |
784 | set_task_state(current, TASK_UNINTERRUPTIBLE); | |
785 | dm_bufio_unlock(c); | |
786 | ||
787 | io_schedule(); | |
788 | ||
95d402f0 MP |
789 | remove_wait_queue(&c->free_buffer_wait, &wait); |
790 | ||
791 | dm_bufio_lock(c); | |
792 | } | |
793 | ||
a66cc28f MP |
794 | enum new_flag { |
795 | NF_FRESH = 0, | |
796 | NF_READ = 1, | |
797 | NF_GET = 2, | |
798 | NF_PREFETCH = 3 | |
799 | }; | |
800 | ||
95d402f0 MP |
801 | /* |
802 | * Allocate a new buffer. If the allocation is not possible, wait until | |
803 | * some other thread frees a buffer. | |
804 | * | |
805 | * May drop the lock and regain it. | |
806 | */ | |
a66cc28f | 807 | static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf) |
95d402f0 MP |
808 | { |
809 | struct dm_buffer *b; | |
810 | ||
811 | /* | |
812 | * dm-bufio is resistant to allocation failures (it just keeps | |
813 | * one buffer reserved in cases all the allocations fail). | |
814 | * So set flags to not try too hard: | |
815 | * GFP_NOIO: don't recurse into the I/O layer | |
816 | * __GFP_NORETRY: don't retry and rather return failure | |
817 | * __GFP_NOMEMALLOC: don't use emergency reserves | |
818 | * __GFP_NOWARN: don't print a warning in case of failure | |
819 | * | |
820 | * For debugging, if we set the cache size to 1, no new buffers will | |
821 | * be allocated. | |
822 | */ | |
823 | while (1) { | |
824 | if (dm_bufio_cache_size_latch != 1) { | |
825 | b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); | |
826 | if (b) | |
827 | return b; | |
828 | } | |
829 | ||
a66cc28f MP |
830 | if (nf == NF_PREFETCH) |
831 | return NULL; | |
832 | ||
95d402f0 MP |
833 | if (!list_empty(&c->reserved_buffers)) { |
834 | b = list_entry(c->reserved_buffers.next, | |
835 | struct dm_buffer, lru_list); | |
836 | list_del(&b->lru_list); | |
837 | c->need_reserved_buffers++; | |
838 | ||
839 | return b; | |
840 | } | |
841 | ||
842 | b = __get_unclaimed_buffer(c); | |
843 | if (b) | |
844 | return b; | |
845 | ||
846 | __wait_for_free_buffer(c); | |
847 | } | |
848 | } | |
849 | ||
a66cc28f | 850 | static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf) |
95d402f0 | 851 | { |
a66cc28f MP |
852 | struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf); |
853 | ||
854 | if (!b) | |
855 | return NULL; | |
95d402f0 MP |
856 | |
857 | if (c->alloc_callback) | |
858 | c->alloc_callback(b); | |
859 | ||
860 | return b; | |
861 | } | |
862 | ||
863 | /* | |
864 | * Free a buffer and wake other threads waiting for free buffers. | |
865 | */ | |
866 | static void __free_buffer_wake(struct dm_buffer *b) | |
867 | { | |
868 | struct dm_bufio_client *c = b->c; | |
869 | ||
870 | if (!c->need_reserved_buffers) | |
871 | free_buffer(b); | |
872 | else { | |
873 | list_add(&b->lru_list, &c->reserved_buffers); | |
874 | c->need_reserved_buffers--; | |
875 | } | |
876 | ||
877 | wake_up(&c->free_buffer_wait); | |
878 | } | |
879 | ||
2480945c MP |
880 | static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait, |
881 | struct list_head *write_list) | |
95d402f0 MP |
882 | { |
883 | struct dm_buffer *b, *tmp; | |
884 | ||
885 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) { | |
886 | BUG_ON(test_bit(B_READING, &b->state)); | |
887 | ||
888 | if (!test_bit(B_DIRTY, &b->state) && | |
889 | !test_bit(B_WRITING, &b->state)) { | |
890 | __relink_lru(b, LIST_CLEAN); | |
891 | continue; | |
892 | } | |
893 | ||
894 | if (no_wait && test_bit(B_WRITING, &b->state)) | |
895 | return; | |
896 | ||
2480945c | 897 | __write_dirty_buffer(b, write_list); |
95d402f0 MP |
898 | dm_bufio_cond_resched(); |
899 | } | |
900 | } | |
901 | ||
902 | /* | |
903 | * Get writeback threshold and buffer limit for a given client. | |
904 | */ | |
905 | static void __get_memory_limit(struct dm_bufio_client *c, | |
906 | unsigned long *threshold_buffers, | |
907 | unsigned long *limit_buffers) | |
908 | { | |
909 | unsigned long buffers; | |
910 | ||
fe5fe906 | 911 | if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) { |
95d402f0 MP |
912 | mutex_lock(&dm_bufio_clients_lock); |
913 | __cache_size_refresh(); | |
914 | mutex_unlock(&dm_bufio_clients_lock); | |
915 | } | |
916 | ||
917 | buffers = dm_bufio_cache_size_per_client >> | |
918 | (c->sectors_per_block_bits + SECTOR_SHIFT); | |
919 | ||
55b082e6 MP |
920 | if (buffers < c->minimum_buffers) |
921 | buffers = c->minimum_buffers; | |
95d402f0 MP |
922 | |
923 | *limit_buffers = buffers; | |
924 | *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100; | |
925 | } | |
926 | ||
927 | /* | |
928 | * Check if we're over watermark. | |
929 | * If we are over threshold_buffers, start freeing buffers. | |
930 | * If we're over "limit_buffers", block until we get under the limit. | |
931 | */ | |
2480945c MP |
932 | static void __check_watermark(struct dm_bufio_client *c, |
933 | struct list_head *write_list) | |
95d402f0 MP |
934 | { |
935 | unsigned long threshold_buffers, limit_buffers; | |
936 | ||
937 | __get_memory_limit(c, &threshold_buffers, &limit_buffers); | |
938 | ||
939 | while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] > | |
940 | limit_buffers) { | |
941 | ||
942 | struct dm_buffer *b = __get_unclaimed_buffer(c); | |
943 | ||
944 | if (!b) | |
945 | return; | |
946 | ||
947 | __free_buffer_wake(b); | |
948 | dm_bufio_cond_resched(); | |
949 | } | |
950 | ||
951 | if (c->n_buffers[LIST_DIRTY] > threshold_buffers) | |
2480945c | 952 | __write_dirty_buffers_async(c, 1, write_list); |
95d402f0 MP |
953 | } |
954 | ||
95d402f0 MP |
955 | /*---------------------------------------------------------------- |
956 | * Getting a buffer | |
957 | *--------------------------------------------------------------*/ | |
958 | ||
95d402f0 | 959 | static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block, |
2480945c MP |
960 | enum new_flag nf, int *need_submit, |
961 | struct list_head *write_list) | |
95d402f0 MP |
962 | { |
963 | struct dm_buffer *b, *new_b = NULL; | |
964 | ||
965 | *need_submit = 0; | |
966 | ||
967 | b = __find(c, block); | |
a66cc28f MP |
968 | if (b) |
969 | goto found_buffer; | |
95d402f0 MP |
970 | |
971 | if (nf == NF_GET) | |
972 | return NULL; | |
973 | ||
a66cc28f MP |
974 | new_b = __alloc_buffer_wait(c, nf); |
975 | if (!new_b) | |
976 | return NULL; | |
95d402f0 MP |
977 | |
978 | /* | |
979 | * We've had a period where the mutex was unlocked, so need to | |
980 | * recheck the hash table. | |
981 | */ | |
982 | b = __find(c, block); | |
983 | if (b) { | |
984 | __free_buffer_wake(new_b); | |
a66cc28f | 985 | goto found_buffer; |
95d402f0 MP |
986 | } |
987 | ||
2480945c | 988 | __check_watermark(c, write_list); |
95d402f0 MP |
989 | |
990 | b = new_b; | |
991 | b->hold_count = 1; | |
992 | b->read_error = 0; | |
993 | b->write_error = 0; | |
994 | __link_buffer(b, block, LIST_CLEAN); | |
995 | ||
996 | if (nf == NF_FRESH) { | |
997 | b->state = 0; | |
998 | return b; | |
999 | } | |
1000 | ||
1001 | b->state = 1 << B_READING; | |
1002 | *need_submit = 1; | |
1003 | ||
1004 | return b; | |
a66cc28f MP |
1005 | |
1006 | found_buffer: | |
1007 | if (nf == NF_PREFETCH) | |
1008 | return NULL; | |
1009 | /* | |
1010 | * Note: it is essential that we don't wait for the buffer to be | |
1011 | * read if dm_bufio_get function is used. Both dm_bufio_get and | |
1012 | * dm_bufio_prefetch can be used in the driver request routine. | |
1013 | * If the user called both dm_bufio_prefetch and dm_bufio_get on | |
1014 | * the same buffer, it would deadlock if we waited. | |
1015 | */ | |
1016 | if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state))) | |
1017 | return NULL; | |
1018 | ||
1019 | b->hold_count++; | |
1020 | __relink_lru(b, test_bit(B_DIRTY, &b->state) || | |
1021 | test_bit(B_WRITING, &b->state)); | |
1022 | return b; | |
95d402f0 MP |
1023 | } |
1024 | ||
1025 | /* | |
1026 | * The endio routine for reading: set the error, clear the bit and wake up | |
1027 | * anyone waiting on the buffer. | |
1028 | */ | |
1029 | static void read_endio(struct bio *bio, int error) | |
1030 | { | |
1031 | struct dm_buffer *b = container_of(bio, struct dm_buffer, bio); | |
1032 | ||
1033 | b->read_error = error; | |
1034 | ||
1035 | BUG_ON(!test_bit(B_READING, &b->state)); | |
1036 | ||
4e857c58 | 1037 | smp_mb__before_atomic(); |
95d402f0 | 1038 | clear_bit(B_READING, &b->state); |
4e857c58 | 1039 | smp_mb__after_atomic(); |
95d402f0 MP |
1040 | |
1041 | wake_up_bit(&b->state, B_READING); | |
1042 | } | |
1043 | ||
1044 | /* | |
1045 | * A common routine for dm_bufio_new and dm_bufio_read. Operation of these | |
1046 | * functions is similar except that dm_bufio_new doesn't read the | |
1047 | * buffer from the disk (assuming that the caller overwrites all the data | |
1048 | * and uses dm_bufio_mark_buffer_dirty to write new data back). | |
1049 | */ | |
1050 | static void *new_read(struct dm_bufio_client *c, sector_t block, | |
1051 | enum new_flag nf, struct dm_buffer **bp) | |
1052 | { | |
1053 | int need_submit; | |
1054 | struct dm_buffer *b; | |
1055 | ||
2480945c MP |
1056 | LIST_HEAD(write_list); |
1057 | ||
95d402f0 | 1058 | dm_bufio_lock(c); |
2480945c | 1059 | b = __bufio_new(c, block, nf, &need_submit, &write_list); |
95d402f0 MP |
1060 | dm_bufio_unlock(c); |
1061 | ||
2480945c MP |
1062 | __flush_write_list(&write_list); |
1063 | ||
a66cc28f | 1064 | if (!b) |
95d402f0 MP |
1065 | return b; |
1066 | ||
1067 | if (need_submit) | |
1068 | submit_io(b, READ, b->block, read_endio); | |
1069 | ||
74316201 | 1070 | wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE); |
95d402f0 MP |
1071 | |
1072 | if (b->read_error) { | |
1073 | int error = b->read_error; | |
1074 | ||
1075 | dm_bufio_release(b); | |
1076 | ||
1077 | return ERR_PTR(error); | |
1078 | } | |
1079 | ||
1080 | *bp = b; | |
1081 | ||
1082 | return b->data; | |
1083 | } | |
1084 | ||
1085 | void *dm_bufio_get(struct dm_bufio_client *c, sector_t block, | |
1086 | struct dm_buffer **bp) | |
1087 | { | |
1088 | return new_read(c, block, NF_GET, bp); | |
1089 | } | |
1090 | EXPORT_SYMBOL_GPL(dm_bufio_get); | |
1091 | ||
1092 | void *dm_bufio_read(struct dm_bufio_client *c, sector_t block, | |
1093 | struct dm_buffer **bp) | |
1094 | { | |
1095 | BUG_ON(dm_bufio_in_request()); | |
1096 | ||
1097 | return new_read(c, block, NF_READ, bp); | |
1098 | } | |
1099 | EXPORT_SYMBOL_GPL(dm_bufio_read); | |
1100 | ||
1101 | void *dm_bufio_new(struct dm_bufio_client *c, sector_t block, | |
1102 | struct dm_buffer **bp) | |
1103 | { | |
1104 | BUG_ON(dm_bufio_in_request()); | |
1105 | ||
1106 | return new_read(c, block, NF_FRESH, bp); | |
1107 | } | |
1108 | EXPORT_SYMBOL_GPL(dm_bufio_new); | |
1109 | ||
a66cc28f MP |
1110 | void dm_bufio_prefetch(struct dm_bufio_client *c, |
1111 | sector_t block, unsigned n_blocks) | |
1112 | { | |
1113 | struct blk_plug plug; | |
1114 | ||
2480945c MP |
1115 | LIST_HEAD(write_list); |
1116 | ||
3b6b7813 MP |
1117 | BUG_ON(dm_bufio_in_request()); |
1118 | ||
a66cc28f MP |
1119 | blk_start_plug(&plug); |
1120 | dm_bufio_lock(c); | |
1121 | ||
1122 | for (; n_blocks--; block++) { | |
1123 | int need_submit; | |
1124 | struct dm_buffer *b; | |
2480945c MP |
1125 | b = __bufio_new(c, block, NF_PREFETCH, &need_submit, |
1126 | &write_list); | |
1127 | if (unlikely(!list_empty(&write_list))) { | |
1128 | dm_bufio_unlock(c); | |
1129 | blk_finish_plug(&plug); | |
1130 | __flush_write_list(&write_list); | |
1131 | blk_start_plug(&plug); | |
1132 | dm_bufio_lock(c); | |
1133 | } | |
a66cc28f MP |
1134 | if (unlikely(b != NULL)) { |
1135 | dm_bufio_unlock(c); | |
1136 | ||
1137 | if (need_submit) | |
1138 | submit_io(b, READ, b->block, read_endio); | |
1139 | dm_bufio_release(b); | |
1140 | ||
1141 | dm_bufio_cond_resched(); | |
1142 | ||
1143 | if (!n_blocks) | |
1144 | goto flush_plug; | |
1145 | dm_bufio_lock(c); | |
1146 | } | |
a66cc28f MP |
1147 | } |
1148 | ||
1149 | dm_bufio_unlock(c); | |
1150 | ||
1151 | flush_plug: | |
1152 | blk_finish_plug(&plug); | |
1153 | } | |
1154 | EXPORT_SYMBOL_GPL(dm_bufio_prefetch); | |
1155 | ||
95d402f0 MP |
1156 | void dm_bufio_release(struct dm_buffer *b) |
1157 | { | |
1158 | struct dm_bufio_client *c = b->c; | |
1159 | ||
1160 | dm_bufio_lock(c); | |
1161 | ||
95d402f0 MP |
1162 | BUG_ON(!b->hold_count); |
1163 | ||
1164 | b->hold_count--; | |
1165 | if (!b->hold_count) { | |
1166 | wake_up(&c->free_buffer_wait); | |
1167 | ||
1168 | /* | |
1169 | * If there were errors on the buffer, and the buffer is not | |
1170 | * to be written, free the buffer. There is no point in caching | |
1171 | * invalid buffer. | |
1172 | */ | |
1173 | if ((b->read_error || b->write_error) && | |
a66cc28f | 1174 | !test_bit(B_READING, &b->state) && |
95d402f0 MP |
1175 | !test_bit(B_WRITING, &b->state) && |
1176 | !test_bit(B_DIRTY, &b->state)) { | |
1177 | __unlink_buffer(b); | |
1178 | __free_buffer_wake(b); | |
1179 | } | |
1180 | } | |
1181 | ||
1182 | dm_bufio_unlock(c); | |
1183 | } | |
1184 | EXPORT_SYMBOL_GPL(dm_bufio_release); | |
1185 | ||
1186 | void dm_bufio_mark_buffer_dirty(struct dm_buffer *b) | |
1187 | { | |
1188 | struct dm_bufio_client *c = b->c; | |
1189 | ||
1190 | dm_bufio_lock(c); | |
1191 | ||
a66cc28f MP |
1192 | BUG_ON(test_bit(B_READING, &b->state)); |
1193 | ||
95d402f0 MP |
1194 | if (!test_and_set_bit(B_DIRTY, &b->state)) |
1195 | __relink_lru(b, LIST_DIRTY); | |
1196 | ||
1197 | dm_bufio_unlock(c); | |
1198 | } | |
1199 | EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty); | |
1200 | ||
1201 | void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c) | |
1202 | { | |
2480945c MP |
1203 | LIST_HEAD(write_list); |
1204 | ||
95d402f0 MP |
1205 | BUG_ON(dm_bufio_in_request()); |
1206 | ||
1207 | dm_bufio_lock(c); | |
2480945c | 1208 | __write_dirty_buffers_async(c, 0, &write_list); |
95d402f0 | 1209 | dm_bufio_unlock(c); |
2480945c | 1210 | __flush_write_list(&write_list); |
95d402f0 MP |
1211 | } |
1212 | EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async); | |
1213 | ||
1214 | /* | |
1215 | * For performance, it is essential that the buffers are written asynchronously | |
1216 | * and simultaneously (so that the block layer can merge the writes) and then | |
1217 | * waited upon. | |
1218 | * | |
1219 | * Finally, we flush hardware disk cache. | |
1220 | */ | |
1221 | int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c) | |
1222 | { | |
1223 | int a, f; | |
1224 | unsigned long buffers_processed = 0; | |
1225 | struct dm_buffer *b, *tmp; | |
1226 | ||
2480945c MP |
1227 | LIST_HEAD(write_list); |
1228 | ||
1229 | dm_bufio_lock(c); | |
1230 | __write_dirty_buffers_async(c, 0, &write_list); | |
1231 | dm_bufio_unlock(c); | |
1232 | __flush_write_list(&write_list); | |
95d402f0 | 1233 | dm_bufio_lock(c); |
95d402f0 MP |
1234 | |
1235 | again: | |
1236 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) { | |
1237 | int dropped_lock = 0; | |
1238 | ||
1239 | if (buffers_processed < c->n_buffers[LIST_DIRTY]) | |
1240 | buffers_processed++; | |
1241 | ||
1242 | BUG_ON(test_bit(B_READING, &b->state)); | |
1243 | ||
1244 | if (test_bit(B_WRITING, &b->state)) { | |
1245 | if (buffers_processed < c->n_buffers[LIST_DIRTY]) { | |
1246 | dropped_lock = 1; | |
1247 | b->hold_count++; | |
1248 | dm_bufio_unlock(c); | |
74316201 N |
1249 | wait_on_bit_io(&b->state, B_WRITING, |
1250 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1251 | dm_bufio_lock(c); |
1252 | b->hold_count--; | |
1253 | } else | |
74316201 N |
1254 | wait_on_bit_io(&b->state, B_WRITING, |
1255 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1256 | } |
1257 | ||
1258 | if (!test_bit(B_DIRTY, &b->state) && | |
1259 | !test_bit(B_WRITING, &b->state)) | |
1260 | __relink_lru(b, LIST_CLEAN); | |
1261 | ||
1262 | dm_bufio_cond_resched(); | |
1263 | ||
1264 | /* | |
1265 | * If we dropped the lock, the list is no longer consistent, | |
1266 | * so we must restart the search. | |
1267 | * | |
1268 | * In the most common case, the buffer just processed is | |
1269 | * relinked to the clean list, so we won't loop scanning the | |
1270 | * same buffer again and again. | |
1271 | * | |
1272 | * This may livelock if there is another thread simultaneously | |
1273 | * dirtying buffers, so we count the number of buffers walked | |
1274 | * and if it exceeds the total number of buffers, it means that | |
1275 | * someone is doing some writes simultaneously with us. In | |
1276 | * this case, stop, dropping the lock. | |
1277 | */ | |
1278 | if (dropped_lock) | |
1279 | goto again; | |
1280 | } | |
1281 | wake_up(&c->free_buffer_wait); | |
1282 | dm_bufio_unlock(c); | |
1283 | ||
1284 | a = xchg(&c->async_write_error, 0); | |
1285 | f = dm_bufio_issue_flush(c); | |
1286 | if (a) | |
1287 | return a; | |
1288 | ||
1289 | return f; | |
1290 | } | |
1291 | EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers); | |
1292 | ||
1293 | /* | |
1294 | * Use dm-io to send and empty barrier flush the device. | |
1295 | */ | |
1296 | int dm_bufio_issue_flush(struct dm_bufio_client *c) | |
1297 | { | |
1298 | struct dm_io_request io_req = { | |
3daec3b4 | 1299 | .bi_rw = WRITE_FLUSH, |
95d402f0 MP |
1300 | .mem.type = DM_IO_KMEM, |
1301 | .mem.ptr.addr = NULL, | |
1302 | .client = c->dm_io, | |
1303 | }; | |
1304 | struct dm_io_region io_reg = { | |
1305 | .bdev = c->bdev, | |
1306 | .sector = 0, | |
1307 | .count = 0, | |
1308 | }; | |
1309 | ||
1310 | BUG_ON(dm_bufio_in_request()); | |
1311 | ||
1312 | return dm_io(&io_req, 1, &io_reg, NULL); | |
1313 | } | |
1314 | EXPORT_SYMBOL_GPL(dm_bufio_issue_flush); | |
1315 | ||
1316 | /* | |
1317 | * We first delete any other buffer that may be at that new location. | |
1318 | * | |
1319 | * Then, we write the buffer to the original location if it was dirty. | |
1320 | * | |
1321 | * Then, if we are the only one who is holding the buffer, relink the buffer | |
1322 | * in the hash queue for the new location. | |
1323 | * | |
1324 | * If there was someone else holding the buffer, we write it to the new | |
1325 | * location but not relink it, because that other user needs to have the buffer | |
1326 | * at the same place. | |
1327 | */ | |
1328 | void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block) | |
1329 | { | |
1330 | struct dm_bufio_client *c = b->c; | |
1331 | struct dm_buffer *new; | |
1332 | ||
1333 | BUG_ON(dm_bufio_in_request()); | |
1334 | ||
1335 | dm_bufio_lock(c); | |
1336 | ||
1337 | retry: | |
1338 | new = __find(c, new_block); | |
1339 | if (new) { | |
1340 | if (new->hold_count) { | |
1341 | __wait_for_free_buffer(c); | |
1342 | goto retry; | |
1343 | } | |
1344 | ||
1345 | /* | |
1346 | * FIXME: Is there any point waiting for a write that's going | |
1347 | * to be overwritten in a bit? | |
1348 | */ | |
1349 | __make_buffer_clean(new); | |
1350 | __unlink_buffer(new); | |
1351 | __free_buffer_wake(new); | |
1352 | } | |
1353 | ||
1354 | BUG_ON(!b->hold_count); | |
1355 | BUG_ON(test_bit(B_READING, &b->state)); | |
1356 | ||
2480945c | 1357 | __write_dirty_buffer(b, NULL); |
95d402f0 | 1358 | if (b->hold_count == 1) { |
74316201 N |
1359 | wait_on_bit_io(&b->state, B_WRITING, |
1360 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1361 | set_bit(B_DIRTY, &b->state); |
1362 | __unlink_buffer(b); | |
1363 | __link_buffer(b, new_block, LIST_DIRTY); | |
1364 | } else { | |
1365 | sector_t old_block; | |
74316201 N |
1366 | wait_on_bit_lock_io(&b->state, B_WRITING, |
1367 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1368 | /* |
1369 | * Relink buffer to "new_block" so that write_callback | |
1370 | * sees "new_block" as a block number. | |
1371 | * After the write, link the buffer back to old_block. | |
1372 | * All this must be done in bufio lock, so that block number | |
1373 | * change isn't visible to other threads. | |
1374 | */ | |
1375 | old_block = b->block; | |
1376 | __unlink_buffer(b); | |
1377 | __link_buffer(b, new_block, b->list_mode); | |
1378 | submit_io(b, WRITE, new_block, write_endio); | |
74316201 N |
1379 | wait_on_bit_io(&b->state, B_WRITING, |
1380 | TASK_UNINTERRUPTIBLE); | |
95d402f0 MP |
1381 | __unlink_buffer(b); |
1382 | __link_buffer(b, old_block, b->list_mode); | |
1383 | } | |
1384 | ||
1385 | dm_bufio_unlock(c); | |
1386 | dm_bufio_release(b); | |
1387 | } | |
1388 | EXPORT_SYMBOL_GPL(dm_bufio_release_move); | |
1389 | ||
55494bf2 MP |
1390 | /* |
1391 | * Free the given buffer. | |
1392 | * | |
1393 | * This is just a hint, if the buffer is in use or dirty, this function | |
1394 | * does nothing. | |
1395 | */ | |
1396 | void dm_bufio_forget(struct dm_bufio_client *c, sector_t block) | |
1397 | { | |
1398 | struct dm_buffer *b; | |
1399 | ||
1400 | dm_bufio_lock(c); | |
1401 | ||
1402 | b = __find(c, block); | |
1403 | if (b && likely(!b->hold_count) && likely(!b->state)) { | |
1404 | __unlink_buffer(b); | |
1405 | __free_buffer_wake(b); | |
1406 | } | |
1407 | ||
1408 | dm_bufio_unlock(c); | |
1409 | } | |
1410 | EXPORT_SYMBOL(dm_bufio_forget); | |
1411 | ||
55b082e6 MP |
1412 | void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n) |
1413 | { | |
1414 | c->minimum_buffers = n; | |
1415 | } | |
1416 | EXPORT_SYMBOL(dm_bufio_set_minimum_buffers); | |
1417 | ||
95d402f0 MP |
1418 | unsigned dm_bufio_get_block_size(struct dm_bufio_client *c) |
1419 | { | |
1420 | return c->block_size; | |
1421 | } | |
1422 | EXPORT_SYMBOL_GPL(dm_bufio_get_block_size); | |
1423 | ||
1424 | sector_t dm_bufio_get_device_size(struct dm_bufio_client *c) | |
1425 | { | |
1426 | return i_size_read(c->bdev->bd_inode) >> | |
1427 | (SECTOR_SHIFT + c->sectors_per_block_bits); | |
1428 | } | |
1429 | EXPORT_SYMBOL_GPL(dm_bufio_get_device_size); | |
1430 | ||
1431 | sector_t dm_bufio_get_block_number(struct dm_buffer *b) | |
1432 | { | |
1433 | return b->block; | |
1434 | } | |
1435 | EXPORT_SYMBOL_GPL(dm_bufio_get_block_number); | |
1436 | ||
1437 | void *dm_bufio_get_block_data(struct dm_buffer *b) | |
1438 | { | |
1439 | return b->data; | |
1440 | } | |
1441 | EXPORT_SYMBOL_GPL(dm_bufio_get_block_data); | |
1442 | ||
1443 | void *dm_bufio_get_aux_data(struct dm_buffer *b) | |
1444 | { | |
1445 | return b + 1; | |
1446 | } | |
1447 | EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data); | |
1448 | ||
1449 | struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b) | |
1450 | { | |
1451 | return b->c; | |
1452 | } | |
1453 | EXPORT_SYMBOL_GPL(dm_bufio_get_client); | |
1454 | ||
1455 | static void drop_buffers(struct dm_bufio_client *c) | |
1456 | { | |
1457 | struct dm_buffer *b; | |
1458 | int i; | |
1459 | ||
1460 | BUG_ON(dm_bufio_in_request()); | |
1461 | ||
1462 | /* | |
1463 | * An optimization so that the buffers are not written one-by-one. | |
1464 | */ | |
1465 | dm_bufio_write_dirty_buffers_async(c); | |
1466 | ||
1467 | dm_bufio_lock(c); | |
1468 | ||
1469 | while ((b = __get_unclaimed_buffer(c))) | |
1470 | __free_buffer_wake(b); | |
1471 | ||
1472 | for (i = 0; i < LIST_SIZE; i++) | |
1473 | list_for_each_entry(b, &c->lru[i], lru_list) | |
1474 | DMERR("leaked buffer %llx, hold count %u, list %d", | |
1475 | (unsigned long long)b->block, b->hold_count, i); | |
1476 | ||
1477 | for (i = 0; i < LIST_SIZE; i++) | |
1478 | BUG_ON(!list_empty(&c->lru[i])); | |
1479 | ||
1480 | dm_bufio_unlock(c); | |
1481 | } | |
1482 | ||
1483 | /* | |
33096a78 JT |
1484 | * We may not be able to evict this buffer if IO pending or the client |
1485 | * is still using it. Caller is expected to know buffer is too old. | |
1486 | * | |
9d28eb12 MP |
1487 | * And if GFP_NOFS is used, we must not do any I/O because we hold |
1488 | * dm_bufio_clients_lock and we would risk deadlock if the I/O gets | |
1489 | * rerouted to different bufio client. | |
95d402f0 | 1490 | */ |
33096a78 | 1491 | static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp) |
95d402f0 | 1492 | { |
9d28eb12 | 1493 | if (!(gfp & __GFP_FS)) { |
95d402f0 MP |
1494 | if (test_bit(B_READING, &b->state) || |
1495 | test_bit(B_WRITING, &b->state) || | |
1496 | test_bit(B_DIRTY, &b->state)) | |
33096a78 | 1497 | return false; |
95d402f0 MP |
1498 | } |
1499 | ||
1500 | if (b->hold_count) | |
33096a78 | 1501 | return false; |
95d402f0 MP |
1502 | |
1503 | __make_buffer_clean(b); | |
1504 | __unlink_buffer(b); | |
1505 | __free_buffer_wake(b); | |
1506 | ||
33096a78 | 1507 | return true; |
95d402f0 MP |
1508 | } |
1509 | ||
33096a78 JT |
1510 | static unsigned get_retain_buffers(struct dm_bufio_client *c) |
1511 | { | |
1512 | unsigned retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes); | |
1513 | return retain_bytes / c->block_size; | |
1514 | } | |
1515 | ||
1516 | static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan, | |
1517 | gfp_t gfp_mask) | |
95d402f0 MP |
1518 | { |
1519 | int l; | |
1520 | struct dm_buffer *b, *tmp; | |
33096a78 JT |
1521 | unsigned long freed = 0; |
1522 | unsigned long count = nr_to_scan; | |
1523 | unsigned retain_target = get_retain_buffers(c); | |
95d402f0 MP |
1524 | |
1525 | for (l = 0; l < LIST_SIZE; l++) { | |
7dc19d5a | 1526 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) { |
33096a78 JT |
1527 | if (__try_evict_buffer(b, gfp_mask)) |
1528 | freed++; | |
1529 | if (!--nr_to_scan || ((count - freed) <= retain_target)) | |
0e825862 MP |
1530 | return freed; |
1531 | dm_bufio_cond_resched(); | |
7dc19d5a | 1532 | } |
95d402f0 | 1533 | } |
7dc19d5a | 1534 | return freed; |
95d402f0 MP |
1535 | } |
1536 | ||
7dc19d5a DC |
1537 | static unsigned long |
1538 | dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) | |
95d402f0 | 1539 | { |
7dc19d5a DC |
1540 | struct dm_bufio_client *c; |
1541 | unsigned long freed; | |
95d402f0 | 1542 | |
7dc19d5a | 1543 | c = container_of(shrink, struct dm_bufio_client, shrinker); |
9d28eb12 | 1544 | if (sc->gfp_mask & __GFP_FS) |
95d402f0 MP |
1545 | dm_bufio_lock(c); |
1546 | else if (!dm_bufio_trylock(c)) | |
7dc19d5a | 1547 | return SHRINK_STOP; |
95d402f0 | 1548 | |
7dc19d5a DC |
1549 | freed = __scan(c, sc->nr_to_scan, sc->gfp_mask); |
1550 | dm_bufio_unlock(c); | |
1551 | return freed; | |
1552 | } | |
95d402f0 | 1553 | |
7dc19d5a DC |
1554 | static unsigned long |
1555 | dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc) | |
1556 | { | |
1557 | struct dm_bufio_client *c; | |
1558 | unsigned long count; | |
95d402f0 | 1559 | |
7dc19d5a | 1560 | c = container_of(shrink, struct dm_bufio_client, shrinker); |
9d28eb12 | 1561 | if (sc->gfp_mask & __GFP_FS) |
7dc19d5a DC |
1562 | dm_bufio_lock(c); |
1563 | else if (!dm_bufio_trylock(c)) | |
1564 | return 0; | |
95d402f0 | 1565 | |
7dc19d5a DC |
1566 | count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY]; |
1567 | dm_bufio_unlock(c); | |
1568 | return count; | |
95d402f0 MP |
1569 | } |
1570 | ||
1571 | /* | |
1572 | * Create the buffering interface | |
1573 | */ | |
1574 | struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size, | |
1575 | unsigned reserved_buffers, unsigned aux_size, | |
1576 | void (*alloc_callback)(struct dm_buffer *), | |
1577 | void (*write_callback)(struct dm_buffer *)) | |
1578 | { | |
1579 | int r; | |
1580 | struct dm_bufio_client *c; | |
1581 | unsigned i; | |
1582 | ||
1583 | BUG_ON(block_size < 1 << SECTOR_SHIFT || | |
1584 | (block_size & (block_size - 1))); | |
1585 | ||
d8c712ea | 1586 | c = kzalloc(sizeof(*c), GFP_KERNEL); |
95d402f0 MP |
1587 | if (!c) { |
1588 | r = -ENOMEM; | |
1589 | goto bad_client; | |
1590 | } | |
4e420c45 | 1591 | c->buffer_tree = RB_ROOT; |
95d402f0 MP |
1592 | |
1593 | c->bdev = bdev; | |
1594 | c->block_size = block_size; | |
1595 | c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT; | |
1596 | c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ? | |
1597 | ffs(block_size) - 1 - PAGE_SHIFT : 0; | |
1598 | c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ? | |
1599 | PAGE_SHIFT - (ffs(block_size) - 1) : 0); | |
1600 | ||
1601 | c->aux_size = aux_size; | |
1602 | c->alloc_callback = alloc_callback; | |
1603 | c->write_callback = write_callback; | |
1604 | ||
1605 | for (i = 0; i < LIST_SIZE; i++) { | |
1606 | INIT_LIST_HEAD(&c->lru[i]); | |
1607 | c->n_buffers[i] = 0; | |
1608 | } | |
1609 | ||
95d402f0 MP |
1610 | mutex_init(&c->lock); |
1611 | INIT_LIST_HEAD(&c->reserved_buffers); | |
1612 | c->need_reserved_buffers = reserved_buffers; | |
1613 | ||
55b082e6 MP |
1614 | c->minimum_buffers = DM_BUFIO_MIN_BUFFERS; |
1615 | ||
95d402f0 MP |
1616 | init_waitqueue_head(&c->free_buffer_wait); |
1617 | c->async_write_error = 0; | |
1618 | ||
1619 | c->dm_io = dm_io_client_create(); | |
1620 | if (IS_ERR(c->dm_io)) { | |
1621 | r = PTR_ERR(c->dm_io); | |
1622 | goto bad_dm_io; | |
1623 | } | |
1624 | ||
1625 | mutex_lock(&dm_bufio_clients_lock); | |
1626 | if (c->blocks_per_page_bits) { | |
1627 | if (!DM_BUFIO_CACHE_NAME(c)) { | |
1628 | DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size); | |
1629 | if (!DM_BUFIO_CACHE_NAME(c)) { | |
1630 | r = -ENOMEM; | |
1631 | mutex_unlock(&dm_bufio_clients_lock); | |
1632 | goto bad_cache; | |
1633 | } | |
1634 | } | |
1635 | ||
1636 | if (!DM_BUFIO_CACHE(c)) { | |
1637 | DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c), | |
1638 | c->block_size, | |
1639 | c->block_size, 0, NULL); | |
1640 | if (!DM_BUFIO_CACHE(c)) { | |
1641 | r = -ENOMEM; | |
1642 | mutex_unlock(&dm_bufio_clients_lock); | |
1643 | goto bad_cache; | |
1644 | } | |
1645 | } | |
1646 | } | |
1647 | mutex_unlock(&dm_bufio_clients_lock); | |
1648 | ||
1649 | while (c->need_reserved_buffers) { | |
1650 | struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL); | |
1651 | ||
1652 | if (!b) { | |
1653 | r = -ENOMEM; | |
1654 | goto bad_buffer; | |
1655 | } | |
1656 | __free_buffer_wake(b); | |
1657 | } | |
1658 | ||
1659 | mutex_lock(&dm_bufio_clients_lock); | |
1660 | dm_bufio_client_count++; | |
1661 | list_add(&c->client_list, &dm_bufio_all_clients); | |
1662 | __cache_size_refresh(); | |
1663 | mutex_unlock(&dm_bufio_clients_lock); | |
1664 | ||
7dc19d5a DC |
1665 | c->shrinker.count_objects = dm_bufio_shrink_count; |
1666 | c->shrinker.scan_objects = dm_bufio_shrink_scan; | |
95d402f0 MP |
1667 | c->shrinker.seeks = 1; |
1668 | c->shrinker.batch = 0; | |
1669 | register_shrinker(&c->shrinker); | |
1670 | ||
1671 | return c; | |
1672 | ||
1673 | bad_buffer: | |
1674 | bad_cache: | |
1675 | while (!list_empty(&c->reserved_buffers)) { | |
1676 | struct dm_buffer *b = list_entry(c->reserved_buffers.next, | |
1677 | struct dm_buffer, lru_list); | |
1678 | list_del(&b->lru_list); | |
1679 | free_buffer(b); | |
1680 | } | |
1681 | dm_io_client_destroy(c->dm_io); | |
1682 | bad_dm_io: | |
95d402f0 MP |
1683 | kfree(c); |
1684 | bad_client: | |
1685 | return ERR_PTR(r); | |
1686 | } | |
1687 | EXPORT_SYMBOL_GPL(dm_bufio_client_create); | |
1688 | ||
1689 | /* | |
1690 | * Free the buffering interface. | |
1691 | * It is required that there are no references on any buffers. | |
1692 | */ | |
1693 | void dm_bufio_client_destroy(struct dm_bufio_client *c) | |
1694 | { | |
1695 | unsigned i; | |
1696 | ||
1697 | drop_buffers(c); | |
1698 | ||
1699 | unregister_shrinker(&c->shrinker); | |
1700 | ||
1701 | mutex_lock(&dm_bufio_clients_lock); | |
1702 | ||
1703 | list_del(&c->client_list); | |
1704 | dm_bufio_client_count--; | |
1705 | __cache_size_refresh(); | |
1706 | ||
1707 | mutex_unlock(&dm_bufio_clients_lock); | |
1708 | ||
4e420c45 | 1709 | BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree)); |
95d402f0 MP |
1710 | BUG_ON(c->need_reserved_buffers); |
1711 | ||
1712 | while (!list_empty(&c->reserved_buffers)) { | |
1713 | struct dm_buffer *b = list_entry(c->reserved_buffers.next, | |
1714 | struct dm_buffer, lru_list); | |
1715 | list_del(&b->lru_list); | |
1716 | free_buffer(b); | |
1717 | } | |
1718 | ||
1719 | for (i = 0; i < LIST_SIZE; i++) | |
1720 | if (c->n_buffers[i]) | |
1721 | DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]); | |
1722 | ||
1723 | for (i = 0; i < LIST_SIZE; i++) | |
1724 | BUG_ON(c->n_buffers[i]); | |
1725 | ||
1726 | dm_io_client_destroy(c->dm_io); | |
95d402f0 MP |
1727 | kfree(c); |
1728 | } | |
1729 | EXPORT_SYMBOL_GPL(dm_bufio_client_destroy); | |
1730 | ||
33096a78 | 1731 | static unsigned get_max_age_hz(void) |
95d402f0 | 1732 | { |
33096a78 | 1733 | unsigned max_age = ACCESS_ONCE(dm_bufio_max_age); |
95d402f0 | 1734 | |
33096a78 JT |
1735 | if (max_age > UINT_MAX / HZ) |
1736 | max_age = UINT_MAX / HZ; | |
95d402f0 | 1737 | |
33096a78 JT |
1738 | return max_age * HZ; |
1739 | } | |
95d402f0 | 1740 | |
33096a78 JT |
1741 | static bool older_than(struct dm_buffer *b, unsigned long age_hz) |
1742 | { | |
f495339c | 1743 | return time_after_eq(jiffies, b->last_accessed + age_hz); |
33096a78 JT |
1744 | } |
1745 | ||
1746 | static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz) | |
1747 | { | |
1748 | struct dm_buffer *b, *tmp; | |
1749 | unsigned retain_target = get_retain_buffers(c); | |
1750 | unsigned count; | |
1751 | ||
1752 | dm_bufio_lock(c); | |
1753 | ||
1754 | count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY]; | |
1755 | list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) { | |
1756 | if (count <= retain_target) | |
1757 | break; | |
1758 | ||
1759 | if (!older_than(b, age_hz)) | |
1760 | break; | |
1761 | ||
1762 | if (__try_evict_buffer(b, 0)) | |
1763 | count--; | |
95d402f0 | 1764 | |
95d402f0 MP |
1765 | dm_bufio_cond_resched(); |
1766 | } | |
33096a78 JT |
1767 | |
1768 | dm_bufio_unlock(c); | |
1769 | } | |
1770 | ||
1771 | static void cleanup_old_buffers(void) | |
1772 | { | |
1773 | unsigned long max_age_hz = get_max_age_hz(); | |
1774 | struct dm_bufio_client *c; | |
1775 | ||
1776 | mutex_lock(&dm_bufio_clients_lock); | |
1777 | ||
1778 | list_for_each_entry(c, &dm_bufio_all_clients, client_list) | |
1779 | __evict_old_buffers(c, max_age_hz); | |
1780 | ||
95d402f0 MP |
1781 | mutex_unlock(&dm_bufio_clients_lock); |
1782 | } | |
1783 | ||
1784 | static struct workqueue_struct *dm_bufio_wq; | |
1785 | static struct delayed_work dm_bufio_work; | |
1786 | ||
1787 | static void work_fn(struct work_struct *w) | |
1788 | { | |
1789 | cleanup_old_buffers(); | |
1790 | ||
1791 | queue_delayed_work(dm_bufio_wq, &dm_bufio_work, | |
1792 | DM_BUFIO_WORK_TIMER_SECS * HZ); | |
1793 | } | |
1794 | ||
1795 | /*---------------------------------------------------------------- | |
1796 | * Module setup | |
1797 | *--------------------------------------------------------------*/ | |
1798 | ||
1799 | /* | |
1800 | * This is called only once for the whole dm_bufio module. | |
1801 | * It initializes memory limit. | |
1802 | */ | |
1803 | static int __init dm_bufio_init(void) | |
1804 | { | |
1805 | __u64 mem; | |
1806 | ||
4cb57ab4 MP |
1807 | dm_bufio_allocated_kmem_cache = 0; |
1808 | dm_bufio_allocated_get_free_pages = 0; | |
1809 | dm_bufio_allocated_vmalloc = 0; | |
1810 | dm_bufio_current_allocated = 0; | |
1811 | ||
95d402f0 MP |
1812 | memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches); |
1813 | memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names); | |
1814 | ||
1815 | mem = (__u64)((totalram_pages - totalhigh_pages) * | |
1816 | DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT; | |
1817 | ||
1818 | if (mem > ULONG_MAX) | |
1819 | mem = ULONG_MAX; | |
1820 | ||
1821 | #ifdef CONFIG_MMU | |
1822 | /* | |
1823 | * Get the size of vmalloc space the same way as VMALLOC_TOTAL | |
1824 | * in fs/proc/internal.h | |
1825 | */ | |
1826 | if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100) | |
1827 | mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100; | |
1828 | #endif | |
1829 | ||
1830 | dm_bufio_default_cache_size = mem; | |
1831 | ||
1832 | mutex_lock(&dm_bufio_clients_lock); | |
1833 | __cache_size_refresh(); | |
1834 | mutex_unlock(&dm_bufio_clients_lock); | |
1835 | ||
1836 | dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache"); | |
1837 | if (!dm_bufio_wq) | |
1838 | return -ENOMEM; | |
1839 | ||
1840 | INIT_DELAYED_WORK(&dm_bufio_work, work_fn); | |
1841 | queue_delayed_work(dm_bufio_wq, &dm_bufio_work, | |
1842 | DM_BUFIO_WORK_TIMER_SECS * HZ); | |
1843 | ||
1844 | return 0; | |
1845 | } | |
1846 | ||
1847 | /* | |
1848 | * This is called once when unloading the dm_bufio module. | |
1849 | */ | |
1850 | static void __exit dm_bufio_exit(void) | |
1851 | { | |
1852 | int bug = 0; | |
1853 | int i; | |
1854 | ||
1855 | cancel_delayed_work_sync(&dm_bufio_work); | |
1856 | destroy_workqueue(dm_bufio_wq); | |
1857 | ||
1858 | for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) { | |
1859 | struct kmem_cache *kc = dm_bufio_caches[i]; | |
1860 | ||
1861 | if (kc) | |
1862 | kmem_cache_destroy(kc); | |
1863 | } | |
1864 | ||
1865 | for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++) | |
1866 | kfree(dm_bufio_cache_names[i]); | |
1867 | ||
1868 | if (dm_bufio_client_count) { | |
1869 | DMCRIT("%s: dm_bufio_client_count leaked: %d", | |
1870 | __func__, dm_bufio_client_count); | |
1871 | bug = 1; | |
1872 | } | |
1873 | ||
1874 | if (dm_bufio_current_allocated) { | |
1875 | DMCRIT("%s: dm_bufio_current_allocated leaked: %lu", | |
1876 | __func__, dm_bufio_current_allocated); | |
1877 | bug = 1; | |
1878 | } | |
1879 | ||
1880 | if (dm_bufio_allocated_get_free_pages) { | |
1881 | DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu", | |
1882 | __func__, dm_bufio_allocated_get_free_pages); | |
1883 | bug = 1; | |
1884 | } | |
1885 | ||
1886 | if (dm_bufio_allocated_vmalloc) { | |
1887 | DMCRIT("%s: dm_bufio_vmalloc leaked: %lu", | |
1888 | __func__, dm_bufio_allocated_vmalloc); | |
1889 | bug = 1; | |
1890 | } | |
1891 | ||
1892 | if (bug) | |
1893 | BUG(); | |
1894 | } | |
1895 | ||
1896 | module_init(dm_bufio_init) | |
1897 | module_exit(dm_bufio_exit) | |
1898 | ||
1899 | module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR); | |
1900 | MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache"); | |
1901 | ||
1902 | module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR); | |
1903 | MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds"); | |
33096a78 JT |
1904 | |
1905 | module_param_named(retain_bytes, dm_bufio_retain_bytes, uint, S_IRUGO | S_IWUSR); | |
1906 | MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory"); | |
95d402f0 MP |
1907 | |
1908 | module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR); | |
1909 | MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory"); | |
1910 | ||
1911 | module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO); | |
1912 | MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc"); | |
1913 | ||
1914 | module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO); | |
1915 | MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages"); | |
1916 | ||
1917 | module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO); | |
1918 | MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc"); | |
1919 | ||
1920 | module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO); | |
1921 | MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache"); | |
1922 | ||
1923 | MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>"); | |
1924 | MODULE_DESCRIPTION(DM_NAME " buffered I/O library"); | |
1925 | MODULE_LICENSE("GPL"); |