blob: 75b51dfa5e0396dd2fa17756819bd678fda00208 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/fs/buffer.c
3 *
4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
5 */
6
7/*
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9 *
10 * Removed a lot of unnecessary code and simplified things now that
11 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12 *
13 * Speed up hash, lru, and free list operations. Use gfp() for allocating
14 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
15 *
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17 *
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19 */
20
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/kernel.h>
22#include <linux/syscalls.h>
23#include <linux/fs.h>
24#include <linux/mm.h>
25#include <linux/percpu.h>
26#include <linux/slab.h>
Randy Dunlap16f7e0f2006-01-11 12:17:46 -080027#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/blkdev.h>
29#include <linux/file.h>
30#include <linux/quotaops.h>
31#include <linux/highmem.h>
32#include <linux/module.h>
33#include <linux/writeback.h>
34#include <linux/hash.h>
35#include <linux/suspend.h>
36#include <linux/buffer_head.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080037#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/bio.h>
39#include <linux/notifier.h>
40#include <linux/cpu.h>
41#include <linux/bitops.h>
42#include <linux/mpage.h>
Ingo Molnarfb1c8f92005-09-10 00:25:56 -070043#include <linux/bit_spinlock.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
45static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49inline void
50init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51{
52 bh->b_end_io = handler;
53 bh->b_private = private;
54}
55
56static int sync_buffer(void *word)
57{
58 struct block_device *bd;
59 struct buffer_head *bh
60 = container_of(word, struct buffer_head, b_state);
61
62 smp_mb();
63 bd = bh->b_bdev;
64 if (bd)
65 blk_run_address_space(bd->bd_inode->i_mapping);
66 io_schedule();
67 return 0;
68}
69
70void fastcall __lock_buffer(struct buffer_head *bh)
71{
72 wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73 TASK_UNINTERRUPTIBLE);
74}
75EXPORT_SYMBOL(__lock_buffer);
76
77void fastcall unlock_buffer(struct buffer_head *bh)
78{
Nick Piggin72ed3d02007-02-10 01:46:22 -080079 smp_mb__before_clear_bit();
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 clear_buffer_locked(bh);
81 smp_mb__after_clear_bit();
82 wake_up_bit(&bh->b_state, BH_Lock);
83}
84
85/*
86 * Block until a buffer comes unlocked. This doesn't stop it
87 * from becoming locked again - you have to lock it yourself
88 * if you want to preserve its state.
89 */
90void __wait_on_buffer(struct buffer_head * bh)
91{
92 wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
93}
94
95static void
96__clear_page_buffers(struct page *page)
97{
98 ClearPagePrivate(page);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -070099 set_page_private(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 page_cache_release(page);
101}
102
103static void buffer_io_error(struct buffer_head *bh)
104{
105 char b[BDEVNAME_SIZE];
106
107 printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108 bdevname(bh->b_bdev, b),
109 (unsigned long long)bh->b_blocknr);
110}
111
112/*
113 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
114 * unlock the buffer. This is what ll_rw_block uses too.
115 */
116void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
117{
118 if (uptodate) {
119 set_buffer_uptodate(bh);
120 } else {
121 /* This happens, due to failed READA attempts. */
122 clear_buffer_uptodate(bh);
123 }
124 unlock_buffer(bh);
125 put_bh(bh);
126}
127
128void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
129{
130 char b[BDEVNAME_SIZE];
131
132 if (uptodate) {
133 set_buffer_uptodate(bh);
134 } else {
135 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
136 buffer_io_error(bh);
137 printk(KERN_WARNING "lost page write due to "
138 "I/O error on %s\n",
139 bdevname(bh->b_bdev, b));
140 }
141 set_buffer_write_io_error(bh);
142 clear_buffer_uptodate(bh);
143 }
144 unlock_buffer(bh);
145 put_bh(bh);
146}
147
148/*
149 * Write out and wait upon all the dirty data associated with a block
150 * device via its mapping. Does not take the superblock lock.
151 */
152int sync_blockdev(struct block_device *bdev)
153{
154 int ret = 0;
155
OGAWA Hirofumi28fd1292006-01-08 01:02:14 -0800156 if (bdev)
157 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 return ret;
159}
160EXPORT_SYMBOL(sync_blockdev);
161
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162/*
163 * Write out and wait upon all dirty data associated with this
164 * device. Filesystem data as well as the underlying block
165 * device. Takes the superblock lock.
166 */
167int fsync_bdev(struct block_device *bdev)
168{
169 struct super_block *sb = get_super(bdev);
170 if (sb) {
171 int res = fsync_super(sb);
172 drop_super(sb);
173 return res;
174 }
175 return sync_blockdev(bdev);
176}
177
178/**
179 * freeze_bdev -- lock a filesystem and force it into a consistent state
180 * @bdev: blockdevice to lock
181 *
David Chinnerf73ca1b2007-01-10 23:15:41 -0800182 * This takes the block device bd_mount_sem to make sure no new mounts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 * happen on bdev until thaw_bdev() is called.
184 * If a superblock is found on this device, we take the s_umount semaphore
185 * on it to make sure nobody unmounts until the snapshot creation is done.
186 */
187struct super_block *freeze_bdev(struct block_device *bdev)
188{
189 struct super_block *sb;
190
David Chinnerf73ca1b2007-01-10 23:15:41 -0800191 down(&bdev->bd_mount_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 sb = get_super(bdev);
193 if (sb && !(sb->s_flags & MS_RDONLY)) {
194 sb->s_frozen = SB_FREEZE_WRITE;
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700195 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196
OGAWA Hirofumid25b9a12006-03-25 03:07:44 -0800197 __fsync_super(sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198
199 sb->s_frozen = SB_FREEZE_TRANS;
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700200 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201
202 sync_blockdev(sb->s_bdev);
203
204 if (sb->s_op->write_super_lockfs)
205 sb->s_op->write_super_lockfs(sb);
206 }
207
208 sync_blockdev(bdev);
209 return sb; /* thaw_bdev releases s->s_umount and bd_mount_sem */
210}
211EXPORT_SYMBOL(freeze_bdev);
212
213/**
214 * thaw_bdev -- unlock filesystem
215 * @bdev: blockdevice to unlock
216 * @sb: associated superblock
217 *
218 * Unlocks the filesystem and marks it writeable again after freeze_bdev().
219 */
220void thaw_bdev(struct block_device *bdev, struct super_block *sb)
221{
222 if (sb) {
223 BUG_ON(sb->s_bdev != bdev);
224
225 if (sb->s_op->unlockfs)
226 sb->s_op->unlockfs(sb);
227 sb->s_frozen = SB_UNFROZEN;
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700228 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 wake_up(&sb->s_wait_unfrozen);
230 drop_super(sb);
231 }
232
David Chinnerf73ca1b2007-01-10 23:15:41 -0800233 up(&bdev->bd_mount_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234}
235EXPORT_SYMBOL(thaw_bdev);
236
237/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 * Various filesystems appear to want __find_get_block to be non-blocking.
239 * But it's the page lock which protects the buffers. To get around this,
240 * we get exclusion from try_to_free_buffers with the blockdev mapping's
241 * private_lock.
242 *
243 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
244 * may be quite high. This code could TryLock the page, and if that
245 * succeeds, there is no need to take private_lock. (But if
246 * private_lock is contended then so is mapping->tree_lock).
247 */
248static struct buffer_head *
Coywolf Qi Hunt385fd4c2005-11-07 00:59:39 -0800249__find_get_block_slow(struct block_device *bdev, sector_t block)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250{
251 struct inode *bd_inode = bdev->bd_inode;
252 struct address_space *bd_mapping = bd_inode->i_mapping;
253 struct buffer_head *ret = NULL;
254 pgoff_t index;
255 struct buffer_head *bh;
256 struct buffer_head *head;
257 struct page *page;
258 int all_mapped = 1;
259
260 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
261 page = find_get_page(bd_mapping, index);
262 if (!page)
263 goto out;
264
265 spin_lock(&bd_mapping->private_lock);
266 if (!page_has_buffers(page))
267 goto out_unlock;
268 head = page_buffers(page);
269 bh = head;
270 do {
271 if (bh->b_blocknr == block) {
272 ret = bh;
273 get_bh(bh);
274 goto out_unlock;
275 }
276 if (!buffer_mapped(bh))
277 all_mapped = 0;
278 bh = bh->b_this_page;
279 } while (bh != head);
280
281 /* we might be here because some of the buffers on this page are
282 * not mapped. This is due to various races between
283 * file io on the block device and getblk. It gets dealt with
284 * elsewhere, don't buffer_error if we had some unmapped buffers
285 */
286 if (all_mapped) {
287 printk("__find_get_block_slow() failed. "
288 "block=%llu, b_blocknr=%llu\n",
Badari Pulavarty205f87f2006-03-26 01:38:00 -0800289 (unsigned long long)block,
290 (unsigned long long)bh->b_blocknr);
291 printk("b_state=0x%08lx, b_size=%zu\n",
292 bh->b_state, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
294 }
295out_unlock:
296 spin_unlock(&bd_mapping->private_lock);
297 page_cache_release(page);
298out:
299 return ret;
300}
301
302/* If invalidate_buffers() will trash dirty buffers, it means some kind
303 of fs corruption is going on. Trashing dirty data always imply losing
304 information that was supposed to be just stored on the physical layer
305 by the user.
306
307 Thus invalidate_buffers in general usage is not allwowed to trash
308 dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
309 be preserved. These buffers are simply skipped.
310
311 We also skip buffers which are still in use. For example this can
312 happen if a userspace program is reading the block device.
313
314 NOTE: In the case where the user removed a removable-media-disk even if
315 there's still dirty data not synced on disk (due a bug in the device driver
316 or due an error of the user), by not destroying the dirty buffers we could
317 generate corruption also on the next media inserted, thus a parameter is
318 necessary to handle this case in the most safe way possible (trying
319 to not corrupt also the new disk inserted with the data belonging to
320 the old now corrupted disk). Also for the ramdisk the natural thing
321 to do in order to release the ramdisk memory is to destroy dirty buffers.
322
323 These are two special cases. Normal usage imply the device driver
324 to issue a sync on the device (without waiting I/O completion) and
325 then an invalidate_buffers call that doesn't trash dirty buffers.
326
327 For handling cache coherency with the blkdev pagecache the 'update' case
328 is been introduced. It is needed to re-read from disk any pinned
329 buffer. NOTE: re-reading from disk is destructive so we can do it only
330 when we assume nobody is changing the buffercache under our I/O and when
331 we think the disk contains more recent information than the buffercache.
332 The update == 1 pass marks the buffers we need to update, the update == 2
333 pass does the actual I/O. */
Peter Zijlstraf98393a2007-05-06 14:49:54 -0700334void invalidate_bdev(struct block_device *bdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335{
Andrew Morton0e1dfc62006-07-30 03:03:28 -0700336 struct address_space *mapping = bdev->bd_inode->i_mapping;
337
338 if (mapping->nrpages == 0)
339 return;
340
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 invalidate_bh_lrus();
Andrew Mortonfc0ecff2007-02-10 01:45:39 -0800342 invalidate_mapping_pages(mapping, 0, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343}
344
345/*
346 * Kick pdflush then try to free up some ZONE_NORMAL memory.
347 */
348static void free_more_memory(void)
349{
350 struct zone **zones;
351 pg_data_t *pgdat;
352
Pekka J Enberg687a21c2005-06-28 20:44:55 -0700353 wakeup_pdflush(1024);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 yield();
355
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -0800356 for_each_online_pgdat(pgdat) {
Al Viroaf4ca452005-10-21 02:55:38 -0400357 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358 if (*zones)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700359 try_to_free_pages(zones, 0, GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 }
361}
362
363/*
364 * I/O completion handler for block_read_full_page() - pages
365 * which come unlocked at the end of I/O.
366 */
367static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
368{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 unsigned long flags;
Nick Piggina3972202005-07-07 17:56:56 -0700370 struct buffer_head *first;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 struct buffer_head *tmp;
372 struct page *page;
373 int page_uptodate = 1;
374
375 BUG_ON(!buffer_async_read(bh));
376
377 page = bh->b_page;
378 if (uptodate) {
379 set_buffer_uptodate(bh);
380 } else {
381 clear_buffer_uptodate(bh);
382 if (printk_ratelimit())
383 buffer_io_error(bh);
384 SetPageError(page);
385 }
386
387 /*
388 * Be _very_ careful from here on. Bad things can happen if
389 * two buffer heads end IO at almost the same time and both
390 * decide that the page is now completely done.
391 */
Nick Piggina3972202005-07-07 17:56:56 -0700392 first = page_buffers(page);
393 local_irq_save(flags);
394 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 clear_buffer_async_read(bh);
396 unlock_buffer(bh);
397 tmp = bh;
398 do {
399 if (!buffer_uptodate(tmp))
400 page_uptodate = 0;
401 if (buffer_async_read(tmp)) {
402 BUG_ON(!buffer_locked(tmp));
403 goto still_busy;
404 }
405 tmp = tmp->b_this_page;
406 } while (tmp != bh);
Nick Piggina3972202005-07-07 17:56:56 -0700407 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
408 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /*
411 * If none of the buffers had errors and they are all
412 * uptodate then we can set the page uptodate.
413 */
414 if (page_uptodate && !PageError(page))
415 SetPageUptodate(page);
416 unlock_page(page);
417 return;
418
419still_busy:
Nick Piggina3972202005-07-07 17:56:56 -0700420 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
421 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 return;
423}
424
425/*
426 * Completion handler for block_write_full_page() - pages which are unlocked
427 * during I/O, and which have PageWriteback cleared upon I/O completion.
428 */
Adrian Bunkb6cd0b72006-06-27 02:53:54 -0700429static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430{
431 char b[BDEVNAME_SIZE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 unsigned long flags;
Nick Piggina3972202005-07-07 17:56:56 -0700433 struct buffer_head *first;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 struct buffer_head *tmp;
435 struct page *page;
436
437 BUG_ON(!buffer_async_write(bh));
438
439 page = bh->b_page;
440 if (uptodate) {
441 set_buffer_uptodate(bh);
442 } else {
443 if (printk_ratelimit()) {
444 buffer_io_error(bh);
445 printk(KERN_WARNING "lost page write due to "
446 "I/O error on %s\n",
447 bdevname(bh->b_bdev, b));
448 }
449 set_bit(AS_EIO, &page->mapping->flags);
Jan Kara58ff4072006-10-17 00:10:19 -0700450 set_buffer_write_io_error(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 clear_buffer_uptodate(bh);
452 SetPageError(page);
453 }
454
Nick Piggina3972202005-07-07 17:56:56 -0700455 first = page_buffers(page);
456 local_irq_save(flags);
457 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 clear_buffer_async_write(bh);
460 unlock_buffer(bh);
461 tmp = bh->b_this_page;
462 while (tmp != bh) {
463 if (buffer_async_write(tmp)) {
464 BUG_ON(!buffer_locked(tmp));
465 goto still_busy;
466 }
467 tmp = tmp->b_this_page;
468 }
Nick Piggina3972202005-07-07 17:56:56 -0700469 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
470 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 end_page_writeback(page);
472 return;
473
474still_busy:
Nick Piggina3972202005-07-07 17:56:56 -0700475 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
476 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 return;
478}
479
480/*
481 * If a page's buffers are under async readin (end_buffer_async_read
482 * completion) then there is a possibility that another thread of
483 * control could lock one of the buffers after it has completed
484 * but while some of the other buffers have not completed. This
485 * locked buffer would confuse end_buffer_async_read() into not unlocking
486 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
487 * that this buffer is not under async I/O.
488 *
489 * The page comes unlocked when it has no locked buffer_async buffers
490 * left.
491 *
492 * PageLocked prevents anyone starting new async I/O reads any of
493 * the buffers.
494 *
495 * PageWriteback is used to prevent simultaneous writeout of the same
496 * page.
497 *
498 * PageLocked prevents anyone from starting writeback of a page which is
499 * under read I/O (PageWriteback is only ever set against a locked page).
500 */
501static void mark_buffer_async_read(struct buffer_head *bh)
502{
503 bh->b_end_io = end_buffer_async_read;
504 set_buffer_async_read(bh);
505}
506
507void mark_buffer_async_write(struct buffer_head *bh)
508{
509 bh->b_end_io = end_buffer_async_write;
510 set_buffer_async_write(bh);
511}
512EXPORT_SYMBOL(mark_buffer_async_write);
513
514
515/*
516 * fs/buffer.c contains helper functions for buffer-backed address space's
517 * fsync functions. A common requirement for buffer-based filesystems is
518 * that certain data from the backing blockdev needs to be written out for
519 * a successful fsync(). For example, ext2 indirect blocks need to be
520 * written back and waited upon before fsync() returns.
521 *
522 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
523 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
524 * management of a list of dependent buffers at ->i_mapping->private_list.
525 *
526 * Locking is a little subtle: try_to_free_buffers() will remove buffers
527 * from their controlling inode's queue when they are being freed. But
528 * try_to_free_buffers() will be operating against the *blockdev* mapping
529 * at the time, not against the S_ISREG file which depends on those buffers.
530 * So the locking for private_list is via the private_lock in the address_space
531 * which backs the buffers. Which is different from the address_space
532 * against which the buffers are listed. So for a particular address_space,
533 * mapping->private_lock does *not* protect mapping->private_list! In fact,
534 * mapping->private_list will always be protected by the backing blockdev's
535 * ->private_lock.
536 *
537 * Which introduces a requirement: all buffers on an address_space's
538 * ->private_list must be from the same address_space: the blockdev's.
539 *
540 * address_spaces which do not place buffers at ->private_list via these
541 * utility functions are free to use private_lock and private_list for
542 * whatever they want. The only requirement is that list_empty(private_list)
543 * be true at clear_inode() time.
544 *
545 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
546 * filesystems should do that. invalidate_inode_buffers() should just go
547 * BUG_ON(!list_empty).
548 *
549 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
550 * take an address_space, not an inode. And it should be called
551 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
552 * queued up.
553 *
554 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
555 * list if it is already on a list. Because if the buffer is on a list,
556 * it *must* already be on the right one. If not, the filesystem is being
557 * silly. This will save a ton of locking. But first we have to ensure
558 * that buffers are taken *off* the old inode's list when they are freed
559 * (presumably in truncate). That requires careful auditing of all
560 * filesystems (do it inside bforget()). It could also be done by bringing
561 * b_inode back.
562 */
563
564/*
565 * The buffer's backing address_space's private_lock must be held
566 */
567static inline void __remove_assoc_queue(struct buffer_head *bh)
568{
569 list_del_init(&bh->b_assoc_buffers);
Jan Kara58ff4072006-10-17 00:10:19 -0700570 WARN_ON(!bh->b_assoc_map);
571 if (buffer_write_io_error(bh))
572 set_bit(AS_EIO, &bh->b_assoc_map->flags);
573 bh->b_assoc_map = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
576int inode_has_buffers(struct inode *inode)
577{
578 return !list_empty(&inode->i_data.private_list);
579}
580
581/*
582 * osync is designed to support O_SYNC io. It waits synchronously for
583 * all already-submitted IO to complete, but does not queue any new
584 * writes to the disk.
585 *
586 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
587 * you dirty the buffers, and then use osync_inode_buffers to wait for
588 * completion. Any other dirty buffers which are not yet queued for
589 * write will not be flushed to disk by the osync.
590 */
591static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
592{
593 struct buffer_head *bh;
594 struct list_head *p;
595 int err = 0;
596
597 spin_lock(lock);
598repeat:
599 list_for_each_prev(p, list) {
600 bh = BH_ENTRY(p);
601 if (buffer_locked(bh)) {
602 get_bh(bh);
603 spin_unlock(lock);
604 wait_on_buffer(bh);
605 if (!buffer_uptodate(bh))
606 err = -EIO;
607 brelse(bh);
608 spin_lock(lock);
609 goto repeat;
610 }
611 }
612 spin_unlock(lock);
613 return err;
614}
615
616/**
617 * sync_mapping_buffers - write out and wait upon a mapping's "associated"
618 * buffers
Martin Waitz67be2dd2005-05-01 08:59:26 -0700619 * @mapping: the mapping which wants those buffers written
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 *
621 * Starts I/O against the buffers at mapping->private_list, and waits upon
622 * that I/O.
623 *
Martin Waitz67be2dd2005-05-01 08:59:26 -0700624 * Basically, this is a convenience function for fsync().
625 * @mapping is a file or directory which needs those buffers to be written for
626 * a successful fsync().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 */
628int sync_mapping_buffers(struct address_space *mapping)
629{
630 struct address_space *buffer_mapping = mapping->assoc_mapping;
631
632 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
633 return 0;
634
635 return fsync_buffers_list(&buffer_mapping->private_lock,
636 &mapping->private_list);
637}
638EXPORT_SYMBOL(sync_mapping_buffers);
639
640/*
641 * Called when we've recently written block `bblock', and it is known that
642 * `bblock' was for a buffer_boundary() buffer. This means that the block at
643 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
644 * dirty, schedule it for IO. So that indirects merge nicely with their data.
645 */
646void write_boundary_block(struct block_device *bdev,
647 sector_t bblock, unsigned blocksize)
648{
649 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
650 if (bh) {
651 if (buffer_dirty(bh))
652 ll_rw_block(WRITE, 1, &bh);
653 put_bh(bh);
654 }
655}
656
657void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
658{
659 struct address_space *mapping = inode->i_mapping;
660 struct address_space *buffer_mapping = bh->b_page->mapping;
661
662 mark_buffer_dirty(bh);
663 if (!mapping->assoc_mapping) {
664 mapping->assoc_mapping = buffer_mapping;
665 } else {
Eric Sesterhenne827f922006-03-26 18:24:46 +0200666 BUG_ON(mapping->assoc_mapping != buffer_mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 }
668 if (list_empty(&bh->b_assoc_buffers)) {
669 spin_lock(&buffer_mapping->private_lock);
670 list_move_tail(&bh->b_assoc_buffers,
671 &mapping->private_list);
Jan Kara58ff4072006-10-17 00:10:19 -0700672 bh->b_assoc_map = mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 spin_unlock(&buffer_mapping->private_lock);
674 }
675}
676EXPORT_SYMBOL(mark_buffer_dirty_inode);
677
678/*
Nick Piggin787d2212007-07-17 04:03:34 -0700679 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
680 * dirty.
681 *
682 * If warn is true, then emit a warning if the page is not uptodate and has
683 * not been truncated.
684 */
685static int __set_page_dirty(struct page *page,
686 struct address_space *mapping, int warn)
687{
688 if (unlikely(!mapping))
689 return !TestSetPageDirty(page);
690
691 if (TestSetPageDirty(page))
692 return 0;
693
694 write_lock_irq(&mapping->tree_lock);
695 if (page->mapping) { /* Race with truncate? */
696 WARN_ON_ONCE(warn && !PageUptodate(page));
697
698 if (mapping_cap_account_dirty(mapping)) {
699 __inc_zone_page_state(page, NR_FILE_DIRTY);
700 task_io_account_write(PAGE_CACHE_SIZE);
701 }
702 radix_tree_tag_set(&mapping->page_tree,
703 page_index(page), PAGECACHE_TAG_DIRTY);
704 }
705 write_unlock_irq(&mapping->tree_lock);
706 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
707
708 return 1;
709}
710
711/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 * Add a page to the dirty page list.
713 *
714 * It is a sad fact of life that this function is called from several places
715 * deeply under spinlocking. It may not sleep.
716 *
717 * If the page has buffers, the uptodate buffers are set dirty, to preserve
718 * dirty-state coherency between the page and the buffers. It the page does
719 * not have buffers then when they are later attached they will all be set
720 * dirty.
721 *
722 * The buffers are dirtied before the page is dirtied. There's a small race
723 * window in which a writepage caller may see the page cleanness but not the
724 * buffer dirtiness. That's fine. If this code were to set the page dirty
725 * before the buffers, a concurrent writepage caller could clear the page dirty
726 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
727 * page on the dirty page list.
728 *
729 * We use private_lock to lock against try_to_free_buffers while using the
730 * page's buffer list. Also use this to protect against clean buffers being
731 * added to the page after it was set dirty.
732 *
733 * FIXME: may need to call ->reservepage here as well. That's rather up to the
734 * address_space though.
735 */
736int __set_page_dirty_buffers(struct page *page)
737{
Nick Piggin787d2212007-07-17 04:03:34 -0700738 struct address_space *mapping = page_mapping(page);
Nick Pigginebf7a222006-10-10 04:36:54 +0200739
740 if (unlikely(!mapping))
741 return !TestSetPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742
743 spin_lock(&mapping->private_lock);
744 if (page_has_buffers(page)) {
745 struct buffer_head *head = page_buffers(page);
746 struct buffer_head *bh = head;
747
748 do {
749 set_buffer_dirty(bh);
750 bh = bh->b_this_page;
751 } while (bh != head);
752 }
753 spin_unlock(&mapping->private_lock);
754
Nick Piggin787d2212007-07-17 04:03:34 -0700755 return __set_page_dirty(page, mapping, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756}
757EXPORT_SYMBOL(__set_page_dirty_buffers);
758
759/*
760 * Write out and wait upon a list of buffers.
761 *
762 * We have conflicting pressures: we want to make sure that all
763 * initially dirty buffers get waited on, but that any subsequently
764 * dirtied buffers don't. After all, we don't want fsync to last
765 * forever if somebody is actively writing to the file.
766 *
767 * Do this in two main stages: first we copy dirty buffers to a
768 * temporary inode list, queueing the writes as we go. Then we clean
769 * up, waiting for those writes to complete.
770 *
771 * During this second stage, any subsequent updates to the file may end
772 * up refiling the buffer on the original inode's dirty list again, so
773 * there is a chance we will end up with a buffer queued for write but
774 * not yet completed on that list. So, as a final cleanup we go through
775 * the osync code to catch these locked, dirty buffers without requeuing
776 * any newly dirty buffers for write.
777 */
778static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
779{
780 struct buffer_head *bh;
781 struct list_head tmp;
782 int err = 0, err2;
783
784 INIT_LIST_HEAD(&tmp);
785
786 spin_lock(lock);
787 while (!list_empty(list)) {
788 bh = BH_ENTRY(list->next);
Jan Kara58ff4072006-10-17 00:10:19 -0700789 __remove_assoc_queue(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 if (buffer_dirty(bh) || buffer_locked(bh)) {
791 list_add(&bh->b_assoc_buffers, &tmp);
792 if (buffer_dirty(bh)) {
793 get_bh(bh);
794 spin_unlock(lock);
795 /*
796 * Ensure any pending I/O completes so that
797 * ll_rw_block() actually writes the current
798 * contents - it is a noop if I/O is still in
799 * flight on potentially older contents.
800 */
Jan Karaa7662232005-09-06 15:19:10 -0700801 ll_rw_block(SWRITE, 1, &bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 brelse(bh);
803 spin_lock(lock);
804 }
805 }
806 }
807
808 while (!list_empty(&tmp)) {
809 bh = BH_ENTRY(tmp.prev);
Jan Kara58ff4072006-10-17 00:10:19 -0700810 list_del_init(&bh->b_assoc_buffers);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 get_bh(bh);
812 spin_unlock(lock);
813 wait_on_buffer(bh);
814 if (!buffer_uptodate(bh))
815 err = -EIO;
816 brelse(bh);
817 spin_lock(lock);
818 }
819
820 spin_unlock(lock);
821 err2 = osync_buffers_list(lock, list);
822 if (err)
823 return err;
824 else
825 return err2;
826}
827
828/*
829 * Invalidate any and all dirty buffers on a given inode. We are
830 * probably unmounting the fs, but that doesn't mean we have already
831 * done a sync(). Just drop the buffers from the inode list.
832 *
833 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
834 * assumes that all the buffers are against the blockdev. Not true
835 * for reiserfs.
836 */
837void invalidate_inode_buffers(struct inode *inode)
838{
839 if (inode_has_buffers(inode)) {
840 struct address_space *mapping = &inode->i_data;
841 struct list_head *list = &mapping->private_list;
842 struct address_space *buffer_mapping = mapping->assoc_mapping;
843
844 spin_lock(&buffer_mapping->private_lock);
845 while (!list_empty(list))
846 __remove_assoc_queue(BH_ENTRY(list->next));
847 spin_unlock(&buffer_mapping->private_lock);
848 }
849}
850
851/*
852 * Remove any clean buffers from the inode's buffer list. This is called
853 * when we're trying to free the inode itself. Those buffers can pin it.
854 *
855 * Returns true if all buffers were removed.
856 */
857int remove_inode_buffers(struct inode *inode)
858{
859 int ret = 1;
860
861 if (inode_has_buffers(inode)) {
862 struct address_space *mapping = &inode->i_data;
863 struct list_head *list = &mapping->private_list;
864 struct address_space *buffer_mapping = mapping->assoc_mapping;
865
866 spin_lock(&buffer_mapping->private_lock);
867 while (!list_empty(list)) {
868 struct buffer_head *bh = BH_ENTRY(list->next);
869 if (buffer_dirty(bh)) {
870 ret = 0;
871 break;
872 }
873 __remove_assoc_queue(bh);
874 }
875 spin_unlock(&buffer_mapping->private_lock);
876 }
877 return ret;
878}
879
880/*
881 * Create the appropriate buffers when given a page for data area and
882 * the size of each buffer.. Use the bh->b_this_page linked list to
883 * follow the buffers created. Return NULL if unable to create more
884 * buffers.
885 *
886 * The retry flag is used to differentiate async IO (paging, swapping)
887 * which may not fail from ordinary buffer allocations.
888 */
889struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
890 int retry)
891{
892 struct buffer_head *bh, *head;
893 long offset;
894
895try_again:
896 head = NULL;
897 offset = PAGE_SIZE;
898 while ((offset -= size) >= 0) {
899 bh = alloc_buffer_head(GFP_NOFS);
900 if (!bh)
901 goto no_grow;
902
903 bh->b_bdev = NULL;
904 bh->b_this_page = head;
905 bh->b_blocknr = -1;
906 head = bh;
907
908 bh->b_state = 0;
909 atomic_set(&bh->b_count, 0);
Chris Masonfc5cd582006-02-01 03:06:48 -0800910 bh->b_private = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 bh->b_size = size;
912
913 /* Link the buffer to its page */
914 set_bh_page(bh, page, offset);
915
Nathan Scott01ffe332006-01-17 09:02:07 +1100916 init_buffer(bh, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 }
918 return head;
919/*
920 * In case anything failed, we just free everything we got.
921 */
922no_grow:
923 if (head) {
924 do {
925 bh = head;
926 head = head->b_this_page;
927 free_buffer_head(bh);
928 } while (head);
929 }
930
931 /*
932 * Return failure for non-async IO requests. Async IO requests
933 * are not allowed to fail, so we have to wait until buffer heads
934 * become available. But we don't want tasks sleeping with
935 * partially complete buffers, so all were released above.
936 */
937 if (!retry)
938 return NULL;
939
940 /* We're _really_ low on memory. Now we just
941 * wait for old buffer heads to become free due to
942 * finishing IO. Since this is an async request and
943 * the reserve list is empty, we're sure there are
944 * async buffer heads in use.
945 */
946 free_more_memory();
947 goto try_again;
948}
949EXPORT_SYMBOL_GPL(alloc_page_buffers);
950
951static inline void
952link_dev_buffers(struct page *page, struct buffer_head *head)
953{
954 struct buffer_head *bh, *tail;
955
956 bh = head;
957 do {
958 tail = bh;
959 bh = bh->b_this_page;
960 } while (bh);
961 tail->b_this_page = head;
962 attach_page_buffers(page, head);
963}
964
965/*
966 * Initialise the state of a blockdev page's buffers.
967 */
968static void
969init_page_buffers(struct page *page, struct block_device *bdev,
970 sector_t block, int size)
971{
972 struct buffer_head *head = page_buffers(page);
973 struct buffer_head *bh = head;
974 int uptodate = PageUptodate(page);
975
976 do {
977 if (!buffer_mapped(bh)) {
978 init_buffer(bh, NULL, NULL);
979 bh->b_bdev = bdev;
980 bh->b_blocknr = block;
981 if (uptodate)
982 set_buffer_uptodate(bh);
983 set_buffer_mapped(bh);
984 }
985 block++;
986 bh = bh->b_this_page;
987 } while (bh != head);
988}
989
990/*
991 * Create the page-cache page that contains the requested block.
992 *
993 * This is user purely for blockdev mappings.
994 */
995static struct page *
996grow_dev_page(struct block_device *bdev, sector_t block,
997 pgoff_t index, int size)
998{
999 struct inode *inode = bdev->bd_inode;
1000 struct page *page;
1001 struct buffer_head *bh;
1002
Christoph Lameterea125892007-05-16 22:11:21 -07001003 page = find_or_create_page(inode->i_mapping, index,
Mel Gorman769848c2007-07-17 04:03:05 -07001004 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 if (!page)
1006 return NULL;
1007
Eric Sesterhenne827f922006-03-26 18:24:46 +02001008 BUG_ON(!PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009
1010 if (page_has_buffers(page)) {
1011 bh = page_buffers(page);
1012 if (bh->b_size == size) {
1013 init_page_buffers(page, bdev, block, size);
1014 return page;
1015 }
1016 if (!try_to_free_buffers(page))
1017 goto failed;
1018 }
1019
1020 /*
1021 * Allocate some buffers for this page
1022 */
1023 bh = alloc_page_buffers(page, size, 0);
1024 if (!bh)
1025 goto failed;
1026
1027 /*
1028 * Link the page to the buffers and initialise them. Take the
1029 * lock to be atomic wrt __find_get_block(), which does not
1030 * run under the page lock.
1031 */
1032 spin_lock(&inode->i_mapping->private_lock);
1033 link_dev_buffers(page, bh);
1034 init_page_buffers(page, bdev, block, size);
1035 spin_unlock(&inode->i_mapping->private_lock);
1036 return page;
1037
1038failed:
1039 BUG();
1040 unlock_page(page);
1041 page_cache_release(page);
1042 return NULL;
1043}
1044
1045/*
1046 * Create buffers for the specified block device block's page. If
1047 * that page was dirty, the buffers are set dirty also.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08001049static int
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050grow_buffers(struct block_device *bdev, sector_t block, int size)
1051{
1052 struct page *page;
1053 pgoff_t index;
1054 int sizebits;
1055
1056 sizebits = -1;
1057 do {
1058 sizebits++;
1059 } while ((size << sizebits) < PAGE_SIZE);
1060
1061 index = block >> sizebits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062
Andrew Mortone5657932006-10-11 01:21:46 -07001063 /*
1064 * Check for a block which wants to lie outside our maximum possible
1065 * pagecache index. (this comparison is done using sector_t types).
1066 */
1067 if (unlikely(index != block >> sizebits)) {
1068 char b[BDEVNAME_SIZE];
1069
1070 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1071 "device %s\n",
1072 __FUNCTION__, (unsigned long long)block,
1073 bdevname(bdev, b));
1074 return -EIO;
1075 }
1076 block = index << sizebits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 /* Create a page with the proper size buffers.. */
1078 page = grow_dev_page(bdev, block, index, size);
1079 if (!page)
1080 return 0;
1081 unlock_page(page);
1082 page_cache_release(page);
1083 return 1;
1084}
1085
Adrian Bunk75c96f82005-05-05 16:16:09 -07001086static struct buffer_head *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087__getblk_slow(struct block_device *bdev, sector_t block, int size)
1088{
1089 /* Size must be multiple of hard sectorsize */
1090 if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1091 (size < 512 || size > PAGE_SIZE))) {
1092 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1093 size);
1094 printk(KERN_ERR "hardsect size: %d\n",
1095 bdev_hardsect_size(bdev));
1096
1097 dump_stack();
1098 return NULL;
1099 }
1100
1101 for (;;) {
1102 struct buffer_head * bh;
Andrew Mortone5657932006-10-11 01:21:46 -07001103 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104
1105 bh = __find_get_block(bdev, block, size);
1106 if (bh)
1107 return bh;
1108
Andrew Mortone5657932006-10-11 01:21:46 -07001109 ret = grow_buffers(bdev, block, size);
1110 if (ret < 0)
1111 return NULL;
1112 if (ret == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 free_more_memory();
1114 }
1115}
1116
1117/*
1118 * The relationship between dirty buffers and dirty pages:
1119 *
1120 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1121 * the page is tagged dirty in its radix tree.
1122 *
1123 * At all times, the dirtiness of the buffers represents the dirtiness of
1124 * subsections of the page. If the page has buffers, the page dirty bit is
1125 * merely a hint about the true dirty state.
1126 *
1127 * When a page is set dirty in its entirety, all its buffers are marked dirty
1128 * (if the page has buffers).
1129 *
1130 * When a buffer is marked dirty, its page is dirtied, but the page's other
1131 * buffers are not.
1132 *
1133 * Also. When blockdev buffers are explicitly read with bread(), they
1134 * individually become uptodate. But their backing page remains not
1135 * uptodate - even if all of its buffers are uptodate. A subsequent
1136 * block_read_full_page() against that page will discover all the uptodate
1137 * buffers, will set the page uptodate and will perform no I/O.
1138 */
1139
1140/**
1141 * mark_buffer_dirty - mark a buffer_head as needing writeout
Martin Waitz67be2dd2005-05-01 08:59:26 -07001142 * @bh: the buffer_head to mark dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 *
1144 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1145 * backing page dirty, then tag the page as dirty in its address_space's radix
1146 * tree and then attach the address_space's inode to its superblock's dirty
1147 * inode list.
1148 *
1149 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1150 * mapping->tree_lock and the global inode_lock.
1151 */
1152void fastcall mark_buffer_dirty(struct buffer_head *bh)
1153{
Nick Piggin787d2212007-07-17 04:03:34 -07001154 WARN_ON_ONCE(!buffer_uptodate(bh));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
Nick Piggin787d2212007-07-17 04:03:34 -07001156 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157}
1158
1159/*
1160 * Decrement a buffer_head's reference count. If all buffers against a page
1161 * have zero reference count, are clean and unlocked, and if the page is clean
1162 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1163 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1164 * a page but it ends up not being freed, and buffers may later be reattached).
1165 */
1166void __brelse(struct buffer_head * buf)
1167{
1168 if (atomic_read(&buf->b_count)) {
1169 put_bh(buf);
1170 return;
1171 }
1172 printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1173 WARN_ON(1);
1174}
1175
1176/*
1177 * bforget() is like brelse(), except it discards any
1178 * potentially dirty data.
1179 */
1180void __bforget(struct buffer_head *bh)
1181{
1182 clear_buffer_dirty(bh);
1183 if (!list_empty(&bh->b_assoc_buffers)) {
1184 struct address_space *buffer_mapping = bh->b_page->mapping;
1185
1186 spin_lock(&buffer_mapping->private_lock);
1187 list_del_init(&bh->b_assoc_buffers);
Jan Kara58ff4072006-10-17 00:10:19 -07001188 bh->b_assoc_map = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 spin_unlock(&buffer_mapping->private_lock);
1190 }
1191 __brelse(bh);
1192}
1193
1194static struct buffer_head *__bread_slow(struct buffer_head *bh)
1195{
1196 lock_buffer(bh);
1197 if (buffer_uptodate(bh)) {
1198 unlock_buffer(bh);
1199 return bh;
1200 } else {
1201 get_bh(bh);
1202 bh->b_end_io = end_buffer_read_sync;
1203 submit_bh(READ, bh);
1204 wait_on_buffer(bh);
1205 if (buffer_uptodate(bh))
1206 return bh;
1207 }
1208 brelse(bh);
1209 return NULL;
1210}
1211
1212/*
1213 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1214 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1215 * refcount elevated by one when they're in an LRU. A buffer can only appear
1216 * once in a particular CPU's LRU. A single buffer can be present in multiple
1217 * CPU's LRUs at the same time.
1218 *
1219 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1220 * sb_find_get_block().
1221 *
1222 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1223 * a local interrupt disable for that.
1224 */
1225
1226#define BH_LRU_SIZE 8
1227
1228struct bh_lru {
1229 struct buffer_head *bhs[BH_LRU_SIZE];
1230};
1231
1232static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1233
1234#ifdef CONFIG_SMP
1235#define bh_lru_lock() local_irq_disable()
1236#define bh_lru_unlock() local_irq_enable()
1237#else
1238#define bh_lru_lock() preempt_disable()
1239#define bh_lru_unlock() preempt_enable()
1240#endif
1241
1242static inline void check_irqs_on(void)
1243{
1244#ifdef irqs_disabled
1245 BUG_ON(irqs_disabled());
1246#endif
1247}
1248
1249/*
1250 * The LRU management algorithm is dopey-but-simple. Sorry.
1251 */
1252static void bh_lru_install(struct buffer_head *bh)
1253{
1254 struct buffer_head *evictee = NULL;
1255 struct bh_lru *lru;
1256
1257 check_irqs_on();
1258 bh_lru_lock();
1259 lru = &__get_cpu_var(bh_lrus);
1260 if (lru->bhs[0] != bh) {
1261 struct buffer_head *bhs[BH_LRU_SIZE];
1262 int in;
1263 int out = 0;
1264
1265 get_bh(bh);
1266 bhs[out++] = bh;
1267 for (in = 0; in < BH_LRU_SIZE; in++) {
1268 struct buffer_head *bh2 = lru->bhs[in];
1269
1270 if (bh2 == bh) {
1271 __brelse(bh2);
1272 } else {
1273 if (out >= BH_LRU_SIZE) {
1274 BUG_ON(evictee != NULL);
1275 evictee = bh2;
1276 } else {
1277 bhs[out++] = bh2;
1278 }
1279 }
1280 }
1281 while (out < BH_LRU_SIZE)
1282 bhs[out++] = NULL;
1283 memcpy(lru->bhs, bhs, sizeof(bhs));
1284 }
1285 bh_lru_unlock();
1286
1287 if (evictee)
1288 __brelse(evictee);
1289}
1290
1291/*
1292 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1293 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08001294static struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001295lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296{
1297 struct buffer_head *ret = NULL;
1298 struct bh_lru *lru;
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001299 unsigned int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300
1301 check_irqs_on();
1302 bh_lru_lock();
1303 lru = &__get_cpu_var(bh_lrus);
1304 for (i = 0; i < BH_LRU_SIZE; i++) {
1305 struct buffer_head *bh = lru->bhs[i];
1306
1307 if (bh && bh->b_bdev == bdev &&
1308 bh->b_blocknr == block && bh->b_size == size) {
1309 if (i) {
1310 while (i) {
1311 lru->bhs[i] = lru->bhs[i - 1];
1312 i--;
1313 }
1314 lru->bhs[0] = bh;
1315 }
1316 get_bh(bh);
1317 ret = bh;
1318 break;
1319 }
1320 }
1321 bh_lru_unlock();
1322 return ret;
1323}
1324
1325/*
1326 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1327 * it in the LRU and mark it as accessed. If it is not present then return
1328 * NULL
1329 */
1330struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001331__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332{
1333 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1334
1335 if (bh == NULL) {
Coywolf Qi Hunt385fd4c2005-11-07 00:59:39 -08001336 bh = __find_get_block_slow(bdev, block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 if (bh)
1338 bh_lru_install(bh);
1339 }
1340 if (bh)
1341 touch_buffer(bh);
1342 return bh;
1343}
1344EXPORT_SYMBOL(__find_get_block);
1345
1346/*
1347 * __getblk will locate (and, if necessary, create) the buffer_head
1348 * which corresponds to the passed block_device, block and size. The
1349 * returned buffer has its reference count incremented.
1350 *
1351 * __getblk() cannot fail - it just keeps trying. If you pass it an
1352 * illegal block number, __getblk() will happily return a buffer_head
1353 * which represents the non-existent block. Very weird.
1354 *
1355 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1356 * attempt is failing. FIXME, perhaps?
1357 */
1358struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001359__getblk(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360{
1361 struct buffer_head *bh = __find_get_block(bdev, block, size);
1362
1363 might_sleep();
1364 if (bh == NULL)
1365 bh = __getblk_slow(bdev, block, size);
1366 return bh;
1367}
1368EXPORT_SYMBOL(__getblk);
1369
1370/*
1371 * Do async read-ahead on a buffer..
1372 */
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001373void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374{
1375 struct buffer_head *bh = __getblk(bdev, block, size);
Andrew Mortona3e713b2005-10-30 15:03:15 -08001376 if (likely(bh)) {
1377 ll_rw_block(READA, 1, &bh);
1378 brelse(bh);
1379 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381EXPORT_SYMBOL(__breadahead);
1382
1383/**
1384 * __bread() - reads a specified block and returns the bh
Martin Waitz67be2dd2005-05-01 08:59:26 -07001385 * @bdev: the block_device to read from
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 * @block: number of block
1387 * @size: size (in bytes) to read
1388 *
1389 * Reads a specified block, and returns buffer head that contains it.
1390 * It returns NULL if the block was unreadable.
1391 */
1392struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001393__bread(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394{
1395 struct buffer_head *bh = __getblk(bdev, block, size);
1396
Andrew Mortona3e713b2005-10-30 15:03:15 -08001397 if (likely(bh) && !buffer_uptodate(bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 bh = __bread_slow(bh);
1399 return bh;
1400}
1401EXPORT_SYMBOL(__bread);
1402
1403/*
1404 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1405 * This doesn't race because it runs in each cpu either in irq
1406 * or with preempt disabled.
1407 */
1408static void invalidate_bh_lru(void *arg)
1409{
1410 struct bh_lru *b = &get_cpu_var(bh_lrus);
1411 int i;
1412
1413 for (i = 0; i < BH_LRU_SIZE; i++) {
1414 brelse(b->bhs[i]);
1415 b->bhs[i] = NULL;
1416 }
1417 put_cpu_var(bh_lrus);
1418}
1419
Peter Zijlstraf9a14392007-05-06 14:49:55 -07001420void invalidate_bh_lrus(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421{
1422 on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1423}
1424
1425void set_bh_page(struct buffer_head *bh,
1426 struct page *page, unsigned long offset)
1427{
1428 bh->b_page = page;
Eric Sesterhenne827f922006-03-26 18:24:46 +02001429 BUG_ON(offset >= PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 if (PageHighMem(page))
1431 /*
1432 * This catches illegal uses and preserves the offset:
1433 */
1434 bh->b_data = (char *)(0 + offset);
1435 else
1436 bh->b_data = page_address(page) + offset;
1437}
1438EXPORT_SYMBOL(set_bh_page);
1439
1440/*
1441 * Called when truncating a buffer on a page completely.
1442 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08001443static void discard_buffer(struct buffer_head * bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444{
1445 lock_buffer(bh);
1446 clear_buffer_dirty(bh);
1447 bh->b_bdev = NULL;
1448 clear_buffer_mapped(bh);
1449 clear_buffer_req(bh);
1450 clear_buffer_new(bh);
1451 clear_buffer_delay(bh);
David Chinner33a266d2007-02-12 00:51:41 -08001452 clear_buffer_unwritten(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 unlock_buffer(bh);
1454}
1455
1456/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 * block_invalidatepage - invalidate part of all of a buffer-backed page
1458 *
1459 * @page: the page which is affected
1460 * @offset: the index of the truncation point
1461 *
1462 * block_invalidatepage() is called when all or part of the page has become
1463 * invalidatedby a truncate operation.
1464 *
1465 * block_invalidatepage() does not have to release all buffers, but it must
1466 * ensure that no dirty buffer is left outside @offset and that no I/O
1467 * is underway against any of the blocks which are outside the truncation
1468 * point. Because the caller is about to free (and possibly reuse) those
1469 * blocks on-disk.
1470 */
NeilBrown2ff28e22006-03-26 01:37:18 -08001471void block_invalidatepage(struct page *page, unsigned long offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
1473 struct buffer_head *head, *bh, *next;
1474 unsigned int curr_off = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
1476 BUG_ON(!PageLocked(page));
1477 if (!page_has_buffers(page))
1478 goto out;
1479
1480 head = page_buffers(page);
1481 bh = head;
1482 do {
1483 unsigned int next_off = curr_off + bh->b_size;
1484 next = bh->b_this_page;
1485
1486 /*
1487 * is this block fully invalidated?
1488 */
1489 if (offset <= curr_off)
1490 discard_buffer(bh);
1491 curr_off = next_off;
1492 bh = next;
1493 } while (bh != head);
1494
1495 /*
1496 * We release buffers only if the entire page is being invalidated.
1497 * The get_block cached value has been unconditionally invalidated,
1498 * so real IO is not possible anymore.
1499 */
1500 if (offset == 0)
NeilBrown2ff28e22006-03-26 01:37:18 -08001501 try_to_release_page(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502out:
NeilBrown2ff28e22006-03-26 01:37:18 -08001503 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504}
1505EXPORT_SYMBOL(block_invalidatepage);
1506
1507/*
1508 * We attach and possibly dirty the buffers atomically wrt
1509 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1510 * is already excluded via the page lock.
1511 */
1512void create_empty_buffers(struct page *page,
1513 unsigned long blocksize, unsigned long b_state)
1514{
1515 struct buffer_head *bh, *head, *tail;
1516
1517 head = alloc_page_buffers(page, blocksize, 1);
1518 bh = head;
1519 do {
1520 bh->b_state |= b_state;
1521 tail = bh;
1522 bh = bh->b_this_page;
1523 } while (bh);
1524 tail->b_this_page = head;
1525
1526 spin_lock(&page->mapping->private_lock);
1527 if (PageUptodate(page) || PageDirty(page)) {
1528 bh = head;
1529 do {
1530 if (PageDirty(page))
1531 set_buffer_dirty(bh);
1532 if (PageUptodate(page))
1533 set_buffer_uptodate(bh);
1534 bh = bh->b_this_page;
1535 } while (bh != head);
1536 }
1537 attach_page_buffers(page, head);
1538 spin_unlock(&page->mapping->private_lock);
1539}
1540EXPORT_SYMBOL(create_empty_buffers);
1541
1542/*
1543 * We are taking a block for data and we don't want any output from any
1544 * buffer-cache aliases starting from return from that function and
1545 * until the moment when something will explicitly mark the buffer
1546 * dirty (hopefully that will not happen until we will free that block ;-)
1547 * We don't even need to mark it not-uptodate - nobody can expect
1548 * anything from a newly allocated buffer anyway. We used to used
1549 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1550 * don't want to mark the alias unmapped, for example - it would confuse
1551 * anyone who might pick it with bread() afterwards...
1552 *
1553 * Also.. Note that bforget() doesn't lock the buffer. So there can
1554 * be writeout I/O going on against recently-freed buffers. We don't
1555 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1556 * only if we really need to. That happens here.
1557 */
1558void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1559{
1560 struct buffer_head *old_bh;
1561
1562 might_sleep();
1563
Coywolf Qi Hunt385fd4c2005-11-07 00:59:39 -08001564 old_bh = __find_get_block_slow(bdev, block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 if (old_bh) {
1566 clear_buffer_dirty(old_bh);
1567 wait_on_buffer(old_bh);
1568 clear_buffer_req(old_bh);
1569 __brelse(old_bh);
1570 }
1571}
1572EXPORT_SYMBOL(unmap_underlying_metadata);
1573
1574/*
1575 * NOTE! All mapped/uptodate combinations are valid:
1576 *
1577 * Mapped Uptodate Meaning
1578 *
1579 * No No "unknown" - must do get_block()
1580 * No Yes "hole" - zero-filled
1581 * Yes No "allocated" - allocated on disk, not read in
1582 * Yes Yes "valid" - allocated and up-to-date in memory.
1583 *
1584 * "Dirty" is valid only with the last case (mapped+uptodate).
1585 */
1586
1587/*
1588 * While block_write_full_page is writing back the dirty buffers under
1589 * the page lock, whoever dirtied the buffers may decide to clean them
1590 * again at any time. We handle that by only looking at the buffer
1591 * state inside lock_buffer().
1592 *
1593 * If block_write_full_page() is called for regular writeback
1594 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1595 * locked buffer. This only can happen if someone has written the buffer
1596 * directly, with submit_bh(). At the address_space level PageWriteback
1597 * prevents this contention from occurring.
1598 */
1599static int __block_write_full_page(struct inode *inode, struct page *page,
1600 get_block_t *get_block, struct writeback_control *wbc)
1601{
1602 int err;
1603 sector_t block;
1604 sector_t last_block;
Andrew Mortonf0fbd5f2005-05-05 16:15:48 -07001605 struct buffer_head *bh, *head;
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001606 const unsigned blocksize = 1 << inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607 int nr_underway = 0;
1608
1609 BUG_ON(!PageLocked(page));
1610
1611 last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1612
1613 if (!page_has_buffers(page)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001614 create_empty_buffers(page, blocksize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 (1 << BH_Dirty)|(1 << BH_Uptodate));
1616 }
1617
1618 /*
1619 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1620 * here, and the (potentially unmapped) buffers may become dirty at
1621 * any time. If a buffer becomes dirty here after we've inspected it
1622 * then we just miss that fact, and the page stays dirty.
1623 *
1624 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1625 * handle that here by just cleaning them.
1626 */
1627
Andrew Morton54b21a72006-01-08 01:03:05 -08001628 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 head = page_buffers(page);
1630 bh = head;
1631
1632 /*
1633 * Get all the dirty buffers mapped to disk addresses and
1634 * handle any aliases from the underlying blockdev's mapping.
1635 */
1636 do {
1637 if (block > last_block) {
1638 /*
1639 * mapped buffers outside i_size will occur, because
1640 * this page can be outside i_size when there is a
1641 * truncate in progress.
1642 */
1643 /*
1644 * The buffer was zeroed by block_write_full_page()
1645 */
1646 clear_buffer_dirty(bh);
1647 set_buffer_uptodate(bh);
1648 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001649 WARN_ON(bh->b_size != blocksize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 err = get_block(inode, block, bh, 1);
1651 if (err)
1652 goto recover;
1653 if (buffer_new(bh)) {
1654 /* blockdev mappings never come here */
1655 clear_buffer_new(bh);
1656 unmap_underlying_metadata(bh->b_bdev,
1657 bh->b_blocknr);
1658 }
1659 }
1660 bh = bh->b_this_page;
1661 block++;
1662 } while (bh != head);
1663
1664 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 if (!buffer_mapped(bh))
1666 continue;
1667 /*
1668 * If it's a fully non-blocking write attempt and we cannot
1669 * lock the buffer then redirty the page. Note that this can
1670 * potentially cause a busy-wait loop from pdflush and kswapd
1671 * activity, but those code paths have their own higher-level
1672 * throttling.
1673 */
1674 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1675 lock_buffer(bh);
1676 } else if (test_set_buffer_locked(bh)) {
1677 redirty_page_for_writepage(wbc, page);
1678 continue;
1679 }
1680 if (test_clear_buffer_dirty(bh)) {
1681 mark_buffer_async_write(bh);
1682 } else {
1683 unlock_buffer(bh);
1684 }
1685 } while ((bh = bh->b_this_page) != head);
1686
1687 /*
1688 * The page and its buffers are protected by PageWriteback(), so we can
1689 * drop the bh refcounts early.
1690 */
1691 BUG_ON(PageWriteback(page));
1692 set_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693
1694 do {
1695 struct buffer_head *next = bh->b_this_page;
1696 if (buffer_async_write(bh)) {
1697 submit_bh(WRITE, bh);
1698 nr_underway++;
1699 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 bh = next;
1701 } while (bh != head);
Andrew Morton05937ba2005-05-05 16:15:47 -07001702 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703
1704 err = 0;
1705done:
1706 if (nr_underway == 0) {
1707 /*
1708 * The page was marked dirty, but the buffers were
1709 * clean. Someone wrote them back by hand with
1710 * ll_rw_block/submit_bh. A rare case.
1711 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 end_page_writeback(page);
Nick Piggin3d67f2d2007-05-06 14:49:05 -07001713
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 /*
1715 * The page and buffer_heads can be released at any time from
1716 * here on.
1717 */
1718 wbc->pages_skipped++; /* We didn't write this page */
1719 }
1720 return err;
1721
1722recover:
1723 /*
1724 * ENOSPC, or some other error. We may already have added some
1725 * blocks to the file, so we need to write these out to avoid
1726 * exposing stale data.
1727 * The page is currently locked and not marked for writeback
1728 */
1729 bh = head;
1730 /* Recovery: lock and submit the mapped buffers */
1731 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1733 lock_buffer(bh);
1734 mark_buffer_async_write(bh);
1735 } else {
1736 /*
1737 * The buffer may have been set dirty during
1738 * attachment to a dirty page.
1739 */
1740 clear_buffer_dirty(bh);
1741 }
1742 } while ((bh = bh->b_this_page) != head);
1743 SetPageError(page);
1744 BUG_ON(PageWriteback(page));
Andrew Morton7e4c3692007-05-08 00:23:27 -07001745 mapping_set_error(page->mapping, err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 set_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 do {
1748 struct buffer_head *next = bh->b_this_page;
1749 if (buffer_async_write(bh)) {
1750 clear_buffer_dirty(bh);
1751 submit_bh(WRITE, bh);
1752 nr_underway++;
1753 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 bh = next;
1755 } while (bh != head);
Nick Pigginffda9d32007-02-20 13:57:54 -08001756 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 goto done;
1758}
1759
1760static int __block_prepare_write(struct inode *inode, struct page *page,
1761 unsigned from, unsigned to, get_block_t *get_block)
1762{
1763 unsigned block_start, block_end;
1764 sector_t block;
1765 int err = 0;
1766 unsigned blocksize, bbits;
1767 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1768
1769 BUG_ON(!PageLocked(page));
1770 BUG_ON(from > PAGE_CACHE_SIZE);
1771 BUG_ON(to > PAGE_CACHE_SIZE);
1772 BUG_ON(from > to);
1773
1774 blocksize = 1 << inode->i_blkbits;
1775 if (!page_has_buffers(page))
1776 create_empty_buffers(page, blocksize, 0);
1777 head = page_buffers(page);
1778
1779 bbits = inode->i_blkbits;
1780 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1781
1782 for(bh = head, block_start = 0; bh != head || !block_start;
1783 block++, block_start=block_end, bh = bh->b_this_page) {
1784 block_end = block_start + blocksize;
1785 if (block_end <= from || block_start >= to) {
1786 if (PageUptodate(page)) {
1787 if (!buffer_uptodate(bh))
1788 set_buffer_uptodate(bh);
1789 }
1790 continue;
1791 }
1792 if (buffer_new(bh))
1793 clear_buffer_new(bh);
1794 if (!buffer_mapped(bh)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001795 WARN_ON(bh->b_size != blocksize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 err = get_block(inode, block, bh, 1);
1797 if (err)
Nick Pigginf3ddbdc2005-05-05 16:15:45 -07001798 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 if (buffer_new(bh)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 unmap_underlying_metadata(bh->b_bdev,
1801 bh->b_blocknr);
1802 if (PageUptodate(page)) {
1803 set_buffer_uptodate(bh);
1804 continue;
1805 }
1806 if (block_end > to || block_start < from) {
1807 void *kaddr;
1808
1809 kaddr = kmap_atomic(page, KM_USER0);
1810 if (block_end > to)
1811 memset(kaddr+to, 0,
1812 block_end-to);
1813 if (block_start < from)
1814 memset(kaddr+block_start,
1815 0, from-block_start);
1816 flush_dcache_page(page);
1817 kunmap_atomic(kaddr, KM_USER0);
1818 }
1819 continue;
1820 }
1821 }
1822 if (PageUptodate(page)) {
1823 if (!buffer_uptodate(bh))
1824 set_buffer_uptodate(bh);
1825 continue;
1826 }
1827 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
David Chinner33a266d2007-02-12 00:51:41 -08001828 !buffer_unwritten(bh) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 (block_start < from || block_end > to)) {
1830 ll_rw_block(READ, 1, &bh);
1831 *wait_bh++=bh;
1832 }
1833 }
1834 /*
1835 * If we issued read requests - let them complete.
1836 */
1837 while(wait_bh > wait) {
1838 wait_on_buffer(*--wait_bh);
1839 if (!buffer_uptodate(*wait_bh))
Nick Pigginf3ddbdc2005-05-05 16:15:45 -07001840 err = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 }
Anton Altaparmakov152becd2005-06-23 00:10:21 -07001842 if (!err) {
1843 bh = head;
1844 do {
1845 if (buffer_new(bh))
1846 clear_buffer_new(bh);
1847 } while ((bh = bh->b_this_page) != head);
1848 return 0;
1849 }
Nick Pigginf3ddbdc2005-05-05 16:15:45 -07001850 /* Error case: */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 /*
1852 * Zero out any newly allocated blocks to avoid exposing stale
1853 * data. If BH_New is set, we know that the block was newly
1854 * allocated in the above loop.
1855 */
1856 bh = head;
1857 block_start = 0;
1858 do {
1859 block_end = block_start+blocksize;
1860 if (block_end <= from)
1861 goto next_bh;
1862 if (block_start >= to)
1863 break;
1864 if (buffer_new(bh)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 clear_buffer_new(bh);
Nate Diller01f27052007-05-09 02:35:07 -07001866 zero_user_page(page, block_start, bh->b_size, KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 set_buffer_uptodate(bh);
1868 mark_buffer_dirty(bh);
1869 }
1870next_bh:
1871 block_start = block_end;
1872 bh = bh->b_this_page;
1873 } while (bh != head);
1874 return err;
1875}
1876
1877static int __block_commit_write(struct inode *inode, struct page *page,
1878 unsigned from, unsigned to)
1879{
1880 unsigned block_start, block_end;
1881 int partial = 0;
1882 unsigned blocksize;
1883 struct buffer_head *bh, *head;
1884
1885 blocksize = 1 << inode->i_blkbits;
1886
1887 for(bh = head = page_buffers(page), block_start = 0;
1888 bh != head || !block_start;
1889 block_start=block_end, bh = bh->b_this_page) {
1890 block_end = block_start + blocksize;
1891 if (block_end <= from || block_start >= to) {
1892 if (!buffer_uptodate(bh))
1893 partial = 1;
1894 } else {
1895 set_buffer_uptodate(bh);
1896 mark_buffer_dirty(bh);
1897 }
1898 }
1899
1900 /*
1901 * If this is a partial write which happened to make all buffers
1902 * uptodate then we can optimize away a bogus readpage() for
1903 * the next read(). Here we 'discover' whether the page went
1904 * uptodate as a result of this (potentially partial) write.
1905 */
1906 if (!partial)
1907 SetPageUptodate(page);
1908 return 0;
1909}
1910
1911/*
1912 * Generic "read page" function for block devices that have the normal
1913 * get_block functionality. This is most of the block device filesystems.
1914 * Reads the page asynchronously --- the unlock_buffer() and
1915 * set/clear_buffer_uptodate() functions propagate buffer state into the
1916 * page struct once IO has completed.
1917 */
1918int block_read_full_page(struct page *page, get_block_t *get_block)
1919{
1920 struct inode *inode = page->mapping->host;
1921 sector_t iblock, lblock;
1922 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
1923 unsigned int blocksize;
1924 int nr, i;
1925 int fully_mapped = 1;
1926
Matt Mackallcd7619d2005-05-01 08:59:01 -07001927 BUG_ON(!PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 blocksize = 1 << inode->i_blkbits;
1929 if (!page_has_buffers(page))
1930 create_empty_buffers(page, blocksize, 0);
1931 head = page_buffers(page);
1932
1933 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1934 lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
1935 bh = head;
1936 nr = 0;
1937 i = 0;
1938
1939 do {
1940 if (buffer_uptodate(bh))
1941 continue;
1942
1943 if (!buffer_mapped(bh)) {
Andrew Mortonc64610b2005-05-16 21:53:49 -07001944 int err = 0;
1945
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 fully_mapped = 0;
1947 if (iblock < lblock) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001948 WARN_ON(bh->b_size != blocksize);
Andrew Mortonc64610b2005-05-16 21:53:49 -07001949 err = get_block(inode, iblock, bh, 0);
1950 if (err)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 SetPageError(page);
1952 }
1953 if (!buffer_mapped(bh)) {
Nate Diller01f27052007-05-09 02:35:07 -07001954 zero_user_page(page, i * blocksize, blocksize,
1955 KM_USER0);
Andrew Mortonc64610b2005-05-16 21:53:49 -07001956 if (!err)
1957 set_buffer_uptodate(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 continue;
1959 }
1960 /*
1961 * get_block() might have updated the buffer
1962 * synchronously
1963 */
1964 if (buffer_uptodate(bh))
1965 continue;
1966 }
1967 arr[nr++] = bh;
1968 } while (i++, iblock++, (bh = bh->b_this_page) != head);
1969
1970 if (fully_mapped)
1971 SetPageMappedToDisk(page);
1972
1973 if (!nr) {
1974 /*
1975 * All buffers are uptodate - we can set the page uptodate
1976 * as well. But not if get_block() returned an error.
1977 */
1978 if (!PageError(page))
1979 SetPageUptodate(page);
1980 unlock_page(page);
1981 return 0;
1982 }
1983
1984 /* Stage two: lock the buffers */
1985 for (i = 0; i < nr; i++) {
1986 bh = arr[i];
1987 lock_buffer(bh);
1988 mark_buffer_async_read(bh);
1989 }
1990
1991 /*
1992 * Stage 3: start the IO. Check for uptodateness
1993 * inside the buffer lock in case another process reading
1994 * the underlying blockdev brought it uptodate (the sct fix).
1995 */
1996 for (i = 0; i < nr; i++) {
1997 bh = arr[i];
1998 if (buffer_uptodate(bh))
1999 end_buffer_async_read(bh, 1);
2000 else
2001 submit_bh(READ, bh);
2002 }
2003 return 0;
2004}
2005
2006/* utility function for filesystems that need to do work on expanding
2007 * truncates. Uses prepare/commit_write to allow the filesystem to
2008 * deal with the hole.
2009 */
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002010static int __generic_cont_expand(struct inode *inode, loff_t size,
2011 pgoff_t index, unsigned int offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012{
2013 struct address_space *mapping = inode->i_mapping;
2014 struct page *page;
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002015 unsigned long limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016 int err;
2017
2018 err = -EFBIG;
2019 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2020 if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2021 send_sig(SIGXFSZ, current, 0);
2022 goto out;
2023 }
2024 if (size > inode->i_sb->s_maxbytes)
2025 goto out;
2026
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 err = -ENOMEM;
2028 page = grab_cache_page(mapping, index);
2029 if (!page)
2030 goto out;
2031 err = mapping->a_ops->prepare_write(NULL, page, offset, offset);
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002032 if (err) {
2033 /*
2034 * ->prepare_write() may have instantiated a few blocks
2035 * outside i_size. Trim these off again.
2036 */
2037 unlock_page(page);
2038 page_cache_release(page);
2039 vmtruncate(inode, inode->i_size);
2040 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 }
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002042
2043 err = mapping->a_ops->commit_write(NULL, page, offset, offset);
2044
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 unlock_page(page);
2046 page_cache_release(page);
2047 if (err > 0)
2048 err = 0;
2049out:
2050 return err;
2051}
2052
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002053int generic_cont_expand(struct inode *inode, loff_t size)
2054{
2055 pgoff_t index;
2056 unsigned int offset;
2057
2058 offset = (size & (PAGE_CACHE_SIZE - 1)); /* Within page */
2059
2060 /* ugh. in prepare/commit_write, if from==to==start of block, we
2061 ** skip the prepare. make sure we never send an offset for the start
2062 ** of a block
2063 */
2064 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2065 /* caller must handle this extra byte. */
2066 offset++;
2067 }
2068 index = size >> PAGE_CACHE_SHIFT;
2069
2070 return __generic_cont_expand(inode, size, index, offset);
2071}
2072
2073int generic_cont_expand_simple(struct inode *inode, loff_t size)
2074{
2075 loff_t pos = size - 1;
2076 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2077 unsigned int offset = (pos & (PAGE_CACHE_SIZE - 1)) + 1;
2078
2079 /* prepare/commit_write can handle even if from==to==start of block. */
2080 return __generic_cont_expand(inode, size, index, offset);
2081}
2082
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083/*
2084 * For moronic filesystems that do not allow holes in file.
2085 * We may have to extend the file.
2086 */
2087
2088int cont_prepare_write(struct page *page, unsigned offset,
2089 unsigned to, get_block_t *get_block, loff_t *bytes)
2090{
2091 struct address_space *mapping = page->mapping;
2092 struct inode *inode = mapping->host;
2093 struct page *new_page;
2094 pgoff_t pgpos;
2095 long status;
2096 unsigned zerofrom;
2097 unsigned blocksize = 1 << inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098
2099 while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
2100 status = -ENOMEM;
2101 new_page = grab_cache_page(mapping, pgpos);
2102 if (!new_page)
2103 goto out;
2104 /* we might sleep */
2105 if (*bytes>>PAGE_CACHE_SHIFT != pgpos) {
2106 unlock_page(new_page);
2107 page_cache_release(new_page);
2108 continue;
2109 }
2110 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2111 if (zerofrom & (blocksize-1)) {
2112 *bytes |= (blocksize-1);
2113 (*bytes)++;
2114 }
2115 status = __block_prepare_write(inode, new_page, zerofrom,
2116 PAGE_CACHE_SIZE, get_block);
2117 if (status)
2118 goto out_unmap;
OGAWA Hirofumiff1be9a2007-05-20 23:39:40 +09002119 zero_user_page(new_page, zerofrom, PAGE_CACHE_SIZE - zerofrom,
Nate Diller01f27052007-05-09 02:35:07 -07002120 KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 generic_commit_write(NULL, new_page, zerofrom, PAGE_CACHE_SIZE);
2122 unlock_page(new_page);
2123 page_cache_release(new_page);
2124 }
2125
2126 if (page->index < pgpos) {
2127 /* completely inside the area */
2128 zerofrom = offset;
2129 } else {
2130 /* page covers the boundary, find the boundary offset */
2131 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2132
2133 /* if we will expand the thing last block will be filled */
2134 if (to > zerofrom && (zerofrom & (blocksize-1))) {
2135 *bytes |= (blocksize-1);
2136 (*bytes)++;
2137 }
2138
2139 /* starting below the boundary? Nothing to zero out */
2140 if (offset <= zerofrom)
2141 zerofrom = offset;
2142 }
2143 status = __block_prepare_write(inode, page, zerofrom, to, get_block);
2144 if (status)
2145 goto out1;
2146 if (zerofrom < offset) {
Nate Diller01f27052007-05-09 02:35:07 -07002147 zero_user_page(page, zerofrom, offset - zerofrom, KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 __block_commit_write(inode, page, zerofrom, offset);
2149 }
2150 return 0;
2151out1:
2152 ClearPageUptodate(page);
2153 return status;
2154
2155out_unmap:
2156 ClearPageUptodate(new_page);
2157 unlock_page(new_page);
2158 page_cache_release(new_page);
2159out:
2160 return status;
2161}
2162
2163int block_prepare_write(struct page *page, unsigned from, unsigned to,
2164 get_block_t *get_block)
2165{
2166 struct inode *inode = page->mapping->host;
2167 int err = __block_prepare_write(inode, page, from, to, get_block);
2168 if (err)
2169 ClearPageUptodate(page);
2170 return err;
2171}
2172
2173int block_commit_write(struct page *page, unsigned from, unsigned to)
2174{
2175 struct inode *inode = page->mapping->host;
2176 __block_commit_write(inode,page,from,to);
2177 return 0;
2178}
2179
2180int generic_commit_write(struct file *file, struct page *page,
2181 unsigned from, unsigned to)
2182{
2183 struct inode *inode = page->mapping->host;
2184 loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2185 __block_commit_write(inode,page,from,to);
2186 /*
2187 * No need to use i_size_read() here, the i_size
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08002188 * cannot change under us because we hold i_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 */
2190 if (pos > inode->i_size) {
2191 i_size_write(inode, pos);
2192 mark_inode_dirty(inode);
2193 }
2194 return 0;
2195}
2196
David Chinner54171692007-07-19 17:39:55 +10002197/*
2198 * block_page_mkwrite() is not allowed to change the file size as it gets
2199 * called from a page fault handler when a page is first dirtied. Hence we must
2200 * be careful to check for EOF conditions here. We set the page up correctly
2201 * for a written page which means we get ENOSPC checking when writing into
2202 * holes and correct delalloc and unwritten extent mapping on filesystems that
2203 * support these features.
2204 *
2205 * We are not allowed to take the i_mutex here so we have to play games to
2206 * protect against truncate races as the page could now be beyond EOF. Because
2207 * vmtruncate() writes the inode size before removing pages, once we have the
2208 * page lock we can determine safely if the page is beyond EOF. If it is not
2209 * beyond EOF, then the page is guaranteed safe against truncation until we
2210 * unlock the page.
2211 */
2212int
2213block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2214 get_block_t get_block)
2215{
2216 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2217 unsigned long end;
2218 loff_t size;
2219 int ret = -EINVAL;
2220
2221 lock_page(page);
2222 size = i_size_read(inode);
2223 if ((page->mapping != inode->i_mapping) ||
Nick Piggin18336332007-07-20 00:31:45 -07002224 (page_offset(page) > size)) {
David Chinner54171692007-07-19 17:39:55 +10002225 /* page got truncated out from underneath us */
2226 goto out_unlock;
2227 }
2228
2229 /* page is wholly or partially inside EOF */
2230 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2231 end = size & ~PAGE_CACHE_MASK;
2232 else
2233 end = PAGE_CACHE_SIZE;
2234
2235 ret = block_prepare_write(page, 0, end, get_block);
2236 if (!ret)
2237 ret = block_commit_write(page, 0, end);
2238
2239out_unlock:
2240 unlock_page(page);
2241 return ret;
2242}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243
2244/*
2245 * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2246 * immediately, while under the page lock. So it needs a special end_io
2247 * handler which does not touch the bh after unlocking it.
2248 *
2249 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
2250 * a race there is benign: unlock_buffer() only use the bh's address for
2251 * hashing after unlocking the buffer, so it doesn't actually touch the bh
2252 * itself.
2253 */
2254static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2255{
2256 if (uptodate) {
2257 set_buffer_uptodate(bh);
2258 } else {
2259 /* This happens, due to failed READA attempts. */
2260 clear_buffer_uptodate(bh);
2261 }
2262 unlock_buffer(bh);
2263}
2264
2265/*
2266 * On entry, the page is fully not uptodate.
2267 * On exit the page is fully uptodate in the areas outside (from,to)
2268 */
2269int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2270 get_block_t *get_block)
2271{
2272 struct inode *inode = page->mapping->host;
2273 const unsigned blkbits = inode->i_blkbits;
2274 const unsigned blocksize = 1 << blkbits;
2275 struct buffer_head map_bh;
2276 struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
2277 unsigned block_in_page;
2278 unsigned block_start;
2279 sector_t block_in_file;
2280 char *kaddr;
2281 int nr_reads = 0;
2282 int i;
2283 int ret = 0;
2284 int is_mapped_to_disk = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
2286 if (PageMappedToDisk(page))
2287 return 0;
2288
2289 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2290 map_bh.b_page = page;
2291
2292 /*
2293 * We loop across all blocks in the page, whether or not they are
2294 * part of the affected region. This is so we can discover if the
2295 * page is fully mapped-to-disk.
2296 */
2297 for (block_start = 0, block_in_page = 0;
2298 block_start < PAGE_CACHE_SIZE;
2299 block_in_page++, block_start += blocksize) {
2300 unsigned block_end = block_start + blocksize;
2301 int create;
2302
2303 map_bh.b_state = 0;
2304 create = 1;
2305 if (block_start >= to)
2306 create = 0;
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08002307 map_bh.b_size = blocksize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 ret = get_block(inode, block_in_file + block_in_page,
2309 &map_bh, create);
2310 if (ret)
2311 goto failed;
2312 if (!buffer_mapped(&map_bh))
2313 is_mapped_to_disk = 0;
2314 if (buffer_new(&map_bh))
2315 unmap_underlying_metadata(map_bh.b_bdev,
2316 map_bh.b_blocknr);
2317 if (PageUptodate(page))
2318 continue;
2319 if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
2320 kaddr = kmap_atomic(page, KM_USER0);
Nick Piggin22c8ca72007-02-20 13:58:09 -08002321 if (block_start < from)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 memset(kaddr+block_start, 0, from-block_start);
Nick Piggin22c8ca72007-02-20 13:58:09 -08002323 if (block_end > to)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 memset(kaddr + to, 0, block_end - to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 flush_dcache_page(page);
2326 kunmap_atomic(kaddr, KM_USER0);
2327 continue;
2328 }
2329 if (buffer_uptodate(&map_bh))
2330 continue; /* reiserfs does this */
2331 if (block_start < from || block_end > to) {
2332 struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
2333
2334 if (!bh) {
2335 ret = -ENOMEM;
2336 goto failed;
2337 }
2338 bh->b_state = map_bh.b_state;
2339 atomic_set(&bh->b_count, 0);
2340 bh->b_this_page = NULL;
2341 bh->b_page = page;
2342 bh->b_blocknr = map_bh.b_blocknr;
2343 bh->b_size = blocksize;
2344 bh->b_data = (char *)(long)block_start;
2345 bh->b_bdev = map_bh.b_bdev;
2346 bh->b_private = NULL;
2347 read_bh[nr_reads++] = bh;
2348 }
2349 }
2350
2351 if (nr_reads) {
2352 struct buffer_head *bh;
2353
2354 /*
2355 * The page is locked, so these buffers are protected from
2356 * any VM or truncate activity. Hence we don't need to care
2357 * for the buffer_head refcounts.
2358 */
2359 for (i = 0; i < nr_reads; i++) {
2360 bh = read_bh[i];
2361 lock_buffer(bh);
2362 bh->b_end_io = end_buffer_read_nobh;
2363 submit_bh(READ, bh);
2364 }
2365 for (i = 0; i < nr_reads; i++) {
2366 bh = read_bh[i];
2367 wait_on_buffer(bh);
2368 if (!buffer_uptodate(bh))
2369 ret = -EIO;
2370 free_buffer_head(bh);
2371 read_bh[i] = NULL;
2372 }
2373 if (ret)
2374 goto failed;
2375 }
2376
2377 if (is_mapped_to_disk)
2378 SetPageMappedToDisk(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379
2380 return 0;
2381
2382failed:
2383 for (i = 0; i < nr_reads; i++) {
2384 if (read_bh[i])
2385 free_buffer_head(read_bh[i]);
2386 }
2387
2388 /*
2389 * Error recovery is pretty slack. Clear the page and mark it dirty
2390 * so we'll later zero out any blocks which _were_ allocated.
2391 */
Nate Diller01f27052007-05-09 02:35:07 -07002392 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 SetPageUptodate(page);
2394 set_page_dirty(page);
2395 return ret;
2396}
2397EXPORT_SYMBOL(nobh_prepare_write);
2398
Dave Kleikamp57bf63d2007-03-06 01:42:12 -08002399/*
2400 * Make sure any changes to nobh_commit_write() are reflected in
2401 * nobh_truncate_page(), since it doesn't call commit_write().
2402 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403int nobh_commit_write(struct file *file, struct page *page,
2404 unsigned from, unsigned to)
2405{
2406 struct inode *inode = page->mapping->host;
2407 loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2408
Nick Piggin22c8ca72007-02-20 13:58:09 -08002409 SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 set_page_dirty(page);
2411 if (pos > inode->i_size) {
2412 i_size_write(inode, pos);
2413 mark_inode_dirty(inode);
2414 }
2415 return 0;
2416}
2417EXPORT_SYMBOL(nobh_commit_write);
2418
2419/*
2420 * nobh_writepage() - based on block_full_write_page() except
2421 * that it tries to operate without attaching bufferheads to
2422 * the page.
2423 */
2424int nobh_writepage(struct page *page, get_block_t *get_block,
2425 struct writeback_control *wbc)
2426{
2427 struct inode * const inode = page->mapping->host;
2428 loff_t i_size = i_size_read(inode);
2429 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2430 unsigned offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 int ret;
2432
2433 /* Is the page fully inside i_size? */
2434 if (page->index < end_index)
2435 goto out;
2436
2437 /* Is the page fully outside i_size? (truncate in progress) */
2438 offset = i_size & (PAGE_CACHE_SIZE-1);
2439 if (page->index >= end_index+1 || !offset) {
2440 /*
2441 * The page may have dirty, unmapped buffers. For example,
2442 * they may have been added in ext3_writepage(). Make them
2443 * freeable here, so the page does not leak.
2444 */
2445#if 0
2446 /* Not really sure about this - do we need this ? */
2447 if (page->mapping->a_ops->invalidatepage)
2448 page->mapping->a_ops->invalidatepage(page, offset);
2449#endif
2450 unlock_page(page);
2451 return 0; /* don't care */
2452 }
2453
2454 /*
2455 * The page straddles i_size. It must be zeroed out on each and every
2456 * writepage invocation because it may be mmapped. "A file is mapped
2457 * in multiples of the page size. For a file that is not a multiple of
2458 * the page size, the remaining memory is zeroed when mapped, and
2459 * writes to that region are not written out to the file."
2460 */
Nate Diller01f27052007-05-09 02:35:07 -07002461 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462out:
2463 ret = mpage_writepage(page, get_block, wbc);
2464 if (ret == -EAGAIN)
2465 ret = __block_write_full_page(inode, page, get_block, wbc);
2466 return ret;
2467}
2468EXPORT_SYMBOL(nobh_writepage);
2469
2470/*
2471 * This function assumes that ->prepare_write() uses nobh_prepare_write().
2472 */
2473int nobh_truncate_page(struct address_space *mapping, loff_t from)
2474{
2475 struct inode *inode = mapping->host;
2476 unsigned blocksize = 1 << inode->i_blkbits;
2477 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2478 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2479 unsigned to;
2480 struct page *page;
Christoph Hellwigf5e54d62006-06-28 04:26:44 -07002481 const struct address_space_operations *a_ops = mapping->a_ops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 int ret = 0;
2483
2484 if ((offset & (blocksize - 1)) == 0)
2485 goto out;
2486
2487 ret = -ENOMEM;
2488 page = grab_cache_page(mapping, index);
2489 if (!page)
2490 goto out;
2491
2492 to = (offset + blocksize) & ~(blocksize - 1);
2493 ret = a_ops->prepare_write(NULL, page, offset, to);
2494 if (ret == 0) {
Nate Diller01f27052007-05-09 02:35:07 -07002495 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset,
2496 KM_USER0);
Dave Kleikamp57bf63d2007-03-06 01:42:12 -08002497 /*
2498 * It would be more correct to call aops->commit_write()
2499 * here, but this is more efficient.
2500 */
2501 SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 set_page_dirty(page);
2503 }
2504 unlock_page(page);
2505 page_cache_release(page);
2506out:
2507 return ret;
2508}
2509EXPORT_SYMBOL(nobh_truncate_page);
2510
2511int block_truncate_page(struct address_space *mapping,
2512 loff_t from, get_block_t *get_block)
2513{
2514 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2515 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2516 unsigned blocksize;
Andrew Morton54b21a72006-01-08 01:03:05 -08002517 sector_t iblock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 unsigned length, pos;
2519 struct inode *inode = mapping->host;
2520 struct page *page;
2521 struct buffer_head *bh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 int err;
2523
2524 blocksize = 1 << inode->i_blkbits;
2525 length = offset & (blocksize - 1);
2526
2527 /* Block boundary? Nothing to do */
2528 if (!length)
2529 return 0;
2530
2531 length = blocksize - length;
Andrew Morton54b21a72006-01-08 01:03:05 -08002532 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
2534 page = grab_cache_page(mapping, index);
2535 err = -ENOMEM;
2536 if (!page)
2537 goto out;
2538
2539 if (!page_has_buffers(page))
2540 create_empty_buffers(page, blocksize, 0);
2541
2542 /* Find the buffer that contains "offset" */
2543 bh = page_buffers(page);
2544 pos = blocksize;
2545 while (offset >= pos) {
2546 bh = bh->b_this_page;
2547 iblock++;
2548 pos += blocksize;
2549 }
2550
2551 err = 0;
2552 if (!buffer_mapped(bh)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08002553 WARN_ON(bh->b_size != blocksize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 err = get_block(inode, iblock, bh, 0);
2555 if (err)
2556 goto unlock;
2557 /* unmapped? It's a hole - nothing to do */
2558 if (!buffer_mapped(bh))
2559 goto unlock;
2560 }
2561
2562 /* Ok, it's mapped. Make sure it's up-to-date */
2563 if (PageUptodate(page))
2564 set_buffer_uptodate(bh);
2565
David Chinner33a266d2007-02-12 00:51:41 -08002566 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 err = -EIO;
2568 ll_rw_block(READ, 1, &bh);
2569 wait_on_buffer(bh);
2570 /* Uhhuh. Read error. Complain and punt. */
2571 if (!buffer_uptodate(bh))
2572 goto unlock;
2573 }
2574
Nate Diller01f27052007-05-09 02:35:07 -07002575 zero_user_page(page, offset, length, KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 mark_buffer_dirty(bh);
2577 err = 0;
2578
2579unlock:
2580 unlock_page(page);
2581 page_cache_release(page);
2582out:
2583 return err;
2584}
2585
2586/*
2587 * The generic ->writepage function for buffer-backed address_spaces
2588 */
2589int block_write_full_page(struct page *page, get_block_t *get_block,
2590 struct writeback_control *wbc)
2591{
2592 struct inode * const inode = page->mapping->host;
2593 loff_t i_size = i_size_read(inode);
2594 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2595 unsigned offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596
2597 /* Is the page fully inside i_size? */
2598 if (page->index < end_index)
2599 return __block_write_full_page(inode, page, get_block, wbc);
2600
2601 /* Is the page fully outside i_size? (truncate in progress) */
2602 offset = i_size & (PAGE_CACHE_SIZE-1);
2603 if (page->index >= end_index+1 || !offset) {
2604 /*
2605 * The page may have dirty, unmapped buffers. For example,
2606 * they may have been added in ext3_writepage(). Make them
2607 * freeable here, so the page does not leak.
2608 */
Jan Karaaaa40592005-10-30 15:00:16 -08002609 do_invalidatepage(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 unlock_page(page);
2611 return 0; /* don't care */
2612 }
2613
2614 /*
2615 * The page straddles i_size. It must be zeroed out on each and every
2616 * writepage invokation because it may be mmapped. "A file is mapped
2617 * in multiples of the page size. For a file that is not a multiple of
2618 * the page size, the remaining memory is zeroed when mapped, and
2619 * writes to that region are not written out to the file."
2620 */
Nate Diller01f27052007-05-09 02:35:07 -07002621 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 return __block_write_full_page(inode, page, get_block, wbc);
2623}
2624
2625sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2626 get_block_t *get_block)
2627{
2628 struct buffer_head tmp;
2629 struct inode *inode = mapping->host;
2630 tmp.b_state = 0;
2631 tmp.b_blocknr = 0;
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08002632 tmp.b_size = 1 << inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 get_block(inode, block, &tmp, 0);
2634 return tmp.b_blocknr;
2635}
2636
NeilBrown6712ecf2007-09-27 12:47:43 +02002637static void end_bio_bh_io_sync(struct bio *bio, int err)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638{
2639 struct buffer_head *bh = bio->bi_private;
2640
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 if (err == -EOPNOTSUPP) {
2642 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2643 set_bit(BH_Eopnotsupp, &bh->b_state);
2644 }
2645
2646 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2647 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648}
2649
2650int submit_bh(int rw, struct buffer_head * bh)
2651{
2652 struct bio *bio;
2653 int ret = 0;
2654
2655 BUG_ON(!buffer_locked(bh));
2656 BUG_ON(!buffer_mapped(bh));
2657 BUG_ON(!bh->b_end_io);
2658
2659 if (buffer_ordered(bh) && (rw == WRITE))
2660 rw = WRITE_BARRIER;
2661
2662 /*
2663 * Only clear out a write error when rewriting, should this
2664 * include WRITE_SYNC as well?
2665 */
2666 if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2667 clear_buffer_write_io_error(bh);
2668
2669 /*
2670 * from here on down, it's all bio -- do the initial mapping,
2671 * submit_bio -> generic_make_request may further map this bio around
2672 */
2673 bio = bio_alloc(GFP_NOIO, 1);
2674
2675 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2676 bio->bi_bdev = bh->b_bdev;
2677 bio->bi_io_vec[0].bv_page = bh->b_page;
2678 bio->bi_io_vec[0].bv_len = bh->b_size;
2679 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2680
2681 bio->bi_vcnt = 1;
2682 bio->bi_idx = 0;
2683 bio->bi_size = bh->b_size;
2684
2685 bio->bi_end_io = end_bio_bh_io_sync;
2686 bio->bi_private = bh;
2687
2688 bio_get(bio);
2689 submit_bio(rw, bio);
2690
2691 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2692 ret = -EOPNOTSUPP;
2693
2694 bio_put(bio);
2695 return ret;
2696}
2697
2698/**
2699 * ll_rw_block: low-level access to block devices (DEPRECATED)
Jan Karaa7662232005-09-06 15:19:10 -07002700 * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 * @nr: number of &struct buffer_heads in the array
2702 * @bhs: array of pointers to &struct buffer_head
2703 *
Jan Karaa7662232005-09-06 15:19:10 -07002704 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2705 * requests an I/O operation on them, either a %READ or a %WRITE. The third
2706 * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2707 * are sent to disk. The fourth %READA option is described in the documentation
2708 * for generic_make_request() which ll_rw_block() calls.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 *
2710 * This function drops any buffer that it cannot get a lock on (with the
Jan Karaa7662232005-09-06 15:19:10 -07002711 * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2712 * clean when doing a write request, and any buffer that appears to be
2713 * up-to-date when doing read request. Further it marks as clean buffers that
2714 * are processed for writing (the buffer cache won't assume that they are
2715 * actually clean until the buffer gets unlocked).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 *
2717 * ll_rw_block sets b_end_io to simple completion handler that marks
2718 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2719 * any waiters.
2720 *
2721 * All of the buffers must be for the same device, and must also be a
2722 * multiple of the current approved size for the device.
2723 */
2724void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2725{
2726 int i;
2727
2728 for (i = 0; i < nr; i++) {
2729 struct buffer_head *bh = bhs[i];
2730
Jan Karaa7662232005-09-06 15:19:10 -07002731 if (rw == SWRITE)
2732 lock_buffer(bh);
2733 else if (test_set_buffer_locked(bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 continue;
2735
Jan Karaa7662232005-09-06 15:19:10 -07002736 if (rw == WRITE || rw == SWRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 if (test_clear_buffer_dirty(bh)) {
akpm@osdl.org76c30732005-04-16 15:24:07 -07002738 bh->b_end_io = end_buffer_write_sync;
OGAWA Hirofumie60e5c52006-02-03 03:04:43 -08002739 get_bh(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 submit_bh(WRITE, bh);
2741 continue;
2742 }
2743 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 if (!buffer_uptodate(bh)) {
akpm@osdl.org76c30732005-04-16 15:24:07 -07002745 bh->b_end_io = end_buffer_read_sync;
OGAWA Hirofumie60e5c52006-02-03 03:04:43 -08002746 get_bh(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 submit_bh(rw, bh);
2748 continue;
2749 }
2750 }
2751 unlock_buffer(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 }
2753}
2754
2755/*
2756 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2757 * and then start new I/O and then wait upon it. The caller must have a ref on
2758 * the buffer_head.
2759 */
2760int sync_dirty_buffer(struct buffer_head *bh)
2761{
2762 int ret = 0;
2763
2764 WARN_ON(atomic_read(&bh->b_count) < 1);
2765 lock_buffer(bh);
2766 if (test_clear_buffer_dirty(bh)) {
2767 get_bh(bh);
2768 bh->b_end_io = end_buffer_write_sync;
2769 ret = submit_bh(WRITE, bh);
2770 wait_on_buffer(bh);
2771 if (buffer_eopnotsupp(bh)) {
2772 clear_buffer_eopnotsupp(bh);
2773 ret = -EOPNOTSUPP;
2774 }
2775 if (!ret && !buffer_uptodate(bh))
2776 ret = -EIO;
2777 } else {
2778 unlock_buffer(bh);
2779 }
2780 return ret;
2781}
2782
2783/*
2784 * try_to_free_buffers() checks if all the buffers on this particular page
2785 * are unused, and releases them if so.
2786 *
2787 * Exclusion against try_to_free_buffers may be obtained by either
2788 * locking the page or by holding its mapping's private_lock.
2789 *
2790 * If the page is dirty but all the buffers are clean then we need to
2791 * be sure to mark the page clean as well. This is because the page
2792 * may be against a block device, and a later reattachment of buffers
2793 * to a dirty page will set *all* buffers dirty. Which would corrupt
2794 * filesystem data on the same device.
2795 *
2796 * The same applies to regular filesystem pages: if all the buffers are
2797 * clean then we set the page clean and proceed. To do that, we require
2798 * total exclusion from __set_page_dirty_buffers(). That is obtained with
2799 * private_lock.
2800 *
2801 * try_to_free_buffers() is non-blocking.
2802 */
2803static inline int buffer_busy(struct buffer_head *bh)
2804{
2805 return atomic_read(&bh->b_count) |
2806 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2807}
2808
2809static int
2810drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2811{
2812 struct buffer_head *head = page_buffers(page);
2813 struct buffer_head *bh;
2814
2815 bh = head;
2816 do {
akpm@osdl.orgde7d5a32005-05-01 08:58:39 -07002817 if (buffer_write_io_error(bh) && page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 set_bit(AS_EIO, &page->mapping->flags);
2819 if (buffer_busy(bh))
2820 goto failed;
2821 bh = bh->b_this_page;
2822 } while (bh != head);
2823
2824 do {
2825 struct buffer_head *next = bh->b_this_page;
2826
2827 if (!list_empty(&bh->b_assoc_buffers))
2828 __remove_assoc_queue(bh);
2829 bh = next;
2830 } while (bh != head);
2831 *buffers_to_free = head;
2832 __clear_page_buffers(page);
2833 return 1;
2834failed:
2835 return 0;
2836}
2837
2838int try_to_free_buffers(struct page *page)
2839{
2840 struct address_space * const mapping = page->mapping;
2841 struct buffer_head *buffers_to_free = NULL;
2842 int ret = 0;
2843
2844 BUG_ON(!PageLocked(page));
Linus Torvaldsecdfc972007-01-26 12:47:06 -08002845 if (PageWriteback(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 return 0;
2847
2848 if (mapping == NULL) { /* can this still happen? */
2849 ret = drop_buffers(page, &buffers_to_free);
2850 goto out;
2851 }
2852
2853 spin_lock(&mapping->private_lock);
2854 ret = drop_buffers(page, &buffers_to_free);
Linus Torvaldsecdfc972007-01-26 12:47:06 -08002855
2856 /*
2857 * If the filesystem writes its buffers by hand (eg ext3)
2858 * then we can have clean buffers against a dirty page. We
2859 * clean the page here; otherwise the VM will never notice
2860 * that the filesystem did any IO at all.
2861 *
2862 * Also, during truncate, discard_buffer will have marked all
2863 * the page's buffers clean. We discover that here and clean
2864 * the page also.
Nick Piggin87df7242007-01-30 14:36:27 +11002865 *
2866 * private_lock must be held over this entire operation in order
2867 * to synchronise against __set_page_dirty_buffers and prevent the
2868 * dirty bit from being lost.
Linus Torvaldsecdfc972007-01-26 12:47:06 -08002869 */
2870 if (ret)
2871 cancel_dirty_page(page, PAGE_CACHE_SIZE);
Nick Piggin87df7242007-01-30 14:36:27 +11002872 spin_unlock(&mapping->private_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873out:
2874 if (buffers_to_free) {
2875 struct buffer_head *bh = buffers_to_free;
2876
2877 do {
2878 struct buffer_head *next = bh->b_this_page;
2879 free_buffer_head(bh);
2880 bh = next;
2881 } while (bh != buffers_to_free);
2882 }
2883 return ret;
2884}
2885EXPORT_SYMBOL(try_to_free_buffers);
2886
NeilBrown3978d712006-03-26 01:37:17 -08002887void block_sync_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888{
2889 struct address_space *mapping;
2890
2891 smp_mb();
2892 mapping = page_mapping(page);
2893 if (mapping)
2894 blk_run_backing_dev(mapping->backing_dev_info, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895}
2896
2897/*
2898 * There are no bdflush tunables left. But distributions are
2899 * still running obsolete flush daemons, so we terminate them here.
2900 *
2901 * Use of bdflush() is deprecated and will be removed in a future kernel.
2902 * The `pdflush' kernel threads fully replace bdflush daemons and this call.
2903 */
2904asmlinkage long sys_bdflush(int func, long data)
2905{
2906 static int msg_count;
2907
2908 if (!capable(CAP_SYS_ADMIN))
2909 return -EPERM;
2910
2911 if (msg_count < 5) {
2912 msg_count++;
2913 printk(KERN_INFO
2914 "warning: process `%s' used the obsolete bdflush"
2915 " system call\n", current->comm);
2916 printk(KERN_INFO "Fix your initscripts?\n");
2917 }
2918
2919 if (func == 1)
2920 do_exit(0);
2921 return 0;
2922}
2923
2924/*
2925 * Buffer-head allocation
2926 */
Christoph Lametere18b8902006-12-06 20:33:20 -08002927static struct kmem_cache *bh_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928
2929/*
2930 * Once the number of bh's in the machine exceeds this level, we start
2931 * stripping them in writeback.
2932 */
2933static int max_buffer_heads;
2934
2935int buffer_heads_over_limit;
2936
2937struct bh_accounting {
2938 int nr; /* Number of live bh's */
2939 int ratelimit; /* Limit cacheline bouncing */
2940};
2941
2942static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2943
2944static void recalc_bh_state(void)
2945{
2946 int i;
2947 int tot = 0;
2948
2949 if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
2950 return;
2951 __get_cpu_var(bh_accounting).ratelimit = 0;
Eric Dumazet8a143422006-03-24 03:18:10 -08002952 for_each_online_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 tot += per_cpu(bh_accounting, i).nr;
2954 buffer_heads_over_limit = (tot > max_buffer_heads);
2955}
2956
Al Virodd0fc662005-10-07 07:46:04 +01002957struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958{
Christoph Lametera35afb82007-05-16 22:10:57 -07002959 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 if (ret) {
Christoph Lametera35afb82007-05-16 22:10:57 -07002961 INIT_LIST_HEAD(&ret->b_assoc_buffers);
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07002962 get_cpu_var(bh_accounting).nr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 recalc_bh_state();
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07002964 put_cpu_var(bh_accounting);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 }
2966 return ret;
2967}
2968EXPORT_SYMBOL(alloc_buffer_head);
2969
2970void free_buffer_head(struct buffer_head *bh)
2971{
2972 BUG_ON(!list_empty(&bh->b_assoc_buffers));
2973 kmem_cache_free(bh_cachep, bh);
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07002974 get_cpu_var(bh_accounting).nr--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 recalc_bh_state();
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07002976 put_cpu_var(bh_accounting);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977}
2978EXPORT_SYMBOL(free_buffer_head);
2979
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980static void buffer_exit_cpu(int cpu)
2981{
2982 int i;
2983 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
2984
2985 for (i = 0; i < BH_LRU_SIZE; i++) {
2986 brelse(b->bhs[i]);
2987 b->bhs[i] = NULL;
2988 }
Eric Dumazet8a143422006-03-24 03:18:10 -08002989 get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
2990 per_cpu(bh_accounting, cpu).nr = 0;
2991 put_cpu_var(bh_accounting);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992}
2993
2994static int buffer_cpu_notify(struct notifier_block *self,
2995 unsigned long action, void *hcpu)
2996{
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002997 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 buffer_exit_cpu((unsigned long)hcpu);
2999 return NOTIFY_OK;
3000}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001
3002void __init buffer_init(void)
3003{
3004 int nrpages;
3005
Christoph Lametera35afb82007-05-16 22:10:57 -07003006 bh_cachep = KMEM_CACHE(buffer_head,
3007 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008
3009 /*
3010 * Limit the bh occupancy to 10% of ZONE_NORMAL
3011 */
3012 nrpages = (nr_free_buffer_pages() * 10) / 100;
3013 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3014 hotcpu_notifier(buffer_cpu_notify, 0);
3015}
3016
3017EXPORT_SYMBOL(__bforget);
3018EXPORT_SYMBOL(__brelse);
3019EXPORT_SYMBOL(__wait_on_buffer);
3020EXPORT_SYMBOL(block_commit_write);
3021EXPORT_SYMBOL(block_prepare_write);
David Chinner54171692007-07-19 17:39:55 +10003022EXPORT_SYMBOL(block_page_mkwrite);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023EXPORT_SYMBOL(block_read_full_page);
3024EXPORT_SYMBOL(block_sync_page);
3025EXPORT_SYMBOL(block_truncate_page);
3026EXPORT_SYMBOL(block_write_full_page);
3027EXPORT_SYMBOL(cont_prepare_write);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028EXPORT_SYMBOL(end_buffer_read_sync);
3029EXPORT_SYMBOL(end_buffer_write_sync);
3030EXPORT_SYMBOL(file_fsync);
3031EXPORT_SYMBOL(fsync_bdev);
3032EXPORT_SYMBOL(generic_block_bmap);
3033EXPORT_SYMBOL(generic_commit_write);
3034EXPORT_SYMBOL(generic_cont_expand);
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08003035EXPORT_SYMBOL(generic_cont_expand_simple);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036EXPORT_SYMBOL(init_buffer);
3037EXPORT_SYMBOL(invalidate_bdev);
3038EXPORT_SYMBOL(ll_rw_block);
3039EXPORT_SYMBOL(mark_buffer_dirty);
3040EXPORT_SYMBOL(submit_bh);
3041EXPORT_SYMBOL(sync_dirty_buffer);
3042EXPORT_SYMBOL(unlock_buffer);