blob: a073f3f4f013c5c349afede7edb715672bfd4928 [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
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -080070void __lock_buffer(struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -070071{
72 wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73 TASK_UNINTERRUPTIBLE);
74}
75EXPORT_SYMBOL(__lock_buffer);
76
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -080077void unlock_buffer(struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -070078{
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/*
Dmitry Monakhov68671f32007-10-16 01:24:47 -0700113 * End-of-IO handler helper function which does not touch the bh after
114 * unlocking it.
115 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116 * a race there is benign: unlock_buffer() only use the bh's address for
117 * hashing after unlocking the buffer, so it doesn't actually touch the bh
118 * itself.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 */
Dmitry Monakhov68671f32007-10-16 01:24:47 -0700120static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121{
122 if (uptodate) {
123 set_buffer_uptodate(bh);
124 } else {
125 /* This happens, due to failed READA attempts. */
126 clear_buffer_uptodate(bh);
127 }
128 unlock_buffer(bh);
Dmitry Monakhov68671f32007-10-16 01:24:47 -0700129}
130
131/*
132 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
133 * unlock the buffer. This is what ll_rw_block uses too.
134 */
135void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
136{
137 __end_buffer_read_notouch(bh, uptodate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 put_bh(bh);
139}
140
141void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
142{
143 char b[BDEVNAME_SIZE];
144
145 if (uptodate) {
146 set_buffer_uptodate(bh);
147 } else {
148 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
149 buffer_io_error(bh);
150 printk(KERN_WARNING "lost page write due to "
151 "I/O error on %s\n",
152 bdevname(bh->b_bdev, b));
153 }
154 set_buffer_write_io_error(bh);
155 clear_buffer_uptodate(bh);
156 }
157 unlock_buffer(bh);
158 put_bh(bh);
159}
160
161/*
162 * Write out and wait upon all the dirty data associated with a block
163 * device via its mapping. Does not take the superblock lock.
164 */
165int sync_blockdev(struct block_device *bdev)
166{
167 int ret = 0;
168
OGAWA Hirofumi28fd1292006-01-08 01:02:14 -0800169 if (bdev)
170 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 return ret;
172}
173EXPORT_SYMBOL(sync_blockdev);
174
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175/*
176 * Write out and wait upon all dirty data associated with this
177 * device. Filesystem data as well as the underlying block
178 * device. Takes the superblock lock.
179 */
180int fsync_bdev(struct block_device *bdev)
181{
182 struct super_block *sb = get_super(bdev);
183 if (sb) {
184 int res = fsync_super(sb);
185 drop_super(sb);
186 return res;
187 }
188 return sync_blockdev(bdev);
189}
190
191/**
192 * freeze_bdev -- lock a filesystem and force it into a consistent state
193 * @bdev: blockdevice to lock
194 *
David Chinnerf73ca1b2007-01-10 23:15:41 -0800195 * This takes the block device bd_mount_sem to make sure no new mounts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196 * happen on bdev until thaw_bdev() is called.
197 * If a superblock is found on this device, we take the s_umount semaphore
198 * on it to make sure nobody unmounts until the snapshot creation is done.
199 */
200struct super_block *freeze_bdev(struct block_device *bdev)
201{
202 struct super_block *sb;
203
David Chinnerf73ca1b2007-01-10 23:15:41 -0800204 down(&bdev->bd_mount_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 sb = get_super(bdev);
206 if (sb && !(sb->s_flags & MS_RDONLY)) {
207 sb->s_frozen = SB_FREEZE_WRITE;
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700208 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209
OGAWA Hirofumid25b9a12006-03-25 03:07:44 -0800210 __fsync_super(sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
212 sb->s_frozen = SB_FREEZE_TRANS;
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700213 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
215 sync_blockdev(sb->s_bdev);
216
217 if (sb->s_op->write_super_lockfs)
218 sb->s_op->write_super_lockfs(sb);
219 }
220
221 sync_blockdev(bdev);
222 return sb; /* thaw_bdev releases s->s_umount and bd_mount_sem */
223}
224EXPORT_SYMBOL(freeze_bdev);
225
226/**
227 * thaw_bdev -- unlock filesystem
228 * @bdev: blockdevice to unlock
229 * @sb: associated superblock
230 *
231 * Unlocks the filesystem and marks it writeable again after freeze_bdev().
232 */
233void thaw_bdev(struct block_device *bdev, struct super_block *sb)
234{
235 if (sb) {
236 BUG_ON(sb->s_bdev != bdev);
237
238 if (sb->s_op->unlockfs)
239 sb->s_op->unlockfs(sb);
240 sb->s_frozen = SB_UNFROZEN;
akpm@osdl.orgd59dd462005-05-01 08:58:47 -0700241 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 wake_up(&sb->s_wait_unfrozen);
243 drop_super(sb);
244 }
245
David Chinnerf73ca1b2007-01-10 23:15:41 -0800246 up(&bdev->bd_mount_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247}
248EXPORT_SYMBOL(thaw_bdev);
249
250/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 * Various filesystems appear to want __find_get_block to be non-blocking.
252 * But it's the page lock which protects the buffers. To get around this,
253 * we get exclusion from try_to_free_buffers with the blockdev mapping's
254 * private_lock.
255 *
256 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257 * may be quite high. This code could TryLock the page, and if that
258 * succeeds, there is no need to take private_lock. (But if
259 * private_lock is contended then so is mapping->tree_lock).
260 */
261static struct buffer_head *
Coywolf Qi Hunt385fd4c2005-11-07 00:59:39 -0800262__find_get_block_slow(struct block_device *bdev, sector_t block)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263{
264 struct inode *bd_inode = bdev->bd_inode;
265 struct address_space *bd_mapping = bd_inode->i_mapping;
266 struct buffer_head *ret = NULL;
267 pgoff_t index;
268 struct buffer_head *bh;
269 struct buffer_head *head;
270 struct page *page;
271 int all_mapped = 1;
272
273 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274 page = find_get_page(bd_mapping, index);
275 if (!page)
276 goto out;
277
278 spin_lock(&bd_mapping->private_lock);
279 if (!page_has_buffers(page))
280 goto out_unlock;
281 head = page_buffers(page);
282 bh = head;
283 do {
284 if (bh->b_blocknr == block) {
285 ret = bh;
286 get_bh(bh);
287 goto out_unlock;
288 }
289 if (!buffer_mapped(bh))
290 all_mapped = 0;
291 bh = bh->b_this_page;
292 } while (bh != head);
293
294 /* we might be here because some of the buffers on this page are
295 * not mapped. This is due to various races between
296 * file io on the block device and getblk. It gets dealt with
297 * elsewhere, don't buffer_error if we had some unmapped buffers
298 */
299 if (all_mapped) {
300 printk("__find_get_block_slow() failed. "
301 "block=%llu, b_blocknr=%llu\n",
Badari Pulavarty205f87f2006-03-26 01:38:00 -0800302 (unsigned long long)block,
303 (unsigned long long)bh->b_blocknr);
304 printk("b_state=0x%08lx, b_size=%zu\n",
305 bh->b_state, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
307 }
308out_unlock:
309 spin_unlock(&bd_mapping->private_lock);
310 page_cache_release(page);
311out:
312 return ret;
313}
314
315/* If invalidate_buffers() will trash dirty buffers, it means some kind
316 of fs corruption is going on. Trashing dirty data always imply losing
317 information that was supposed to be just stored on the physical layer
318 by the user.
319
320 Thus invalidate_buffers in general usage is not allwowed to trash
321 dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322 be preserved. These buffers are simply skipped.
323
324 We also skip buffers which are still in use. For example this can
325 happen if a userspace program is reading the block device.
326
327 NOTE: In the case where the user removed a removable-media-disk even if
328 there's still dirty data not synced on disk (due a bug in the device driver
329 or due an error of the user), by not destroying the dirty buffers we could
330 generate corruption also on the next media inserted, thus a parameter is
331 necessary to handle this case in the most safe way possible (trying
332 to not corrupt also the new disk inserted with the data belonging to
333 the old now corrupted disk). Also for the ramdisk the natural thing
334 to do in order to release the ramdisk memory is to destroy dirty buffers.
335
336 These are two special cases. Normal usage imply the device driver
337 to issue a sync on the device (without waiting I/O completion) and
338 then an invalidate_buffers call that doesn't trash dirty buffers.
339
340 For handling cache coherency with the blkdev pagecache the 'update' case
341 is been introduced. It is needed to re-read from disk any pinned
342 buffer. NOTE: re-reading from disk is destructive so we can do it only
343 when we assume nobody is changing the buffercache under our I/O and when
344 we think the disk contains more recent information than the buffercache.
345 The update == 1 pass marks the buffers we need to update, the update == 2
346 pass does the actual I/O. */
Peter Zijlstraf98393a2007-05-06 14:49:54 -0700347void invalidate_bdev(struct block_device *bdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348{
Andrew Morton0e1dfc62006-07-30 03:03:28 -0700349 struct address_space *mapping = bdev->bd_inode->i_mapping;
350
351 if (mapping->nrpages == 0)
352 return;
353
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 invalidate_bh_lrus();
Andrew Mortonfc0ecff2007-02-10 01:45:39 -0800355 invalidate_mapping_pages(mapping, 0, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356}
357
358/*
359 * Kick pdflush then try to free up some ZONE_NORMAL memory.
360 */
361static void free_more_memory(void)
362{
Mel Gorman19770b32008-04-28 02:12:18 -0700363 struct zone *zone;
Mel Gorman0e884602008-04-28 02:12:14 -0700364 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
Pekka J Enberg687a21c2005-06-28 20:44:55 -0700366 wakeup_pdflush(1024);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 yield();
368
Mel Gorman0e884602008-04-28 02:12:14 -0700369 for_each_online_node(nid) {
Mel Gorman19770b32008-04-28 02:12:18 -0700370 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
371 gfp_zone(GFP_NOFS), NULL,
372 &zone);
373 if (zone)
Mel Gorman54a6eb52008-04-28 02:12:16 -0700374 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
375 GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 }
377}
378
379/*
380 * I/O completion handler for block_read_full_page() - pages
381 * which come unlocked at the end of I/O.
382 */
383static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
384{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 unsigned long flags;
Nick Piggina3972202005-07-07 17:56:56 -0700386 struct buffer_head *first;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 struct buffer_head *tmp;
388 struct page *page;
389 int page_uptodate = 1;
390
391 BUG_ON(!buffer_async_read(bh));
392
393 page = bh->b_page;
394 if (uptodate) {
395 set_buffer_uptodate(bh);
396 } else {
397 clear_buffer_uptodate(bh);
398 if (printk_ratelimit())
399 buffer_io_error(bh);
400 SetPageError(page);
401 }
402
403 /*
404 * Be _very_ careful from here on. Bad things can happen if
405 * two buffer heads end IO at almost the same time and both
406 * decide that the page is now completely done.
407 */
Nick Piggina3972202005-07-07 17:56:56 -0700408 first = page_buffers(page);
409 local_irq_save(flags);
410 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 clear_buffer_async_read(bh);
412 unlock_buffer(bh);
413 tmp = bh;
414 do {
415 if (!buffer_uptodate(tmp))
416 page_uptodate = 0;
417 if (buffer_async_read(tmp)) {
418 BUG_ON(!buffer_locked(tmp));
419 goto still_busy;
420 }
421 tmp = tmp->b_this_page;
422 } while (tmp != bh);
Nick Piggina3972202005-07-07 17:56:56 -0700423 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
424 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
426 /*
427 * If none of the buffers had errors and they are all
428 * uptodate then we can set the page uptodate.
429 */
430 if (page_uptodate && !PageError(page))
431 SetPageUptodate(page);
432 unlock_page(page);
433 return;
434
435still_busy:
Nick Piggina3972202005-07-07 17:56:56 -0700436 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
437 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 return;
439}
440
441/*
442 * Completion handler for block_write_full_page() - pages which are unlocked
443 * during I/O, and which have PageWriteback cleared upon I/O completion.
444 */
Adrian Bunkb6cd0b72006-06-27 02:53:54 -0700445static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446{
447 char b[BDEVNAME_SIZE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 unsigned long flags;
Nick Piggina3972202005-07-07 17:56:56 -0700449 struct buffer_head *first;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 struct buffer_head *tmp;
451 struct page *page;
452
453 BUG_ON(!buffer_async_write(bh));
454
455 page = bh->b_page;
456 if (uptodate) {
457 set_buffer_uptodate(bh);
458 } else {
459 if (printk_ratelimit()) {
460 buffer_io_error(bh);
461 printk(KERN_WARNING "lost page write due to "
462 "I/O error on %s\n",
463 bdevname(bh->b_bdev, b));
464 }
465 set_bit(AS_EIO, &page->mapping->flags);
Jan Kara58ff4072006-10-17 00:10:19 -0700466 set_buffer_write_io_error(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 clear_buffer_uptodate(bh);
468 SetPageError(page);
469 }
470
Nick Piggina3972202005-07-07 17:56:56 -0700471 first = page_buffers(page);
472 local_irq_save(flags);
473 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
474
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 clear_buffer_async_write(bh);
476 unlock_buffer(bh);
477 tmp = bh->b_this_page;
478 while (tmp != bh) {
479 if (buffer_async_write(tmp)) {
480 BUG_ON(!buffer_locked(tmp));
481 goto still_busy;
482 }
483 tmp = tmp->b_this_page;
484 }
Nick Piggina3972202005-07-07 17:56:56 -0700485 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
486 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 end_page_writeback(page);
488 return;
489
490still_busy:
Nick Piggina3972202005-07-07 17:56:56 -0700491 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
492 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 return;
494}
495
496/*
497 * If a page's buffers are under async readin (end_buffer_async_read
498 * completion) then there is a possibility that another thread of
499 * control could lock one of the buffers after it has completed
500 * but while some of the other buffers have not completed. This
501 * locked buffer would confuse end_buffer_async_read() into not unlocking
502 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
503 * that this buffer is not under async I/O.
504 *
505 * The page comes unlocked when it has no locked buffer_async buffers
506 * left.
507 *
508 * PageLocked prevents anyone starting new async I/O reads any of
509 * the buffers.
510 *
511 * PageWriteback is used to prevent simultaneous writeout of the same
512 * page.
513 *
514 * PageLocked prevents anyone from starting writeback of a page which is
515 * under read I/O (PageWriteback is only ever set against a locked page).
516 */
517static void mark_buffer_async_read(struct buffer_head *bh)
518{
519 bh->b_end_io = end_buffer_async_read;
520 set_buffer_async_read(bh);
521}
522
523void mark_buffer_async_write(struct buffer_head *bh)
524{
525 bh->b_end_io = end_buffer_async_write;
526 set_buffer_async_write(bh);
527}
528EXPORT_SYMBOL(mark_buffer_async_write);
529
530
531/*
532 * fs/buffer.c contains helper functions for buffer-backed address space's
533 * fsync functions. A common requirement for buffer-based filesystems is
534 * that certain data from the backing blockdev needs to be written out for
535 * a successful fsync(). For example, ext2 indirect blocks need to be
536 * written back and waited upon before fsync() returns.
537 *
538 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
539 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
540 * management of a list of dependent buffers at ->i_mapping->private_list.
541 *
542 * Locking is a little subtle: try_to_free_buffers() will remove buffers
543 * from their controlling inode's queue when they are being freed. But
544 * try_to_free_buffers() will be operating against the *blockdev* mapping
545 * at the time, not against the S_ISREG file which depends on those buffers.
546 * So the locking for private_list is via the private_lock in the address_space
547 * which backs the buffers. Which is different from the address_space
548 * against which the buffers are listed. So for a particular address_space,
549 * mapping->private_lock does *not* protect mapping->private_list! In fact,
550 * mapping->private_list will always be protected by the backing blockdev's
551 * ->private_lock.
552 *
553 * Which introduces a requirement: all buffers on an address_space's
554 * ->private_list must be from the same address_space: the blockdev's.
555 *
556 * address_spaces which do not place buffers at ->private_list via these
557 * utility functions are free to use private_lock and private_list for
558 * whatever they want. The only requirement is that list_empty(private_list)
559 * be true at clear_inode() time.
560 *
561 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
562 * filesystems should do that. invalidate_inode_buffers() should just go
563 * BUG_ON(!list_empty).
564 *
565 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
566 * take an address_space, not an inode. And it should be called
567 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
568 * queued up.
569 *
570 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
571 * list if it is already on a list. Because if the buffer is on a list,
572 * it *must* already be on the right one. If not, the filesystem is being
573 * silly. This will save a ton of locking. But first we have to ensure
574 * that buffers are taken *off* the old inode's list when they are freed
575 * (presumably in truncate). That requires careful auditing of all
576 * filesystems (do it inside bforget()). It could also be done by bringing
577 * b_inode back.
578 */
579
580/*
581 * The buffer's backing address_space's private_lock must be held
582 */
583static inline void __remove_assoc_queue(struct buffer_head *bh)
584{
585 list_del_init(&bh->b_assoc_buffers);
Jan Kara58ff4072006-10-17 00:10:19 -0700586 WARN_ON(!bh->b_assoc_map);
587 if (buffer_write_io_error(bh))
588 set_bit(AS_EIO, &bh->b_assoc_map->flags);
589 bh->b_assoc_map = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590}
591
592int inode_has_buffers(struct inode *inode)
593{
594 return !list_empty(&inode->i_data.private_list);
595}
596
597/*
598 * osync is designed to support O_SYNC io. It waits synchronously for
599 * all already-submitted IO to complete, but does not queue any new
600 * writes to the disk.
601 *
602 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
603 * you dirty the buffers, and then use osync_inode_buffers to wait for
604 * completion. Any other dirty buffers which are not yet queued for
605 * write will not be flushed to disk by the osync.
606 */
607static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
608{
609 struct buffer_head *bh;
610 struct list_head *p;
611 int err = 0;
612
613 spin_lock(lock);
614repeat:
615 list_for_each_prev(p, list) {
616 bh = BH_ENTRY(p);
617 if (buffer_locked(bh)) {
618 get_bh(bh);
619 spin_unlock(lock);
620 wait_on_buffer(bh);
621 if (!buffer_uptodate(bh))
622 err = -EIO;
623 brelse(bh);
624 spin_lock(lock);
625 goto repeat;
626 }
627 }
628 spin_unlock(lock);
629 return err;
630}
631
632/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800633 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
Martin Waitz67be2dd2005-05-01 08:59:26 -0700634 * @mapping: the mapping which wants those buffers written
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 *
636 * Starts I/O against the buffers at mapping->private_list, and waits upon
637 * that I/O.
638 *
Martin Waitz67be2dd2005-05-01 08:59:26 -0700639 * Basically, this is a convenience function for fsync().
640 * @mapping is a file or directory which needs those buffers to be written for
641 * a successful fsync().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 */
643int sync_mapping_buffers(struct address_space *mapping)
644{
645 struct address_space *buffer_mapping = mapping->assoc_mapping;
646
647 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
648 return 0;
649
650 return fsync_buffers_list(&buffer_mapping->private_lock,
651 &mapping->private_list);
652}
653EXPORT_SYMBOL(sync_mapping_buffers);
654
655/*
656 * Called when we've recently written block `bblock', and it is known that
657 * `bblock' was for a buffer_boundary() buffer. This means that the block at
658 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
659 * dirty, schedule it for IO. So that indirects merge nicely with their data.
660 */
661void write_boundary_block(struct block_device *bdev,
662 sector_t bblock, unsigned blocksize)
663{
664 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
665 if (bh) {
666 if (buffer_dirty(bh))
667 ll_rw_block(WRITE, 1, &bh);
668 put_bh(bh);
669 }
670}
671
672void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
673{
674 struct address_space *mapping = inode->i_mapping;
675 struct address_space *buffer_mapping = bh->b_page->mapping;
676
677 mark_buffer_dirty(bh);
678 if (!mapping->assoc_mapping) {
679 mapping->assoc_mapping = buffer_mapping;
680 } else {
Eric Sesterhenne827f922006-03-26 18:24:46 +0200681 BUG_ON(mapping->assoc_mapping != buffer_mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 }
Jan Kara535ee2f2008-02-08 04:21:59 -0800683 if (!bh->b_assoc_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 spin_lock(&buffer_mapping->private_lock);
685 list_move_tail(&bh->b_assoc_buffers,
686 &mapping->private_list);
Jan Kara58ff4072006-10-17 00:10:19 -0700687 bh->b_assoc_map = mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 spin_unlock(&buffer_mapping->private_lock);
689 }
690}
691EXPORT_SYMBOL(mark_buffer_dirty_inode);
692
693/*
Nick Piggin787d2212007-07-17 04:03:34 -0700694 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
695 * dirty.
696 *
697 * If warn is true, then emit a warning if the page is not uptodate and has
698 * not been truncated.
699 */
700static int __set_page_dirty(struct page *page,
701 struct address_space *mapping, int warn)
702{
703 if (unlikely(!mapping))
704 return !TestSetPageDirty(page);
705
706 if (TestSetPageDirty(page))
707 return 0;
708
709 write_lock_irq(&mapping->tree_lock);
710 if (page->mapping) { /* Race with truncate? */
711 WARN_ON_ONCE(warn && !PageUptodate(page));
712
713 if (mapping_cap_account_dirty(mapping)) {
714 __inc_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -0700715 __inc_bdi_stat(mapping->backing_dev_info,
716 BDI_RECLAIMABLE);
Nick Piggin787d2212007-07-17 04:03:34 -0700717 task_io_account_write(PAGE_CACHE_SIZE);
718 }
719 radix_tree_tag_set(&mapping->page_tree,
720 page_index(page), PAGECACHE_TAG_DIRTY);
721 }
722 write_unlock_irq(&mapping->tree_lock);
723 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
724
725 return 1;
726}
727
728/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 * Add a page to the dirty page list.
730 *
731 * It is a sad fact of life that this function is called from several places
732 * deeply under spinlocking. It may not sleep.
733 *
734 * If the page has buffers, the uptodate buffers are set dirty, to preserve
735 * dirty-state coherency between the page and the buffers. It the page does
736 * not have buffers then when they are later attached they will all be set
737 * dirty.
738 *
739 * The buffers are dirtied before the page is dirtied. There's a small race
740 * window in which a writepage caller may see the page cleanness but not the
741 * buffer dirtiness. That's fine. If this code were to set the page dirty
742 * before the buffers, a concurrent writepage caller could clear the page dirty
743 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
744 * page on the dirty page list.
745 *
746 * We use private_lock to lock against try_to_free_buffers while using the
747 * page's buffer list. Also use this to protect against clean buffers being
748 * added to the page after it was set dirty.
749 *
750 * FIXME: may need to call ->reservepage here as well. That's rather up to the
751 * address_space though.
752 */
753int __set_page_dirty_buffers(struct page *page)
754{
Nick Piggin787d2212007-07-17 04:03:34 -0700755 struct address_space *mapping = page_mapping(page);
Nick Pigginebf7a222006-10-10 04:36:54 +0200756
757 if (unlikely(!mapping))
758 return !TestSetPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
760 spin_lock(&mapping->private_lock);
761 if (page_has_buffers(page)) {
762 struct buffer_head *head = page_buffers(page);
763 struct buffer_head *bh = head;
764
765 do {
766 set_buffer_dirty(bh);
767 bh = bh->b_this_page;
768 } while (bh != head);
769 }
770 spin_unlock(&mapping->private_lock);
771
Nick Piggin787d2212007-07-17 04:03:34 -0700772 return __set_page_dirty(page, mapping, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773}
774EXPORT_SYMBOL(__set_page_dirty_buffers);
775
776/*
777 * Write out and wait upon a list of buffers.
778 *
779 * We have conflicting pressures: we want to make sure that all
780 * initially dirty buffers get waited on, but that any subsequently
781 * dirtied buffers don't. After all, we don't want fsync to last
782 * forever if somebody is actively writing to the file.
783 *
784 * Do this in two main stages: first we copy dirty buffers to a
785 * temporary inode list, queueing the writes as we go. Then we clean
786 * up, waiting for those writes to complete.
787 *
788 * During this second stage, any subsequent updates to the file may end
789 * up refiling the buffer on the original inode's dirty list again, so
790 * there is a chance we will end up with a buffer queued for write but
791 * not yet completed on that list. So, as a final cleanup we go through
792 * the osync code to catch these locked, dirty buffers without requeuing
793 * any newly dirty buffers for write.
794 */
795static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
796{
797 struct buffer_head *bh;
798 struct list_head tmp;
Jan Kara535ee2f2008-02-08 04:21:59 -0800799 struct address_space *mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 int err = 0, err2;
801
802 INIT_LIST_HEAD(&tmp);
803
804 spin_lock(lock);
805 while (!list_empty(list)) {
806 bh = BH_ENTRY(list->next);
Jan Kara535ee2f2008-02-08 04:21:59 -0800807 mapping = bh->b_assoc_map;
Jan Kara58ff4072006-10-17 00:10:19 -0700808 __remove_assoc_queue(bh);
Jan Kara535ee2f2008-02-08 04:21:59 -0800809 /* Avoid race with mark_buffer_dirty_inode() which does
810 * a lockless check and we rely on seeing the dirty bit */
811 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 if (buffer_dirty(bh) || buffer_locked(bh)) {
813 list_add(&bh->b_assoc_buffers, &tmp);
Jan Kara535ee2f2008-02-08 04:21:59 -0800814 bh->b_assoc_map = mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 if (buffer_dirty(bh)) {
816 get_bh(bh);
817 spin_unlock(lock);
818 /*
819 * Ensure any pending I/O completes so that
820 * ll_rw_block() actually writes the current
821 * contents - it is a noop if I/O is still in
822 * flight on potentially older contents.
823 */
Jan Karaa7662232005-09-06 15:19:10 -0700824 ll_rw_block(SWRITE, 1, &bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 brelse(bh);
826 spin_lock(lock);
827 }
828 }
829 }
830
831 while (!list_empty(&tmp)) {
832 bh = BH_ENTRY(tmp.prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 get_bh(bh);
Jan Kara535ee2f2008-02-08 04:21:59 -0800834 mapping = bh->b_assoc_map;
835 __remove_assoc_queue(bh);
836 /* Avoid race with mark_buffer_dirty_inode() which does
837 * a lockless check and we rely on seeing the dirty bit */
838 smp_mb();
839 if (buffer_dirty(bh)) {
840 list_add(&bh->b_assoc_buffers,
Jan Karae3892292008-03-04 14:28:33 -0800841 &mapping->private_list);
Jan Kara535ee2f2008-02-08 04:21:59 -0800842 bh->b_assoc_map = mapping;
843 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 spin_unlock(lock);
845 wait_on_buffer(bh);
846 if (!buffer_uptodate(bh))
847 err = -EIO;
848 brelse(bh);
849 spin_lock(lock);
850 }
851
852 spin_unlock(lock);
853 err2 = osync_buffers_list(lock, list);
854 if (err)
855 return err;
856 else
857 return err2;
858}
859
860/*
861 * Invalidate any and all dirty buffers on a given inode. We are
862 * probably unmounting the fs, but that doesn't mean we have already
863 * done a sync(). Just drop the buffers from the inode list.
864 *
865 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
866 * assumes that all the buffers are against the blockdev. Not true
867 * for reiserfs.
868 */
869void invalidate_inode_buffers(struct inode *inode)
870{
871 if (inode_has_buffers(inode)) {
872 struct address_space *mapping = &inode->i_data;
873 struct list_head *list = &mapping->private_list;
874 struct address_space *buffer_mapping = mapping->assoc_mapping;
875
876 spin_lock(&buffer_mapping->private_lock);
877 while (!list_empty(list))
878 __remove_assoc_queue(BH_ENTRY(list->next));
879 spin_unlock(&buffer_mapping->private_lock);
880 }
881}
882
883/*
884 * Remove any clean buffers from the inode's buffer list. This is called
885 * when we're trying to free the inode itself. Those buffers can pin it.
886 *
887 * Returns true if all buffers were removed.
888 */
889int remove_inode_buffers(struct inode *inode)
890{
891 int ret = 1;
892
893 if (inode_has_buffers(inode)) {
894 struct address_space *mapping = &inode->i_data;
895 struct list_head *list = &mapping->private_list;
896 struct address_space *buffer_mapping = mapping->assoc_mapping;
897
898 spin_lock(&buffer_mapping->private_lock);
899 while (!list_empty(list)) {
900 struct buffer_head *bh = BH_ENTRY(list->next);
901 if (buffer_dirty(bh)) {
902 ret = 0;
903 break;
904 }
905 __remove_assoc_queue(bh);
906 }
907 spin_unlock(&buffer_mapping->private_lock);
908 }
909 return ret;
910}
911
912/*
913 * Create the appropriate buffers when given a page for data area and
914 * the size of each buffer.. Use the bh->b_this_page linked list to
915 * follow the buffers created. Return NULL if unable to create more
916 * buffers.
917 *
918 * The retry flag is used to differentiate async IO (paging, swapping)
919 * which may not fail from ordinary buffer allocations.
920 */
921struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
922 int retry)
923{
924 struct buffer_head *bh, *head;
925 long offset;
926
927try_again:
928 head = NULL;
929 offset = PAGE_SIZE;
930 while ((offset -= size) >= 0) {
931 bh = alloc_buffer_head(GFP_NOFS);
932 if (!bh)
933 goto no_grow;
934
935 bh->b_bdev = NULL;
936 bh->b_this_page = head;
937 bh->b_blocknr = -1;
938 head = bh;
939
940 bh->b_state = 0;
941 atomic_set(&bh->b_count, 0);
Chris Masonfc5cd582006-02-01 03:06:48 -0800942 bh->b_private = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 bh->b_size = size;
944
945 /* Link the buffer to its page */
946 set_bh_page(bh, page, offset);
947
Nathan Scott01ffe332006-01-17 09:02:07 +1100948 init_buffer(bh, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 }
950 return head;
951/*
952 * In case anything failed, we just free everything we got.
953 */
954no_grow:
955 if (head) {
956 do {
957 bh = head;
958 head = head->b_this_page;
959 free_buffer_head(bh);
960 } while (head);
961 }
962
963 /*
964 * Return failure for non-async IO requests. Async IO requests
965 * are not allowed to fail, so we have to wait until buffer heads
966 * become available. But we don't want tasks sleeping with
967 * partially complete buffers, so all were released above.
968 */
969 if (!retry)
970 return NULL;
971
972 /* We're _really_ low on memory. Now we just
973 * wait for old buffer heads to become free due to
974 * finishing IO. Since this is an async request and
975 * the reserve list is empty, we're sure there are
976 * async buffer heads in use.
977 */
978 free_more_memory();
979 goto try_again;
980}
981EXPORT_SYMBOL_GPL(alloc_page_buffers);
982
983static inline void
984link_dev_buffers(struct page *page, struct buffer_head *head)
985{
986 struct buffer_head *bh, *tail;
987
988 bh = head;
989 do {
990 tail = bh;
991 bh = bh->b_this_page;
992 } while (bh);
993 tail->b_this_page = head;
994 attach_page_buffers(page, head);
995}
996
997/*
998 * Initialise the state of a blockdev page's buffers.
999 */
1000static void
1001init_page_buffers(struct page *page, struct block_device *bdev,
1002 sector_t block, int size)
1003{
1004 struct buffer_head *head = page_buffers(page);
1005 struct buffer_head *bh = head;
1006 int uptodate = PageUptodate(page);
1007
1008 do {
1009 if (!buffer_mapped(bh)) {
1010 init_buffer(bh, NULL, NULL);
1011 bh->b_bdev = bdev;
1012 bh->b_blocknr = block;
1013 if (uptodate)
1014 set_buffer_uptodate(bh);
1015 set_buffer_mapped(bh);
1016 }
1017 block++;
1018 bh = bh->b_this_page;
1019 } while (bh != head);
1020}
1021
1022/*
1023 * Create the page-cache page that contains the requested block.
1024 *
1025 * This is user purely for blockdev mappings.
1026 */
1027static struct page *
1028grow_dev_page(struct block_device *bdev, sector_t block,
1029 pgoff_t index, int size)
1030{
1031 struct inode *inode = bdev->bd_inode;
1032 struct page *page;
1033 struct buffer_head *bh;
1034
Christoph Lameterea125892007-05-16 22:11:21 -07001035 page = find_or_create_page(inode->i_mapping, index,
Mel Gorman769848c2007-07-17 04:03:05 -07001036 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 if (!page)
1038 return NULL;
1039
Eric Sesterhenne827f922006-03-26 18:24:46 +02001040 BUG_ON(!PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041
1042 if (page_has_buffers(page)) {
1043 bh = page_buffers(page);
1044 if (bh->b_size == size) {
1045 init_page_buffers(page, bdev, block, size);
1046 return page;
1047 }
1048 if (!try_to_free_buffers(page))
1049 goto failed;
1050 }
1051
1052 /*
1053 * Allocate some buffers for this page
1054 */
1055 bh = alloc_page_buffers(page, size, 0);
1056 if (!bh)
1057 goto failed;
1058
1059 /*
1060 * Link the page to the buffers and initialise them. Take the
1061 * lock to be atomic wrt __find_get_block(), which does not
1062 * run under the page lock.
1063 */
1064 spin_lock(&inode->i_mapping->private_lock);
1065 link_dev_buffers(page, bh);
1066 init_page_buffers(page, bdev, block, size);
1067 spin_unlock(&inode->i_mapping->private_lock);
1068 return page;
1069
1070failed:
1071 BUG();
1072 unlock_page(page);
1073 page_cache_release(page);
1074 return NULL;
1075}
1076
1077/*
1078 * Create buffers for the specified block device block's page. If
1079 * that page was dirty, the buffers are set dirty also.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08001081static int
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082grow_buffers(struct block_device *bdev, sector_t block, int size)
1083{
1084 struct page *page;
1085 pgoff_t index;
1086 int sizebits;
1087
1088 sizebits = -1;
1089 do {
1090 sizebits++;
1091 } while ((size << sizebits) < PAGE_SIZE);
1092
1093 index = block >> sizebits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094
Andrew Mortone5657932006-10-11 01:21:46 -07001095 /*
1096 * Check for a block which wants to lie outside our maximum possible
1097 * pagecache index. (this comparison is done using sector_t types).
1098 */
1099 if (unlikely(index != block >> sizebits)) {
1100 char b[BDEVNAME_SIZE];
1101
1102 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1103 "device %s\n",
Harvey Harrison8e24eea2008-04-30 00:55:09 -07001104 __func__, (unsigned long long)block,
Andrew Mortone5657932006-10-11 01:21:46 -07001105 bdevname(bdev, b));
1106 return -EIO;
1107 }
1108 block = index << sizebits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 /* Create a page with the proper size buffers.. */
1110 page = grow_dev_page(bdev, block, index, size);
1111 if (!page)
1112 return 0;
1113 unlock_page(page);
1114 page_cache_release(page);
1115 return 1;
1116}
1117
Adrian Bunk75c96f82005-05-05 16:16:09 -07001118static struct buffer_head *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119__getblk_slow(struct block_device *bdev, sector_t block, int size)
1120{
1121 /* Size must be multiple of hard sectorsize */
1122 if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1123 (size < 512 || size > PAGE_SIZE))) {
1124 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1125 size);
1126 printk(KERN_ERR "hardsect size: %d\n",
1127 bdev_hardsect_size(bdev));
1128
1129 dump_stack();
1130 return NULL;
1131 }
1132
1133 for (;;) {
1134 struct buffer_head * bh;
Andrew Mortone5657932006-10-11 01:21:46 -07001135 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
1137 bh = __find_get_block(bdev, block, size);
1138 if (bh)
1139 return bh;
1140
Andrew Mortone5657932006-10-11 01:21:46 -07001141 ret = grow_buffers(bdev, block, size);
1142 if (ret < 0)
1143 return NULL;
1144 if (ret == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 free_more_memory();
1146 }
1147}
1148
1149/*
1150 * The relationship between dirty buffers and dirty pages:
1151 *
1152 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1153 * the page is tagged dirty in its radix tree.
1154 *
1155 * At all times, the dirtiness of the buffers represents the dirtiness of
1156 * subsections of the page. If the page has buffers, the page dirty bit is
1157 * merely a hint about the true dirty state.
1158 *
1159 * When a page is set dirty in its entirety, all its buffers are marked dirty
1160 * (if the page has buffers).
1161 *
1162 * When a buffer is marked dirty, its page is dirtied, but the page's other
1163 * buffers are not.
1164 *
1165 * Also. When blockdev buffers are explicitly read with bread(), they
1166 * individually become uptodate. But their backing page remains not
1167 * uptodate - even if all of its buffers are uptodate. A subsequent
1168 * block_read_full_page() against that page will discover all the uptodate
1169 * buffers, will set the page uptodate and will perform no I/O.
1170 */
1171
1172/**
1173 * mark_buffer_dirty - mark a buffer_head as needing writeout
Martin Waitz67be2dd2005-05-01 08:59:26 -07001174 * @bh: the buffer_head to mark dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 *
1176 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1177 * backing page dirty, then tag the page as dirty in its address_space's radix
1178 * tree and then attach the address_space's inode to its superblock's dirty
1179 * inode list.
1180 *
1181 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1182 * mapping->tree_lock and the global inode_lock.
1183 */
Harvey Harrisonfc9b52c2008-02-08 04:19:52 -08001184void mark_buffer_dirty(struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185{
Nick Piggin787d2212007-07-17 04:03:34 -07001186 WARN_ON_ONCE(!buffer_uptodate(bh));
Linus Torvalds1be62dc2008-04-04 14:38:17 -07001187
1188 /*
1189 * Very *carefully* optimize the it-is-already-dirty case.
1190 *
1191 * Don't let the final "is it dirty" escape to before we
1192 * perhaps modified the buffer.
1193 */
1194 if (buffer_dirty(bh)) {
1195 smp_mb();
1196 if (buffer_dirty(bh))
1197 return;
1198 }
1199
1200 if (!test_set_buffer_dirty(bh))
Nick Piggin787d2212007-07-17 04:03:34 -07001201 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202}
1203
1204/*
1205 * Decrement a buffer_head's reference count. If all buffers against a page
1206 * have zero reference count, are clean and unlocked, and if the page is clean
1207 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1208 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1209 * a page but it ends up not being freed, and buffers may later be reattached).
1210 */
1211void __brelse(struct buffer_head * buf)
1212{
1213 if (atomic_read(&buf->b_count)) {
1214 put_bh(buf);
1215 return;
1216 }
1217 printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1218 WARN_ON(1);
1219}
1220
1221/*
1222 * bforget() is like brelse(), except it discards any
1223 * potentially dirty data.
1224 */
1225void __bforget(struct buffer_head *bh)
1226{
1227 clear_buffer_dirty(bh);
Jan Kara535ee2f2008-02-08 04:21:59 -08001228 if (bh->b_assoc_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 struct address_space *buffer_mapping = bh->b_page->mapping;
1230
1231 spin_lock(&buffer_mapping->private_lock);
1232 list_del_init(&bh->b_assoc_buffers);
Jan Kara58ff4072006-10-17 00:10:19 -07001233 bh->b_assoc_map = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 spin_unlock(&buffer_mapping->private_lock);
1235 }
1236 __brelse(bh);
1237}
1238
1239static struct buffer_head *__bread_slow(struct buffer_head *bh)
1240{
1241 lock_buffer(bh);
1242 if (buffer_uptodate(bh)) {
1243 unlock_buffer(bh);
1244 return bh;
1245 } else {
1246 get_bh(bh);
1247 bh->b_end_io = end_buffer_read_sync;
1248 submit_bh(READ, bh);
1249 wait_on_buffer(bh);
1250 if (buffer_uptodate(bh))
1251 return bh;
1252 }
1253 brelse(bh);
1254 return NULL;
1255}
1256
1257/*
1258 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1259 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1260 * refcount elevated by one when they're in an LRU. A buffer can only appear
1261 * once in a particular CPU's LRU. A single buffer can be present in multiple
1262 * CPU's LRUs at the same time.
1263 *
1264 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1265 * sb_find_get_block().
1266 *
1267 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1268 * a local interrupt disable for that.
1269 */
1270
1271#define BH_LRU_SIZE 8
1272
1273struct bh_lru {
1274 struct buffer_head *bhs[BH_LRU_SIZE];
1275};
1276
1277static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1278
1279#ifdef CONFIG_SMP
1280#define bh_lru_lock() local_irq_disable()
1281#define bh_lru_unlock() local_irq_enable()
1282#else
1283#define bh_lru_lock() preempt_disable()
1284#define bh_lru_unlock() preempt_enable()
1285#endif
1286
1287static inline void check_irqs_on(void)
1288{
1289#ifdef irqs_disabled
1290 BUG_ON(irqs_disabled());
1291#endif
1292}
1293
1294/*
1295 * The LRU management algorithm is dopey-but-simple. Sorry.
1296 */
1297static void bh_lru_install(struct buffer_head *bh)
1298{
1299 struct buffer_head *evictee = NULL;
1300 struct bh_lru *lru;
1301
1302 check_irqs_on();
1303 bh_lru_lock();
1304 lru = &__get_cpu_var(bh_lrus);
1305 if (lru->bhs[0] != bh) {
1306 struct buffer_head *bhs[BH_LRU_SIZE];
1307 int in;
1308 int out = 0;
1309
1310 get_bh(bh);
1311 bhs[out++] = bh;
1312 for (in = 0; in < BH_LRU_SIZE; in++) {
1313 struct buffer_head *bh2 = lru->bhs[in];
1314
1315 if (bh2 == bh) {
1316 __brelse(bh2);
1317 } else {
1318 if (out >= BH_LRU_SIZE) {
1319 BUG_ON(evictee != NULL);
1320 evictee = bh2;
1321 } else {
1322 bhs[out++] = bh2;
1323 }
1324 }
1325 }
1326 while (out < BH_LRU_SIZE)
1327 bhs[out++] = NULL;
1328 memcpy(lru->bhs, bhs, sizeof(bhs));
1329 }
1330 bh_lru_unlock();
1331
1332 if (evictee)
1333 __brelse(evictee);
1334}
1335
1336/*
1337 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1338 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08001339static struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001340lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341{
1342 struct buffer_head *ret = NULL;
1343 struct bh_lru *lru;
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001344 unsigned int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345
1346 check_irqs_on();
1347 bh_lru_lock();
1348 lru = &__get_cpu_var(bh_lrus);
1349 for (i = 0; i < BH_LRU_SIZE; i++) {
1350 struct buffer_head *bh = lru->bhs[i];
1351
1352 if (bh && bh->b_bdev == bdev &&
1353 bh->b_blocknr == block && bh->b_size == size) {
1354 if (i) {
1355 while (i) {
1356 lru->bhs[i] = lru->bhs[i - 1];
1357 i--;
1358 }
1359 lru->bhs[0] = bh;
1360 }
1361 get_bh(bh);
1362 ret = bh;
1363 break;
1364 }
1365 }
1366 bh_lru_unlock();
1367 return ret;
1368}
1369
1370/*
1371 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1372 * it in the LRU and mark it as accessed. If it is not present then return
1373 * NULL
1374 */
1375struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001376__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377{
1378 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1379
1380 if (bh == NULL) {
Coywolf Qi Hunt385fd4c2005-11-07 00:59:39 -08001381 bh = __find_get_block_slow(bdev, block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 if (bh)
1383 bh_lru_install(bh);
1384 }
1385 if (bh)
1386 touch_buffer(bh);
1387 return bh;
1388}
1389EXPORT_SYMBOL(__find_get_block);
1390
1391/*
1392 * __getblk will locate (and, if necessary, create) the buffer_head
1393 * which corresponds to the passed block_device, block and size. The
1394 * returned buffer has its reference count incremented.
1395 *
1396 * __getblk() cannot fail - it just keeps trying. If you pass it an
1397 * illegal block number, __getblk() will happily return a buffer_head
1398 * which represents the non-existent block. Very weird.
1399 *
1400 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1401 * attempt is failing. FIXME, perhaps?
1402 */
1403struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001404__getblk(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405{
1406 struct buffer_head *bh = __find_get_block(bdev, block, size);
1407
1408 might_sleep();
1409 if (bh == NULL)
1410 bh = __getblk_slow(bdev, block, size);
1411 return bh;
1412}
1413EXPORT_SYMBOL(__getblk);
1414
1415/*
1416 * Do async read-ahead on a buffer..
1417 */
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001418void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419{
1420 struct buffer_head *bh = __getblk(bdev, block, size);
Andrew Mortona3e713b2005-10-30 15:03:15 -08001421 if (likely(bh)) {
1422 ll_rw_block(READA, 1, &bh);
1423 brelse(bh);
1424 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425}
1426EXPORT_SYMBOL(__breadahead);
1427
1428/**
1429 * __bread() - reads a specified block and returns the bh
Martin Waitz67be2dd2005-05-01 08:59:26 -07001430 * @bdev: the block_device to read from
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 * @block: number of block
1432 * @size: size (in bytes) to read
1433 *
1434 * Reads a specified block, and returns buffer head that contains it.
1435 * It returns NULL if the block was unreadable.
1436 */
1437struct buffer_head *
Tomasz Kvarsin3991d3b2007-02-12 00:52:14 -08001438__bread(struct block_device *bdev, sector_t block, unsigned size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439{
1440 struct buffer_head *bh = __getblk(bdev, block, size);
1441
Andrew Mortona3e713b2005-10-30 15:03:15 -08001442 if (likely(bh) && !buffer_uptodate(bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 bh = __bread_slow(bh);
1444 return bh;
1445}
1446EXPORT_SYMBOL(__bread);
1447
1448/*
1449 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1450 * This doesn't race because it runs in each cpu either in irq
1451 * or with preempt disabled.
1452 */
1453static void invalidate_bh_lru(void *arg)
1454{
1455 struct bh_lru *b = &get_cpu_var(bh_lrus);
1456 int i;
1457
1458 for (i = 0; i < BH_LRU_SIZE; i++) {
1459 brelse(b->bhs[i]);
1460 b->bhs[i] = NULL;
1461 }
1462 put_cpu_var(bh_lrus);
1463}
1464
Peter Zijlstraf9a14392007-05-06 14:49:55 -07001465void invalidate_bh_lrus(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466{
1467 on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1468}
Nick Piggin9db55792008-02-08 04:19:49 -08001469EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470
1471void set_bh_page(struct buffer_head *bh,
1472 struct page *page, unsigned long offset)
1473{
1474 bh->b_page = page;
Eric Sesterhenne827f922006-03-26 18:24:46 +02001475 BUG_ON(offset >= PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 if (PageHighMem(page))
1477 /*
1478 * This catches illegal uses and preserves the offset:
1479 */
1480 bh->b_data = (char *)(0 + offset);
1481 else
1482 bh->b_data = page_address(page) + offset;
1483}
1484EXPORT_SYMBOL(set_bh_page);
1485
1486/*
1487 * Called when truncating a buffer on a page completely.
1488 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08001489static void discard_buffer(struct buffer_head * bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490{
1491 lock_buffer(bh);
1492 clear_buffer_dirty(bh);
1493 bh->b_bdev = NULL;
1494 clear_buffer_mapped(bh);
1495 clear_buffer_req(bh);
1496 clear_buffer_new(bh);
1497 clear_buffer_delay(bh);
David Chinner33a266d2007-02-12 00:51:41 -08001498 clear_buffer_unwritten(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 unlock_buffer(bh);
1500}
1501
1502/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 * block_invalidatepage - invalidate part of all of a buffer-backed page
1504 *
1505 * @page: the page which is affected
1506 * @offset: the index of the truncation point
1507 *
1508 * block_invalidatepage() is called when all or part of the page has become
1509 * invalidatedby a truncate operation.
1510 *
1511 * block_invalidatepage() does not have to release all buffers, but it must
1512 * ensure that no dirty buffer is left outside @offset and that no I/O
1513 * is underway against any of the blocks which are outside the truncation
1514 * point. Because the caller is about to free (and possibly reuse) those
1515 * blocks on-disk.
1516 */
NeilBrown2ff28e22006-03-26 01:37:18 -08001517void block_invalidatepage(struct page *page, unsigned long offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518{
1519 struct buffer_head *head, *bh, *next;
1520 unsigned int curr_off = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
1522 BUG_ON(!PageLocked(page));
1523 if (!page_has_buffers(page))
1524 goto out;
1525
1526 head = page_buffers(page);
1527 bh = head;
1528 do {
1529 unsigned int next_off = curr_off + bh->b_size;
1530 next = bh->b_this_page;
1531
1532 /*
1533 * is this block fully invalidated?
1534 */
1535 if (offset <= curr_off)
1536 discard_buffer(bh);
1537 curr_off = next_off;
1538 bh = next;
1539 } while (bh != head);
1540
1541 /*
1542 * We release buffers only if the entire page is being invalidated.
1543 * The get_block cached value has been unconditionally invalidated,
1544 * so real IO is not possible anymore.
1545 */
1546 if (offset == 0)
NeilBrown2ff28e22006-03-26 01:37:18 -08001547 try_to_release_page(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548out:
NeilBrown2ff28e22006-03-26 01:37:18 -08001549 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550}
1551EXPORT_SYMBOL(block_invalidatepage);
1552
1553/*
1554 * We attach and possibly dirty the buffers atomically wrt
1555 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1556 * is already excluded via the page lock.
1557 */
1558void create_empty_buffers(struct page *page,
1559 unsigned long blocksize, unsigned long b_state)
1560{
1561 struct buffer_head *bh, *head, *tail;
1562
1563 head = alloc_page_buffers(page, blocksize, 1);
1564 bh = head;
1565 do {
1566 bh->b_state |= b_state;
1567 tail = bh;
1568 bh = bh->b_this_page;
1569 } while (bh);
1570 tail->b_this_page = head;
1571
1572 spin_lock(&page->mapping->private_lock);
1573 if (PageUptodate(page) || PageDirty(page)) {
1574 bh = head;
1575 do {
1576 if (PageDirty(page))
1577 set_buffer_dirty(bh);
1578 if (PageUptodate(page))
1579 set_buffer_uptodate(bh);
1580 bh = bh->b_this_page;
1581 } while (bh != head);
1582 }
1583 attach_page_buffers(page, head);
1584 spin_unlock(&page->mapping->private_lock);
1585}
1586EXPORT_SYMBOL(create_empty_buffers);
1587
1588/*
1589 * We are taking a block for data and we don't want any output from any
1590 * buffer-cache aliases starting from return from that function and
1591 * until the moment when something will explicitly mark the buffer
1592 * dirty (hopefully that will not happen until we will free that block ;-)
1593 * We don't even need to mark it not-uptodate - nobody can expect
1594 * anything from a newly allocated buffer anyway. We used to used
1595 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1596 * don't want to mark the alias unmapped, for example - it would confuse
1597 * anyone who might pick it with bread() afterwards...
1598 *
1599 * Also.. Note that bforget() doesn't lock the buffer. So there can
1600 * be writeout I/O going on against recently-freed buffers. We don't
1601 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1602 * only if we really need to. That happens here.
1603 */
1604void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1605{
1606 struct buffer_head *old_bh;
1607
1608 might_sleep();
1609
Coywolf Qi Hunt385fd4c2005-11-07 00:59:39 -08001610 old_bh = __find_get_block_slow(bdev, block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 if (old_bh) {
1612 clear_buffer_dirty(old_bh);
1613 wait_on_buffer(old_bh);
1614 clear_buffer_req(old_bh);
1615 __brelse(old_bh);
1616 }
1617}
1618EXPORT_SYMBOL(unmap_underlying_metadata);
1619
1620/*
1621 * NOTE! All mapped/uptodate combinations are valid:
1622 *
1623 * Mapped Uptodate Meaning
1624 *
1625 * No No "unknown" - must do get_block()
1626 * No Yes "hole" - zero-filled
1627 * Yes No "allocated" - allocated on disk, not read in
1628 * Yes Yes "valid" - allocated and up-to-date in memory.
1629 *
1630 * "Dirty" is valid only with the last case (mapped+uptodate).
1631 */
1632
1633/*
1634 * While block_write_full_page is writing back the dirty buffers under
1635 * the page lock, whoever dirtied the buffers may decide to clean them
1636 * again at any time. We handle that by only looking at the buffer
1637 * state inside lock_buffer().
1638 *
1639 * If block_write_full_page() is called for regular writeback
1640 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1641 * locked buffer. This only can happen if someone has written the buffer
1642 * directly, with submit_bh(). At the address_space level PageWriteback
1643 * prevents this contention from occurring.
1644 */
1645static int __block_write_full_page(struct inode *inode, struct page *page,
1646 get_block_t *get_block, struct writeback_control *wbc)
1647{
1648 int err;
1649 sector_t block;
1650 sector_t last_block;
Andrew Mortonf0fbd5f2005-05-05 16:15:48 -07001651 struct buffer_head *bh, *head;
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001652 const unsigned blocksize = 1 << inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 int nr_underway = 0;
1654
1655 BUG_ON(!PageLocked(page));
1656
1657 last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1658
1659 if (!page_has_buffers(page)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001660 create_empty_buffers(page, blocksize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 (1 << BH_Dirty)|(1 << BH_Uptodate));
1662 }
1663
1664 /*
1665 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1666 * here, and the (potentially unmapped) buffers may become dirty at
1667 * any time. If a buffer becomes dirty here after we've inspected it
1668 * then we just miss that fact, and the page stays dirty.
1669 *
1670 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1671 * handle that here by just cleaning them.
1672 */
1673
Andrew Morton54b21a72006-01-08 01:03:05 -08001674 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 head = page_buffers(page);
1676 bh = head;
1677
1678 /*
1679 * Get all the dirty buffers mapped to disk addresses and
1680 * handle any aliases from the underlying blockdev's mapping.
1681 */
1682 do {
1683 if (block > last_block) {
1684 /*
1685 * mapped buffers outside i_size will occur, because
1686 * this page can be outside i_size when there is a
1687 * truncate in progress.
1688 */
1689 /*
1690 * The buffer was zeroed by block_write_full_page()
1691 */
1692 clear_buffer_dirty(bh);
1693 set_buffer_uptodate(bh);
1694 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001695 WARN_ON(bh->b_size != blocksize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 err = get_block(inode, block, bh, 1);
1697 if (err)
1698 goto recover;
1699 if (buffer_new(bh)) {
1700 /* blockdev mappings never come here */
1701 clear_buffer_new(bh);
1702 unmap_underlying_metadata(bh->b_bdev,
1703 bh->b_blocknr);
1704 }
1705 }
1706 bh = bh->b_this_page;
1707 block++;
1708 } while (bh != head);
1709
1710 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 if (!buffer_mapped(bh))
1712 continue;
1713 /*
1714 * If it's a fully non-blocking write attempt and we cannot
1715 * lock the buffer then redirty the page. Note that this can
1716 * potentially cause a busy-wait loop from pdflush and kswapd
1717 * activity, but those code paths have their own higher-level
1718 * throttling.
1719 */
1720 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1721 lock_buffer(bh);
1722 } else if (test_set_buffer_locked(bh)) {
1723 redirty_page_for_writepage(wbc, page);
1724 continue;
1725 }
1726 if (test_clear_buffer_dirty(bh)) {
1727 mark_buffer_async_write(bh);
1728 } else {
1729 unlock_buffer(bh);
1730 }
1731 } while ((bh = bh->b_this_page) != head);
1732
1733 /*
1734 * The page and its buffers are protected by PageWriteback(), so we can
1735 * drop the bh refcounts early.
1736 */
1737 BUG_ON(PageWriteback(page));
1738 set_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739
1740 do {
1741 struct buffer_head *next = bh->b_this_page;
1742 if (buffer_async_write(bh)) {
1743 submit_bh(WRITE, bh);
1744 nr_underway++;
1745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 bh = next;
1747 } while (bh != head);
Andrew Morton05937ba2005-05-05 16:15:47 -07001748 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749
1750 err = 0;
1751done:
1752 if (nr_underway == 0) {
1753 /*
1754 * The page was marked dirty, but the buffers were
1755 * clean. Someone wrote them back by hand with
1756 * ll_rw_block/submit_bh. A rare case.
1757 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 end_page_writeback(page);
Nick Piggin3d67f2d2007-05-06 14:49:05 -07001759
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 /*
1761 * The page and buffer_heads can be released at any time from
1762 * here on.
1763 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 }
1765 return err;
1766
1767recover:
1768 /*
1769 * ENOSPC, or some other error. We may already have added some
1770 * blocks to the file, so we need to write these out to avoid
1771 * exposing stale data.
1772 * The page is currently locked and not marked for writeback
1773 */
1774 bh = head;
1775 /* Recovery: lock and submit the mapped buffers */
1776 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1778 lock_buffer(bh);
1779 mark_buffer_async_write(bh);
1780 } else {
1781 /*
1782 * The buffer may have been set dirty during
1783 * attachment to a dirty page.
1784 */
1785 clear_buffer_dirty(bh);
1786 }
1787 } while ((bh = bh->b_this_page) != head);
1788 SetPageError(page);
1789 BUG_ON(PageWriteback(page));
Andrew Morton7e4c3692007-05-08 00:23:27 -07001790 mapping_set_error(page->mapping, err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 set_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 do {
1793 struct buffer_head *next = bh->b_this_page;
1794 if (buffer_async_write(bh)) {
1795 clear_buffer_dirty(bh);
1796 submit_bh(WRITE, bh);
1797 nr_underway++;
1798 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 bh = next;
1800 } while (bh != head);
Nick Pigginffda9d32007-02-20 13:57:54 -08001801 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 goto done;
1803}
1804
Nick Pigginafddba42007-10-16 01:25:01 -07001805/*
1806 * If a page has any new buffers, zero them out here, and mark them uptodate
1807 * and dirty so they'll be written out (in order to prevent uninitialised
1808 * block data from leaking). And clear the new bit.
1809 */
1810void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1811{
1812 unsigned int block_start, block_end;
1813 struct buffer_head *head, *bh;
1814
1815 BUG_ON(!PageLocked(page));
1816 if (!page_has_buffers(page))
1817 return;
1818
1819 bh = head = page_buffers(page);
1820 block_start = 0;
1821 do {
1822 block_end = block_start + bh->b_size;
1823
1824 if (buffer_new(bh)) {
1825 if (block_end > from && block_start < to) {
1826 if (!PageUptodate(page)) {
1827 unsigned start, size;
1828
1829 start = max(from, block_start);
1830 size = min(to, block_end) - start;
1831
Christoph Lametereebd2aa2008-02-04 22:28:29 -08001832 zero_user(page, start, size);
Nick Pigginafddba42007-10-16 01:25:01 -07001833 set_buffer_uptodate(bh);
1834 }
1835
1836 clear_buffer_new(bh);
1837 mark_buffer_dirty(bh);
1838 }
1839 }
1840
1841 block_start = block_end;
1842 bh = bh->b_this_page;
1843 } while (bh != head);
1844}
1845EXPORT_SYMBOL(page_zero_new_buffers);
1846
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847static int __block_prepare_write(struct inode *inode, struct page *page,
1848 unsigned from, unsigned to, get_block_t *get_block)
1849{
1850 unsigned block_start, block_end;
1851 sector_t block;
1852 int err = 0;
1853 unsigned blocksize, bbits;
1854 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1855
1856 BUG_ON(!PageLocked(page));
1857 BUG_ON(from > PAGE_CACHE_SIZE);
1858 BUG_ON(to > PAGE_CACHE_SIZE);
1859 BUG_ON(from > to);
1860
1861 blocksize = 1 << inode->i_blkbits;
1862 if (!page_has_buffers(page))
1863 create_empty_buffers(page, blocksize, 0);
1864 head = page_buffers(page);
1865
1866 bbits = inode->i_blkbits;
1867 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1868
1869 for(bh = head, block_start = 0; bh != head || !block_start;
1870 block++, block_start=block_end, bh = bh->b_this_page) {
1871 block_end = block_start + blocksize;
1872 if (block_end <= from || block_start >= to) {
1873 if (PageUptodate(page)) {
1874 if (!buffer_uptodate(bh))
1875 set_buffer_uptodate(bh);
1876 }
1877 continue;
1878 }
1879 if (buffer_new(bh))
1880 clear_buffer_new(bh);
1881 if (!buffer_mapped(bh)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08001882 WARN_ON(bh->b_size != blocksize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 err = get_block(inode, block, bh, 1);
1884 if (err)
Nick Pigginf3ddbdc2005-05-05 16:15:45 -07001885 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 if (buffer_new(bh)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 unmap_underlying_metadata(bh->b_bdev,
1888 bh->b_blocknr);
1889 if (PageUptodate(page)) {
Nick Piggin637aff42007-10-16 01:25:00 -07001890 clear_buffer_new(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 set_buffer_uptodate(bh);
Nick Piggin637aff42007-10-16 01:25:00 -07001892 mark_buffer_dirty(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 continue;
1894 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -08001895 if (block_end > to || block_start < from)
1896 zero_user_segments(page,
1897 to, block_end,
1898 block_start, from);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 continue;
1900 }
1901 }
1902 if (PageUptodate(page)) {
1903 if (!buffer_uptodate(bh))
1904 set_buffer_uptodate(bh);
1905 continue;
1906 }
1907 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
David Chinner33a266d2007-02-12 00:51:41 -08001908 !buffer_unwritten(bh) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 (block_start < from || block_end > to)) {
1910 ll_rw_block(READ, 1, &bh);
1911 *wait_bh++=bh;
1912 }
1913 }
1914 /*
1915 * If we issued read requests - let them complete.
1916 */
1917 while(wait_bh > wait) {
1918 wait_on_buffer(*--wait_bh);
1919 if (!buffer_uptodate(*wait_bh))
Nick Pigginf3ddbdc2005-05-05 16:15:45 -07001920 err = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 }
Nick Pigginafddba42007-10-16 01:25:01 -07001922 if (unlikely(err))
1923 page_zero_new_buffers(page, from, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 return err;
1925}
1926
1927static int __block_commit_write(struct inode *inode, struct page *page,
1928 unsigned from, unsigned to)
1929{
1930 unsigned block_start, block_end;
1931 int partial = 0;
1932 unsigned blocksize;
1933 struct buffer_head *bh, *head;
1934
1935 blocksize = 1 << inode->i_blkbits;
1936
1937 for(bh = head = page_buffers(page), block_start = 0;
1938 bh != head || !block_start;
1939 block_start=block_end, bh = bh->b_this_page) {
1940 block_end = block_start + blocksize;
1941 if (block_end <= from || block_start >= to) {
1942 if (!buffer_uptodate(bh))
1943 partial = 1;
1944 } else {
1945 set_buffer_uptodate(bh);
1946 mark_buffer_dirty(bh);
1947 }
Nick Pigginafddba42007-10-16 01:25:01 -07001948 clear_buffer_new(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 }
1950
1951 /*
1952 * If this is a partial write which happened to make all buffers
1953 * uptodate then we can optimize away a bogus readpage() for
1954 * the next read(). Here we 'discover' whether the page went
1955 * uptodate as a result of this (potentially partial) write.
1956 */
1957 if (!partial)
1958 SetPageUptodate(page);
1959 return 0;
1960}
1961
1962/*
Nick Pigginafddba42007-10-16 01:25:01 -07001963 * block_write_begin takes care of the basic task of block allocation and
1964 * bringing partial write blocks uptodate first.
1965 *
1966 * If *pagep is not NULL, then block_write_begin uses the locked page
1967 * at *pagep rather than allocating its own. In this case, the page will
1968 * not be unlocked or deallocated on failure.
1969 */
1970int block_write_begin(struct file *file, struct address_space *mapping,
1971 loff_t pos, unsigned len, unsigned flags,
1972 struct page **pagep, void **fsdata,
1973 get_block_t *get_block)
1974{
1975 struct inode *inode = mapping->host;
1976 int status = 0;
1977 struct page *page;
1978 pgoff_t index;
1979 unsigned start, end;
1980 int ownpage = 0;
1981
1982 index = pos >> PAGE_CACHE_SHIFT;
1983 start = pos & (PAGE_CACHE_SIZE - 1);
1984 end = start + len;
1985
1986 page = *pagep;
1987 if (page == NULL) {
1988 ownpage = 1;
1989 page = __grab_cache_page(mapping, index);
1990 if (!page) {
1991 status = -ENOMEM;
1992 goto out;
1993 }
1994 *pagep = page;
1995 } else
1996 BUG_ON(!PageLocked(page));
1997
1998 status = __block_prepare_write(inode, page, start, end, get_block);
1999 if (unlikely(status)) {
2000 ClearPageUptodate(page);
2001
2002 if (ownpage) {
2003 unlock_page(page);
2004 page_cache_release(page);
2005 *pagep = NULL;
2006
2007 /*
2008 * prepare_write() may have instantiated a few blocks
2009 * outside i_size. Trim these off again. Don't need
2010 * i_size_read because we hold i_mutex.
2011 */
2012 if (pos + len > inode->i_size)
2013 vmtruncate(inode, inode->i_size);
2014 }
2015 goto out;
2016 }
2017
2018out:
2019 return status;
2020}
2021EXPORT_SYMBOL(block_write_begin);
2022
2023int block_write_end(struct file *file, struct address_space *mapping,
2024 loff_t pos, unsigned len, unsigned copied,
2025 struct page *page, void *fsdata)
2026{
2027 struct inode *inode = mapping->host;
2028 unsigned start;
2029
2030 start = pos & (PAGE_CACHE_SIZE - 1);
2031
2032 if (unlikely(copied < len)) {
2033 /*
2034 * The buffers that were written will now be uptodate, so we
2035 * don't have to worry about a readpage reading them and
2036 * overwriting a partial write. However if we have encountered
2037 * a short write and only partially written into a buffer, it
2038 * will not be marked uptodate, so a readpage might come in and
2039 * destroy our partial write.
2040 *
2041 * Do the simplest thing, and just treat any short write to a
2042 * non uptodate page as a zero-length write, and force the
2043 * caller to redo the whole thing.
2044 */
2045 if (!PageUptodate(page))
2046 copied = 0;
2047
2048 page_zero_new_buffers(page, start+copied, start+len);
2049 }
2050 flush_dcache_page(page);
2051
2052 /* This could be a short (even 0-length) commit */
2053 __block_commit_write(inode, page, start, start+copied);
2054
2055 return copied;
2056}
2057EXPORT_SYMBOL(block_write_end);
2058
2059int generic_write_end(struct file *file, struct address_space *mapping,
2060 loff_t pos, unsigned len, unsigned copied,
2061 struct page *page, void *fsdata)
2062{
2063 struct inode *inode = mapping->host;
2064
2065 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2066
2067 /*
2068 * No need to use i_size_read() here, the i_size
2069 * cannot change under us because we hold i_mutex.
2070 *
2071 * But it's important to update i_size while still holding page lock:
2072 * page writeout could otherwise come in and zero beyond i_size.
2073 */
2074 if (pos+copied > inode->i_size) {
2075 i_size_write(inode, pos+copied);
2076 mark_inode_dirty(inode);
2077 }
2078
2079 unlock_page(page);
2080 page_cache_release(page);
2081
2082 return copied;
2083}
2084EXPORT_SYMBOL(generic_write_end);
2085
2086/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 * Generic "read page" function for block devices that have the normal
2088 * get_block functionality. This is most of the block device filesystems.
2089 * Reads the page asynchronously --- the unlock_buffer() and
2090 * set/clear_buffer_uptodate() functions propagate buffer state into the
2091 * page struct once IO has completed.
2092 */
2093int block_read_full_page(struct page *page, get_block_t *get_block)
2094{
2095 struct inode *inode = page->mapping->host;
2096 sector_t iblock, lblock;
2097 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2098 unsigned int blocksize;
2099 int nr, i;
2100 int fully_mapped = 1;
2101
Matt Mackallcd7619d2005-05-01 08:59:01 -07002102 BUG_ON(!PageLocked(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 blocksize = 1 << inode->i_blkbits;
2104 if (!page_has_buffers(page))
2105 create_empty_buffers(page, blocksize, 0);
2106 head = page_buffers(page);
2107
2108 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2109 lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2110 bh = head;
2111 nr = 0;
2112 i = 0;
2113
2114 do {
2115 if (buffer_uptodate(bh))
2116 continue;
2117
2118 if (!buffer_mapped(bh)) {
Andrew Mortonc64610b2005-05-16 21:53:49 -07002119 int err = 0;
2120
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 fully_mapped = 0;
2122 if (iblock < lblock) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08002123 WARN_ON(bh->b_size != blocksize);
Andrew Mortonc64610b2005-05-16 21:53:49 -07002124 err = get_block(inode, iblock, bh, 0);
2125 if (err)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 SetPageError(page);
2127 }
2128 if (!buffer_mapped(bh)) {
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002129 zero_user(page, i * blocksize, blocksize);
Andrew Mortonc64610b2005-05-16 21:53:49 -07002130 if (!err)
2131 set_buffer_uptodate(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 continue;
2133 }
2134 /*
2135 * get_block() might have updated the buffer
2136 * synchronously
2137 */
2138 if (buffer_uptodate(bh))
2139 continue;
2140 }
2141 arr[nr++] = bh;
2142 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2143
2144 if (fully_mapped)
2145 SetPageMappedToDisk(page);
2146
2147 if (!nr) {
2148 /*
2149 * All buffers are uptodate - we can set the page uptodate
2150 * as well. But not if get_block() returned an error.
2151 */
2152 if (!PageError(page))
2153 SetPageUptodate(page);
2154 unlock_page(page);
2155 return 0;
2156 }
2157
2158 /* Stage two: lock the buffers */
2159 for (i = 0; i < nr; i++) {
2160 bh = arr[i];
2161 lock_buffer(bh);
2162 mark_buffer_async_read(bh);
2163 }
2164
2165 /*
2166 * Stage 3: start the IO. Check for uptodateness
2167 * inside the buffer lock in case another process reading
2168 * the underlying blockdev brought it uptodate (the sct fix).
2169 */
2170 for (i = 0; i < nr; i++) {
2171 bh = arr[i];
2172 if (buffer_uptodate(bh))
2173 end_buffer_async_read(bh, 1);
2174 else
2175 submit_bh(READ, bh);
2176 }
2177 return 0;
2178}
2179
2180/* utility function for filesystems that need to do work on expanding
Nick Piggin89e10782007-10-16 01:25:07 -07002181 * truncates. Uses filesystem pagecache writes to allow the filesystem to
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 * deal with the hole.
2183 */
Nick Piggin89e10782007-10-16 01:25:07 -07002184int generic_cont_expand_simple(struct inode *inode, loff_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185{
2186 struct address_space *mapping = inode->i_mapping;
2187 struct page *page;
Nick Piggin89e10782007-10-16 01:25:07 -07002188 void *fsdata;
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002189 unsigned long limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 int err;
2191
2192 err = -EFBIG;
2193 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2194 if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2195 send_sig(SIGXFSZ, current, 0);
2196 goto out;
2197 }
2198 if (size > inode->i_sb->s_maxbytes)
2199 goto out;
2200
Nick Piggin89e10782007-10-16 01:25:07 -07002201 err = pagecache_write_begin(NULL, mapping, size, 0,
2202 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2203 &page, &fsdata);
2204 if (err)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 goto out;
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002206
Nick Piggin89e10782007-10-16 01:25:07 -07002207 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2208 BUG_ON(err > 0);
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002209
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210out:
2211 return err;
2212}
2213
Adrian Bunkf1e3af72008-04-29 00:59:01 -07002214static int cont_expand_zero(struct file *file, struct address_space *mapping,
2215 loff_t pos, loff_t *bytes)
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002216{
Nick Piggin89e10782007-10-16 01:25:07 -07002217 struct inode *inode = mapping->host;
2218 unsigned blocksize = 1 << inode->i_blkbits;
2219 struct page *page;
2220 void *fsdata;
2221 pgoff_t index, curidx;
2222 loff_t curpos;
2223 unsigned zerofrom, offset, len;
2224 int err = 0;
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002225
Nick Piggin89e10782007-10-16 01:25:07 -07002226 index = pos >> PAGE_CACHE_SHIFT;
2227 offset = pos & ~PAGE_CACHE_MASK;
2228
2229 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2230 zerofrom = curpos & ~PAGE_CACHE_MASK;
2231 if (zerofrom & (blocksize-1)) {
2232 *bytes |= (blocksize-1);
2233 (*bytes)++;
2234 }
2235 len = PAGE_CACHE_SIZE - zerofrom;
2236
2237 err = pagecache_write_begin(file, mapping, curpos, len,
2238 AOP_FLAG_UNINTERRUPTIBLE,
2239 &page, &fsdata);
2240 if (err)
2241 goto out;
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002242 zero_user(page, zerofrom, len);
Nick Piggin89e10782007-10-16 01:25:07 -07002243 err = pagecache_write_end(file, mapping, curpos, len, len,
2244 page, fsdata);
2245 if (err < 0)
2246 goto out;
2247 BUG_ON(err != len);
2248 err = 0;
OGAWA Hirofumi061e9742008-04-28 02:16:28 -07002249
2250 balance_dirty_pages_ratelimited(mapping);
Nick Piggin89e10782007-10-16 01:25:07 -07002251 }
2252
2253 /* page covers the boundary, find the boundary offset */
2254 if (index == curidx) {
2255 zerofrom = curpos & ~PAGE_CACHE_MASK;
2256 /* if we will expand the thing last block will be filled */
2257 if (offset <= zerofrom) {
2258 goto out;
2259 }
2260 if (zerofrom & (blocksize-1)) {
2261 *bytes |= (blocksize-1);
2262 (*bytes)++;
2263 }
2264 len = offset - zerofrom;
2265
2266 err = pagecache_write_begin(file, mapping, curpos, len,
2267 AOP_FLAG_UNINTERRUPTIBLE,
2268 &page, &fsdata);
2269 if (err)
2270 goto out;
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002271 zero_user(page, zerofrom, len);
Nick Piggin89e10782007-10-16 01:25:07 -07002272 err = pagecache_write_end(file, mapping, curpos, len, len,
2273 page, fsdata);
2274 if (err < 0)
2275 goto out;
2276 BUG_ON(err != len);
2277 err = 0;
2278 }
2279out:
2280 return err;
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08002281}
2282
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283/*
2284 * For moronic filesystems that do not allow holes in file.
2285 * We may have to extend the file.
2286 */
Nick Piggin89e10782007-10-16 01:25:07 -07002287int cont_write_begin(struct file *file, struct address_space *mapping,
2288 loff_t pos, unsigned len, unsigned flags,
2289 struct page **pagep, void **fsdata,
2290 get_block_t *get_block, loff_t *bytes)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 struct inode *inode = mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 unsigned blocksize = 1 << inode->i_blkbits;
Nick Piggin89e10782007-10-16 01:25:07 -07002294 unsigned zerofrom;
2295 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296
Nick Piggin89e10782007-10-16 01:25:07 -07002297 err = cont_expand_zero(file, mapping, pos, bytes);
2298 if (err)
2299 goto out;
2300
2301 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2302 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2303 *bytes |= (blocksize-1);
2304 (*bytes)++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 }
2306
Nick Piggin89e10782007-10-16 01:25:07 -07002307 *pagep = NULL;
2308 err = block_write_begin(file, mapping, pos, len,
2309 flags, pagep, fsdata, get_block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310out:
Nick Piggin89e10782007-10-16 01:25:07 -07002311 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312}
2313
2314int block_prepare_write(struct page *page, unsigned from, unsigned to,
2315 get_block_t *get_block)
2316{
2317 struct inode *inode = page->mapping->host;
2318 int err = __block_prepare_write(inode, page, from, to, get_block);
2319 if (err)
2320 ClearPageUptodate(page);
2321 return err;
2322}
2323
2324int block_commit_write(struct page *page, unsigned from, unsigned to)
2325{
2326 struct inode *inode = page->mapping->host;
2327 __block_commit_write(inode,page,from,to);
2328 return 0;
2329}
2330
David Chinner54171692007-07-19 17:39:55 +10002331/*
2332 * block_page_mkwrite() is not allowed to change the file size as it gets
2333 * called from a page fault handler when a page is first dirtied. Hence we must
2334 * be careful to check for EOF conditions here. We set the page up correctly
2335 * for a written page which means we get ENOSPC checking when writing into
2336 * holes and correct delalloc and unwritten extent mapping on filesystems that
2337 * support these features.
2338 *
2339 * We are not allowed to take the i_mutex here so we have to play games to
2340 * protect against truncate races as the page could now be beyond EOF. Because
2341 * vmtruncate() writes the inode size before removing pages, once we have the
2342 * page lock we can determine safely if the page is beyond EOF. If it is not
2343 * beyond EOF, then the page is guaranteed safe against truncation until we
2344 * unlock the page.
2345 */
2346int
2347block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2348 get_block_t get_block)
2349{
2350 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2351 unsigned long end;
2352 loff_t size;
2353 int ret = -EINVAL;
2354
2355 lock_page(page);
2356 size = i_size_read(inode);
2357 if ((page->mapping != inode->i_mapping) ||
Nick Piggin18336332007-07-20 00:31:45 -07002358 (page_offset(page) > size)) {
David Chinner54171692007-07-19 17:39:55 +10002359 /* page got truncated out from underneath us */
2360 goto out_unlock;
2361 }
2362
2363 /* page is wholly or partially inside EOF */
2364 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2365 end = size & ~PAGE_CACHE_MASK;
2366 else
2367 end = PAGE_CACHE_SIZE;
2368
2369 ret = block_prepare_write(page, 0, end, get_block);
2370 if (!ret)
2371 ret = block_commit_write(page, 0, end);
2372
2373out_unlock:
2374 unlock_page(page);
2375 return ret;
2376}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377
2378/*
Nick Piggin03158cd2007-10-16 01:25:25 -07002379 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 * immediately, while under the page lock. So it needs a special end_io
2381 * handler which does not touch the bh after unlocking it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 */
2383static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2384{
Dmitry Monakhov68671f32007-10-16 01:24:47 -07002385 __end_buffer_read_notouch(bh, uptodate);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386}
2387
2388/*
Nick Piggin03158cd2007-10-16 01:25:25 -07002389 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2390 * the page (converting it to circular linked list and taking care of page
2391 * dirty races).
2392 */
2393static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2394{
2395 struct buffer_head *bh;
2396
2397 BUG_ON(!PageLocked(page));
2398
2399 spin_lock(&page->mapping->private_lock);
2400 bh = head;
2401 do {
2402 if (PageDirty(page))
2403 set_buffer_dirty(bh);
2404 if (!bh->b_this_page)
2405 bh->b_this_page = head;
2406 bh = bh->b_this_page;
2407 } while (bh != head);
2408 attach_page_buffers(page, head);
2409 spin_unlock(&page->mapping->private_lock);
2410}
2411
2412/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 * On entry, the page is fully not uptodate.
2414 * On exit the page is fully uptodate in the areas outside (from,to)
2415 */
Nick Piggin03158cd2007-10-16 01:25:25 -07002416int nobh_write_begin(struct file *file, struct address_space *mapping,
2417 loff_t pos, unsigned len, unsigned flags,
2418 struct page **pagep, void **fsdata,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 get_block_t *get_block)
2420{
Nick Piggin03158cd2007-10-16 01:25:25 -07002421 struct inode *inode = mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 const unsigned blkbits = inode->i_blkbits;
2423 const unsigned blocksize = 1 << blkbits;
Nick Piggina4b06722007-10-16 01:24:48 -07002424 struct buffer_head *head, *bh;
Nick Piggin03158cd2007-10-16 01:25:25 -07002425 struct page *page;
2426 pgoff_t index;
2427 unsigned from, to;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 unsigned block_in_page;
Nick Piggina4b06722007-10-16 01:24:48 -07002429 unsigned block_start, block_end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 sector_t block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 int nr_reads = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 int ret = 0;
2433 int is_mapped_to_disk = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434
Nick Piggin03158cd2007-10-16 01:25:25 -07002435 index = pos >> PAGE_CACHE_SHIFT;
2436 from = pos & (PAGE_CACHE_SIZE - 1);
2437 to = from + len;
2438
2439 page = __grab_cache_page(mapping, index);
2440 if (!page)
2441 return -ENOMEM;
2442 *pagep = page;
2443 *fsdata = NULL;
2444
2445 if (page_has_buffers(page)) {
2446 unlock_page(page);
2447 page_cache_release(page);
2448 *pagep = NULL;
2449 return block_write_begin(file, mapping, pos, len, flags, pagep,
2450 fsdata, get_block);
2451 }
Nick Piggina4b06722007-10-16 01:24:48 -07002452
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 if (PageMappedToDisk(page))
2454 return 0;
2455
Nick Piggina4b06722007-10-16 01:24:48 -07002456 /*
2457 * Allocate buffers so that we can keep track of state, and potentially
2458 * attach them to the page if an error occurs. In the common case of
2459 * no error, they will just be freed again without ever being attached
2460 * to the page (which is all OK, because we're under the page lock).
2461 *
2462 * Be careful: the buffer linked list is a NULL terminated one, rather
2463 * than the circular one we're used to.
2464 */
2465 head = alloc_page_buffers(page, blocksize, 0);
Nick Piggin03158cd2007-10-16 01:25:25 -07002466 if (!head) {
2467 ret = -ENOMEM;
2468 goto out_release;
2469 }
Nick Piggina4b06722007-10-16 01:24:48 -07002470
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472
2473 /*
2474 * We loop across all blocks in the page, whether or not they are
2475 * part of the affected region. This is so we can discover if the
2476 * page is fully mapped-to-disk.
2477 */
Nick Piggina4b06722007-10-16 01:24:48 -07002478 for (block_start = 0, block_in_page = 0, bh = head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 block_start < PAGE_CACHE_SIZE;
Nick Piggina4b06722007-10-16 01:24:48 -07002480 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 int create;
2482
Nick Piggina4b06722007-10-16 01:24:48 -07002483 block_end = block_start + blocksize;
2484 bh->b_state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 create = 1;
2486 if (block_start >= to)
2487 create = 0;
2488 ret = get_block(inode, block_in_file + block_in_page,
Nick Piggina4b06722007-10-16 01:24:48 -07002489 bh, create);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 if (ret)
2491 goto failed;
Nick Piggina4b06722007-10-16 01:24:48 -07002492 if (!buffer_mapped(bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 is_mapped_to_disk = 0;
Nick Piggina4b06722007-10-16 01:24:48 -07002494 if (buffer_new(bh))
2495 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2496 if (PageUptodate(page)) {
2497 set_buffer_uptodate(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 continue;
Nick Piggina4b06722007-10-16 01:24:48 -07002499 }
2500 if (buffer_new(bh) || !buffer_mapped(bh)) {
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002501 zero_user_segments(page, block_start, from,
2502 to, block_end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 continue;
2504 }
Nick Piggina4b06722007-10-16 01:24:48 -07002505 if (buffer_uptodate(bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 continue; /* reiserfs does this */
2507 if (block_start < from || block_end > to) {
Nick Piggina4b06722007-10-16 01:24:48 -07002508 lock_buffer(bh);
2509 bh->b_end_io = end_buffer_read_nobh;
2510 submit_bh(READ, bh);
2511 nr_reads++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 }
2513 }
2514
2515 if (nr_reads) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 /*
2517 * The page is locked, so these buffers are protected from
2518 * any VM or truncate activity. Hence we don't need to care
2519 * for the buffer_head refcounts.
2520 */
Nick Piggina4b06722007-10-16 01:24:48 -07002521 for (bh = head; bh; bh = bh->b_this_page) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 wait_on_buffer(bh);
2523 if (!buffer_uptodate(bh))
2524 ret = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 }
2526 if (ret)
2527 goto failed;
2528 }
2529
2530 if (is_mapped_to_disk)
2531 SetPageMappedToDisk(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532
Nick Piggin03158cd2007-10-16 01:25:25 -07002533 *fsdata = head; /* to be released by nobh_write_end */
Nick Piggina4b06722007-10-16 01:24:48 -07002534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 return 0;
2536
2537failed:
Nick Piggin03158cd2007-10-16 01:25:25 -07002538 BUG_ON(!ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 /*
Nick Piggina4b06722007-10-16 01:24:48 -07002540 * Error recovery is a bit difficult. We need to zero out blocks that
2541 * were newly allocated, and dirty them to ensure they get written out.
2542 * Buffers need to be attached to the page at this point, otherwise
2543 * the handling of potential IO errors during writeout would be hard
2544 * (could try doing synchronous writeout, but what if that fails too?)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 */
Nick Piggin03158cd2007-10-16 01:25:25 -07002546 attach_nobh_buffers(page, head);
2547 page_zero_new_buffers(page, from, to);
Nick Piggina4b06722007-10-16 01:24:48 -07002548
Nick Piggin03158cd2007-10-16 01:25:25 -07002549out_release:
2550 unlock_page(page);
2551 page_cache_release(page);
2552 *pagep = NULL;
Nick Piggina4b06722007-10-16 01:24:48 -07002553
Nick Piggin03158cd2007-10-16 01:25:25 -07002554 if (pos + len > inode->i_size)
2555 vmtruncate(inode, inode->i_size);
Nick Piggina4b06722007-10-16 01:24:48 -07002556
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 return ret;
2558}
Nick Piggin03158cd2007-10-16 01:25:25 -07002559EXPORT_SYMBOL(nobh_write_begin);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560
Nick Piggin03158cd2007-10-16 01:25:25 -07002561int nobh_write_end(struct file *file, struct address_space *mapping,
2562 loff_t pos, unsigned len, unsigned copied,
2563 struct page *page, void *fsdata)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564{
2565 struct inode *inode = page->mapping->host;
Nick Pigginefdc3132007-10-21 06:57:41 +02002566 struct buffer_head *head = fsdata;
Nick Piggin03158cd2007-10-16 01:25:25 -07002567 struct buffer_head *bh;
Dmitri Monakhov5b41e742008-03-28 14:15:52 -07002568 BUG_ON(fsdata != NULL && page_has_buffers(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569
Dmitri Monakhov5b41e742008-03-28 14:15:52 -07002570 if (unlikely(copied < len) && !page_has_buffers(page))
2571 attach_nobh_buffers(page, head);
2572 if (page_has_buffers(page))
2573 return generic_write_end(file, mapping, pos, len,
2574 copied, page, fsdata);
Nick Piggina4b06722007-10-16 01:24:48 -07002575
Nick Piggin22c8ca72007-02-20 13:58:09 -08002576 SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 set_page_dirty(page);
Nick Piggin03158cd2007-10-16 01:25:25 -07002578 if (pos+copied > inode->i_size) {
2579 i_size_write(inode, pos+copied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 mark_inode_dirty(inode);
2581 }
Nick Piggin03158cd2007-10-16 01:25:25 -07002582
2583 unlock_page(page);
2584 page_cache_release(page);
2585
Nick Piggin03158cd2007-10-16 01:25:25 -07002586 while (head) {
2587 bh = head;
2588 head = head->b_this_page;
2589 free_buffer_head(bh);
2590 }
2591
2592 return copied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593}
Nick Piggin03158cd2007-10-16 01:25:25 -07002594EXPORT_SYMBOL(nobh_write_end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595
2596/*
2597 * nobh_writepage() - based on block_full_write_page() except
2598 * that it tries to operate without attaching bufferheads to
2599 * the page.
2600 */
2601int nobh_writepage(struct page *page, get_block_t *get_block,
2602 struct writeback_control *wbc)
2603{
2604 struct inode * const inode = page->mapping->host;
2605 loff_t i_size = i_size_read(inode);
2606 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2607 unsigned offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 int ret;
2609
2610 /* Is the page fully inside i_size? */
2611 if (page->index < end_index)
2612 goto out;
2613
2614 /* Is the page fully outside i_size? (truncate in progress) */
2615 offset = i_size & (PAGE_CACHE_SIZE-1);
2616 if (page->index >= end_index+1 || !offset) {
2617 /*
2618 * The page may have dirty, unmapped buffers. For example,
2619 * they may have been added in ext3_writepage(). Make them
2620 * freeable here, so the page does not leak.
2621 */
2622#if 0
2623 /* Not really sure about this - do we need this ? */
2624 if (page->mapping->a_ops->invalidatepage)
2625 page->mapping->a_ops->invalidatepage(page, offset);
2626#endif
2627 unlock_page(page);
2628 return 0; /* don't care */
2629 }
2630
2631 /*
2632 * The page straddles i_size. It must be zeroed out on each and every
2633 * writepage invocation because it may be mmapped. "A file is mapped
2634 * in multiples of the page size. For a file that is not a multiple of
2635 * the page size, the remaining memory is zeroed when mapped, and
2636 * writes to that region are not written out to the file."
2637 */
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002638 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639out:
2640 ret = mpage_writepage(page, get_block, wbc);
2641 if (ret == -EAGAIN)
2642 ret = __block_write_full_page(inode, page, get_block, wbc);
2643 return ret;
2644}
2645EXPORT_SYMBOL(nobh_writepage);
2646
Nick Piggin03158cd2007-10-16 01:25:25 -07002647int nobh_truncate_page(struct address_space *mapping,
2648 loff_t from, get_block_t *get_block)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2651 unsigned offset = from & (PAGE_CACHE_SIZE-1);
Nick Piggin03158cd2007-10-16 01:25:25 -07002652 unsigned blocksize;
2653 sector_t iblock;
2654 unsigned length, pos;
2655 struct inode *inode = mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 struct page *page;
Nick Piggin03158cd2007-10-16 01:25:25 -07002657 struct buffer_head map_bh;
2658 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
Nick Piggin03158cd2007-10-16 01:25:25 -07002660 blocksize = 1 << inode->i_blkbits;
2661 length = offset & (blocksize - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662
Nick Piggin03158cd2007-10-16 01:25:25 -07002663 /* Block boundary? Nothing to do */
2664 if (!length)
2665 return 0;
2666
2667 length = blocksize - length;
2668 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2669
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 page = grab_cache_page(mapping, index);
Nick Piggin03158cd2007-10-16 01:25:25 -07002671 err = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 if (!page)
2673 goto out;
2674
Nick Piggin03158cd2007-10-16 01:25:25 -07002675 if (page_has_buffers(page)) {
2676has_buffers:
2677 unlock_page(page);
2678 page_cache_release(page);
2679 return block_truncate_page(mapping, from, get_block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 }
Nick Piggin03158cd2007-10-16 01:25:25 -07002681
2682 /* Find the buffer that contains "offset" */
2683 pos = blocksize;
2684 while (offset >= pos) {
2685 iblock++;
2686 pos += blocksize;
2687 }
2688
2689 err = get_block(inode, iblock, &map_bh, 0);
2690 if (err)
2691 goto unlock;
2692 /* unmapped? It's a hole - nothing to do */
2693 if (!buffer_mapped(&map_bh))
2694 goto unlock;
2695
2696 /* Ok, it's mapped. Make sure it's up-to-date */
2697 if (!PageUptodate(page)) {
2698 err = mapping->a_ops->readpage(NULL, page);
2699 if (err) {
2700 page_cache_release(page);
2701 goto out;
2702 }
2703 lock_page(page);
2704 if (!PageUptodate(page)) {
2705 err = -EIO;
2706 goto unlock;
2707 }
2708 if (page_has_buffers(page))
2709 goto has_buffers;
2710 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002711 zero_user(page, offset, length);
Nick Piggin03158cd2007-10-16 01:25:25 -07002712 set_page_dirty(page);
2713 err = 0;
2714
2715unlock:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 unlock_page(page);
2717 page_cache_release(page);
2718out:
Nick Piggin03158cd2007-10-16 01:25:25 -07002719 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720}
2721EXPORT_SYMBOL(nobh_truncate_page);
2722
2723int block_truncate_page(struct address_space *mapping,
2724 loff_t from, get_block_t *get_block)
2725{
2726 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2727 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2728 unsigned blocksize;
Andrew Morton54b21a72006-01-08 01:03:05 -08002729 sector_t iblock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 unsigned length, pos;
2731 struct inode *inode = mapping->host;
2732 struct page *page;
2733 struct buffer_head *bh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 int err;
2735
2736 blocksize = 1 << inode->i_blkbits;
2737 length = offset & (blocksize - 1);
2738
2739 /* Block boundary? Nothing to do */
2740 if (!length)
2741 return 0;
2742
2743 length = blocksize - length;
Andrew Morton54b21a72006-01-08 01:03:05 -08002744 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
2746 page = grab_cache_page(mapping, index);
2747 err = -ENOMEM;
2748 if (!page)
2749 goto out;
2750
2751 if (!page_has_buffers(page))
2752 create_empty_buffers(page, blocksize, 0);
2753
2754 /* Find the buffer that contains "offset" */
2755 bh = page_buffers(page);
2756 pos = blocksize;
2757 while (offset >= pos) {
2758 bh = bh->b_this_page;
2759 iblock++;
2760 pos += blocksize;
2761 }
2762
2763 err = 0;
2764 if (!buffer_mapped(bh)) {
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08002765 WARN_ON(bh->b_size != blocksize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 err = get_block(inode, iblock, bh, 0);
2767 if (err)
2768 goto unlock;
2769 /* unmapped? It's a hole - nothing to do */
2770 if (!buffer_mapped(bh))
2771 goto unlock;
2772 }
2773
2774 /* Ok, it's mapped. Make sure it's up-to-date */
2775 if (PageUptodate(page))
2776 set_buffer_uptodate(bh);
2777
David Chinner33a266d2007-02-12 00:51:41 -08002778 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 err = -EIO;
2780 ll_rw_block(READ, 1, &bh);
2781 wait_on_buffer(bh);
2782 /* Uhhuh. Read error. Complain and punt. */
2783 if (!buffer_uptodate(bh))
2784 goto unlock;
2785 }
2786
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002787 zero_user(page, offset, length);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 mark_buffer_dirty(bh);
2789 err = 0;
2790
2791unlock:
2792 unlock_page(page);
2793 page_cache_release(page);
2794out:
2795 return err;
2796}
2797
2798/*
2799 * The generic ->writepage function for buffer-backed address_spaces
2800 */
2801int block_write_full_page(struct page *page, get_block_t *get_block,
2802 struct writeback_control *wbc)
2803{
2804 struct inode * const inode = page->mapping->host;
2805 loff_t i_size = i_size_read(inode);
2806 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2807 unsigned offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808
2809 /* Is the page fully inside i_size? */
2810 if (page->index < end_index)
2811 return __block_write_full_page(inode, page, get_block, wbc);
2812
2813 /* Is the page fully outside i_size? (truncate in progress) */
2814 offset = i_size & (PAGE_CACHE_SIZE-1);
2815 if (page->index >= end_index+1 || !offset) {
2816 /*
2817 * The page may have dirty, unmapped buffers. For example,
2818 * they may have been added in ext3_writepage(). Make them
2819 * freeable here, so the page does not leak.
2820 */
Jan Karaaaa40592005-10-30 15:00:16 -08002821 do_invalidatepage(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 unlock_page(page);
2823 return 0; /* don't care */
2824 }
2825
2826 /*
2827 * The page straddles i_size. It must be zeroed out on each and every
2828 * writepage invokation because it may be mmapped. "A file is mapped
2829 * in multiples of the page size. For a file that is not a multiple of
2830 * the page size, the remaining memory is zeroed when mapped, and
2831 * writes to that region are not written out to the file."
2832 */
Christoph Lametereebd2aa2008-02-04 22:28:29 -08002833 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 return __block_write_full_page(inode, page, get_block, wbc);
2835}
2836
2837sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2838 get_block_t *get_block)
2839{
2840 struct buffer_head tmp;
2841 struct inode *inode = mapping->host;
2842 tmp.b_state = 0;
2843 tmp.b_blocknr = 0;
Badari Pulavartyb0cf2322006-03-26 01:38:00 -08002844 tmp.b_size = 1 << inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 get_block(inode, block, &tmp, 0);
2846 return tmp.b_blocknr;
2847}
2848
NeilBrown6712ecf2007-09-27 12:47:43 +02002849static void end_bio_bh_io_sync(struct bio *bio, int err)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850{
2851 struct buffer_head *bh = bio->bi_private;
2852
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 if (err == -EOPNOTSUPP) {
2854 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2855 set_bit(BH_Eopnotsupp, &bh->b_state);
2856 }
2857
2858 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2859 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860}
2861
2862int submit_bh(int rw, struct buffer_head * bh)
2863{
2864 struct bio *bio;
2865 int ret = 0;
2866
2867 BUG_ON(!buffer_locked(bh));
2868 BUG_ON(!buffer_mapped(bh));
2869 BUG_ON(!bh->b_end_io);
2870
2871 if (buffer_ordered(bh) && (rw == WRITE))
2872 rw = WRITE_BARRIER;
2873
2874 /*
2875 * Only clear out a write error when rewriting, should this
2876 * include WRITE_SYNC as well?
2877 */
2878 if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2879 clear_buffer_write_io_error(bh);
2880
2881 /*
2882 * from here on down, it's all bio -- do the initial mapping,
2883 * submit_bio -> generic_make_request may further map this bio around
2884 */
2885 bio = bio_alloc(GFP_NOIO, 1);
2886
2887 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2888 bio->bi_bdev = bh->b_bdev;
2889 bio->bi_io_vec[0].bv_page = bh->b_page;
2890 bio->bi_io_vec[0].bv_len = bh->b_size;
2891 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2892
2893 bio->bi_vcnt = 1;
2894 bio->bi_idx = 0;
2895 bio->bi_size = bh->b_size;
2896
2897 bio->bi_end_io = end_bio_bh_io_sync;
2898 bio->bi_private = bh;
2899
2900 bio_get(bio);
2901 submit_bio(rw, bio);
2902
2903 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2904 ret = -EOPNOTSUPP;
2905
2906 bio_put(bio);
2907 return ret;
2908}
2909
2910/**
2911 * ll_rw_block: low-level access to block devices (DEPRECATED)
Jan Karaa7662232005-09-06 15:19:10 -07002912 * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 * @nr: number of &struct buffer_heads in the array
2914 * @bhs: array of pointers to &struct buffer_head
2915 *
Jan Karaa7662232005-09-06 15:19:10 -07002916 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2917 * requests an I/O operation on them, either a %READ or a %WRITE. The third
2918 * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2919 * are sent to disk. The fourth %READA option is described in the documentation
2920 * for generic_make_request() which ll_rw_block() calls.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 *
2922 * This function drops any buffer that it cannot get a lock on (with the
Jan Karaa7662232005-09-06 15:19:10 -07002923 * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2924 * clean when doing a write request, and any buffer that appears to be
2925 * up-to-date when doing read request. Further it marks as clean buffers that
2926 * are processed for writing (the buffer cache won't assume that they are
2927 * actually clean until the buffer gets unlocked).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 *
2929 * ll_rw_block sets b_end_io to simple completion handler that marks
2930 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2931 * any waiters.
2932 *
2933 * All of the buffers must be for the same device, and must also be a
2934 * multiple of the current approved size for the device.
2935 */
2936void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2937{
2938 int i;
2939
2940 for (i = 0; i < nr; i++) {
2941 struct buffer_head *bh = bhs[i];
2942
Jan Karaa7662232005-09-06 15:19:10 -07002943 if (rw == SWRITE)
2944 lock_buffer(bh);
2945 else if (test_set_buffer_locked(bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 continue;
2947
Jan Karaa7662232005-09-06 15:19:10 -07002948 if (rw == WRITE || rw == SWRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 if (test_clear_buffer_dirty(bh)) {
akpm@osdl.org76c30732005-04-16 15:24:07 -07002950 bh->b_end_io = end_buffer_write_sync;
OGAWA Hirofumie60e5c52006-02-03 03:04:43 -08002951 get_bh(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 submit_bh(WRITE, bh);
2953 continue;
2954 }
2955 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 if (!buffer_uptodate(bh)) {
akpm@osdl.org76c30732005-04-16 15:24:07 -07002957 bh->b_end_io = end_buffer_read_sync;
OGAWA Hirofumie60e5c52006-02-03 03:04:43 -08002958 get_bh(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 submit_bh(rw, bh);
2960 continue;
2961 }
2962 }
2963 unlock_buffer(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 }
2965}
2966
2967/*
2968 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2969 * and then start new I/O and then wait upon it. The caller must have a ref on
2970 * the buffer_head.
2971 */
2972int sync_dirty_buffer(struct buffer_head *bh)
2973{
2974 int ret = 0;
2975
2976 WARN_ON(atomic_read(&bh->b_count) < 1);
2977 lock_buffer(bh);
2978 if (test_clear_buffer_dirty(bh)) {
2979 get_bh(bh);
2980 bh->b_end_io = end_buffer_write_sync;
2981 ret = submit_bh(WRITE, bh);
2982 wait_on_buffer(bh);
2983 if (buffer_eopnotsupp(bh)) {
2984 clear_buffer_eopnotsupp(bh);
2985 ret = -EOPNOTSUPP;
2986 }
2987 if (!ret && !buffer_uptodate(bh))
2988 ret = -EIO;
2989 } else {
2990 unlock_buffer(bh);
2991 }
2992 return ret;
2993}
2994
2995/*
2996 * try_to_free_buffers() checks if all the buffers on this particular page
2997 * are unused, and releases them if so.
2998 *
2999 * Exclusion against try_to_free_buffers may be obtained by either
3000 * locking the page or by holding its mapping's private_lock.
3001 *
3002 * If the page is dirty but all the buffers are clean then we need to
3003 * be sure to mark the page clean as well. This is because the page
3004 * may be against a block device, and a later reattachment of buffers
3005 * to a dirty page will set *all* buffers dirty. Which would corrupt
3006 * filesystem data on the same device.
3007 *
3008 * The same applies to regular filesystem pages: if all the buffers are
3009 * clean then we set the page clean and proceed. To do that, we require
3010 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3011 * private_lock.
3012 *
3013 * try_to_free_buffers() is non-blocking.
3014 */
3015static inline int buffer_busy(struct buffer_head *bh)
3016{
3017 return atomic_read(&bh->b_count) |
3018 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3019}
3020
3021static int
3022drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3023{
3024 struct buffer_head *head = page_buffers(page);
3025 struct buffer_head *bh;
3026
3027 bh = head;
3028 do {
akpm@osdl.orgde7d5a32005-05-01 08:58:39 -07003029 if (buffer_write_io_error(bh) && page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 set_bit(AS_EIO, &page->mapping->flags);
3031 if (buffer_busy(bh))
3032 goto failed;
3033 bh = bh->b_this_page;
3034 } while (bh != head);
3035
3036 do {
3037 struct buffer_head *next = bh->b_this_page;
3038
Jan Kara535ee2f2008-02-08 04:21:59 -08003039 if (bh->b_assoc_map)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 __remove_assoc_queue(bh);
3041 bh = next;
3042 } while (bh != head);
3043 *buffers_to_free = head;
3044 __clear_page_buffers(page);
3045 return 1;
3046failed:
3047 return 0;
3048}
3049
3050int try_to_free_buffers(struct page *page)
3051{
3052 struct address_space * const mapping = page->mapping;
3053 struct buffer_head *buffers_to_free = NULL;
3054 int ret = 0;
3055
3056 BUG_ON(!PageLocked(page));
Linus Torvaldsecdfc972007-01-26 12:47:06 -08003057 if (PageWriteback(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058 return 0;
3059
3060 if (mapping == NULL) { /* can this still happen? */
3061 ret = drop_buffers(page, &buffers_to_free);
3062 goto out;
3063 }
3064
3065 spin_lock(&mapping->private_lock);
3066 ret = drop_buffers(page, &buffers_to_free);
Linus Torvaldsecdfc972007-01-26 12:47:06 -08003067
3068 /*
3069 * If the filesystem writes its buffers by hand (eg ext3)
3070 * then we can have clean buffers against a dirty page. We
3071 * clean the page here; otherwise the VM will never notice
3072 * that the filesystem did any IO at all.
3073 *
3074 * Also, during truncate, discard_buffer will have marked all
3075 * the page's buffers clean. We discover that here and clean
3076 * the page also.
Nick Piggin87df7242007-01-30 14:36:27 +11003077 *
3078 * private_lock must be held over this entire operation in order
3079 * to synchronise against __set_page_dirty_buffers and prevent the
3080 * dirty bit from being lost.
Linus Torvaldsecdfc972007-01-26 12:47:06 -08003081 */
3082 if (ret)
3083 cancel_dirty_page(page, PAGE_CACHE_SIZE);
Nick Piggin87df7242007-01-30 14:36:27 +11003084 spin_unlock(&mapping->private_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085out:
3086 if (buffers_to_free) {
3087 struct buffer_head *bh = buffers_to_free;
3088
3089 do {
3090 struct buffer_head *next = bh->b_this_page;
3091 free_buffer_head(bh);
3092 bh = next;
3093 } while (bh != buffers_to_free);
3094 }
3095 return ret;
3096}
3097EXPORT_SYMBOL(try_to_free_buffers);
3098
NeilBrown3978d712006-03-26 01:37:17 -08003099void block_sync_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100{
3101 struct address_space *mapping;
3102
3103 smp_mb();
3104 mapping = page_mapping(page);
3105 if (mapping)
3106 blk_run_backing_dev(mapping->backing_dev_info, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107}
3108
3109/*
3110 * There are no bdflush tunables left. But distributions are
3111 * still running obsolete flush daemons, so we terminate them here.
3112 *
3113 * Use of bdflush() is deprecated and will be removed in a future kernel.
3114 * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3115 */
3116asmlinkage long sys_bdflush(int func, long data)
3117{
3118 static int msg_count;
3119
3120 if (!capable(CAP_SYS_ADMIN))
3121 return -EPERM;
3122
3123 if (msg_count < 5) {
3124 msg_count++;
3125 printk(KERN_INFO
3126 "warning: process `%s' used the obsolete bdflush"
3127 " system call\n", current->comm);
3128 printk(KERN_INFO "Fix your initscripts?\n");
3129 }
3130
3131 if (func == 1)
3132 do_exit(0);
3133 return 0;
3134}
3135
3136/*
3137 * Buffer-head allocation
3138 */
Christoph Lametere18b8902006-12-06 20:33:20 -08003139static struct kmem_cache *bh_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140
3141/*
3142 * Once the number of bh's in the machine exceeds this level, we start
3143 * stripping them in writeback.
3144 */
3145static int max_buffer_heads;
3146
3147int buffer_heads_over_limit;
3148
3149struct bh_accounting {
3150 int nr; /* Number of live bh's */
3151 int ratelimit; /* Limit cacheline bouncing */
3152};
3153
3154static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3155
3156static void recalc_bh_state(void)
3157{
3158 int i;
3159 int tot = 0;
3160
3161 if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3162 return;
3163 __get_cpu_var(bh_accounting).ratelimit = 0;
Eric Dumazet8a143422006-03-24 03:18:10 -08003164 for_each_online_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 tot += per_cpu(bh_accounting, i).nr;
3166 buffer_heads_over_limit = (tot > max_buffer_heads);
3167}
3168
Al Virodd0fc662005-10-07 07:46:04 +01003169struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170{
Christoph Lameter488514d2008-04-28 02:12:05 -07003171 struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 if (ret) {
Christoph Lametera35afb82007-05-16 22:10:57 -07003173 INIT_LIST_HEAD(&ret->b_assoc_buffers);
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07003174 get_cpu_var(bh_accounting).nr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 recalc_bh_state();
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07003176 put_cpu_var(bh_accounting);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 }
3178 return ret;
3179}
3180EXPORT_SYMBOL(alloc_buffer_head);
3181
3182void free_buffer_head(struct buffer_head *bh)
3183{
3184 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3185 kmem_cache_free(bh_cachep, bh);
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07003186 get_cpu_var(bh_accounting).nr--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 recalc_bh_state();
Coywolf Qi Hunt736c7b82005-09-06 15:18:17 -07003188 put_cpu_var(bh_accounting);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189}
3190EXPORT_SYMBOL(free_buffer_head);
3191
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192static void buffer_exit_cpu(int cpu)
3193{
3194 int i;
3195 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3196
3197 for (i = 0; i < BH_LRU_SIZE; i++) {
3198 brelse(b->bhs[i]);
3199 b->bhs[i] = NULL;
3200 }
Eric Dumazet8a143422006-03-24 03:18:10 -08003201 get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3202 per_cpu(bh_accounting, cpu).nr = 0;
3203 put_cpu_var(bh_accounting);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204}
3205
3206static int buffer_cpu_notify(struct notifier_block *self,
3207 unsigned long action, void *hcpu)
3208{
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003209 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 buffer_exit_cpu((unsigned long)hcpu);
3211 return NOTIFY_OK;
3212}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213
Aneesh Kumar K.V389d1b02008-01-28 23:58:26 -05003214/**
Randy Dunlapa6b91912008-03-19 17:01:00 -07003215 * bh_uptodate_or_lock - Test whether the buffer is uptodate
Aneesh Kumar K.V389d1b02008-01-28 23:58:26 -05003216 * @bh: struct buffer_head
3217 *
3218 * Return true if the buffer is up-to-date and false,
3219 * with the buffer locked, if not.
3220 */
3221int bh_uptodate_or_lock(struct buffer_head *bh)
3222{
3223 if (!buffer_uptodate(bh)) {
3224 lock_buffer(bh);
3225 if (!buffer_uptodate(bh))
3226 return 0;
3227 unlock_buffer(bh);
3228 }
3229 return 1;
3230}
3231EXPORT_SYMBOL(bh_uptodate_or_lock);
3232
3233/**
Randy Dunlapa6b91912008-03-19 17:01:00 -07003234 * bh_submit_read - Submit a locked buffer for reading
Aneesh Kumar K.V389d1b02008-01-28 23:58:26 -05003235 * @bh: struct buffer_head
3236 *
3237 * Returns zero on success and -EIO on error.
3238 */
3239int bh_submit_read(struct buffer_head *bh)
3240{
3241 BUG_ON(!buffer_locked(bh));
3242
3243 if (buffer_uptodate(bh)) {
3244 unlock_buffer(bh);
3245 return 0;
3246 }
3247
3248 get_bh(bh);
3249 bh->b_end_io = end_buffer_read_sync;
3250 submit_bh(READ, bh);
3251 wait_on_buffer(bh);
3252 if (buffer_uptodate(bh))
3253 return 0;
3254 return -EIO;
3255}
3256EXPORT_SYMBOL(bh_submit_read);
3257
Christoph Lameterb98938c2008-02-04 22:28:36 -08003258static void
3259init_buffer_head(struct kmem_cache *cachep, void *data)
3260{
3261 struct buffer_head *bh = data;
3262
3263 memset(bh, 0, sizeof(*bh));
3264 INIT_LIST_HEAD(&bh->b_assoc_buffers);
3265}
3266
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267void __init buffer_init(void)
3268{
3269 int nrpages;
3270
Christoph Lameterb98938c2008-02-04 22:28:36 -08003271 bh_cachep = kmem_cache_create("buffer_head",
3272 sizeof(struct buffer_head), 0,
3273 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3274 SLAB_MEM_SPREAD),
3275 init_buffer_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276
3277 /*
3278 * Limit the bh occupancy to 10% of ZONE_NORMAL
3279 */
3280 nrpages = (nr_free_buffer_pages() * 10) / 100;
3281 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3282 hotcpu_notifier(buffer_cpu_notify, 0);
3283}
3284
3285EXPORT_SYMBOL(__bforget);
3286EXPORT_SYMBOL(__brelse);
3287EXPORT_SYMBOL(__wait_on_buffer);
3288EXPORT_SYMBOL(block_commit_write);
3289EXPORT_SYMBOL(block_prepare_write);
David Chinner54171692007-07-19 17:39:55 +10003290EXPORT_SYMBOL(block_page_mkwrite);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291EXPORT_SYMBOL(block_read_full_page);
3292EXPORT_SYMBOL(block_sync_page);
3293EXPORT_SYMBOL(block_truncate_page);
3294EXPORT_SYMBOL(block_write_full_page);
Nick Piggin89e10782007-10-16 01:25:07 -07003295EXPORT_SYMBOL(cont_write_begin);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296EXPORT_SYMBOL(end_buffer_read_sync);
3297EXPORT_SYMBOL(end_buffer_write_sync);
3298EXPORT_SYMBOL(file_fsync);
3299EXPORT_SYMBOL(fsync_bdev);
3300EXPORT_SYMBOL(generic_block_bmap);
OGAWA Hirofumi05eb0b52006-01-08 01:02:13 -08003301EXPORT_SYMBOL(generic_cont_expand_simple);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302EXPORT_SYMBOL(init_buffer);
3303EXPORT_SYMBOL(invalidate_bdev);
3304EXPORT_SYMBOL(ll_rw_block);
3305EXPORT_SYMBOL(mark_buffer_dirty);
3306EXPORT_SYMBOL(submit_bh);
3307EXPORT_SYMBOL(sync_dirty_buffer);
3308EXPORT_SYMBOL(unlock_buffer);