blob: ae45f77765c0ee65a647fc81c899315de2f0862f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
11 * 10Apr2002 akpm@zip.com.au
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020017#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
19#include <linux/sched.h>
20#include <linux/fs.h>
21#include <linux/mm.h>
22#include <linux/writeback.h>
23#include <linux/blkdev.h>
24#include <linux/backing-dev.h>
25#include <linux/buffer_head.h>
David Howells07f3f052006-09-30 20:52:18 +020026#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070027
Adrian Bunkf11b00f2008-04-29 00:58:56 -070028
29/**
30 * writeback_acquire - attempt to get exclusive writeback access to a device
31 * @bdi: the device's backing_dev_info structure
32 *
33 * It is a waste of resources to have more than one pdflush thread blocked on
34 * a single request queue. Exclusion at the request_queue level is obtained
35 * via a flag in the request_queue's backing_dev_info.state.
36 *
37 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
38 * unless they implement their own. Which is somewhat inefficient, as this
39 * may prevent concurrent writeback against multiple devices.
40 */
41static int writeback_acquire(struct backing_dev_info *bdi)
42{
43 return !test_and_set_bit(BDI_pdflush, &bdi->state);
44}
45
46/**
47 * writeback_in_progress - determine whether there is writeback in progress
48 * @bdi: the device's backing_dev_info structure.
49 *
50 * Determine whether there is writeback in progress against a backing device.
51 */
52int writeback_in_progress(struct backing_dev_info *bdi)
53{
54 return test_bit(BDI_pdflush, &bdi->state);
55}
56
57/**
58 * writeback_release - relinquish exclusive writeback access against a device.
59 * @bdi: the device's backing_dev_info structure
60 */
61static void writeback_release(struct backing_dev_info *bdi)
62{
63 BUG_ON(!writeback_in_progress(bdi));
64 clear_bit(BDI_pdflush, &bdi->state);
65}
66
Linus Torvalds1da177e2005-04-16 15:20:36 -070067/**
68 * __mark_inode_dirty - internal function
69 * @inode: inode to mark
70 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
71 * Mark an inode as dirty. Callers should use mark_inode_dirty or
72 * mark_inode_dirty_sync.
73 *
74 * Put the inode on the super block's dirty list.
75 *
76 * CAREFUL! We mark it dirty unconditionally, but move it onto the
77 * dirty list only if it is hashed or if it refers to a blockdev.
78 * If it was not hashed, it will never be added to the dirty list
79 * even if it is later hashed, as it will have been marked dirty already.
80 *
81 * In short, make sure you hash any inodes _before_ you start marking
82 * them dirty.
83 *
84 * This function *must* be atomic for the I_DIRTY_PAGES case -
85 * set_page_dirty() is called under spinlock in several places.
86 *
87 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
88 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
89 * the kernel-internal blockdev inode represents the dirtying time of the
90 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
91 * page->mapping->host, so the page-dirtying time is recorded in the internal
92 * blockdev inode.
93 */
94void __mark_inode_dirty(struct inode *inode, int flags)
95{
96 struct super_block *sb = inode->i_sb;
97
98 /*
99 * Don't do this for I_DIRTY_PAGES - that doesn't actually
100 * dirty the inode itself
101 */
102 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
103 if (sb->s_op->dirty_inode)
104 sb->s_op->dirty_inode(inode);
105 }
106
107 /*
108 * make sure that changes are seen by all cpus before we test i_state
109 * -- mikulas
110 */
111 smp_mb();
112
113 /* avoid the locking if we can */
114 if ((inode->i_state & flags) == flags)
115 return;
116
117 if (unlikely(block_dump)) {
118 struct dentry *dentry = NULL;
119 const char *name = "?";
120
121 if (!list_empty(&inode->i_dentry)) {
122 dentry = list_entry(inode->i_dentry.next,
123 struct dentry, d_alias);
124 if (dentry && dentry->d_name.name)
125 name = (const char *) dentry->d_name.name;
126 }
127
128 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev"))
129 printk(KERN_DEBUG
130 "%s(%d): dirtied inode %lu (%s) on %s\n",
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -0700131 current->comm, task_pid_nr(current), inode->i_ino,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 name, inode->i_sb->s_id);
133 }
134
135 spin_lock(&inode_lock);
136 if ((inode->i_state & flags) != flags) {
137 const int was_dirty = inode->i_state & I_DIRTY;
138
139 inode->i_state |= flags;
140
141 /*
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700142 * If the inode is being synced, just update its dirty state.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 * The unlocker will place the inode on the appropriate
144 * superblock list, based upon its state.
145 */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700146 if (inode->i_state & I_SYNC)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 goto out;
148
149 /*
150 * Only add valid (hashed) inodes to the superblock's
151 * dirty list. Add blockdev inodes as well.
152 */
153 if (!S_ISBLK(inode->i_mode)) {
154 if (hlist_unhashed(&inode->i_hash))
155 goto out;
156 }
157 if (inode->i_state & (I_FREEING|I_CLEAR))
158 goto out;
159
160 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700161 * If the inode was already on s_dirty/s_io/s_more_io, don't
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 * reposition it (that would break s_dirty time-ordering).
163 */
164 if (!was_dirty) {
165 inode->dirtied_when = jiffies;
166 list_move(&inode->i_list, &sb->s_dirty);
167 }
168 }
169out:
170 spin_unlock(&inode_lock);
171}
172
173EXPORT_SYMBOL(__mark_inode_dirty);
174
175static int write_inode(struct inode *inode, int sync)
176{
177 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
178 return inode->i_sb->s_op->write_inode(inode, sync);
179 return 0;
180}
181
182/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700183 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
184 * furthest end of its superblock's dirty-inode list.
185 *
186 * Before stamping the inode's ->dirtied_when, we check to see whether it is
187 * already the most-recently-dirtied inode on the s_dirty list. If that is
188 * the case then the inode must have been redirtied while it was being written
189 * out and we don't reset its dirtied_when.
190 */
191static void redirty_tail(struct inode *inode)
192{
193 struct super_block *sb = inode->i_sb;
194
195 if (!list_empty(&sb->s_dirty)) {
196 struct inode *tail_inode;
197
198 tail_inode = list_entry(sb->s_dirty.next, struct inode, i_list);
199 if (!time_after_eq(inode->dirtied_when,
200 tail_inode->dirtied_when))
201 inode->dirtied_when = jiffies;
202 }
203 list_move(&inode->i_list, &sb->s_dirty);
204}
205
206/*
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700207 * requeue inode for re-scanning after sb->s_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700208 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700209static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700210{
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700211 list_move(&inode->i_list, &inode->i_sb->s_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700212}
213
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700214static void inode_sync_complete(struct inode *inode)
215{
216 /*
217 * Prevent speculative execution through spin_unlock(&inode_lock);
218 */
219 smp_mb();
220 wake_up_bit(&inode->i_state, __I_SYNC);
221}
222
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700223/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700224 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
225 */
226static void move_expired_inodes(struct list_head *delaying_queue,
227 struct list_head *dispatch_queue,
228 unsigned long *older_than_this)
229{
230 while (!list_empty(delaying_queue)) {
231 struct inode *inode = list_entry(delaying_queue->prev,
232 struct inode, i_list);
233 if (older_than_this &&
234 time_after(inode->dirtied_when, *older_than_this))
235 break;
236 list_move(&inode->i_list, dispatch_queue);
237 }
238}
239
240/*
241 * Queue all expired dirty inodes for io, eldest first.
242 */
243static void queue_io(struct super_block *sb,
244 unsigned long *older_than_this)
245{
246 list_splice_init(&sb->s_more_io, sb->s_io.prev);
247 move_expired_inodes(&sb->s_dirty, &sb->s_io, older_than_this);
248}
249
Fengguang Wu08d8e972007-10-16 23:30:39 -0700250int sb_has_dirty_inodes(struct super_block *sb)
251{
252 return !list_empty(&sb->s_dirty) ||
253 !list_empty(&sb->s_io) ||
254 !list_empty(&sb->s_more_io);
255}
256EXPORT_SYMBOL(sb_has_dirty_inodes);
257
Fengguang Wu2c136572007-10-16 23:30:39 -0700258/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 * Write a single inode's dirty pages and inode data out to disk.
260 * If `wait' is set, wait on the writeout.
261 *
262 * The whole writeout design is quite complex and fragile. We want to avoid
263 * starvation of particular inodes when others are being redirtied, prevent
264 * livelocks, etc.
265 *
266 * Called under inode_lock.
267 */
268static int
269__sync_single_inode(struct inode *inode, struct writeback_control *wbc)
270{
271 unsigned dirty;
272 struct address_space *mapping = inode->i_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273 int wait = wbc->sync_mode == WB_SYNC_ALL;
274 int ret;
275
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700276 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700278 /* Set I_SYNC, reset I_DIRTY */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 dirty = inode->i_state & I_DIRTY;
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700280 inode->i_state |= I_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 inode->i_state &= ~I_DIRTY;
282
283 spin_unlock(&inode_lock);
284
285 ret = do_writepages(mapping, wbc);
286
287 /* Don't write the inode if only I_DIRTY_PAGES was set */
288 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
289 int err = write_inode(inode, wait);
290 if (ret == 0)
291 ret = err;
292 }
293
294 if (wait) {
295 int err = filemap_fdatawait(mapping);
296 if (ret == 0)
297 ret = err;
298 }
299
300 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700301 inode->i_state &= ~I_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 if (!(inode->i_state & I_FREEING)) {
303 if (!(inode->i_state & I_DIRTY) &&
304 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
305 /*
306 * We didn't write back all the pages. nfs_writepages()
307 * sometimes bales out without doing anything. Redirty
Fengguang Wu2c136572007-10-16 23:30:39 -0700308 * the inode; Move it from s_io onto s_more_io/s_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700309 */
310 /*
311 * akpm: if the caller was the kupdate function we put
312 * this inode at the head of s_dirty so it gets first
313 * consideration. Otherwise, move it to the tail, for
314 * the reasons described there. I'm not really sure
315 * how much sense this makes. Presumably I had a good
316 * reasons for doing it this way, and I'd rather not
317 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 */
319 if (wbc->for_kupdate) {
320 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700321 * For the kupdate function we move the inode
322 * to s_more_io so it will get more writeout as
323 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 */
325 inode->i_state |= I_DIRTY_PAGES;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800326 if (wbc->nr_to_write <= 0) {
327 /*
328 * slice used up: queue for next turn
329 */
330 requeue_io(inode);
331 } else {
332 /*
333 * somehow blocked: retry later
334 */
335 redirty_tail(inode);
336 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 } else {
338 /*
339 * Otherwise fully redirty the inode so that
340 * other inodes on this superblock will get some
341 * writeout. Otherwise heavy writing to one
342 * file would indefinitely suspend writeout of
343 * all the other files.
344 */
345 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700346 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 }
348 } else if (inode->i_state & I_DIRTY) {
349 /*
350 * Someone redirtied the inode while were writing back
351 * the pages.
352 */
Andrew Morton6610a0b2007-10-16 23:30:32 -0700353 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 } else if (atomic_read(&inode->i_count)) {
355 /*
356 * The inode is clean, inuse
357 */
358 list_move(&inode->i_list, &inode_in_use);
359 } else {
360 /*
361 * The inode is clean, unused
362 */
363 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 }
365 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700366 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 return ret;
368}
369
370/*
Andrea Arcangeli7f04c262005-10-30 15:03:05 -0800371 * Write out an inode's dirty pages. Called under inode_lock. Either the
372 * caller has ref on the inode (either via __iget or via syscall against an fd)
373 * or the inode has I_WILL_FREE set (via generic_forget_inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 */
375static int
Andrea Arcangeli7f04c262005-10-30 15:03:05 -0800376__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377{
378 wait_queue_head_t *wqh;
379
Andrea Arcangeli7f04c262005-10-30 15:03:05 -0800380 if (!atomic_read(&inode->i_count))
Andrea Arcangeli659603e2005-10-31 14:08:54 -0800381 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
Andrea Arcangeli7f04c262005-10-30 15:03:05 -0800382 else
383 WARN_ON(inode->i_state & I_WILL_FREE);
384
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700385 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_SYNC)) {
Andrew Morton65cb9b42007-10-16 23:30:37 -0700386 /*
387 * We're skipping this inode because it's locked, and we're not
Fengguang Wu2c136572007-10-16 23:30:39 -0700388 * doing writeback-for-data-integrity. Move it to s_more_io so
389 * that writeback can proceed with the other inodes on s_io.
390 * We'll have another go at writing back this inode when we
391 * completed a full scan of s_io.
Andrew Morton65cb9b42007-10-16 23:30:37 -0700392 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700393 requeue_io(inode);
Qi Yong2d544562008-02-04 22:29:23 -0800394 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 }
396
397 /*
398 * It's a data-integrity sync. We must wait.
399 */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700400 if (inode->i_state & I_SYNC) {
401 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700403 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 spin_unlock(&inode_lock);
406 __wait_on_bit(wqh, &wq, inode_wait,
407 TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700409 } while (inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 }
411 return __sync_single_inode(inode, wbc);
412}
413
414/*
415 * Write out a superblock's list of dirty inodes. A wait will be performed
416 * upon no inodes, all inodes or the final one, depending upon sync_mode.
417 *
418 * If older_than_this is non-NULL, then only write out inodes which
419 * had their first dirtying at a time earlier than *older_than_this.
420 *
421 * If we're a pdlfush thread, then implement pdflush collision avoidance
422 * against the entire list.
423 *
424 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
425 * that it can be located for waiting on in __writeback_single_inode().
426 *
427 * Called under inode_lock.
428 *
429 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
430 * This function assumes that the blockdev superblock's inodes are backed by
431 * a variety of queues, so all inodes are searched. For other superblocks,
432 * assume that all inodes are backed by the same queue.
433 *
434 * FIXME: this linear search could get expensive with many fileystems. But
435 * how to fix? We need to go from an address_space to all inodes which share
436 * a queue with that address_space. (Easy: have a global "dirty superblocks"
437 * list).
438 *
439 * The inodes to be written are parked on sb->s_io. They are moved back onto
440 * sb->s_dirty as they are selected for writing. This way, none can be missed
441 * on the writer throttling path, and we get decent balancing between many
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700442 * throttled threads: we don't want them all piling up on inode_sync_wait.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 */
444static void
445sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
446{
447 const unsigned long start = jiffies; /* livelock avoidance */
448
449 if (!wbc->for_kupdate || list_empty(&sb->s_io))
Fengguang Wu2c136572007-10-16 23:30:39 -0700450 queue_io(sb, wbc->older_than_this);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451
452 while (!list_empty(&sb->s_io)) {
453 struct inode *inode = list_entry(sb->s_io.prev,
454 struct inode, i_list);
455 struct address_space *mapping = inode->i_mapping;
456 struct backing_dev_info *bdi = mapping->backing_dev_info;
457 long pages_skipped;
458
459 if (!bdi_cap_writeback_dirty(bdi)) {
Andrew Morton9852a0e72007-10-16 23:30:33 -0700460 redirty_tail(inode);
David Howells7b0de422006-08-29 19:06:07 +0100461 if (sb_is_blkdev_sb(sb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 /*
463 * Dirty memory-backed blockdev: the ramdisk
464 * driver does this. Skip just this inode
465 */
466 continue;
467 }
468 /*
469 * Dirty memory-backed inode against a filesystem other
470 * than the kernel-internal bdev filesystem. Skip the
471 * entire superblock.
472 */
473 break;
474 }
475
476 if (wbc->nonblocking && bdi_write_congested(bdi)) {
477 wbc->encountered_congestion = 1;
David Howells7b0de422006-08-29 19:06:07 +0100478 if (!sb_is_blkdev_sb(sb))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 break; /* Skip a congested fs */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700480 requeue_io(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 continue; /* Skip a congested blockdev */
482 }
483
484 if (wbc->bdi && bdi != wbc->bdi) {
David Howells7b0de422006-08-29 19:06:07 +0100485 if (!sb_is_blkdev_sb(sb))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 break; /* fs has the wrong queue */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700487 requeue_io(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 continue; /* blockdev has wrong queue */
489 }
490
491 /* Was this inode dirtied after sync_sb_inodes was called? */
492 if (time_after(inode->dirtied_when, start))
493 break;
494
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 /* Is another pdflush already flushing this queue? */
496 if (current_is_pdflush() && !writeback_acquire(bdi))
497 break;
498
499 BUG_ON(inode->i_state & I_FREEING);
500 __iget(inode);
501 pages_skipped = wbc->pages_skipped;
502 __writeback_single_inode(inode, wbc);
503 if (wbc->sync_mode == WB_SYNC_HOLD) {
504 inode->dirtied_when = jiffies;
505 list_move(&inode->i_list, &sb->s_dirty);
506 }
507 if (current_is_pdflush())
508 writeback_release(bdi);
509 if (wbc->pages_skipped != pages_skipped) {
510 /*
511 * writeback is not making progress due to locked
512 * buffers. Skip this inode for now.
513 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700514 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 }
516 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800518 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800520 if (wbc->nr_to_write <= 0) {
521 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800523 }
524 if (!list_empty(&sb->s_more_io))
525 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 }
527 return; /* Leave any unwritten inodes on s_io */
528}
529
530/*
531 * Start writeback of dirty pagecache data against all unlocked inodes.
532 *
533 * Note:
534 * We don't need to grab a reference to superblock here. If it has non-empty
535 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
Fengguang Wu2c136572007-10-16 23:30:39 -0700536 * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
538 * inode from superblock lists we are OK.
539 *
540 * If `older_than_this' is non-zero then only flush inodes which have a
541 * flushtime older than *older_than_this.
542 *
543 * If `bdi' is non-zero then we will scan the first inode against each
544 * superblock until we find the matching ones. One group will be the dirty
545 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
546 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
547 * super-efficient but we're about to do a ton of I/O...
548 */
549void
550writeback_inodes(struct writeback_control *wbc)
551{
552 struct super_block *sb;
553
554 might_sleep();
555 spin_lock(&sb_lock);
556restart:
Akinobu Mita797074e2008-02-06 01:37:08 -0800557 list_for_each_entry_reverse(sb, &super_blocks, s_list) {
Fengguang Wu08d8e972007-10-16 23:30:39 -0700558 if (sb_has_dirty_inodes(sb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 /* we're making our own get_super here */
560 sb->s_count++;
561 spin_unlock(&sb_lock);
562 /*
563 * If we can't get the readlock, there's no sense in
564 * waiting around, most of the time the FS is going to
565 * be unmounted by the time it is released.
566 */
567 if (down_read_trylock(&sb->s_umount)) {
568 if (sb->s_root) {
569 spin_lock(&inode_lock);
570 sync_sb_inodes(sb, wbc);
571 spin_unlock(&inode_lock);
572 }
573 up_read(&sb->s_umount);
574 }
575 spin_lock(&sb_lock);
576 if (__put_super_and_need_restart(sb))
577 goto restart;
578 }
579 if (wbc->nr_to_write <= 0)
580 break;
581 }
582 spin_unlock(&sb_lock);
583}
584
585/*
586 * writeback and wait upon the filesystem's dirty inodes. The caller will
587 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
588 * used to park the written inodes on sb->s_dirty for the wait pass.
589 *
590 * A finite limit is set on the number of pages which will be written.
591 * To prevent infinite livelock of sys_sync().
592 *
593 * We add in the number of potentially dirty inodes, because each inode write
594 * can dirty pagecache in the underlying blockdev.
595 */
596void sync_inodes_sb(struct super_block *sb, int wait)
597{
598 struct writeback_control wbc = {
599 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700600 .range_start = 0,
601 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 };
Christoph Lameterb1e7a8f2006-06-30 01:55:39 -0700603 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
Christoph Lameterfd39fc82006-06-30 01:55:40 -0700604 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
606 wbc.nr_to_write = nr_dirty + nr_unstable +
607 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
608 nr_dirty + nr_unstable;
609 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
610 spin_lock(&inode_lock);
611 sync_sb_inodes(sb, &wbc);
612 spin_unlock(&inode_lock);
613}
614
615/*
616 * Rather lame livelock avoidance.
617 */
618static void set_sb_syncing(int val)
619{
620 struct super_block *sb;
621 spin_lock(&sb_lock);
Akinobu Mita797074e2008-02-06 01:37:08 -0800622 list_for_each_entry_reverse(sb, &super_blocks, s_list)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 sb->s_syncing = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 spin_unlock(&sb_lock);
625}
626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627/**
Martin Waitz67be2dd2005-05-01 08:59:26 -0700628 * sync_inodes - writes all inodes to disk
629 * @wait: wait for completion
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 *
631 * sync_inodes() goes through each super block's dirty inode list, writes the
632 * inodes out, waits on the writeout and puts the inodes back on the normal
633 * list.
634 *
635 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
636 * part of the sync functions is that the blockdev "superblock" is processed
637 * last. This is because the write_inode() function of a typical fs will
638 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
639 * What we want to do is to perform all that dirtying first, and then write
640 * back all those inode blocks via the blockdev mapping in one sweep. So the
641 * additional (somewhat redundant) sync_blockdev() calls here are to make
642 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
643 * outstanding dirty inodes, the writeback goes block-at-a-time within the
644 * filesystem's write_inode(). This is extremely slow.
645 */
Kirill Korotaev618f0632005-06-23 00:09:54 -0700646static void __sync_inodes(int wait)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647{
648 struct super_block *sb;
649
Kirill Korotaev618f0632005-06-23 00:09:54 -0700650 spin_lock(&sb_lock);
651restart:
652 list_for_each_entry(sb, &super_blocks, s_list) {
653 if (sb->s_syncing)
654 continue;
655 sb->s_syncing = 1;
656 sb->s_count++;
657 spin_unlock(&sb_lock);
658 down_read(&sb->s_umount);
659 if (sb->s_root) {
660 sync_inodes_sb(sb, wait);
661 sync_blockdev(sb->s_bdev);
662 }
663 up_read(&sb->s_umount);
664 spin_lock(&sb_lock);
665 if (__put_super_and_need_restart(sb))
666 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 }
Kirill Korotaev618f0632005-06-23 00:09:54 -0700668 spin_unlock(&sb_lock);
669}
670
671void sync_inodes(int wait)
672{
673 set_sb_syncing(0);
674 __sync_inodes(0);
675
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 if (wait) {
677 set_sb_syncing(0);
Kirill Korotaev618f0632005-06-23 00:09:54 -0700678 __sync_inodes(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 }
680}
681
682/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -0800683 * write_inode_now - write an inode to disk
684 * @inode: inode to write to disk
685 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -0800687 * This function commits an inode to disk immediately if it is dirty. This is
688 * primarily needed by knfsd.
689 *
690 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692int write_inode_now(struct inode *inode, int sync)
693{
694 int ret;
695 struct writeback_control wbc = {
696 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -0800697 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700698 .range_start = 0,
699 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 };
701
702 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -0800703 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
705 might_sleep();
706 spin_lock(&inode_lock);
707 ret = __writeback_single_inode(inode, &wbc);
708 spin_unlock(&inode_lock);
709 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700710 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 return ret;
712}
713EXPORT_SYMBOL(write_inode_now);
714
715/**
716 * sync_inode - write an inode and its pages to disk.
717 * @inode: the inode to sync
718 * @wbc: controls the writeback mode
719 *
720 * sync_inode() will write an inode and its pages to disk. It will also
721 * correctly update the inode on its superblock's dirty inode lists and will
722 * update inode->i_state.
723 *
724 * The caller must have a ref on the inode.
725 */
726int sync_inode(struct inode *inode, struct writeback_control *wbc)
727{
728 int ret;
729
730 spin_lock(&inode_lock);
731 ret = __writeback_single_inode(inode, wbc);
732 spin_unlock(&inode_lock);
733 return ret;
734}
735EXPORT_SYMBOL(sync_inode);
736
737/**
738 * generic_osync_inode - flush all dirty data for a given inode to disk
739 * @inode: inode to write
Martin Waitz67be2dd2005-05-01 08:59:26 -0700740 * @mapping: the address_space that should be flushed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 * @what: what to write and wait upon
742 *
743 * This can be called by file_write functions for files which have the
744 * O_SYNC flag set, to flush dirty writes to disk.
745 *
746 * @what is a bitmask, specifying which part of the inode's data should be
Randy Dunlapb8887e62005-11-07 01:01:07 -0800747 * written and waited upon.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 *
749 * OSYNC_DATA: i_mapping's dirty data
750 * OSYNC_METADATA: the buffers at i_mapping->private_list
751 * OSYNC_INODE: the inode itself
752 */
753
754int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
755{
756 int err = 0;
757 int need_write_inode_now = 0;
758 int err2;
759
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 if (what & OSYNC_DATA)
761 err = filemap_fdatawrite(mapping);
762 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
763 err2 = sync_mapping_buffers(mapping);
764 if (!err)
765 err = err2;
766 }
767 if (what & OSYNC_DATA) {
768 err2 = filemap_fdatawait(mapping);
769 if (!err)
770 err = err2;
771 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772
773 spin_lock(&inode_lock);
774 if ((inode->i_state & I_DIRTY) &&
775 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
776 need_write_inode_now = 1;
777 spin_unlock(&inode_lock);
778
779 if (need_write_inode_now) {
780 err2 = write_inode_now(inode, 1);
781 if (!err)
782 err = err2;
783 }
784 else
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700785 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
787 return err;
788}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789EXPORT_SYMBOL(generic_osync_inode);