blob: 15e375bf93e6d4fbf1f4bb977d78c83a62a96f85 [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 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * 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>
Jens Axboe03ba3782009-09-09 09:08:54 +020022#include <linux/kthread.h>
23#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/backing-dev.h>
27#include <linux/buffer_head.h>
David Howells07f3f052006-09-30 20:52:18 +020028#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
Jens Axboe66f3b8e2009-09-02 09:19:46 +020030#define inode_to_bdi(inode) ((inode)->i_mapping->backing_dev_info)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070031
Jens Axboe03ba3782009-09-09 09:08:54 +020032/*
Jens Axboed0bceac2009-05-18 08:20:32 +020033 * We don't actually have pdflush, but this one is exported though /proc...
34 */
35int nr_pdflush_threads;
36
37/*
Jens Axboec4a77a62009-09-16 15:18:25 +020038 * Passed into wb_writeback(), essentially a subset of writeback_control
39 */
40struct wb_writeback_args {
41 long nr_pages;
42 struct super_block *sb;
43 enum writeback_sync_modes sync_mode;
Wu Fengguangd3ddec72009-09-23 20:33:40 +080044 int for_kupdate:1;
45 int range_cyclic:1;
46 int for_background:1;
Jens Axboec4a77a62009-09-16 15:18:25 +020047};
48
49/*
Jens Axboe03ba3782009-09-09 09:08:54 +020050 * Work items for the bdi_writeback threads
Adrian Bunkf11b00f2008-04-29 00:58:56 -070051 */
Jens Axboe03ba3782009-09-09 09:08:54 +020052struct bdi_work {
Jens Axboe8010c3b2009-09-15 20:04:57 +020053 struct list_head list; /* pending work list */
54 struct rcu_head rcu_head; /* for RCU free/clear of work */
Jens Axboe03ba3782009-09-09 09:08:54 +020055
Jens Axboe8010c3b2009-09-15 20:04:57 +020056 unsigned long seen; /* threads that have seen this work */
57 atomic_t pending; /* number of threads still to do work */
Jens Axboe03ba3782009-09-09 09:08:54 +020058
Jens Axboe8010c3b2009-09-15 20:04:57 +020059 struct wb_writeback_args args; /* writeback arguments */
Jens Axboe03ba3782009-09-09 09:08:54 +020060
Jens Axboe8010c3b2009-09-15 20:04:57 +020061 unsigned long state; /* flag bits, see WS_* */
Jens Axboe03ba3782009-09-09 09:08:54 +020062};
63
64enum {
65 WS_USED_B = 0,
66 WS_ONSTACK_B,
67};
68
69#define WS_USED (1 << WS_USED_B)
70#define WS_ONSTACK (1 << WS_ONSTACK_B)
71
72static inline bool bdi_work_on_stack(struct bdi_work *work)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070073{
Jens Axboe03ba3782009-09-09 09:08:54 +020074 return test_bit(WS_ONSTACK_B, &work->state);
75}
76
77static inline void bdi_work_init(struct bdi_work *work,
Jens Axboeb6e51312009-09-16 15:13:54 +020078 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +020079{
80 INIT_RCU_HEAD(&work->rcu_head);
Jens Axboeb6e51312009-09-16 15:13:54 +020081 work->args = *args;
Jens Axboe03ba3782009-09-09 09:08:54 +020082 work->state = WS_USED;
83}
84
Adrian Bunkf11b00f2008-04-29 00:58:56 -070085/**
86 * writeback_in_progress - determine whether there is writeback in progress
87 * @bdi: the device's backing_dev_info structure.
88 *
Jens Axboe03ba3782009-09-09 09:08:54 +020089 * Determine whether there is writeback waiting to be handled against a
90 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070091 */
92int writeback_in_progress(struct backing_dev_info *bdi)
93{
Jens Axboe03ba3782009-09-09 09:08:54 +020094 return !list_empty(&bdi->work_list);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070095}
96
Jens Axboe03ba3782009-09-09 09:08:54 +020097static void bdi_work_clear(struct bdi_work *work)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070098{
Jens Axboe03ba3782009-09-09 09:08:54 +020099 clear_bit(WS_USED_B, &work->state);
100 smp_mb__after_clear_bit();
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200101 /*
102 * work can have disappeared at this point. bit waitq functions
103 * should be able to tolerate this, provided bdi_sched_wait does
104 * not dereference it's pointer argument.
105 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200106 wake_up_bit(&work->state, WS_USED_B);
Adrian Bunkf11b00f2008-04-29 00:58:56 -0700107}
108
Jens Axboe03ba3782009-09-09 09:08:54 +0200109static void bdi_work_free(struct rcu_head *head)
Nick Piggin4195f732009-05-28 09:01:15 +0200110{
Jens Axboe03ba3782009-09-09 09:08:54 +0200111 struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
Nick Piggin4195f732009-05-28 09:01:15 +0200112
Jens Axboe03ba3782009-09-09 09:08:54 +0200113 if (!bdi_work_on_stack(work))
114 kfree(work);
115 else
116 bdi_work_clear(work);
117}
118
119static void wb_work_complete(struct bdi_work *work)
120{
Jens Axboec4a77a62009-09-16 15:18:25 +0200121 const enum writeback_sync_modes sync_mode = work->args.sync_mode;
Nick Piggin77b9d052009-09-15 21:34:51 +0200122 int onstack = bdi_work_on_stack(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200123
124 /*
125 * For allocated work, we can clear the done/seen bit right here.
126 * For on-stack work, we need to postpone both the clear and free
127 * to after the RCU grace period, since the stack could be invalidated
128 * as soon as bdi_work_clear() has done the wakeup.
129 */
Nick Piggin77b9d052009-09-15 21:34:51 +0200130 if (!onstack)
Jens Axboe03ba3782009-09-09 09:08:54 +0200131 bdi_work_clear(work);
Nick Piggin77b9d052009-09-15 21:34:51 +0200132 if (sync_mode == WB_SYNC_NONE || onstack)
Jens Axboe03ba3782009-09-09 09:08:54 +0200133 call_rcu(&work->rcu_head, bdi_work_free);
134}
135
136static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
137{
138 /*
139 * The caller has retrieved the work arguments from this work,
140 * drop our reference. If this is the last ref, delete and free it
141 */
142 if (atomic_dec_and_test(&work->pending)) {
143 struct backing_dev_info *bdi = wb->bdi;
144
145 spin_lock(&bdi->wb_lock);
146 list_del_rcu(&work->list);
147 spin_unlock(&bdi->wb_lock);
148
149 wb_work_complete(work);
Nick Piggin4195f732009-05-28 09:01:15 +0200150 }
151}
152
Jens Axboe03ba3782009-09-09 09:08:54 +0200153static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154{
Jens Axboebcddc3f2009-09-13 20:07:36 +0200155 work->seen = bdi->wb_mask;
156 BUG_ON(!work->seen);
157 atomic_set(&work->pending, bdi->wb_cnt);
158 BUG_ON(!bdi->wb_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Jens Axboebcddc3f2009-09-13 20:07:36 +0200160 /*
Nick Piggindeed62e2009-09-15 21:32:58 +0200161 * list_add_tail_rcu() contains the necessary barriers to
162 * make sure the above stores are seen before the item is
163 * noticed on the list
Jens Axboebcddc3f2009-09-13 20:07:36 +0200164 */
Jens Axboebcddc3f2009-09-13 20:07:36 +0200165 spin_lock(&bdi->wb_lock);
166 list_add_tail_rcu(&work->list, &bdi->work_list);
167 spin_unlock(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200170 * If the default thread isn't there, make sure we add it. When
171 * it gets created and wakes up, we'll run this work.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200173 if (unlikely(list_empty_careful(&bdi->wb_list)))
174 wake_up_process(default_backing_dev_info.wb.task);
175 else {
176 struct bdi_writeback *wb = &bdi->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200178 if (wb->task)
Jens Axboe03ba3782009-09-09 09:08:54 +0200179 wake_up_process(wb->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181}
182
Jens Axboe03ba3782009-09-09 09:08:54 +0200183/*
184 * Used for on-stack allocated work items. The caller needs to wait until
185 * the wb threads have acked the work before it's safe to continue.
186 */
187static void bdi_wait_on_work_clear(struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188{
Jens Axboe03ba3782009-09-09 09:08:54 +0200189 wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
190 TASK_UNINTERRUPTIBLE);
191}
192
Jens Axboef11fcae2009-09-15 09:53:35 +0200193static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
Jens Axboeb6e51312009-09-16 15:13:54 +0200194 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200195{
196 struct bdi_work *work;
197
Jens Axboebcddc3f2009-09-13 20:07:36 +0200198 /*
199 * This is WB_SYNC_NONE writeback, so if allocation fails just
200 * wakeup the thread for old dirty data writeback
201 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200202 work = kmalloc(sizeof(*work), GFP_ATOMIC);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200203 if (work) {
Jens Axboeb6e51312009-09-16 15:13:54 +0200204 bdi_work_init(work, args);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200205 bdi_queue_work(bdi, work);
206 } else {
207 struct bdi_writeback *wb = &bdi->wb;
Jens Axboe03ba3782009-09-09 09:08:54 +0200208
Jens Axboebcddc3f2009-09-13 20:07:36 +0200209 if (wb->task)
210 wake_up_process(wb->task);
211 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200212}
213
Jens Axboeb6e51312009-09-16 15:13:54 +0200214/**
215 * bdi_sync_writeback - start and wait for writeback
216 * @bdi: the backing device to write from
217 * @sb: write inodes from this super_block
218 *
219 * Description:
220 * This does WB_SYNC_ALL data integrity writeback and waits for the
221 * IO to complete. Callers must hold the sb s_umount semaphore for
222 * reading, to avoid having the super disappear before we are done.
223 */
224static void bdi_sync_writeback(struct backing_dev_info *bdi,
225 struct super_block *sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200226{
Jens Axboeb6e51312009-09-16 15:13:54 +0200227 struct wb_writeback_args args = {
228 .sb = sb,
229 .sync_mode = WB_SYNC_ALL,
230 .nr_pages = LONG_MAX,
231 .range_cyclic = 0,
232 };
233 struct bdi_work work;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200234
Jens Axboeb6e51312009-09-16 15:13:54 +0200235 bdi_work_init(&work, &args);
236 work.state |= WS_ONSTACK;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200237
Jens Axboeb6e51312009-09-16 15:13:54 +0200238 bdi_queue_work(bdi, &work);
239 bdi_wait_on_work_clear(&work);
240}
241
242/**
243 * bdi_start_writeback - start writeback
244 * @bdi: the backing device to write from
245 * @nr_pages: the number of pages to write
246 *
247 * Description:
248 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
249 * started when this function returns, we make no guarentees on
250 * completion. Caller need not hold sb s_umount semaphore.
251 *
252 */
253void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
254{
255 struct wb_writeback_args args = {
256 .sync_mode = WB_SYNC_NONE,
257 .nr_pages = nr_pages,
258 .range_cyclic = 1,
259 };
260
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800261 /*
262 * We treat @nr_pages=0 as the special case to do background writeback,
263 * ie. to sync pages until the background dirty threshold is reached.
264 */
265 if (!nr_pages) {
266 args.nr_pages = LONG_MAX;
267 args.for_background = 1;
268 }
269
Jens Axboeb6e51312009-09-16 15:13:54 +0200270 bdi_alloc_queue_work(bdi, &args);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271}
272
273/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700274 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
275 * furthest end of its superblock's dirty-inode list.
276 *
277 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200278 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700279 * the case then the inode must have been redirtied while it was being written
280 * out and we don't reset its dirtied_when.
281 */
282static void redirty_tail(struct inode *inode)
283{
Jens Axboe03ba3782009-09-09 09:08:54 +0200284 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700285
Jens Axboe03ba3782009-09-09 09:08:54 +0200286 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200287 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700288
Jens Axboe03ba3782009-09-09 09:08:54 +0200289 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200290 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700291 inode->dirtied_when = jiffies;
292 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200293 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700294}
295
296/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200297 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700298 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700299static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700300{
Jens Axboe03ba3782009-09-09 09:08:54 +0200301 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
302
303 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700304}
305
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700306static void inode_sync_complete(struct inode *inode)
307{
308 /*
309 * Prevent speculative execution through spin_unlock(&inode_lock);
310 */
311 smp_mb();
312 wake_up_bit(&inode->i_state, __I_SYNC);
313}
314
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700315static bool inode_dirtied_after(struct inode *inode, unsigned long t)
316{
317 bool ret = time_after(inode->dirtied_when, t);
318#ifndef CONFIG_64BIT
319 /*
320 * For inodes being constantly redirtied, dirtied_when can get stuck.
321 * It _appears_ to be in the future, but is actually in distant past.
322 * This test is necessary to prevent such wrapped-around relative times
323 * from permanently stopping the whole pdflush writeback.
324 */
325 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
326#endif
327 return ret;
328}
329
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700330/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700331 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
332 */
333static void move_expired_inodes(struct list_head *delaying_queue,
334 struct list_head *dispatch_queue,
335 unsigned long *older_than_this)
336{
337 while (!list_empty(delaying_queue)) {
338 struct inode *inode = list_entry(delaying_queue->prev,
339 struct inode, i_list);
340 if (older_than_this &&
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700341 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700342 break;
343 list_move(&inode->i_list, dispatch_queue);
344 }
345}
346
347/*
348 * Queue all expired dirty inodes for io, eldest first.
349 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200350static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700351{
Jens Axboe03ba3782009-09-09 09:08:54 +0200352 list_splice_init(&wb->b_more_io, wb->b_io.prev);
353 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200354}
355
Jens Axboe03ba3782009-09-09 09:08:54 +0200356static int write_inode(struct inode *inode, int sync)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200357{
Jens Axboe03ba3782009-09-09 09:08:54 +0200358 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
359 return inode->i_sb->s_op->write_inode(inode, sync);
360 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700361}
362
363/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200364 * Wait for writeback on an inode to complete.
365 */
366static void inode_wait_for_writeback(struct inode *inode)
367{
368 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
369 wait_queue_head_t *wqh;
370
371 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
372 do {
373 spin_unlock(&inode_lock);
374 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
375 spin_lock(&inode_lock);
376 } while (inode->i_state & I_SYNC);
377}
378
379/*
380 * Write out an inode's dirty pages. Called under inode_lock. Either the
381 * caller has ref on the inode (either via __iget or via syscall against an fd)
382 * or the inode has I_WILL_FREE set (via generic_forget_inode)
383 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 * If `wait' is set, wait on the writeout.
385 *
386 * The whole writeout design is quite complex and fragile. We want to avoid
387 * starvation of particular inodes when others are being redirtied, prevent
388 * livelocks, etc.
389 *
390 * Called under inode_lock.
391 */
392static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200393writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 struct address_space *mapping = inode->i_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 int wait = wbc->sync_mode == WB_SYNC_ALL;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200397 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 int ret;
399
Christoph Hellwig01c03192009-06-08 13:35:40 +0200400 if (!atomic_read(&inode->i_count))
401 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
402 else
403 WARN_ON(inode->i_state & I_WILL_FREE);
404
405 if (inode->i_state & I_SYNC) {
406 /*
407 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200408 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200409 * writeback can proceed with the other inodes on s_io.
410 *
411 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200412 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200413 */
414 if (!wait) {
415 requeue_io(inode);
416 return 0;
417 }
418
419 /*
420 * It's a data-integrity sync. We must wait.
421 */
422 inode_wait_for_writeback(inode);
423 }
424
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700425 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700427 /* Set I_SYNC, reset I_DIRTY */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 dirty = inode->i_state & I_DIRTY;
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700429 inode->i_state |= I_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 inode->i_state &= ~I_DIRTY;
431
432 spin_unlock(&inode_lock);
433
434 ret = do_writepages(mapping, wbc);
435
436 /* Don't write the inode if only I_DIRTY_PAGES was set */
437 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
438 int err = write_inode(inode, wait);
439 if (ret == 0)
440 ret = err;
441 }
442
443 if (wait) {
444 int err = filemap_fdatawait(mapping);
445 if (ret == 0)
446 ret = err;
447 }
448
449 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700450 inode->i_state &= ~I_SYNC;
Wu Fengguang84a89242009-06-16 15:33:17 -0700451 if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800452 if (inode->i_state & I_DIRTY) {
453 /*
454 * Someone redirtied the inode while were writing back
455 * the pages.
456 */
457 redirty_tail(inode);
458 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 /*
460 * We didn't write back all the pages. nfs_writepages()
461 * sometimes bales out without doing anything. Redirty
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200462 * the inode; Move it from b_io onto b_more_io/b_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700463 */
464 /*
465 * akpm: if the caller was the kupdate function we put
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200466 * this inode at the head of b_dirty so it gets first
Andrew Morton1b43ef92007-10-16 23:30:35 -0700467 * consideration. Otherwise, move it to the tail, for
468 * the reasons described there. I'm not really sure
469 * how much sense this makes. Presumably I had a good
470 * reasons for doing it this way, and I'd rather not
471 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 */
473 if (wbc->for_kupdate) {
474 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700475 * For the kupdate function we move the inode
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200476 * to b_more_io so it will get more writeout as
Fengguang Wu2c136572007-10-16 23:30:39 -0700477 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 */
479 inode->i_state |= I_DIRTY_PAGES;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800480 if (wbc->nr_to_write <= 0) {
481 /*
482 * slice used up: queue for next turn
483 */
484 requeue_io(inode);
485 } else {
486 /*
487 * somehow blocked: retry later
488 */
489 redirty_tail(inode);
490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 } else {
492 /*
493 * Otherwise fully redirty the inode so that
494 * other inodes on this superblock will get some
495 * writeout. Otherwise heavy writing to one
496 * file would indefinitely suspend writeout of
497 * all the other files.
498 */
499 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700500 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 } else if (atomic_read(&inode->i_count)) {
503 /*
504 * The inode is clean, inuse
505 */
506 list_move(&inode->i_list, &inode_in_use);
507 } else {
508 /*
509 * The inode is clean, unused
510 */
511 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 }
513 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700514 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 return ret;
516}
517
Jens Axboe03ba3782009-09-09 09:08:54 +0200518/*
519 * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
520 * before calling writeback. So make sure that we do pin it, so it doesn't
521 * go away while we are writing inodes from it.
522 *
523 * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
524 * 1 if we failed.
525 */
526static int pin_sb_for_writeback(struct writeback_control *wbc,
527 struct inode *inode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Jens Axboe03ba3782009-09-09 09:08:54 +0200529 struct super_block *sb = inode->i_sb;
530
531 /*
532 * Caller must already hold the ref for this
533 */
534 if (wbc->sync_mode == WB_SYNC_ALL) {
535 WARN_ON(!rwsem_is_locked(&sb->s_umount));
536 return 0;
537 }
538
539 spin_lock(&sb_lock);
540 sb->s_count++;
541 if (down_read_trylock(&sb->s_umount)) {
542 if (sb->s_root) {
543 spin_unlock(&sb_lock);
544 return 0;
545 }
546 /*
547 * umounted, drop rwsem again and fall through to failure
548 */
549 up_read(&sb->s_umount);
550 }
551
552 sb->s_count--;
553 spin_unlock(&sb_lock);
554 return 1;
555}
556
557static void unpin_sb_for_writeback(struct writeback_control *wbc,
558 struct inode *inode)
559{
560 struct super_block *sb = inode->i_sb;
561
562 if (wbc->sync_mode == WB_SYNC_ALL)
563 return;
564
565 up_read(&sb->s_umount);
566 put_super(sb);
567}
568
569static void writeback_inodes_wb(struct bdi_writeback *wb,
570 struct writeback_control *wbc)
571{
572 struct super_block *sb = wbc->sb;
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200573 const int is_blkdev_sb = sb_is_blkdev_sb(sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 const unsigned long start = jiffies; /* livelock avoidance */
575
Hans Reiserae8547b2008-05-07 15:48:57 +0300576 spin_lock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
Jens Axboe03ba3782009-09-09 09:08:54 +0200578 if (!wbc->for_kupdate || list_empty(&wb->b_io))
579 queue_io(wb, wbc->older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200580
Jens Axboe03ba3782009-09-09 09:08:54 +0200581 while (!list_empty(&wb->b_io)) {
582 struct inode *inode = list_entry(wb->b_io.prev,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 struct inode, i_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 long pages_skipped;
585
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200586 /*
587 * super block given and doesn't match, skip this inode
588 */
589 if (sb && sb != inode->i_sb) {
590 redirty_tail(inode);
591 continue;
592 }
593
Jens Axboe03ba3782009-09-09 09:08:54 +0200594 if (!bdi_cap_writeback_dirty(wb->bdi)) {
Andrew Morton9852a0e72007-10-16 23:30:33 -0700595 redirty_tail(inode);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200596 if (is_blkdev_sb) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 /*
598 * Dirty memory-backed blockdev: the ramdisk
599 * driver does this. Skip just this inode
600 */
601 continue;
602 }
603 /*
604 * Dirty memory-backed inode against a filesystem other
605 * than the kernel-internal bdev filesystem. Skip the
606 * entire superblock.
607 */
608 break;
609 }
610
Wu Fengguang84a89242009-06-16 15:33:17 -0700611 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700612 requeue_io(inode);
613 continue;
614 }
615
Jens Axboe03ba3782009-09-09 09:08:54 +0200616 if (wbc->nonblocking && bdi_write_congested(wb->bdi)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 wbc->encountered_congestion = 1;
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200618 if (!is_blkdev_sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 break; /* Skip a congested fs */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700620 requeue_io(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 continue; /* Skip a congested blockdev */
622 }
623
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700624 /*
625 * Was this inode dirtied after sync_sb_inodes was called?
626 * This keeps sync from extra jobs and livelock.
627 */
628 if (inode_dirtied_after(inode, start))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 break;
630
Jens Axboe03ba3782009-09-09 09:08:54 +0200631 if (pin_sb_for_writeback(wbc, inode)) {
632 requeue_io(inode);
633 continue;
634 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635
Wu Fengguang84a89242009-06-16 15:33:17 -0700636 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 __iget(inode);
638 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200639 writeback_single_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200640 unpin_sb_for_writeback(wbc, inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 if (wbc->pages_skipped != pages_skipped) {
642 /*
643 * writeback is not making progress due to locked
644 * buffers. Skip this inode for now.
645 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700646 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 }
648 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800650 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800652 if (wbc->nr_to_write <= 0) {
653 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800655 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200656 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800657 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 }
Nick Piggin38f21972009-01-06 14:40:25 -0800659
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200660 spin_unlock(&inode_lock);
661 /* Leave any unwritten inodes on b_io */
662}
663
Jens Axboe03ba3782009-09-09 09:08:54 +0200664void writeback_inodes_wbc(struct writeback_control *wbc)
665{
666 struct backing_dev_info *bdi = wbc->bdi;
667
668 writeback_inodes_wb(&bdi->wb, wbc);
669}
670
671/*
672 * The maximum number of pages to writeout in a single bdi flush/kupdate
673 * operation. We do this so we don't hold I_SYNC against an inode for
674 * enormous amounts of time, which would block a userspace task which has
675 * been forced to throttle against that inode. Also, the code reevaluates
676 * the dirty each time it has written this many pages.
677 */
678#define MAX_WRITEBACK_PAGES 1024
679
680static inline bool over_bground_thresh(void)
681{
682 unsigned long background_thresh, dirty_thresh;
683
684 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
685
686 return (global_page_state(NR_FILE_DIRTY) +
687 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
688}
689
690/*
691 * Explicit flushing or periodic writeback of "old" data.
692 *
693 * Define "old": the first time one of an inode's pages is dirtied, we mark the
694 * dirtying-time in the inode's address_space. So this periodic writeback code
695 * just walks the superblock inode list, writing back any inodes which are
696 * older than a specific point in time.
697 *
698 * Try to run once per dirty_writeback_interval. But if a writeback event
699 * takes longer than a dirty_writeback_interval interval, then leave a
700 * one-second gap.
701 *
702 * older_than_this takes precedence over nr_to_write. So we'll only write back
703 * all dirty pages if they are all attached to "old" mappings.
704 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200705static long wb_writeback(struct bdi_writeback *wb,
706 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200707{
708 struct writeback_control wbc = {
709 .bdi = wb->bdi,
Jens Axboec4a77a62009-09-16 15:18:25 +0200710 .sb = args->sb,
711 .sync_mode = args->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200712 .older_than_this = NULL,
Jens Axboec4a77a62009-09-16 15:18:25 +0200713 .for_kupdate = args->for_kupdate,
714 .range_cyclic = args->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200715 };
716 unsigned long oldest_jif;
717 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200718 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200719
720 if (wbc.for_kupdate) {
721 wbc.older_than_this = &oldest_jif;
722 oldest_jif = jiffies -
723 msecs_to_jiffies(dirty_expire_interval * 10);
724 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200725 if (!wbc.range_cyclic) {
726 wbc.range_start = 0;
727 wbc.range_end = LLONG_MAX;
728 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200729
730 for (;;) {
731 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800732 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200733 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800734 if (args->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200735 break;
736
737 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800738 * For background writeout, stop when we are below the
739 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200740 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800741 if (args->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200742 break;
743
744 wbc.more_io = 0;
745 wbc.encountered_congestion = 0;
746 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
747 wbc.pages_skipped = 0;
748 writeback_inodes_wb(wb, &wbc);
Jens Axboec4a77a62009-09-16 15:18:25 +0200749 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200750 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
751
752 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200753 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200754 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200755 if (wbc.nr_to_write <= 0)
756 continue;
757 /*
758 * Didn't write everything and we don't have more IO, bail
759 */
760 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200761 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200762 /*
763 * Did we write something? Try for more
764 */
765 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
766 continue;
767 /*
768 * Nothing written. Wait for some inode to
769 * become available for writeback. Otherwise
770 * we'll just busyloop.
771 */
772 spin_lock(&inode_lock);
773 if (!list_empty(&wb->b_more_io)) {
774 inode = list_entry(wb->b_more_io.prev,
775 struct inode, i_list);
776 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200777 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200778 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200779 }
780
781 return wrote;
782}
783
784/*
785 * Return the next bdi_work struct that hasn't been processed by this
Jens Axboe8010c3b2009-09-15 20:04:57 +0200786 * wb thread yet. ->seen is initially set for each thread that exists
787 * for this device, when a thread first notices a piece of work it
788 * clears its bit. Depending on writeback type, the thread will notify
789 * completion on either receiving the work (WB_SYNC_NONE) or after
790 * it is done (WB_SYNC_ALL).
Jens Axboe03ba3782009-09-09 09:08:54 +0200791 */
792static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
793 struct bdi_writeback *wb)
794{
795 struct bdi_work *work, *ret = NULL;
796
797 rcu_read_lock();
798
799 list_for_each_entry_rcu(work, &bdi->work_list, list) {
Nick Piggin77fad5e2009-09-15 21:34:12 +0200800 if (!test_bit(wb->nr, &work->seen))
Jens Axboe03ba3782009-09-09 09:08:54 +0200801 continue;
Nick Piggin77fad5e2009-09-15 21:34:12 +0200802 clear_bit(wb->nr, &work->seen);
Jens Axboe03ba3782009-09-09 09:08:54 +0200803
804 ret = work;
805 break;
806 }
807
808 rcu_read_unlock();
809 return ret;
810}
811
812static long wb_check_old_data_flush(struct bdi_writeback *wb)
813{
814 unsigned long expired;
815 long nr_pages;
816
817 expired = wb->last_old_flush +
818 msecs_to_jiffies(dirty_writeback_interval * 10);
819 if (time_before(jiffies, expired))
820 return 0;
821
822 wb->last_old_flush = jiffies;
823 nr_pages = global_page_state(NR_FILE_DIRTY) +
824 global_page_state(NR_UNSTABLE_NFS) +
825 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
826
Jens Axboec4a77a62009-09-16 15:18:25 +0200827 if (nr_pages) {
828 struct wb_writeback_args args = {
829 .nr_pages = nr_pages,
830 .sync_mode = WB_SYNC_NONE,
831 .for_kupdate = 1,
832 .range_cyclic = 1,
833 };
834
835 return wb_writeback(wb, &args);
836 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200837
838 return 0;
839}
840
841/*
842 * Retrieve work items and do the writeback they describe
843 */
844long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
845{
846 struct backing_dev_info *bdi = wb->bdi;
847 struct bdi_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200848 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200849
850 while ((work = get_next_work_item(bdi, wb)) != NULL) {
Jens Axboec4a77a62009-09-16 15:18:25 +0200851 struct wb_writeback_args args = work->args;
Jens Axboe03ba3782009-09-09 09:08:54 +0200852
853 /*
854 * Override sync mode, in case we must wait for completion
855 */
856 if (force_wait)
Jens Axboec4a77a62009-09-16 15:18:25 +0200857 work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200858
859 /*
860 * If this isn't a data integrity operation, just notify
861 * that we have seen this work and we are now starting it.
862 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200863 if (args.sync_mode == WB_SYNC_NONE)
Jens Axboe03ba3782009-09-09 09:08:54 +0200864 wb_clear_pending(wb, work);
865
Jens Axboec4a77a62009-09-16 15:18:25 +0200866 wrote += wb_writeback(wb, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200867
868 /*
869 * This is a data integrity writeback, so only do the
870 * notification when we have completed the work.
871 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200872 if (args.sync_mode == WB_SYNC_ALL)
Jens Axboe03ba3782009-09-09 09:08:54 +0200873 wb_clear_pending(wb, work);
874 }
875
876 /*
877 * Check for periodic writeback, kupdated() style
878 */
879 wrote += wb_check_old_data_flush(wb);
880
881 return wrote;
882}
883
884/*
885 * Handle writeback of dirty data for the device backed by this bdi. Also
886 * wakes up periodically and does kupdated style flushing.
887 */
888int bdi_writeback_task(struct bdi_writeback *wb)
889{
890 unsigned long last_active = jiffies;
891 unsigned long wait_jiffies = -1UL;
892 long pages_written;
893
894 while (!kthread_should_stop()) {
895 pages_written = wb_do_writeback(wb, 0);
896
897 if (pages_written)
898 last_active = jiffies;
899 else if (wait_jiffies != -1UL) {
900 unsigned long max_idle;
901
902 /*
903 * Longest period of inactivity that we tolerate. If we
904 * see dirty data again later, the task will get
905 * recreated automatically.
906 */
907 max_idle = max(5UL * 60 * HZ, wait_jiffies);
908 if (time_after(jiffies, max_idle + last_active))
909 break;
910 }
911
912 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
Jens Axboe49db0412009-09-15 21:27:40 +0200913 schedule_timeout_interruptible(wait_jiffies);
Jens Axboe03ba3782009-09-09 09:08:54 +0200914 try_to_freeze();
915 }
916
917 return 0;
918}
919
920/*
Jens Axboeb6e51312009-09-16 15:13:54 +0200921 * Schedule writeback for all backing devices. This does WB_SYNC_NONE
922 * writeback, for integrity writeback see bdi_sync_writeback().
Jens Axboe03ba3782009-09-09 09:08:54 +0200923 */
Jens Axboeb6e51312009-09-16 15:13:54 +0200924static void bdi_writeback_all(struct super_block *sb, long nr_pages)
Jens Axboe03ba3782009-09-09 09:08:54 +0200925{
Jens Axboeb6e51312009-09-16 15:13:54 +0200926 struct wb_writeback_args args = {
927 .sb = sb,
928 .nr_pages = nr_pages,
929 .sync_mode = WB_SYNC_NONE,
930 };
Jens Axboe03ba3782009-09-09 09:08:54 +0200931 struct backing_dev_info *bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200932
Jens Axboecfc4ba52009-09-14 13:12:40 +0200933 rcu_read_lock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200934
Jens Axboecfc4ba52009-09-14 13:12:40 +0200935 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200936 if (!bdi_has_dirty_io(bdi))
937 continue;
938
Jens Axboeb6e51312009-09-16 15:13:54 +0200939 bdi_alloc_queue_work(bdi, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200940 }
941
Jens Axboecfc4ba52009-09-14 13:12:40 +0200942 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200943}
944
945/*
946 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
947 * the whole world.
948 */
949void wakeup_flusher_threads(long nr_pages)
950{
Jens Axboe03ba3782009-09-09 09:08:54 +0200951 if (nr_pages == 0)
952 nr_pages = global_page_state(NR_FILE_DIRTY) +
953 global_page_state(NR_UNSTABLE_NFS);
Jens Axboeb6e51312009-09-16 15:13:54 +0200954 bdi_writeback_all(NULL, nr_pages);
Jens Axboe03ba3782009-09-09 09:08:54 +0200955}
956
957static noinline void block_dump___mark_inode_dirty(struct inode *inode)
958{
959 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
960 struct dentry *dentry;
961 const char *name = "?";
962
963 dentry = d_find_alias(inode);
964 if (dentry) {
965 spin_lock(&dentry->d_lock);
966 name = (const char *) dentry->d_name.name;
967 }
968 printk(KERN_DEBUG
969 "%s(%d): dirtied inode %lu (%s) on %s\n",
970 current->comm, task_pid_nr(current), inode->i_ino,
971 name, inode->i_sb->s_id);
972 if (dentry) {
973 spin_unlock(&dentry->d_lock);
974 dput(dentry);
975 }
976 }
977}
978
979/**
980 * __mark_inode_dirty - internal function
981 * @inode: inode to mark
982 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
983 * Mark an inode as dirty. Callers should use mark_inode_dirty or
984 * mark_inode_dirty_sync.
985 *
986 * Put the inode on the super block's dirty list.
987 *
988 * CAREFUL! We mark it dirty unconditionally, but move it onto the
989 * dirty list only if it is hashed or if it refers to a blockdev.
990 * If it was not hashed, it will never be added to the dirty list
991 * even if it is later hashed, as it will have been marked dirty already.
992 *
993 * In short, make sure you hash any inodes _before_ you start marking
994 * them dirty.
995 *
996 * This function *must* be atomic for the I_DIRTY_PAGES case -
997 * set_page_dirty() is called under spinlock in several places.
998 *
999 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1000 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1001 * the kernel-internal blockdev inode represents the dirtying time of the
1002 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1003 * page->mapping->host, so the page-dirtying time is recorded in the internal
1004 * blockdev inode.
1005 */
1006void __mark_inode_dirty(struct inode *inode, int flags)
1007{
1008 struct super_block *sb = inode->i_sb;
1009
1010 /*
1011 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1012 * dirty the inode itself
1013 */
1014 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1015 if (sb->s_op->dirty_inode)
1016 sb->s_op->dirty_inode(inode);
1017 }
1018
1019 /*
1020 * make sure that changes are seen by all cpus before we test i_state
1021 * -- mikulas
1022 */
1023 smp_mb();
1024
1025 /* avoid the locking if we can */
1026 if ((inode->i_state & flags) == flags)
1027 return;
1028
1029 if (unlikely(block_dump))
1030 block_dump___mark_inode_dirty(inode);
1031
1032 spin_lock(&inode_lock);
1033 if ((inode->i_state & flags) != flags) {
1034 const int was_dirty = inode->i_state & I_DIRTY;
1035
1036 inode->i_state |= flags;
1037
1038 /*
1039 * If the inode is being synced, just update its dirty state.
1040 * The unlocker will place the inode on the appropriate
1041 * superblock list, based upon its state.
1042 */
1043 if (inode->i_state & I_SYNC)
1044 goto out;
1045
1046 /*
1047 * Only add valid (hashed) inodes to the superblock's
1048 * dirty list. Add blockdev inodes as well.
1049 */
1050 if (!S_ISBLK(inode->i_mode)) {
1051 if (hlist_unhashed(&inode->i_hash))
1052 goto out;
1053 }
1054 if (inode->i_state & (I_FREEING|I_CLEAR))
1055 goto out;
1056
1057 /*
1058 * If the inode was already on b_dirty/b_io/b_more_io, don't
1059 * reposition it (that would break b_dirty time-ordering).
1060 */
1061 if (!was_dirty) {
1062 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Jens Axboe500b0672009-09-09 09:10:25 +02001063 struct backing_dev_info *bdi = wb->bdi;
1064
1065 if (bdi_cap_writeback_dirty(bdi) &&
1066 !test_bit(BDI_registered, &bdi->state)) {
1067 WARN_ON(1);
1068 printk(KERN_ERR "bdi-%s not registered\n",
1069 bdi->name);
1070 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001071
1072 inode->dirtied_when = jiffies;
1073 list_move(&inode->i_list, &wb->b_dirty);
1074 }
1075 }
1076out:
1077 spin_unlock(&inode_lock);
1078}
1079EXPORT_SYMBOL(__mark_inode_dirty);
1080
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001081/*
1082 * Write out a superblock's list of dirty inodes. A wait will be performed
1083 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1084 *
1085 * If older_than_this is non-NULL, then only write out inodes which
1086 * had their first dirtying at a time earlier than *older_than_this.
1087 *
1088 * If we're a pdlfush thread, then implement pdflush collision avoidance
1089 * against the entire list.
1090 *
1091 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1092 * This function assumes that the blockdev superblock's inodes are backed by
1093 * a variety of queues, so all inodes are searched. For other superblocks,
1094 * assume that all inodes are backed by the same queue.
1095 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001096 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1097 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1098 * on the writer throttling path, and we get decent balancing between many
1099 * throttled threads: we don't want them all piling up on inode_sync_wait.
1100 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001101static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001102{
Jens Axboe03ba3782009-09-09 09:08:54 +02001103 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001104
Jens Axboe03ba3782009-09-09 09:08:54 +02001105 /*
1106 * We need to be protected against the filesystem going from
1107 * r/o to r/w or vice versa.
1108 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001109 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001110
Jens Axboe03ba3782009-09-09 09:08:54 +02001111 spin_lock(&inode_lock);
1112
1113 /*
1114 * Data integrity sync. Must wait for all pages under writeback,
1115 * because there may have been pages dirtied before our sync
1116 * call, but which had writeout started before we write it out.
1117 * In which case, the inode may not be on the dirty list, but
1118 * we still have to wait for that writeout.
1119 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001120 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001121 struct address_space *mapping;
1122
1123 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1124 continue;
1125 mapping = inode->i_mapping;
1126 if (mapping->nrpages == 0)
1127 continue;
1128 __iget(inode);
1129 spin_unlock(&inode_lock);
1130 /*
1131 * We hold a reference to 'inode' so it couldn't have
1132 * been removed from s_inodes list while we dropped the
1133 * inode_lock. We cannot iput the inode now as we can
1134 * be holding the last reference and we cannot iput it
1135 * under inode_lock. So we keep the reference and iput
1136 * it later.
1137 */
1138 iput(old_inode);
1139 old_inode = inode;
1140
1141 filemap_fdatawait(mapping);
1142
1143 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001144
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001145 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001146 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001147 spin_unlock(&inode_lock);
1148 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149}
1150
Jens Axboed8a85592009-09-02 12:34:32 +02001151/**
1152 * writeback_inodes_sb - writeback dirty inodes from given super_block
1153 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 *
Jens Axboed8a85592009-09-02 12:34:32 +02001155 * Start writeback on some inodes on this super_block. No guarantees are made
1156 * on how many (if any) will be written, and this function does not wait
1157 * for IO completion of submitted IO. The number of pages submitted is
1158 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001160void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161{
Jens Axboed8a85592009-09-02 12:34:32 +02001162 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1163 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1164 long nr_to_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
Jens Axboed8a85592009-09-02 12:34:32 +02001166 nr_to_write = nr_dirty + nr_unstable +
Nick Piggin38f21972009-01-06 14:40:25 -08001167 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
Nick Piggin38f21972009-01-06 14:40:25 -08001168
Jens Axboeb6e51312009-09-16 15:13:54 +02001169 bdi_writeback_all(sb, nr_to_write);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170}
Jens Axboed8a85592009-09-02 12:34:32 +02001171EXPORT_SYMBOL(writeback_inodes_sb);
1172
1173/**
1174 * sync_inodes_sb - sync sb inode pages
1175 * @sb: the superblock
1176 *
1177 * This function writes and waits on any dirty inode belonging to this
1178 * super_block. The number of pages synced is returned.
1179 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001180void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001181{
Jens Axboeb6e51312009-09-16 15:13:54 +02001182 bdi_sync_writeback(sb->s_bdi, sb);
1183 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001184}
1185EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001188 * write_inode_now - write an inode to disk
1189 * @inode: inode to write to disk
1190 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001192 * This function commits an inode to disk immediately if it is dirty. This is
1193 * primarily needed by knfsd.
1194 *
1195 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197int write_inode_now(struct inode *inode, int sync)
1198{
1199 int ret;
1200 struct writeback_control wbc = {
1201 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001202 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001203 .range_start = 0,
1204 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 };
1206
1207 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001208 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209
1210 might_sleep();
1211 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001212 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 spin_unlock(&inode_lock);
1214 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001215 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 return ret;
1217}
1218EXPORT_SYMBOL(write_inode_now);
1219
1220/**
1221 * sync_inode - write an inode and its pages to disk.
1222 * @inode: the inode to sync
1223 * @wbc: controls the writeback mode
1224 *
1225 * sync_inode() will write an inode and its pages to disk. It will also
1226 * correctly update the inode on its superblock's dirty inode lists and will
1227 * update inode->i_state.
1228 *
1229 * The caller must have a ref on the inode.
1230 */
1231int sync_inode(struct inode *inode, struct writeback_control *wbc)
1232{
1233 int ret;
1234
1235 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001236 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 spin_unlock(&inode_lock);
1238 return ret;
1239}
1240EXPORT_SYMBOL(sync_inode);