blob: 5e00f1772c20fc6d9e4e70a087800e93b3ae3df2 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 akpm@zip.com.au
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
David Howellscf9a2ae2006-08-29 19:05:54 +010035#include <linux/buffer_head.h>
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
38/*
39 * The maximum number of pages to writeout in a single bdflush/kupdate
Joern Engel1c0eeaf2007-10-16 23:30:44 -070040 * operation. We do this so we don't hold I_SYNC against an inode for
Linus Torvalds1da177e2005-04-16 15:20:36 -070041 * enormous amounts of time, which would block a userspace task which has
42 * been forced to throttle against that inode. Also, the code reevaluates
43 * the dirty each time it has written this many pages.
44 */
45#define MAX_WRITEBACK_PAGES 1024
46
47/*
48 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
49 * will look to see if it needs to force writeback or throttling.
50 */
51static long ratelimit_pages = 32;
52
Linus Torvalds1da177e2005-04-16 15:20:36 -070053/*
54 * When balance_dirty_pages decides that the caller needs to perform some
55 * non-background writeback, this is how many pages it will attempt to write.
56 * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
57 * large amounts of I/O are submitted.
58 */
59static inline long sync_writeback_pages(void)
60{
61 return ratelimit_pages + ratelimit_pages / 2;
62}
63
64/* The following parameters are exported via /proc/sys/vm */
65
66/*
67 * Start background writeback (via pdflush) at this percentage
68 */
Linus Torvalds07db59b2007-04-27 09:10:47 -070069int dirty_background_ratio = 5;
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
71/*
Bron Gondwana195cf452008-02-04 22:29:20 -080072 * free highmem will not be subtracted from the total free memory
73 * for calculating free ratios if vm_highmem_is_dirtyable is true
74 */
75int vm_highmem_is_dirtyable;
76
77/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 * The generator of dirty data starts writeback at this percentage
79 */
Linus Torvalds07db59b2007-04-27 09:10:47 -070080int vm_dirty_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070081
82/*
Coywolf Qi Huntfd5403c2006-04-10 22:54:35 -070083 * The interval between `kupdate'-style writebacks, in jiffies
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 */
Bart Samwelf6ef9432006-03-24 03:15:48 -080085int dirty_writeback_interval = 5 * HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
87/*
Coywolf Qi Huntfd5403c2006-04-10 22:54:35 -070088 * The longest number of jiffies for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 */
Bart Samwelf6ef9432006-03-24 03:15:48 -080090int dirty_expire_interval = 30 * HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
93 * Flag that makes the machine dump writes/reads and block dirtyings.
94 */
95int block_dump;
96
97/*
Bart Samweled5b43f2006-03-24 03:15:49 -080098 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
99 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 */
101int laptop_mode;
102
103EXPORT_SYMBOL(laptop_mode);
104
105/* End of sysctl-exported parameters */
106
107
108static void background_writeout(unsigned long _min_pages);
109
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700111 * Scale the writeback cache size proportional to the relative writeout speeds.
112 *
113 * We do this by keeping a floating proportion between BDIs, based on page
114 * writeback completions [end_page_writeback()]. Those devices that write out
115 * pages fastest will get the larger share, while the slower will get a smaller
116 * share.
117 *
118 * We use page writeout completions because we are interested in getting rid of
119 * dirty pages. Having them written out is the primary goal.
120 *
121 * We introduce a concept of time, a period over which we measure these events,
122 * because demand can/will vary over time. The length of this period itself is
123 * measured in page writeback completions.
124 *
125 */
126static struct prop_descriptor vm_completions;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700127static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700128
129static unsigned long determine_dirtyable_memory(void);
130
131/*
132 * couple the period to the dirty_ratio:
133 *
134 * period/2 ~ roundup_pow_of_two(dirty limit)
135 */
136static int calc_period_shift(void)
137{
138 unsigned long dirty_total;
139
140 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
141 return 2 + ilog2(dirty_total - 1);
142}
143
144/*
145 * update the period when the dirty ratio changes.
146 */
147int dirty_ratio_handler(struct ctl_table *table, int write,
148 struct file *filp, void __user *buffer, size_t *lenp,
149 loff_t *ppos)
150{
151 int old_ratio = vm_dirty_ratio;
152 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
153 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
154 int shift = calc_period_shift();
155 prop_change_shift(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700156 prop_change_shift(&vm_dirties, shift);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700157 }
158 return ret;
159}
160
161/*
162 * Increment the BDI's writeout completion count and the global writeout
163 * completion count. Called from test_clear_page_writeback().
164 */
165static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
166{
167 __prop_inc_percpu(&vm_completions, &bdi->completions);
168}
169
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700170static inline void task_dirty_inc(struct task_struct *tsk)
171{
172 prop_inc_single(&vm_dirties, &tsk->dirties);
173}
174
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700175/*
176 * Obtain an accurate fraction of the BDI's portion.
177 */
178static void bdi_writeout_fraction(struct backing_dev_info *bdi,
179 long *numerator, long *denominator)
180{
181 if (bdi_cap_writeback_dirty(bdi)) {
182 prop_fraction_percpu(&vm_completions, &bdi->completions,
183 numerator, denominator);
184 } else {
185 *numerator = 0;
186 *denominator = 1;
187 }
188}
189
190/*
191 * Clip the earned share of dirty pages to that which is actually available.
192 * This avoids exceeding the total dirty_limit when the floating averages
193 * fluctuate too quickly.
194 */
195static void
196clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
197{
198 long avail_dirty;
199
200 avail_dirty = dirty -
201 (global_page_state(NR_FILE_DIRTY) +
202 global_page_state(NR_WRITEBACK) +
203 global_page_state(NR_UNSTABLE_NFS));
204
205 if (avail_dirty < 0)
206 avail_dirty = 0;
207
208 avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
209 bdi_stat(bdi, BDI_WRITEBACK);
210
211 *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
212}
213
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700214static inline void task_dirties_fraction(struct task_struct *tsk,
215 long *numerator, long *denominator)
216{
217 prop_fraction_single(&vm_dirties, &tsk->dirties,
218 numerator, denominator);
219}
220
221/*
222 * scale the dirty limit
223 *
224 * task specific dirty limit:
225 *
226 * dirty -= (dirty/8) * p_{t}
227 */
Adrian Bunkf61eaf92008-02-04 22:29:08 -0800228static void task_dirty_limit(struct task_struct *tsk, long *pdirty)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700229{
230 long numerator, denominator;
231 long dirty = *pdirty;
232 u64 inv = dirty >> 3;
233
234 task_dirties_fraction(tsk, &numerator, &denominator);
235 inv *= numerator;
236 do_div(inv, denominator);
237
238 dirty -= inv;
239 if (dirty < *pdirty/2)
240 dirty = *pdirty/2;
241
242 *pdirty = dirty;
243}
244
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700245/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 * Work out the current dirty-memory clamping and background writeout
247 * thresholds.
248 *
249 * The main aim here is to lower them aggressively if there is a lot of mapped
250 * memory around. To avoid stressing page reclaim with lots of unreclaimable
251 * pages. It is better to clamp down on writers than to start swapping, and
252 * performing lots of scanning.
253 *
254 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
255 *
256 * We don't permit the clamping level to fall below 5% - that is getting rather
257 * excessive.
258 *
259 * We make sure that the background writeout level is below the adjusted
260 * clamping level.
261 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700262
263static unsigned long highmem_dirtyable_memory(unsigned long total)
264{
265#ifdef CONFIG_HIGHMEM
266 int node;
267 unsigned long x = 0;
268
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700269 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700270 struct zone *z =
271 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
272
273 x += zone_page_state(z, NR_FREE_PAGES)
274 + zone_page_state(z, NR_INACTIVE)
275 + zone_page_state(z, NR_ACTIVE);
276 }
277 /*
278 * Make sure that the number of highmem pages is never larger
279 * than the number of the total dirtyable memory. This can only
280 * occur in very strange VM situations but we want to make sure
281 * that this does not occur.
282 */
283 return min(x, total);
284#else
285 return 0;
286#endif
287}
288
289static unsigned long determine_dirtyable_memory(void)
290{
291 unsigned long x;
292
293 x = global_page_state(NR_FREE_PAGES)
294 + global_page_state(NR_INACTIVE)
295 + global_page_state(NR_ACTIVE);
Bron Gondwana195cf452008-02-04 22:29:20 -0800296
297 if (!vm_highmem_is_dirtyable)
298 x -= highmem_dirtyable_memory(x);
299
Christoph Lameter1b424462007-05-06 14:48:59 -0700300 return x + 1; /* Ensure that we never return 0 */
301}
302
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303static void
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700304get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
305 struct backing_dev_info *bdi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306{
307 int background_ratio; /* Percentages */
308 int dirty_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 long background;
310 long dirty;
Christoph Lameter1b424462007-05-06 14:48:59 -0700311 unsigned long available_memory = determine_dirtyable_memory();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 struct task_struct *tsk;
313
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 dirty_ratio = vm_dirty_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 if (dirty_ratio < 5)
316 dirty_ratio = 5;
317
318 background_ratio = dirty_background_ratio;
319 if (background_ratio >= dirty_ratio)
320 background_ratio = dirty_ratio / 2;
321
322 background = (background_ratio * available_memory) / 100;
323 dirty = (dirty_ratio * available_memory) / 100;
324 tsk = current;
325 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
326 background += background / 4;
327 dirty += dirty / 4;
328 }
329 *pbackground = background;
330 *pdirty = dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700331
332 if (bdi) {
333 u64 bdi_dirty = dirty;
334 long numerator, denominator;
335
336 /*
337 * Calculate this BDI's share of the dirty ratio.
338 */
339 bdi_writeout_fraction(bdi, &numerator, &denominator);
340
341 bdi_dirty *= numerator;
342 do_div(bdi_dirty, denominator);
343
344 *pbdi_dirty = bdi_dirty;
345 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700346 task_dirty_limit(current, pbdi_dirty);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700347 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348}
349
350/*
351 * balance_dirty_pages() must be called by processes which are generating dirty
352 * data. It looks at the number of dirty pages in the machine and will force
353 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
354 * If we're over `background_thresh' then pdflush is woken to perform some
355 * writeout.
356 */
357static void balance_dirty_pages(struct address_space *mapping)
358{
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800359 long nr_reclaimable, bdi_nr_reclaimable;
360 long nr_writeback, bdi_nr_writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 long background_thresh;
362 long dirty_thresh;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700363 long bdi_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 unsigned long pages_written = 0;
365 unsigned long write_chunk = sync_writeback_pages();
366
367 struct backing_dev_info *bdi = mapping->backing_dev_info;
368
369 for (;;) {
370 struct writeback_control wbc = {
371 .bdi = bdi,
372 .sync_mode = WB_SYNC_NONE,
373 .older_than_this = NULL,
374 .nr_to_write = write_chunk,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700375 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 };
377
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700378 get_dirty_limits(&background_thresh, &dirty_thresh,
379 &bdi_thresh, bdi);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800380
381 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
382 global_page_state(NR_UNSTABLE_NFS);
383 nr_writeback = global_page_state(NR_WRITEBACK);
384
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700385 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
386 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800387
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700388 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
389 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800391 /*
392 * Throttle it only when the background writeback cannot
393 * catch-up. This avoids (excessively) small writeouts
394 * when the bdi limits are ramping up.
395 */
396 if (nr_reclaimable + nr_writeback <
397 (background_thresh + dirty_thresh) / 2)
398 break;
399
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700400 if (!bdi->dirty_exceeded)
401 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
403 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
404 * Unstable writes are a feature of certain networked
405 * filesystems (i.e. NFS) in which data may have been
406 * written to the server's write cache, but has not yet
407 * been flushed to permanent storage.
408 */
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700409 if (bdi_nr_reclaimable) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 writeback_inodes(&wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 pages_written += write_chunk - wbc.nr_to_write;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700412 get_dirty_limits(&background_thresh, &dirty_thresh,
413 &bdi_thresh, bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700415
416 /*
417 * In order to avoid the stacked BDI deadlock we need
418 * to ensure we accurately count the 'dirty' pages when
419 * the threshold is low.
420 *
421 * Otherwise it would be possible to get thresh+n pages
422 * reported dirty, even though there are thresh-m pages
423 * actually dirty; with m+n sitting in the percpu
424 * deltas.
425 */
426 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
427 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
428 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
429 } else if (bdi_nr_reclaimable) {
430 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
431 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
432 }
433
434 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
435 break;
436 if (pages_written >= write_chunk)
437 break; /* We've done our duty */
438
Andrew Morton3fcfab12006-10-19 23:28:16 -0700439 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 }
441
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700442 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
443 bdi->dirty_exceeded)
444 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445
446 if (writeback_in_progress(bdi))
447 return; /* pdflush is already working this queue */
448
449 /*
450 * In laptop mode, we wait until hitting the higher threshold before
451 * starting background writeout, and then write out all the way down
452 * to the lower threshold. So slow writers cause minimal disk activity.
453 *
454 * In normal mode, we start background writeout at the lower
455 * background_thresh, to keep the amount of dirty memory low.
456 */
457 if ((laptop_mode && pages_written) ||
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700458 (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
459 + global_page_state(NR_UNSTABLE_NFS)
460 > background_thresh)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 pdflush_operation(background_writeout, 0);
462}
463
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200464void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700465{
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200466 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700467 struct address_space *mapping = page_mapping(page);
468
469 if (mapping)
470 balance_dirty_pages_ratelimited(mapping);
471 }
472}
473
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800475 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -0700476 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +0200477 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 *
479 * Processes which are dirtying memory should call in here once for each page
480 * which was newly dirtied. The function will periodically check the system's
481 * dirty state and will initiate writeback if needed.
482 *
483 * On really big machines, get_writeback_state is expensive, so try to avoid
484 * calling it too often (ratelimiting). But once we're over the dirty memory
485 * limit we decrease the ratelimiting by a lot, to prevent individual processes
486 * from overshooting the limit by (ratelimit_pages) each.
487 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800488void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
489 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490{
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800491 static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
492 unsigned long ratelimit;
493 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494
495 ratelimit = ratelimit_pages;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700496 if (mapping->backing_dev_info->dirty_exceeded)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 ratelimit = 8;
498
499 /*
500 * Check the rate limiting. Also, we do not want to throttle real-time
501 * tasks in balance_dirty_pages(). Period.
502 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800503 preempt_disable();
504 p = &__get_cpu_var(ratelimits);
505 *p += nr_pages_dirtied;
506 if (unlikely(*p >= ratelimit)) {
507 *p = 0;
508 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 balance_dirty_pages(mapping);
510 return;
511 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800512 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513}
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800514EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
Andrew Morton232ea4d2007-02-28 20:13:21 -0800516void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 long background_thresh;
519 long dirty_thresh;
520
521 for ( ; ; ) {
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700522 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523
524 /*
525 * Boost the allowable dirty threshold a bit for page
526 * allocators so they don't get DoS'ed by heavy writers
527 */
528 dirty_thresh += dirty_thresh / 10; /* wheeee... */
529
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700530 if (global_page_state(NR_UNSTABLE_NFS) +
531 global_page_state(NR_WRITEBACK) <= dirty_thresh)
532 break;
Andrew Morton3fcfab12006-10-19 23:28:16 -0700533 congestion_wait(WRITE, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -0700534
535 /*
536 * The caller might hold locks which can prevent IO completion
537 * or progress in the filesystem. So we cannot just sit here
538 * waiting for IO to complete.
539 */
540 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
541 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 }
543}
544
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545/*
546 * writeback at least _min_pages, and keep writing until the amount of dirty
547 * memory is less than the background threshold, or until we're all clean.
548 */
549static void background_writeout(unsigned long _min_pages)
550{
551 long min_pages = _min_pages;
552 struct writeback_control wbc = {
553 .bdi = NULL,
554 .sync_mode = WB_SYNC_NONE,
555 .older_than_this = NULL,
556 .nr_to_write = 0,
557 .nonblocking = 1,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700558 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 };
560
561 for ( ; ; ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 long background_thresh;
563 long dirty_thresh;
564
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700565 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700566 if (global_page_state(NR_FILE_DIRTY) +
567 global_page_state(NR_UNSTABLE_NFS) < background_thresh
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 && min_pages <= 0)
569 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800570 wbc.more_io = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 wbc.encountered_congestion = 0;
572 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
573 wbc.pages_skipped = 0;
574 writeback_inodes(&wbc);
575 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
576 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
577 /* Wrote less than expected */
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800578 if (wbc.encountered_congestion || wbc.more_io)
579 congestion_wait(WRITE, HZ/10);
580 else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 break;
582 }
583 }
584}
585
586/*
587 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
588 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
589 * -1 if all pdflush threads were busy.
590 */
Pekka J Enberg687a21c2005-06-28 20:44:55 -0700591int wakeup_pdflush(long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700593 if (nr_pages == 0)
594 nr_pages = global_page_state(NR_FILE_DIRTY) +
595 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 return pdflush_operation(background_writeout, nr_pages);
597}
598
599static void wb_timer_fn(unsigned long unused);
600static void laptop_timer_fn(unsigned long unused);
601
Ingo Molnar8d06afa2005-09-09 13:10:40 -0700602static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
603static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
605/*
606 * Periodic writeback of "old" data.
607 *
608 * Define "old": the first time one of an inode's pages is dirtied, we mark the
609 * dirtying-time in the inode's address_space. So this periodic writeback code
610 * just walks the superblock inode list, writing back any inodes which are
611 * older than a specific point in time.
612 *
Bart Samwelf6ef9432006-03-24 03:15:48 -0800613 * Try to run once per dirty_writeback_interval. But if a writeback event
614 * takes longer than a dirty_writeback_interval interval, then leave a
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 * one-second gap.
616 *
617 * older_than_this takes precedence over nr_to_write. So we'll only write back
618 * all dirty pages if they are all attached to "old" mappings.
619 */
620static void wb_kupdate(unsigned long arg)
621{
622 unsigned long oldest_jif;
623 unsigned long start_jif;
624 unsigned long next_jif;
625 long nr_to_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 struct writeback_control wbc = {
627 .bdi = NULL,
628 .sync_mode = WB_SYNC_NONE,
629 .older_than_this = &oldest_jif,
630 .nr_to_write = 0,
631 .nonblocking = 1,
632 .for_kupdate = 1,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700633 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 };
635
636 sync_supers();
637
Bart Samwelf6ef9432006-03-24 03:15:48 -0800638 oldest_jif = jiffies - dirty_expire_interval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 start_jif = jiffies;
Bart Samwelf6ef9432006-03-24 03:15:48 -0800640 next_jif = start_jif + dirty_writeback_interval;
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700641 nr_to_write = global_page_state(NR_FILE_DIRTY) +
642 global_page_state(NR_UNSTABLE_NFS) +
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
644 while (nr_to_write > 0) {
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800645 wbc.more_io = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 wbc.encountered_congestion = 0;
647 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
648 writeback_inodes(&wbc);
649 if (wbc.nr_to_write > 0) {
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800650 if (wbc.encountered_congestion || wbc.more_io)
Andrew Morton3fcfab12006-10-19 23:28:16 -0700651 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 else
653 break; /* All the old data is written */
654 }
655 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
656 }
657 if (time_before(next_jif, jiffies + HZ))
658 next_jif = jiffies + HZ;
Bart Samwelf6ef9432006-03-24 03:15:48 -0800659 if (dirty_writeback_interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 mod_timer(&wb_timer, next_jif);
661}
662
663/*
664 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
665 */
666int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Andrew Morton3e733f02007-07-15 23:41:05 -0700667 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668{
Bart Samwelf6ef9432006-03-24 03:15:48 -0800669 proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
Andrew Morton3e733f02007-07-15 23:41:05 -0700670 if (dirty_writeback_interval)
671 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
672 else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 del_timer(&wb_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 return 0;
675}
676
677static void wb_timer_fn(unsigned long unused)
678{
679 if (pdflush_operation(wb_kupdate, 0) < 0)
680 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
681}
682
683static void laptop_flush(unsigned long unused)
684{
685 sys_sync();
686}
687
688static void laptop_timer_fn(unsigned long unused)
689{
690 pdflush_operation(laptop_flush, 0);
691}
692
693/*
694 * We've spun up the disk and we're in laptop mode: schedule writeback
695 * of all dirty data a few seconds from now. If the flush is already scheduled
696 * then push it back - the user is still using the disk.
697 */
698void laptop_io_completion(void)
699{
Bart Samweled5b43f2006-03-24 03:15:49 -0800700 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
702
703/*
704 * We're in laptop mode and we've just synced. The sync's writes will have
705 * caused another writeback to be scheduled by laptop_io_completion.
706 * Nothing needs to be written back anymore, so we unschedule the writeback.
707 */
708void laptop_sync_completion(void)
709{
710 del_timer(&laptop_mode_wb_timer);
711}
712
713/*
714 * If ratelimit_pages is too high then we can get into dirty-data overload
715 * if a large number of processes all perform writes at the same time.
716 * If it is too low then SMP machines will call the (expensive)
717 * get_writeback_state too often.
718 *
719 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
720 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
721 * thresholds before writeback cuts in.
722 *
723 * But the limit should not be set too high. Because it also controls the
724 * amount of memory which the balance_dirty_pages() caller has to write back.
725 * If this is too large then the caller will block on the IO queue all the
726 * time. So limit it to four megabytes - the balance_dirty_pages() caller
727 * will write six megabyte chunks, max.
728 */
729
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700730void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700732 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 if (ratelimit_pages < 16)
734 ratelimit_pages = 16;
735 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
736 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
737}
738
Chandra Seetharaman26c21432006-06-27 02:54:10 -0700739static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
741{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700742 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -0800743 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744}
745
Chandra Seetharaman74b85f32006-06-27 02:54:09 -0700746static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 .notifier_call = ratelimit_handler,
748 .next = NULL,
749};
750
751/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800752 * Called early on to tune the page writeback dirty limits.
753 *
754 * We used to scale dirty pages according to how total memory
755 * related to pages that could be allocated for buffers (by
756 * comparing nr_free_buffer_pages() to vm_total_pages.
757 *
758 * However, that was when we used "dirty_ratio" to scale with
759 * all memory, and we don't do that any more. "dirty_ratio"
760 * is now applied to total non-HIGHPAGE memory (by subtracting
761 * totalhigh_pages from vm_total_pages), and as such we can't
762 * get into the old insane situation any more where we had
763 * large amounts of dirty pages compared to a small amount of
764 * non-HIGHMEM memory.
765 *
766 * But we might still want to scale the dirty_ratio by how
767 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 */
769void __init page_writeback_init(void)
770{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700771 int shift;
772
Bart Samwelf6ef9432006-03-24 03:15:48 -0800773 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700774 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700776
777 shift = calc_period_shift();
778 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700779 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780}
781
David Howells811d7362006-08-29 19:06:09 +0100782/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700783 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
David Howells811d7362006-08-29 19:06:09 +0100784 * @mapping: address space structure to write
785 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700786 * @writepage: function called for each page
787 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +0100788 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700789 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +0100790 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
791 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
792 * and msync() need to guarantee that all the data which was dirty at the time
793 * the call was made get new I/O started against them. If wbc->sync_mode is
794 * WB_SYNC_ALL then we were called for data integrity and we must wait for
795 * existing IO to complete.
David Howells811d7362006-08-29 19:06:09 +0100796 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700797int write_cache_pages(struct address_space *mapping,
798 struct writeback_control *wbc, writepage_t writepage,
799 void *data)
David Howells811d7362006-08-29 19:06:09 +0100800{
801 struct backing_dev_info *bdi = mapping->backing_dev_info;
802 int ret = 0;
803 int done = 0;
David Howells811d7362006-08-29 19:06:09 +0100804 struct pagevec pvec;
805 int nr_pages;
806 pgoff_t index;
807 pgoff_t end; /* Inclusive */
808 int scanned = 0;
809 int range_whole = 0;
810
811 if (wbc->nonblocking && bdi_write_congested(bdi)) {
812 wbc->encountered_congestion = 1;
813 return 0;
814 }
815
David Howells811d7362006-08-29 19:06:09 +0100816 pagevec_init(&pvec, 0);
817 if (wbc->range_cyclic) {
818 index = mapping->writeback_index; /* Start from prev offset */
819 end = -1;
820 } else {
821 index = wbc->range_start >> PAGE_CACHE_SHIFT;
822 end = wbc->range_end >> PAGE_CACHE_SHIFT;
823 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
824 range_whole = 1;
825 scanned = 1;
826 }
827retry:
828 while (!done && (index <= end) &&
829 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
830 PAGECACHE_TAG_DIRTY,
831 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
832 unsigned i;
833
834 scanned = 1;
835 for (i = 0; i < nr_pages; i++) {
836 struct page *page = pvec.pages[i];
837
838 /*
839 * At this point we hold neither mapping->tree_lock nor
840 * lock on the page itself: the page may be truncated or
841 * invalidated (changing page->mapping to NULL), or even
842 * swizzled back from swapper_space to tmpfs file
843 * mapping
844 */
845 lock_page(page);
846
847 if (unlikely(page->mapping != mapping)) {
848 unlock_page(page);
849 continue;
850 }
851
852 if (!wbc->range_cyclic && page->index > end) {
853 done = 1;
854 unlock_page(page);
855 continue;
856 }
857
858 if (wbc->sync_mode != WB_SYNC_NONE)
859 wait_on_page_writeback(page);
860
861 if (PageWriteback(page) ||
862 !clear_page_dirty_for_io(page)) {
863 unlock_page(page);
864 continue;
865 }
866
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700867 ret = (*writepage)(page, wbc, data);
David Howells811d7362006-08-29 19:06:09 +0100868
Andrew Mortone4230032007-10-16 23:26:02 -0700869 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
David Howells811d7362006-08-29 19:06:09 +0100870 unlock_page(page);
Andrew Mortone4230032007-10-16 23:26:02 -0700871 ret = 0;
872 }
David Howells811d7362006-08-29 19:06:09 +0100873 if (ret || (--(wbc->nr_to_write) <= 0))
874 done = 1;
875 if (wbc->nonblocking && bdi_write_congested(bdi)) {
876 wbc->encountered_congestion = 1;
877 done = 1;
878 }
879 }
880 pagevec_release(&pvec);
881 cond_resched();
882 }
883 if (!scanned && !done) {
884 /*
885 * We hit the last page and there is more work to be done: wrap
886 * back to the start of the file
887 */
888 scanned = 1;
889 index = 0;
890 goto retry;
891 }
892 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
893 mapping->writeback_index = index;
894 return ret;
895}
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700896EXPORT_SYMBOL(write_cache_pages);
897
898/*
899 * Function used by generic_writepages to call the real writepage
900 * function and set the mapping flags on error
901 */
902static int __writepage(struct page *page, struct writeback_control *wbc,
903 void *data)
904{
905 struct address_space *mapping = data;
906 int ret = mapping->a_ops->writepage(page, wbc);
907 mapping_set_error(mapping, ret);
908 return ret;
909}
910
911/**
912 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
913 * @mapping: address space structure to write
914 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
915 *
916 * This is a library function, which implements the writepages()
917 * address_space_operation.
918 */
919int generic_writepages(struct address_space *mapping,
920 struct writeback_control *wbc)
921{
922 /* deal with chardevs and other special file */
923 if (!mapping->a_ops->writepage)
924 return 0;
925
926 return write_cache_pages(mapping, wbc, __writepage, mapping);
927}
David Howells811d7362006-08-29 19:06:09 +0100928
929EXPORT_SYMBOL(generic_writepages);
930
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
932{
Andrew Morton22905f72005-11-16 15:07:01 -0800933 int ret;
934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 if (wbc->nr_to_write <= 0)
936 return 0;
Andrew Morton22905f72005-11-16 15:07:01 -0800937 wbc->for_writepages = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700939 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -0800940 else
941 ret = generic_writepages(mapping, wbc);
942 wbc->for_writepages = 0;
943 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944}
945
946/**
947 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -0700948 * @page: the page to write
949 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 *
951 * The page must be locked by the caller and will be unlocked upon return.
952 *
953 * write_one_page() returns a negative error code if I/O failed.
954 */
955int write_one_page(struct page *page, int wait)
956{
957 struct address_space *mapping = page->mapping;
958 int ret = 0;
959 struct writeback_control wbc = {
960 .sync_mode = WB_SYNC_ALL,
961 .nr_to_write = 1,
962 };
963
964 BUG_ON(!PageLocked(page));
965
966 if (wait)
967 wait_on_page_writeback(page);
968
969 if (clear_page_dirty_for_io(page)) {
970 page_cache_get(page);
971 ret = mapping->a_ops->writepage(page, &wbc);
972 if (ret == 0 && wait) {
973 wait_on_page_writeback(page);
974 if (PageError(page))
975 ret = -EIO;
976 }
977 page_cache_release(page);
978 } else {
979 unlock_page(page);
980 }
981 return ret;
982}
983EXPORT_SYMBOL(write_one_page);
984
985/*
Ken Chen76719322007-02-10 01:43:15 -0800986 * For address_spaces which do not use buffers nor write back.
987 */
988int __set_page_dirty_no_writeback(struct page *page)
989{
990 if (!PageDirty(page))
991 SetPageDirty(page);
992 return 0;
993}
994
995/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 * For address_spaces which do not use buffers. Just tag the page as dirty in
997 * its radix tree.
998 *
999 * This is also used when a single buffer is being dirtied: we want to set the
1000 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1001 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1002 *
1003 * Most callers have locked the page, which pins the address_space in memory.
1004 * But zap_pte_range() does not lock the page, however in that case the
1005 * mapping is pinned by the vma's ->vm_file reference.
1006 *
1007 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001008 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 */
1010int __set_page_dirty_nobuffers(struct page *page)
1011{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 if (!TestSetPageDirty(page)) {
1013 struct address_space *mapping = page_mapping(page);
1014 struct address_space *mapping2;
1015
Andrew Morton8c085402006-12-10 02:19:24 -08001016 if (!mapping)
1017 return 1;
1018
1019 write_lock_irq(&mapping->tree_lock);
1020 mapping2 = page_mapping(page);
1021 if (mapping2) { /* Race with truncate? */
1022 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001023 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Andrew Morton55e829a2006-12-10 02:19:27 -08001024 if (mapping_cap_account_dirty(mapping)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001025 __inc_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001026 __inc_bdi_stat(mapping->backing_dev_info,
1027 BDI_RECLAIMABLE);
Andrew Morton55e829a2006-12-10 02:19:27 -08001028 task_io_account_write(PAGE_CACHE_SIZE);
1029 }
Andrew Morton8c085402006-12-10 02:19:24 -08001030 radix_tree_tag_set(&mapping->page_tree,
1031 page_index(page), PAGECACHE_TAG_DIRTY);
1032 }
1033 write_unlock_irq(&mapping->tree_lock);
1034 if (mapping->host) {
1035 /* !PageAnon && !swapper_space */
1036 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001038 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001040 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041}
1042EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1043
1044/*
1045 * When a writepage implementation decides that it doesn't want to write this
1046 * page for some reason, it should redirty the locked page via
1047 * redirty_page_for_writepage() and it should then unlock the page and return 0
1048 */
1049int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1050{
1051 wbc->pages_skipped++;
1052 return __set_page_dirty_nobuffers(page);
1053}
1054EXPORT_SYMBOL(redirty_page_for_writepage);
1055
1056/*
1057 * If the mapping doesn't provide a set_page_dirty a_op, then
1058 * just fall through and assume that it wants buffer_heads.
1059 */
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001060static int __set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061{
1062 struct address_space *mapping = page_mapping(page);
1063
1064 if (likely(mapping)) {
1065 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
David Howells93614012006-09-30 20:45:40 +02001066#ifdef CONFIG_BLOCK
1067 if (!spd)
1068 spd = __set_page_dirty_buffers;
1069#endif
1070 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001072 if (!PageDirty(page)) {
1073 if (!TestSetPageDirty(page))
1074 return 1;
1075 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 return 0;
1077}
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001078
Harvey Harrison920c7a52008-02-04 22:29:26 -08001079int set_page_dirty(struct page *page)
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001080{
1081 int ret = __set_page_dirty(page);
1082 if (ret)
1083 task_dirty_inc(current);
1084 return ret;
1085}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086EXPORT_SYMBOL(set_page_dirty);
1087
1088/*
1089 * set_page_dirty() is racy if the caller has no reference against
1090 * page->mapping->host, and if the page is unlocked. This is because another
1091 * CPU could truncate the page off the mapping and then free the mapping.
1092 *
1093 * Usually, the page _is_ locked, or the caller is a user-space process which
1094 * holds a reference on the inode by having an open file.
1095 *
1096 * In other cases, the page should be locked before running set_page_dirty().
1097 */
1098int set_page_dirty_lock(struct page *page)
1099{
1100 int ret;
1101
Nick Piggindb376482006-09-25 23:31:24 -07001102 lock_page_nosync(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 ret = set_page_dirty(page);
1104 unlock_page(page);
1105 return ret;
1106}
1107EXPORT_SYMBOL(set_page_dirty_lock);
1108
1109/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 * Clear a page's dirty flag, while caring for dirty memory accounting.
1111 * Returns true if the page was previously dirty.
1112 *
1113 * This is for preparing to put the page under writeout. We leave the page
1114 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1115 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1116 * implementation will run either set_page_writeback() or set_page_dirty(),
1117 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1118 * back into sync.
1119 *
1120 * This incoherency between the page's dirty flag and radix-tree tag is
1121 * unfortunate, but it only exists while the page is locked.
1122 */
1123int clear_page_dirty_for_io(struct page *page)
1124{
1125 struct address_space *mapping = page_mapping(page);
1126
Nick Piggin79352892007-07-19 01:47:22 -07001127 BUG_ON(!PageLocked(page));
1128
Fengguang Wufe3cba12007-07-19 01:48:07 -07001129 ClearPageReclaim(page);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001130 if (mapping && mapping_cap_account_dirty(mapping)) {
1131 /*
1132 * Yes, Virginia, this is indeed insane.
1133 *
1134 * We use this sequence to make sure that
1135 * (a) we account for dirty stats properly
1136 * (b) we tell the low-level filesystem to
1137 * mark the whole page dirty if it was
1138 * dirty in a pagetable. Only to then
1139 * (c) clean the page again and return 1 to
1140 * cause the writeback.
1141 *
1142 * This way we avoid all nasty races with the
1143 * dirty bit in multiple places and clearing
1144 * them concurrently from different threads.
1145 *
1146 * Note! Normally the "set_page_dirty(page)"
1147 * has no effect on the actual dirty bit - since
1148 * that will already usually be set. But we
1149 * need the side effects, and it can help us
1150 * avoid races.
1151 *
1152 * We basically use the page "master dirty bit"
1153 * as a serialization point for all the different
1154 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001155 */
1156 if (page_mkclean(page))
1157 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001158 /*
1159 * We carefully synchronise fault handlers against
1160 * installing a dirty pte and marking the page dirty
1161 * at this point. We do this by having them hold the
1162 * page lock at some point after installing their
1163 * pte, but before marking the page dirty.
1164 * Pages are always locked coming in here, so we get
1165 * the desired exclusion. See mm/memory.c:do_wp_page()
1166 * for more comments.
1167 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001168 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001169 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001170 dec_bdi_stat(mapping->backing_dev_info,
1171 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001172 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001174 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001176 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001178EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179
1180int test_clear_page_writeback(struct page *page)
1181{
1182 struct address_space *mapping = page_mapping(page);
1183 int ret;
1184
1185 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001186 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 unsigned long flags;
1188
1189 write_lock_irqsave(&mapping->tree_lock, flags);
1190 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001191 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 radix_tree_tag_clear(&mapping->page_tree,
1193 page_index(page),
1194 PAGECACHE_TAG_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001195 if (bdi_cap_writeback_dirty(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001196 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001197 __bdi_writeout_inc(bdi);
1198 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 write_unlock_irqrestore(&mapping->tree_lock, flags);
1201 } else {
1202 ret = TestClearPageWriteback(page);
1203 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001204 if (ret)
1205 dec_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 return ret;
1207}
1208
1209int test_set_page_writeback(struct page *page)
1210{
1211 struct address_space *mapping = page_mapping(page);
1212 int ret;
1213
1214 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001215 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 unsigned long flags;
1217
1218 write_lock_irqsave(&mapping->tree_lock, flags);
1219 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001220 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 radix_tree_tag_set(&mapping->page_tree,
1222 page_index(page),
1223 PAGECACHE_TAG_WRITEBACK);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001224 if (bdi_cap_writeback_dirty(bdi))
1225 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1226 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 if (!PageDirty(page))
1228 radix_tree_tag_clear(&mapping->page_tree,
1229 page_index(page),
1230 PAGECACHE_TAG_DIRTY);
1231 write_unlock_irqrestore(&mapping->tree_lock, flags);
1232 } else {
1233 ret = TestSetPageWriteback(page);
1234 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001235 if (!ret)
1236 inc_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 return ret;
1238
1239}
1240EXPORT_SYMBOL(test_set_page_writeback);
1241
1242/*
Nick Piggin00128182007-10-16 01:24:40 -07001243 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 * passed tag.
1245 */
1246int mapping_tagged(struct address_space *mapping, int tag)
1247{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 int ret;
Nick Piggin00128182007-10-16 01:24:40 -07001249 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 ret = radix_tree_tagged(&mapping->page_tree, tag);
Nick Piggin00128182007-10-16 01:24:40 -07001251 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 return ret;
1253}
1254EXPORT_SYMBOL(mapping_tagged);