blob: 838a5e31394cbbbe0e9a34b66e1570cf03b5bb86 [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/*
72 * The generator of dirty data starts writeback at this percentage
73 */
Linus Torvalds07db59b2007-04-27 09:10:47 -070074int vm_dirty_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
76/*
Coywolf Qi Huntfd5403c2006-04-10 22:54:35 -070077 * The interval between `kupdate'-style writebacks, in jiffies
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 */
Bart Samwelf6ef9432006-03-24 03:15:48 -080079int dirty_writeback_interval = 5 * HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -070080
81/*
Coywolf Qi Huntfd5403c2006-04-10 22:54:35 -070082 * The longest number of jiffies for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -070083 */
Bart Samwelf6ef9432006-03-24 03:15:48 -080084int dirty_expire_interval = 30 * HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
86/*
87 * Flag that makes the machine dump writes/reads and block dirtyings.
88 */
89int block_dump;
90
91/*
Bart Samweled5b43f2006-03-24 03:15:49 -080092 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
93 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -070094 */
95int laptop_mode;
96
97EXPORT_SYMBOL(laptop_mode);
98
99/* End of sysctl-exported parameters */
100
101
102static void background_writeout(unsigned long _min_pages);
103
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700105 * Scale the writeback cache size proportional to the relative writeout speeds.
106 *
107 * We do this by keeping a floating proportion between BDIs, based on page
108 * writeback completions [end_page_writeback()]. Those devices that write out
109 * pages fastest will get the larger share, while the slower will get a smaller
110 * share.
111 *
112 * We use page writeout completions because we are interested in getting rid of
113 * dirty pages. Having them written out is the primary goal.
114 *
115 * We introduce a concept of time, a period over which we measure these events,
116 * because demand can/will vary over time. The length of this period itself is
117 * measured in page writeback completions.
118 *
119 */
120static struct prop_descriptor vm_completions;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700121static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700122
123static unsigned long determine_dirtyable_memory(void);
124
125/*
126 * couple the period to the dirty_ratio:
127 *
128 * period/2 ~ roundup_pow_of_two(dirty limit)
129 */
130static int calc_period_shift(void)
131{
132 unsigned long dirty_total;
133
134 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
135 return 2 + ilog2(dirty_total - 1);
136}
137
138/*
139 * update the period when the dirty ratio changes.
140 */
141int dirty_ratio_handler(struct ctl_table *table, int write,
142 struct file *filp, void __user *buffer, size_t *lenp,
143 loff_t *ppos)
144{
145 int old_ratio = vm_dirty_ratio;
146 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
147 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
148 int shift = calc_period_shift();
149 prop_change_shift(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700150 prop_change_shift(&vm_dirties, shift);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700151 }
152 return ret;
153}
154
155/*
156 * Increment the BDI's writeout completion count and the global writeout
157 * completion count. Called from test_clear_page_writeback().
158 */
159static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
160{
161 __prop_inc_percpu(&vm_completions, &bdi->completions);
162}
163
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700164static inline void task_dirty_inc(struct task_struct *tsk)
165{
166 prop_inc_single(&vm_dirties, &tsk->dirties);
167}
168
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700169/*
170 * Obtain an accurate fraction of the BDI's portion.
171 */
172static void bdi_writeout_fraction(struct backing_dev_info *bdi,
173 long *numerator, long *denominator)
174{
175 if (bdi_cap_writeback_dirty(bdi)) {
176 prop_fraction_percpu(&vm_completions, &bdi->completions,
177 numerator, denominator);
178 } else {
179 *numerator = 0;
180 *denominator = 1;
181 }
182}
183
184/*
185 * Clip the earned share of dirty pages to that which is actually available.
186 * This avoids exceeding the total dirty_limit when the floating averages
187 * fluctuate too quickly.
188 */
189static void
190clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
191{
192 long avail_dirty;
193
194 avail_dirty = dirty -
195 (global_page_state(NR_FILE_DIRTY) +
196 global_page_state(NR_WRITEBACK) +
197 global_page_state(NR_UNSTABLE_NFS));
198
199 if (avail_dirty < 0)
200 avail_dirty = 0;
201
202 avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
203 bdi_stat(bdi, BDI_WRITEBACK);
204
205 *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
206}
207
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700208static inline void task_dirties_fraction(struct task_struct *tsk,
209 long *numerator, long *denominator)
210{
211 prop_fraction_single(&vm_dirties, &tsk->dirties,
212 numerator, denominator);
213}
214
215/*
216 * scale the dirty limit
217 *
218 * task specific dirty limit:
219 *
220 * dirty -= (dirty/8) * p_{t}
221 */
222void task_dirty_limit(struct task_struct *tsk, long *pdirty)
223{
224 long numerator, denominator;
225 long dirty = *pdirty;
226 u64 inv = dirty >> 3;
227
228 task_dirties_fraction(tsk, &numerator, &denominator);
229 inv *= numerator;
230 do_div(inv, denominator);
231
232 dirty -= inv;
233 if (dirty < *pdirty/2)
234 dirty = *pdirty/2;
235
236 *pdirty = dirty;
237}
238
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700239/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 * Work out the current dirty-memory clamping and background writeout
241 * thresholds.
242 *
243 * The main aim here is to lower them aggressively if there is a lot of mapped
244 * memory around. To avoid stressing page reclaim with lots of unreclaimable
245 * pages. It is better to clamp down on writers than to start swapping, and
246 * performing lots of scanning.
247 *
248 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
249 *
250 * We don't permit the clamping level to fall below 5% - that is getting rather
251 * excessive.
252 *
253 * We make sure that the background writeout level is below the adjusted
254 * clamping level.
255 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700256
257static unsigned long highmem_dirtyable_memory(unsigned long total)
258{
259#ifdef CONFIG_HIGHMEM
260 int node;
261 unsigned long x = 0;
262
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700263 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700264 struct zone *z =
265 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
266
267 x += zone_page_state(z, NR_FREE_PAGES)
268 + zone_page_state(z, NR_INACTIVE)
269 + zone_page_state(z, NR_ACTIVE);
270 }
271 /*
272 * Make sure that the number of highmem pages is never larger
273 * than the number of the total dirtyable memory. This can only
274 * occur in very strange VM situations but we want to make sure
275 * that this does not occur.
276 */
277 return min(x, total);
278#else
279 return 0;
280#endif
281}
282
283static unsigned long determine_dirtyable_memory(void)
284{
285 unsigned long x;
286
287 x = global_page_state(NR_FREE_PAGES)
288 + global_page_state(NR_INACTIVE)
289 + global_page_state(NR_ACTIVE);
290 x -= highmem_dirtyable_memory(x);
291 return x + 1; /* Ensure that we never return 0 */
292}
293
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294static void
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700295get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
296 struct backing_dev_info *bdi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 int background_ratio; /* Percentages */
299 int dirty_ratio;
300 int unmapped_ratio;
301 long background;
302 long dirty;
Christoph Lameter1b424462007-05-06 14:48:59 -0700303 unsigned long available_memory = determine_dirtyable_memory();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 struct task_struct *tsk;
305
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700306 unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) +
307 global_page_state(NR_ANON_PAGES)) * 100) /
Christoph Lameter1b424462007-05-06 14:48:59 -0700308 available_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
310 dirty_ratio = vm_dirty_ratio;
311 if (dirty_ratio > unmapped_ratio / 2)
312 dirty_ratio = unmapped_ratio / 2;
313
314 if (dirty_ratio < 5)
315 dirty_ratio = 5;
316
317 background_ratio = dirty_background_ratio;
318 if (background_ratio >= dirty_ratio)
319 background_ratio = dirty_ratio / 2;
320
321 background = (background_ratio * available_memory) / 100;
322 dirty = (dirty_ratio * available_memory) / 100;
323 tsk = current;
324 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
325 background += background / 4;
326 dirty += dirty / 4;
327 }
328 *pbackground = background;
329 *pdirty = dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700330
331 if (bdi) {
332 u64 bdi_dirty = dirty;
333 long numerator, denominator;
334
335 /*
336 * Calculate this BDI's share of the dirty ratio.
337 */
338 bdi_writeout_fraction(bdi, &numerator, &denominator);
339
340 bdi_dirty *= numerator;
341 do_div(bdi_dirty, denominator);
342
343 *pbdi_dirty = bdi_dirty;
344 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700345 task_dirty_limit(current, pbdi_dirty);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347}
348
349/*
350 * balance_dirty_pages() must be called by processes which are generating dirty
351 * data. It looks at the number of dirty pages in the machine and will force
352 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
353 * If we're over `background_thresh' then pdflush is woken to perform some
354 * writeout.
355 */
356static void balance_dirty_pages(struct address_space *mapping)
357{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700358 long bdi_nr_reclaimable;
359 long bdi_nr_writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 long background_thresh;
361 long dirty_thresh;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700362 long bdi_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 unsigned long pages_written = 0;
364 unsigned long write_chunk = sync_writeback_pages();
365
366 struct backing_dev_info *bdi = mapping->backing_dev_info;
367
368 for (;;) {
369 struct writeback_control wbc = {
370 .bdi = bdi,
371 .sync_mode = WB_SYNC_NONE,
372 .older_than_this = NULL,
373 .nr_to_write = write_chunk,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700374 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 };
376
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700377 get_dirty_limits(&background_thresh, &dirty_thresh,
378 &bdi_thresh, bdi);
379 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
380 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
381 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
382 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700384 if (!bdi->dirty_exceeded)
385 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
387 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
388 * Unstable writes are a feature of certain networked
389 * filesystems (i.e. NFS) in which data may have been
390 * written to the server's write cache, but has not yet
391 * been flushed to permanent storage.
392 */
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700393 if (bdi_nr_reclaimable) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 writeback_inodes(&wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 pages_written += write_chunk - wbc.nr_to_write;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700396 get_dirty_limits(&background_thresh, &dirty_thresh,
397 &bdi_thresh, bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700399
400 /*
401 * In order to avoid the stacked BDI deadlock we need
402 * to ensure we accurately count the 'dirty' pages when
403 * the threshold is low.
404 *
405 * Otherwise it would be possible to get thresh+n pages
406 * reported dirty, even though there are thresh-m pages
407 * actually dirty; with m+n sitting in the percpu
408 * deltas.
409 */
410 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
411 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
412 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
413 } else if (bdi_nr_reclaimable) {
414 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
415 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
416 }
417
418 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
419 break;
420 if (pages_written >= write_chunk)
421 break; /* We've done our duty */
422
Andrew Morton3fcfab12006-10-19 23:28:16 -0700423 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 }
425
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700426 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
427 bdi->dirty_exceeded)
428 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
430 if (writeback_in_progress(bdi))
431 return; /* pdflush is already working this queue */
432
433 /*
434 * In laptop mode, we wait until hitting the higher threshold before
435 * starting background writeout, and then write out all the way down
436 * to the lower threshold. So slow writers cause minimal disk activity.
437 *
438 * In normal mode, we start background writeout at the lower
439 * background_thresh, to keep the amount of dirty memory low.
440 */
441 if ((laptop_mode && pages_written) ||
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700442 (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
443 + global_page_state(NR_UNSTABLE_NFS)
444 > background_thresh)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 pdflush_operation(background_writeout, 0);
446}
447
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200448void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700449{
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200450 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700451 struct address_space *mapping = page_mapping(page);
452
453 if (mapping)
454 balance_dirty_pages_ratelimited(mapping);
455 }
456}
457
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800459 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -0700460 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +0200461 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 *
463 * Processes which are dirtying memory should call in here once for each page
464 * which was newly dirtied. The function will periodically check the system's
465 * dirty state and will initiate writeback if needed.
466 *
467 * On really big machines, get_writeback_state is expensive, so try to avoid
468 * calling it too often (ratelimiting). But once we're over the dirty memory
469 * limit we decrease the ratelimiting by a lot, to prevent individual processes
470 * from overshooting the limit by (ratelimit_pages) each.
471 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800472void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
473 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474{
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800475 static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
476 unsigned long ratelimit;
477 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
479 ratelimit = ratelimit_pages;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700480 if (mapping->backing_dev_info->dirty_exceeded)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 ratelimit = 8;
482
483 /*
484 * Check the rate limiting. Also, we do not want to throttle real-time
485 * tasks in balance_dirty_pages(). Period.
486 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800487 preempt_disable();
488 p = &__get_cpu_var(ratelimits);
489 *p += nr_pages_dirtied;
490 if (unlikely(*p >= ratelimit)) {
491 *p = 0;
492 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 balance_dirty_pages(mapping);
494 return;
495 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800496 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497}
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800498EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
Andrew Morton232ea4d2007-02-28 20:13:21 -0800500void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 long background_thresh;
503 long dirty_thresh;
504
505 for ( ; ; ) {
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700506 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507
508 /*
509 * Boost the allowable dirty threshold a bit for page
510 * allocators so they don't get DoS'ed by heavy writers
511 */
512 dirty_thresh += dirty_thresh / 10; /* wheeee... */
513
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700514 if (global_page_state(NR_UNSTABLE_NFS) +
515 global_page_state(NR_WRITEBACK) <= dirty_thresh)
516 break;
Andrew Morton3fcfab12006-10-19 23:28:16 -0700517 congestion_wait(WRITE, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -0700518
519 /*
520 * The caller might hold locks which can prevent IO completion
521 * or progress in the filesystem. So we cannot just sit here
522 * waiting for IO to complete.
523 */
524 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
525 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 }
527}
528
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529/*
530 * writeback at least _min_pages, and keep writing until the amount of dirty
531 * memory is less than the background threshold, or until we're all clean.
532 */
533static void background_writeout(unsigned long _min_pages)
534{
535 long min_pages = _min_pages;
536 struct writeback_control wbc = {
537 .bdi = NULL,
538 .sync_mode = WB_SYNC_NONE,
539 .older_than_this = NULL,
540 .nr_to_write = 0,
541 .nonblocking = 1,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700542 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 };
544
545 for ( ; ; ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 long background_thresh;
547 long dirty_thresh;
548
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700549 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700550 if (global_page_state(NR_FILE_DIRTY) +
551 global_page_state(NR_UNSTABLE_NFS) < background_thresh
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 && min_pages <= 0)
553 break;
Fengguang Wu2e6883b2007-10-16 23:30:43 -0700554 wbc.more_io = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 wbc.encountered_congestion = 0;
556 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
557 wbc.pages_skipped = 0;
558 writeback_inodes(&wbc);
559 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
560 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
561 /* Wrote less than expected */
Fengguang Wu2e6883b2007-10-16 23:30:43 -0700562 if (wbc.encountered_congestion || wbc.more_io)
563 congestion_wait(WRITE, HZ/10);
564 else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 break;
566 }
567 }
568}
569
570/*
571 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
572 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
573 * -1 if all pdflush threads were busy.
574 */
Pekka J Enberg687a21c2005-06-28 20:44:55 -0700575int wakeup_pdflush(long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576{
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700577 if (nr_pages == 0)
578 nr_pages = global_page_state(NR_FILE_DIRTY) +
579 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 return pdflush_operation(background_writeout, nr_pages);
581}
582
583static void wb_timer_fn(unsigned long unused);
584static void laptop_timer_fn(unsigned long unused);
585
Ingo Molnar8d06afa2005-09-09 13:10:40 -0700586static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
587static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
589/*
590 * Periodic writeback of "old" data.
591 *
592 * Define "old": the first time one of an inode's pages is dirtied, we mark the
593 * dirtying-time in the inode's address_space. So this periodic writeback code
594 * just walks the superblock inode list, writing back any inodes which are
595 * older than a specific point in time.
596 *
Bart Samwelf6ef9432006-03-24 03:15:48 -0800597 * Try to run once per dirty_writeback_interval. But if a writeback event
598 * takes longer than a dirty_writeback_interval interval, then leave a
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 * one-second gap.
600 *
601 * older_than_this takes precedence over nr_to_write. So we'll only write back
602 * all dirty pages if they are all attached to "old" mappings.
603 */
604static void wb_kupdate(unsigned long arg)
605{
606 unsigned long oldest_jif;
607 unsigned long start_jif;
608 unsigned long next_jif;
609 long nr_to_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 struct writeback_control wbc = {
611 .bdi = NULL,
612 .sync_mode = WB_SYNC_NONE,
613 .older_than_this = &oldest_jif,
614 .nr_to_write = 0,
615 .nonblocking = 1,
616 .for_kupdate = 1,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700617 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 };
619
620 sync_supers();
621
Bart Samwelf6ef9432006-03-24 03:15:48 -0800622 oldest_jif = jiffies - dirty_expire_interval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 start_jif = jiffies;
Bart Samwelf6ef9432006-03-24 03:15:48 -0800624 next_jif = start_jif + dirty_writeback_interval;
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700625 nr_to_write = global_page_state(NR_FILE_DIRTY) +
626 global_page_state(NR_UNSTABLE_NFS) +
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
628 while (nr_to_write > 0) {
Fengguang Wu2e6883b2007-10-16 23:30:43 -0700629 wbc.more_io = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 wbc.encountered_congestion = 0;
631 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
632 writeback_inodes(&wbc);
633 if (wbc.nr_to_write > 0) {
Fengguang Wu2e6883b2007-10-16 23:30:43 -0700634 if (wbc.encountered_congestion || wbc.more_io)
Andrew Morton3fcfab12006-10-19 23:28:16 -0700635 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 else
637 break; /* All the old data is written */
638 }
639 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
640 }
641 if (time_before(next_jif, jiffies + HZ))
642 next_jif = jiffies + HZ;
Bart Samwelf6ef9432006-03-24 03:15:48 -0800643 if (dirty_writeback_interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 mod_timer(&wb_timer, next_jif);
645}
646
647/*
648 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
649 */
650int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Andrew Morton3e733f02007-07-15 23:41:05 -0700651 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652{
Bart Samwelf6ef9432006-03-24 03:15:48 -0800653 proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
Andrew Morton3e733f02007-07-15 23:41:05 -0700654 if (dirty_writeback_interval)
655 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
656 else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 del_timer(&wb_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 return 0;
659}
660
661static void wb_timer_fn(unsigned long unused)
662{
663 if (pdflush_operation(wb_kupdate, 0) < 0)
664 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
665}
666
667static void laptop_flush(unsigned long unused)
668{
669 sys_sync();
670}
671
672static void laptop_timer_fn(unsigned long unused)
673{
674 pdflush_operation(laptop_flush, 0);
675}
676
677/*
678 * We've spun up the disk and we're in laptop mode: schedule writeback
679 * of all dirty data a few seconds from now. If the flush is already scheduled
680 * then push it back - the user is still using the disk.
681 */
682void laptop_io_completion(void)
683{
Bart Samweled5b43f2006-03-24 03:15:49 -0800684 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685}
686
687/*
688 * We're in laptop mode and we've just synced. The sync's writes will have
689 * caused another writeback to be scheduled by laptop_io_completion.
690 * Nothing needs to be written back anymore, so we unschedule the writeback.
691 */
692void laptop_sync_completion(void)
693{
694 del_timer(&laptop_mode_wb_timer);
695}
696
697/*
698 * If ratelimit_pages is too high then we can get into dirty-data overload
699 * if a large number of processes all perform writes at the same time.
700 * If it is too low then SMP machines will call the (expensive)
701 * get_writeback_state too often.
702 *
703 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
704 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
705 * thresholds before writeback cuts in.
706 *
707 * But the limit should not be set too high. Because it also controls the
708 * amount of memory which the balance_dirty_pages() caller has to write back.
709 * If this is too large then the caller will block on the IO queue all the
710 * time. So limit it to four megabytes - the balance_dirty_pages() caller
711 * will write six megabyte chunks, max.
712 */
713
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700714void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700716 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 if (ratelimit_pages < 16)
718 ratelimit_pages = 16;
719 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
720 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
721}
722
Chandra Seetharaman26c21432006-06-27 02:54:10 -0700723static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
725{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700726 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -0800727 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728}
729
Chandra Seetharaman74b85f32006-06-27 02:54:09 -0700730static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 .notifier_call = ratelimit_handler,
732 .next = NULL,
733};
734
735/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800736 * Called early on to tune the page writeback dirty limits.
737 *
738 * We used to scale dirty pages according to how total memory
739 * related to pages that could be allocated for buffers (by
740 * comparing nr_free_buffer_pages() to vm_total_pages.
741 *
742 * However, that was when we used "dirty_ratio" to scale with
743 * all memory, and we don't do that any more. "dirty_ratio"
744 * is now applied to total non-HIGHPAGE memory (by subtracting
745 * totalhigh_pages from vm_total_pages), and as such we can't
746 * get into the old insane situation any more where we had
747 * large amounts of dirty pages compared to a small amount of
748 * non-HIGHMEM memory.
749 *
750 * But we might still want to scale the dirty_ratio by how
751 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 */
753void __init page_writeback_init(void)
754{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700755 int shift;
756
Bart Samwelf6ef9432006-03-24 03:15:48 -0800757 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700758 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700760
761 shift = calc_period_shift();
762 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700763 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764}
765
David Howells811d7362006-08-29 19:06:09 +0100766/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700767 * 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 +0100768 * @mapping: address space structure to write
769 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700770 * @writepage: function called for each page
771 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +0100772 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700773 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +0100774 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
775 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
776 * and msync() need to guarantee that all the data which was dirty at the time
777 * the call was made get new I/O started against them. If wbc->sync_mode is
778 * WB_SYNC_ALL then we were called for data integrity and we must wait for
779 * existing IO to complete.
David Howells811d7362006-08-29 19:06:09 +0100780 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700781int write_cache_pages(struct address_space *mapping,
782 struct writeback_control *wbc, writepage_t writepage,
783 void *data)
David Howells811d7362006-08-29 19:06:09 +0100784{
785 struct backing_dev_info *bdi = mapping->backing_dev_info;
786 int ret = 0;
787 int done = 0;
David Howells811d7362006-08-29 19:06:09 +0100788 struct pagevec pvec;
789 int nr_pages;
790 pgoff_t index;
791 pgoff_t end; /* Inclusive */
792 int scanned = 0;
793 int range_whole = 0;
794
795 if (wbc->nonblocking && bdi_write_congested(bdi)) {
796 wbc->encountered_congestion = 1;
797 return 0;
798 }
799
David Howells811d7362006-08-29 19:06:09 +0100800 pagevec_init(&pvec, 0);
801 if (wbc->range_cyclic) {
802 index = mapping->writeback_index; /* Start from prev offset */
803 end = -1;
804 } else {
805 index = wbc->range_start >> PAGE_CACHE_SHIFT;
806 end = wbc->range_end >> PAGE_CACHE_SHIFT;
807 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
808 range_whole = 1;
809 scanned = 1;
810 }
811retry:
812 while (!done && (index <= end) &&
813 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
814 PAGECACHE_TAG_DIRTY,
815 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
816 unsigned i;
817
818 scanned = 1;
819 for (i = 0; i < nr_pages; i++) {
820 struct page *page = pvec.pages[i];
821
822 /*
823 * At this point we hold neither mapping->tree_lock nor
824 * lock on the page itself: the page may be truncated or
825 * invalidated (changing page->mapping to NULL), or even
826 * swizzled back from swapper_space to tmpfs file
827 * mapping
828 */
829 lock_page(page);
830
831 if (unlikely(page->mapping != mapping)) {
832 unlock_page(page);
833 continue;
834 }
835
836 if (!wbc->range_cyclic && page->index > end) {
837 done = 1;
838 unlock_page(page);
839 continue;
840 }
841
842 if (wbc->sync_mode != WB_SYNC_NONE)
843 wait_on_page_writeback(page);
844
845 if (PageWriteback(page) ||
846 !clear_page_dirty_for_io(page)) {
847 unlock_page(page);
848 continue;
849 }
850
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700851 ret = (*writepage)(page, wbc, data);
David Howells811d7362006-08-29 19:06:09 +0100852
Andrew Mortone4230032007-10-16 23:26:02 -0700853 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
David Howells811d7362006-08-29 19:06:09 +0100854 unlock_page(page);
Andrew Mortone4230032007-10-16 23:26:02 -0700855 ret = 0;
856 }
David Howells811d7362006-08-29 19:06:09 +0100857 if (ret || (--(wbc->nr_to_write) <= 0))
858 done = 1;
859 if (wbc->nonblocking && bdi_write_congested(bdi)) {
860 wbc->encountered_congestion = 1;
861 done = 1;
862 }
863 }
864 pagevec_release(&pvec);
865 cond_resched();
866 }
867 if (!scanned && !done) {
868 /*
869 * We hit the last page and there is more work to be done: wrap
870 * back to the start of the file
871 */
872 scanned = 1;
873 index = 0;
874 goto retry;
875 }
876 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
877 mapping->writeback_index = index;
878 return ret;
879}
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700880EXPORT_SYMBOL(write_cache_pages);
881
882/*
883 * Function used by generic_writepages to call the real writepage
884 * function and set the mapping flags on error
885 */
886static int __writepage(struct page *page, struct writeback_control *wbc,
887 void *data)
888{
889 struct address_space *mapping = data;
890 int ret = mapping->a_ops->writepage(page, wbc);
891 mapping_set_error(mapping, ret);
892 return ret;
893}
894
895/**
896 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
897 * @mapping: address space structure to write
898 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
899 *
900 * This is a library function, which implements the writepages()
901 * address_space_operation.
902 */
903int generic_writepages(struct address_space *mapping,
904 struct writeback_control *wbc)
905{
906 /* deal with chardevs and other special file */
907 if (!mapping->a_ops->writepage)
908 return 0;
909
910 return write_cache_pages(mapping, wbc, __writepage, mapping);
911}
David Howells811d7362006-08-29 19:06:09 +0100912
913EXPORT_SYMBOL(generic_writepages);
914
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
916{
Andrew Morton22905f72005-11-16 15:07:01 -0800917 int ret;
918
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 if (wbc->nr_to_write <= 0)
920 return 0;
Andrew Morton22905f72005-11-16 15:07:01 -0800921 wbc->for_writepages = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700923 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -0800924 else
925 ret = generic_writepages(mapping, wbc);
926 wbc->for_writepages = 0;
927 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928}
929
930/**
931 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -0700932 * @page: the page to write
933 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 *
935 * The page must be locked by the caller and will be unlocked upon return.
936 *
937 * write_one_page() returns a negative error code if I/O failed.
938 */
939int write_one_page(struct page *page, int wait)
940{
941 struct address_space *mapping = page->mapping;
942 int ret = 0;
943 struct writeback_control wbc = {
944 .sync_mode = WB_SYNC_ALL,
945 .nr_to_write = 1,
946 };
947
948 BUG_ON(!PageLocked(page));
949
950 if (wait)
951 wait_on_page_writeback(page);
952
953 if (clear_page_dirty_for_io(page)) {
954 page_cache_get(page);
955 ret = mapping->a_ops->writepage(page, &wbc);
956 if (ret == 0 && wait) {
957 wait_on_page_writeback(page);
958 if (PageError(page))
959 ret = -EIO;
960 }
961 page_cache_release(page);
962 } else {
963 unlock_page(page);
964 }
965 return ret;
966}
967EXPORT_SYMBOL(write_one_page);
968
969/*
Ken Chen76719322007-02-10 01:43:15 -0800970 * For address_spaces which do not use buffers nor write back.
971 */
972int __set_page_dirty_no_writeback(struct page *page)
973{
974 if (!PageDirty(page))
975 SetPageDirty(page);
976 return 0;
977}
978
979/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 * For address_spaces which do not use buffers. Just tag the page as dirty in
981 * its radix tree.
982 *
983 * This is also used when a single buffer is being dirtied: we want to set the
984 * page dirty in that case, but not all the buffers. This is a "bottom-up"
985 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
986 *
987 * Most callers have locked the page, which pins the address_space in memory.
988 * But zap_pte_range() does not lock the page, however in that case the
989 * mapping is pinned by the vma's ->vm_file reference.
990 *
991 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +0200992 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 */
994int __set_page_dirty_nobuffers(struct page *page)
995{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 if (!TestSetPageDirty(page)) {
997 struct address_space *mapping = page_mapping(page);
998 struct address_space *mapping2;
999
Andrew Morton8c085402006-12-10 02:19:24 -08001000 if (!mapping)
1001 return 1;
1002
1003 write_lock_irq(&mapping->tree_lock);
1004 mapping2 = page_mapping(page);
1005 if (mapping2) { /* Race with truncate? */
1006 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001007 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Andrew Morton55e829a2006-12-10 02:19:27 -08001008 if (mapping_cap_account_dirty(mapping)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001009 __inc_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001010 __inc_bdi_stat(mapping->backing_dev_info,
1011 BDI_RECLAIMABLE);
Andrew Morton55e829a2006-12-10 02:19:27 -08001012 task_io_account_write(PAGE_CACHE_SIZE);
1013 }
Andrew Morton8c085402006-12-10 02:19:24 -08001014 radix_tree_tag_set(&mapping->page_tree,
1015 page_index(page), PAGECACHE_TAG_DIRTY);
1016 }
1017 write_unlock_irq(&mapping->tree_lock);
1018 if (mapping->host) {
1019 /* !PageAnon && !swapper_space */
1020 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001022 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001024 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025}
1026EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1027
1028/*
1029 * When a writepage implementation decides that it doesn't want to write this
1030 * page for some reason, it should redirty the locked page via
1031 * redirty_page_for_writepage() and it should then unlock the page and return 0
1032 */
1033int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1034{
1035 wbc->pages_skipped++;
1036 return __set_page_dirty_nobuffers(page);
1037}
1038EXPORT_SYMBOL(redirty_page_for_writepage);
1039
1040/*
1041 * If the mapping doesn't provide a set_page_dirty a_op, then
1042 * just fall through and assume that it wants buffer_heads.
1043 */
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001044static int __set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045{
1046 struct address_space *mapping = page_mapping(page);
1047
1048 if (likely(mapping)) {
1049 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
David Howells93614012006-09-30 20:45:40 +02001050#ifdef CONFIG_BLOCK
1051 if (!spd)
1052 spd = __set_page_dirty_buffers;
1053#endif
1054 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001056 if (!PageDirty(page)) {
1057 if (!TestSetPageDirty(page))
1058 return 1;
1059 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 return 0;
1061}
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001062
1063int fastcall set_page_dirty(struct page *page)
1064{
1065 int ret = __set_page_dirty(page);
1066 if (ret)
1067 task_dirty_inc(current);
1068 return ret;
1069}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070EXPORT_SYMBOL(set_page_dirty);
1071
1072/*
1073 * set_page_dirty() is racy if the caller has no reference against
1074 * page->mapping->host, and if the page is unlocked. This is because another
1075 * CPU could truncate the page off the mapping and then free the mapping.
1076 *
1077 * Usually, the page _is_ locked, or the caller is a user-space process which
1078 * holds a reference on the inode by having an open file.
1079 *
1080 * In other cases, the page should be locked before running set_page_dirty().
1081 */
1082int set_page_dirty_lock(struct page *page)
1083{
1084 int ret;
1085
Nick Piggindb376482006-09-25 23:31:24 -07001086 lock_page_nosync(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 ret = set_page_dirty(page);
1088 unlock_page(page);
1089 return ret;
1090}
1091EXPORT_SYMBOL(set_page_dirty_lock);
1092
1093/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 * Clear a page's dirty flag, while caring for dirty memory accounting.
1095 * Returns true if the page was previously dirty.
1096 *
1097 * This is for preparing to put the page under writeout. We leave the page
1098 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1099 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1100 * implementation will run either set_page_writeback() or set_page_dirty(),
1101 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1102 * back into sync.
1103 *
1104 * This incoherency between the page's dirty flag and radix-tree tag is
1105 * unfortunate, but it only exists while the page is locked.
1106 */
1107int clear_page_dirty_for_io(struct page *page)
1108{
1109 struct address_space *mapping = page_mapping(page);
1110
Nick Piggin79352892007-07-19 01:47:22 -07001111 BUG_ON(!PageLocked(page));
1112
Fengguang Wufe3cba12007-07-19 01:48:07 -07001113 ClearPageReclaim(page);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001114 if (mapping && mapping_cap_account_dirty(mapping)) {
1115 /*
1116 * Yes, Virginia, this is indeed insane.
1117 *
1118 * We use this sequence to make sure that
1119 * (a) we account for dirty stats properly
1120 * (b) we tell the low-level filesystem to
1121 * mark the whole page dirty if it was
1122 * dirty in a pagetable. Only to then
1123 * (c) clean the page again and return 1 to
1124 * cause the writeback.
1125 *
1126 * This way we avoid all nasty races with the
1127 * dirty bit in multiple places and clearing
1128 * them concurrently from different threads.
1129 *
1130 * Note! Normally the "set_page_dirty(page)"
1131 * has no effect on the actual dirty bit - since
1132 * that will already usually be set. But we
1133 * need the side effects, and it can help us
1134 * avoid races.
1135 *
1136 * We basically use the page "master dirty bit"
1137 * as a serialization point for all the different
1138 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001139 */
1140 if (page_mkclean(page))
1141 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001142 /*
1143 * We carefully synchronise fault handlers against
1144 * installing a dirty pte and marking the page dirty
1145 * at this point. We do this by having them hold the
1146 * page lock at some point after installing their
1147 * pte, but before marking the page dirty.
1148 * Pages are always locked coming in here, so we get
1149 * the desired exclusion. See mm/memory.c:do_wp_page()
1150 * for more comments.
1151 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001152 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001153 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001154 dec_bdi_stat(mapping->backing_dev_info,
1155 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001156 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001158 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001160 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001162EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
1164int test_clear_page_writeback(struct page *page)
1165{
1166 struct address_space *mapping = page_mapping(page);
1167 int ret;
1168
1169 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001170 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 unsigned long flags;
1172
1173 write_lock_irqsave(&mapping->tree_lock, flags);
1174 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001175 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 radix_tree_tag_clear(&mapping->page_tree,
1177 page_index(page),
1178 PAGECACHE_TAG_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001179 if (bdi_cap_writeback_dirty(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001180 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001181 __bdi_writeout_inc(bdi);
1182 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 write_unlock_irqrestore(&mapping->tree_lock, flags);
1185 } else {
1186 ret = TestClearPageWriteback(page);
1187 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001188 if (ret)
1189 dec_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 return ret;
1191}
1192
1193int test_set_page_writeback(struct page *page)
1194{
1195 struct address_space *mapping = page_mapping(page);
1196 int ret;
1197
1198 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001199 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 unsigned long flags;
1201
1202 write_lock_irqsave(&mapping->tree_lock, flags);
1203 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001204 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 radix_tree_tag_set(&mapping->page_tree,
1206 page_index(page),
1207 PAGECACHE_TAG_WRITEBACK);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001208 if (bdi_cap_writeback_dirty(bdi))
1209 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 if (!PageDirty(page))
1212 radix_tree_tag_clear(&mapping->page_tree,
1213 page_index(page),
1214 PAGECACHE_TAG_DIRTY);
1215 write_unlock_irqrestore(&mapping->tree_lock, flags);
1216 } else {
1217 ret = TestSetPageWriteback(page);
1218 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001219 if (!ret)
1220 inc_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 return ret;
1222
1223}
1224EXPORT_SYMBOL(test_set_page_writeback);
1225
1226/*
Nick Piggin00128182007-10-16 01:24:40 -07001227 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 * passed tag.
1229 */
1230int mapping_tagged(struct address_space *mapping, int tag)
1231{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 int ret;
Nick Piggin00128182007-10-16 01:24:40 -07001233 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 ret = radix_tree_tagged(&mapping->page_tree, tag);
Nick Piggin00128182007-10-16 01:24:40 -07001235 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 return ret;
1237}
1238EXPORT_SYMBOL(mapping_tagged);