mm: prevent balance_dirty_pages() from doing too much work
[linux-2.6.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002    Andrew Morton
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>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #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>
35 #include <linux/buffer_head.h>
36 #include <linux/pagevec.h>
37
38 /*
39  * The maximum number of pages to writeout in a single bdflush/kupdate
40  * operation.  We do this so we don't hold I_SYNC against an inode for
41  * 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  */
51 static long ratelimit_pages = 32;
52
53 /*
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  */
59 static 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  */
69 int dirty_background_ratio = 10;
70
71 /*
72  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
73  * dirty_background_ratio * the amount of dirtyable memory
74  */
75 unsigned long dirty_background_bytes;
76
77 /*
78  * free highmem will not be subtracted from the total free memory
79  * for calculating free ratios if vm_highmem_is_dirtyable is true
80  */
81 int vm_highmem_is_dirtyable;
82
83 /*
84  * The generator of dirty data starts writeback at this percentage
85  */
86 int vm_dirty_ratio = 20;
87
88 /*
89  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
90  * vm_dirty_ratio * the amount of dirtyable memory
91  */
92 unsigned long vm_dirty_bytes;
93
94 /*
95  * The interval between `kupdate'-style writebacks
96  */
97 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
98
99 /*
100  * The longest time for which data is allowed to remain dirty
101  */
102 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
103
104 /*
105  * Flag that makes the machine dump writes/reads and block dirtyings.
106  */
107 int block_dump;
108
109 /*
110  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
111  * a full sync is triggered after this time elapses without any disk activity.
112  */
113 int laptop_mode;
114
115 EXPORT_SYMBOL(laptop_mode);
116
117 /* End of sysctl-exported parameters */
118
119
120 static void background_writeout(unsigned long _min_pages);
121
122 /*
123  * Scale the writeback cache size proportional to the relative writeout speeds.
124  *
125  * We do this by keeping a floating proportion between BDIs, based on page
126  * writeback completions [end_page_writeback()]. Those devices that write out
127  * pages fastest will get the larger share, while the slower will get a smaller
128  * share.
129  *
130  * We use page writeout completions because we are interested in getting rid of
131  * dirty pages. Having them written out is the primary goal.
132  *
133  * We introduce a concept of time, a period over which we measure these events,
134  * because demand can/will vary over time. The length of this period itself is
135  * measured in page writeback completions.
136  *
137  */
138 static struct prop_descriptor vm_completions;
139 static struct prop_descriptor vm_dirties;
140
141 /*
142  * couple the period to the dirty_ratio:
143  *
144  *   period/2 ~ roundup_pow_of_two(dirty limit)
145  */
146 static int calc_period_shift(void)
147 {
148         unsigned long dirty_total;
149
150         if (vm_dirty_bytes)
151                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
152         else
153                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
154                                 100;
155         return 2 + ilog2(dirty_total - 1);
156 }
157
158 /*
159  * update the period when the dirty threshold changes.
160  */
161 static void update_completion_period(void)
162 {
163         int shift = calc_period_shift();
164         prop_change_shift(&vm_completions, shift);
165         prop_change_shift(&vm_dirties, shift);
166 }
167
168 int dirty_background_ratio_handler(struct ctl_table *table, int write,
169                 struct file *filp, void __user *buffer, size_t *lenp,
170                 loff_t *ppos)
171 {
172         int ret;
173
174         ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
175         if (ret == 0 && write)
176                 dirty_background_bytes = 0;
177         return ret;
178 }
179
180 int dirty_background_bytes_handler(struct ctl_table *table, int write,
181                 struct file *filp, void __user *buffer, size_t *lenp,
182                 loff_t *ppos)
183 {
184         int ret;
185
186         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
187         if (ret == 0 && write)
188                 dirty_background_ratio = 0;
189         return ret;
190 }
191
192 int dirty_ratio_handler(struct ctl_table *table, int write,
193                 struct file *filp, void __user *buffer, size_t *lenp,
194                 loff_t *ppos)
195 {
196         int old_ratio = vm_dirty_ratio;
197         int ret;
198
199         ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
200         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
201                 update_completion_period();
202                 vm_dirty_bytes = 0;
203         }
204         return ret;
205 }
206
207
208 int dirty_bytes_handler(struct ctl_table *table, int write,
209                 struct file *filp, void __user *buffer, size_t *lenp,
210                 loff_t *ppos)
211 {
212         unsigned long old_bytes = vm_dirty_bytes;
213         int ret;
214
215         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
216         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
217                 update_completion_period();
218                 vm_dirty_ratio = 0;
219         }
220         return ret;
221 }
222
223 /*
224  * Increment the BDI's writeout completion count and the global writeout
225  * completion count. Called from test_clear_page_writeback().
226  */
227 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
228 {
229         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
230                               bdi->max_prop_frac);
231 }
232
233 void bdi_writeout_inc(struct backing_dev_info *bdi)
234 {
235         unsigned long flags;
236
237         local_irq_save(flags);
238         __bdi_writeout_inc(bdi);
239         local_irq_restore(flags);
240 }
241 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
242
243 void task_dirty_inc(struct task_struct *tsk)
244 {
245         prop_inc_single(&vm_dirties, &tsk->dirties);
246 }
247
248 /*
249  * Obtain an accurate fraction of the BDI's portion.
250  */
251 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
252                 long *numerator, long *denominator)
253 {
254         if (bdi_cap_writeback_dirty(bdi)) {
255                 prop_fraction_percpu(&vm_completions, &bdi->completions,
256                                 numerator, denominator);
257         } else {
258                 *numerator = 0;
259                 *denominator = 1;
260         }
261 }
262
263 /*
264  * Clip the earned share of dirty pages to that which is actually available.
265  * This avoids exceeding the total dirty_limit when the floating averages
266  * fluctuate too quickly.
267  */
268 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
269                 unsigned long dirty, unsigned long *pbdi_dirty)
270 {
271         unsigned long avail_dirty;
272
273         avail_dirty = global_page_state(NR_FILE_DIRTY) +
274                  global_page_state(NR_WRITEBACK) +
275                  global_page_state(NR_UNSTABLE_NFS) +
276                  global_page_state(NR_WRITEBACK_TEMP);
277
278         if (avail_dirty < dirty)
279                 avail_dirty = dirty - avail_dirty;
280         else
281                 avail_dirty = 0;
282
283         avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
284                 bdi_stat(bdi, BDI_WRITEBACK);
285
286         *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
287 }
288
289 static inline void task_dirties_fraction(struct task_struct *tsk,
290                 long *numerator, long *denominator)
291 {
292         prop_fraction_single(&vm_dirties, &tsk->dirties,
293                                 numerator, denominator);
294 }
295
296 /*
297  * scale the dirty limit
298  *
299  * task specific dirty limit:
300  *
301  *   dirty -= (dirty/8) * p_{t}
302  */
303 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
304 {
305         long numerator, denominator;
306         unsigned long dirty = *pdirty;
307         u64 inv = dirty >> 3;
308
309         task_dirties_fraction(tsk, &numerator, &denominator);
310         inv *= numerator;
311         do_div(inv, denominator);
312
313         dirty -= inv;
314         if (dirty < *pdirty/2)
315                 dirty = *pdirty/2;
316
317         *pdirty = dirty;
318 }
319
320 /*
321  *
322  */
323 static DEFINE_SPINLOCK(bdi_lock);
324 static unsigned int bdi_min_ratio;
325
326 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
327 {
328         int ret = 0;
329         unsigned long flags;
330
331         spin_lock_irqsave(&bdi_lock, flags);
332         if (min_ratio > bdi->max_ratio) {
333                 ret = -EINVAL;
334         } else {
335                 min_ratio -= bdi->min_ratio;
336                 if (bdi_min_ratio + min_ratio < 100) {
337                         bdi_min_ratio += min_ratio;
338                         bdi->min_ratio += min_ratio;
339                 } else {
340                         ret = -EINVAL;
341                 }
342         }
343         spin_unlock_irqrestore(&bdi_lock, flags);
344
345         return ret;
346 }
347
348 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
349 {
350         unsigned long flags;
351         int ret = 0;
352
353         if (max_ratio > 100)
354                 return -EINVAL;
355
356         spin_lock_irqsave(&bdi_lock, flags);
357         if (bdi->min_ratio > max_ratio) {
358                 ret = -EINVAL;
359         } else {
360                 bdi->max_ratio = max_ratio;
361                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
362         }
363         spin_unlock_irqrestore(&bdi_lock, flags);
364
365         return ret;
366 }
367 EXPORT_SYMBOL(bdi_set_max_ratio);
368
369 /*
370  * Work out the current dirty-memory clamping and background writeout
371  * thresholds.
372  *
373  * The main aim here is to lower them aggressively if there is a lot of mapped
374  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
375  * pages.  It is better to clamp down on writers than to start swapping, and
376  * performing lots of scanning.
377  *
378  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
379  *
380  * We don't permit the clamping level to fall below 5% - that is getting rather
381  * excessive.
382  *
383  * We make sure that the background writeout level is below the adjusted
384  * clamping level.
385  */
386
387 static unsigned long highmem_dirtyable_memory(unsigned long total)
388 {
389 #ifdef CONFIG_HIGHMEM
390         int node;
391         unsigned long x = 0;
392
393         for_each_node_state(node, N_HIGH_MEMORY) {
394                 struct zone *z =
395                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
396
397                 x += zone_page_state(z, NR_FREE_PAGES) + zone_lru_pages(z);
398         }
399         /*
400          * Make sure that the number of highmem pages is never larger
401          * than the number of the total dirtyable memory. This can only
402          * occur in very strange VM situations but we want to make sure
403          * that this does not occur.
404          */
405         return min(x, total);
406 #else
407         return 0;
408 #endif
409 }
410
411 /**
412  * determine_dirtyable_memory - amount of memory that may be used
413  *
414  * Returns the numebr of pages that can currently be freed and used
415  * by the kernel for direct mappings.
416  */
417 unsigned long determine_dirtyable_memory(void)
418 {
419         unsigned long x;
420
421         x = global_page_state(NR_FREE_PAGES) + global_lru_pages();
422
423         if (!vm_highmem_is_dirtyable)
424                 x -= highmem_dirtyable_memory(x);
425
426         return x + 1;   /* Ensure that we never return 0 */
427 }
428
429 void
430 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
431                  unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
432 {
433         unsigned long background;
434         unsigned long dirty;
435         unsigned long available_memory = determine_dirtyable_memory();
436         struct task_struct *tsk;
437
438         if (vm_dirty_bytes)
439                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
440         else {
441                 int dirty_ratio;
442
443                 dirty_ratio = vm_dirty_ratio;
444                 if (dirty_ratio < 5)
445                         dirty_ratio = 5;
446                 dirty = (dirty_ratio * available_memory) / 100;
447         }
448
449         if (dirty_background_bytes)
450                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
451         else
452                 background = (dirty_background_ratio * available_memory) / 100;
453
454         if (background >= dirty)
455                 background = dirty / 2;
456         tsk = current;
457         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
458                 background += background / 4;
459                 dirty += dirty / 4;
460         }
461         *pbackground = background;
462         *pdirty = dirty;
463
464         if (bdi) {
465                 u64 bdi_dirty;
466                 long numerator, denominator;
467
468                 /*
469                  * Calculate this BDI's share of the dirty ratio.
470                  */
471                 bdi_writeout_fraction(bdi, &numerator, &denominator);
472
473                 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
474                 bdi_dirty *= numerator;
475                 do_div(bdi_dirty, denominator);
476                 bdi_dirty += (dirty * bdi->min_ratio) / 100;
477                 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
478                         bdi_dirty = dirty * bdi->max_ratio / 100;
479
480                 *pbdi_dirty = bdi_dirty;
481                 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
482                 task_dirty_limit(current, pbdi_dirty);
483         }
484 }
485
486 /*
487  * balance_dirty_pages() must be called by processes which are generating dirty
488  * data.  It looks at the number of dirty pages in the machine and will force
489  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
490  * If we're over `background_thresh' then pdflush is woken to perform some
491  * writeout.
492  */
493 static void balance_dirty_pages(struct address_space *mapping)
494 {
495         long nr_reclaimable, bdi_nr_reclaimable;
496         long nr_writeback, bdi_nr_writeback;
497         unsigned long background_thresh;
498         unsigned long dirty_thresh;
499         unsigned long bdi_thresh;
500         unsigned long pages_written = 0;
501         unsigned long write_chunk = sync_writeback_pages();
502
503         struct backing_dev_info *bdi = mapping->backing_dev_info;
504
505         for (;;) {
506                 struct writeback_control wbc = {
507                         .bdi            = bdi,
508                         .sync_mode      = WB_SYNC_NONE,
509                         .older_than_this = NULL,
510                         .nr_to_write    = write_chunk,
511                         .range_cyclic   = 1,
512                 };
513
514                 get_dirty_limits(&background_thresh, &dirty_thresh,
515                                 &bdi_thresh, bdi);
516
517                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
518                                         global_page_state(NR_UNSTABLE_NFS);
519                 nr_writeback = global_page_state(NR_WRITEBACK);
520
521                 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
522                 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
523
524                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
525                         break;
526
527                 /*
528                  * Throttle it only when the background writeback cannot
529                  * catch-up. This avoids (excessively) small writeouts
530                  * when the bdi limits are ramping up.
531                  */
532                 if (nr_reclaimable + nr_writeback <
533                                 (background_thresh + dirty_thresh) / 2)
534                         break;
535
536                 if (!bdi->dirty_exceeded)
537                         bdi->dirty_exceeded = 1;
538
539                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
540                  * Unstable writes are a feature of certain networked
541                  * filesystems (i.e. NFS) in which data may have been
542                  * written to the server's write cache, but has not yet
543                  * been flushed to permanent storage.
544                  * Only move pages to writeback if this bdi is over its
545                  * threshold otherwise wait until the disk writes catch
546                  * up.
547                  */
548                 if (bdi_nr_reclaimable > bdi_thresh) {
549                         writeback_inodes(&wbc);
550                         pages_written += write_chunk - wbc.nr_to_write;
551                         get_dirty_limits(&background_thresh, &dirty_thresh,
552                                        &bdi_thresh, bdi);
553                 }
554
555                 /*
556                  * In order to avoid the stacked BDI deadlock we need
557                  * to ensure we accurately count the 'dirty' pages when
558                  * the threshold is low.
559                  *
560                  * Otherwise it would be possible to get thresh+n pages
561                  * reported dirty, even though there are thresh-m pages
562                  * actually dirty; with m+n sitting in the percpu
563                  * deltas.
564                  */
565                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
566                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
567                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
568                 } else if (bdi_nr_reclaimable) {
569                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
570                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
571                 }
572
573                 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
574                         break;
575                 if (pages_written >= write_chunk)
576                         break;          /* We've done our duty */
577
578                 congestion_wait(WRITE, HZ/10);
579         }
580
581         if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
582                         bdi->dirty_exceeded)
583                 bdi->dirty_exceeded = 0;
584
585         if (writeback_in_progress(bdi))
586                 return;         /* pdflush is already working this queue */
587
588         /*
589          * In laptop mode, we wait until hitting the higher threshold before
590          * starting background writeout, and then write out all the way down
591          * to the lower threshold.  So slow writers cause minimal disk activity.
592          *
593          * In normal mode, we start background writeout at the lower
594          * background_thresh, to keep the amount of dirty memory low.
595          */
596         if ((laptop_mode && pages_written) ||
597                         (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
598                                           + global_page_state(NR_UNSTABLE_NFS)
599                                           > background_thresh)))
600                 pdflush_operation(background_writeout, 0);
601 }
602
603 void set_page_dirty_balance(struct page *page, int page_mkwrite)
604 {
605         if (set_page_dirty(page) || page_mkwrite) {
606                 struct address_space *mapping = page_mapping(page);
607
608                 if (mapping)
609                         balance_dirty_pages_ratelimited(mapping);
610         }
611 }
612
613 /**
614  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
615  * @mapping: address_space which was dirtied
616  * @nr_pages_dirtied: number of pages which the caller has just dirtied
617  *
618  * Processes which are dirtying memory should call in here once for each page
619  * which was newly dirtied.  The function will periodically check the system's
620  * dirty state and will initiate writeback if needed.
621  *
622  * On really big machines, get_writeback_state is expensive, so try to avoid
623  * calling it too often (ratelimiting).  But once we're over the dirty memory
624  * limit we decrease the ratelimiting by a lot, to prevent individual processes
625  * from overshooting the limit by (ratelimit_pages) each.
626  */
627 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
628                                         unsigned long nr_pages_dirtied)
629 {
630         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
631         unsigned long ratelimit;
632         unsigned long *p;
633
634         ratelimit = ratelimit_pages;
635         if (mapping->backing_dev_info->dirty_exceeded)
636                 ratelimit = 8;
637
638         /*
639          * Check the rate limiting. Also, we do not want to throttle real-time
640          * tasks in balance_dirty_pages(). Period.
641          */
642         preempt_disable();
643         p =  &__get_cpu_var(ratelimits);
644         *p += nr_pages_dirtied;
645         if (unlikely(*p >= ratelimit)) {
646                 *p = 0;
647                 preempt_enable();
648                 balance_dirty_pages(mapping);
649                 return;
650         }
651         preempt_enable();
652 }
653 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
654
655 void throttle_vm_writeout(gfp_t gfp_mask)
656 {
657         unsigned long background_thresh;
658         unsigned long dirty_thresh;
659
660         for ( ; ; ) {
661                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
662
663                 /*
664                  * Boost the allowable dirty threshold a bit for page
665                  * allocators so they don't get DoS'ed by heavy writers
666                  */
667                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
668
669                 if (global_page_state(NR_UNSTABLE_NFS) +
670                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
671                                 break;
672                 congestion_wait(WRITE, HZ/10);
673
674                 /*
675                  * The caller might hold locks which can prevent IO completion
676                  * or progress in the filesystem.  So we cannot just sit here
677                  * waiting for IO to complete.
678                  */
679                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
680                         break;
681         }
682 }
683
684 /*
685  * writeback at least _min_pages, and keep writing until the amount of dirty
686  * memory is less than the background threshold, or until we're all clean.
687  */
688 static void background_writeout(unsigned long _min_pages)
689 {
690         long min_pages = _min_pages;
691         struct writeback_control wbc = {
692                 .bdi            = NULL,
693                 .sync_mode      = WB_SYNC_NONE,
694                 .older_than_this = NULL,
695                 .nr_to_write    = 0,
696                 .nonblocking    = 1,
697                 .range_cyclic   = 1,
698         };
699
700         for ( ; ; ) {
701                 unsigned long background_thresh;
702                 unsigned long dirty_thresh;
703
704                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
705                 if (global_page_state(NR_FILE_DIRTY) +
706                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
707                                 && min_pages <= 0)
708                         break;
709                 wbc.more_io = 0;
710                 wbc.encountered_congestion = 0;
711                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
712                 wbc.pages_skipped = 0;
713                 writeback_inodes(&wbc);
714                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
715                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
716                         /* Wrote less than expected */
717                         if (wbc.encountered_congestion || wbc.more_io)
718                                 congestion_wait(WRITE, HZ/10);
719                         else
720                                 break;
721                 }
722         }
723 }
724
725 /*
726  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
727  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
728  * -1 if all pdflush threads were busy.
729  */
730 int wakeup_pdflush(long nr_pages)
731 {
732         if (nr_pages == 0)
733                 nr_pages = global_page_state(NR_FILE_DIRTY) +
734                                 global_page_state(NR_UNSTABLE_NFS);
735         return pdflush_operation(background_writeout, nr_pages);
736 }
737
738 static void wb_timer_fn(unsigned long unused);
739 static void laptop_timer_fn(unsigned long unused);
740
741 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
742 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
743
744 /*
745  * Periodic writeback of "old" data.
746  *
747  * Define "old": the first time one of an inode's pages is dirtied, we mark the
748  * dirtying-time in the inode's address_space.  So this periodic writeback code
749  * just walks the superblock inode list, writing back any inodes which are
750  * older than a specific point in time.
751  *
752  * Try to run once per dirty_writeback_interval.  But if a writeback event
753  * takes longer than a dirty_writeback_interval interval, then leave a
754  * one-second gap.
755  *
756  * older_than_this takes precedence over nr_to_write.  So we'll only write back
757  * all dirty pages if they are all attached to "old" mappings.
758  */
759 static void wb_kupdate(unsigned long arg)
760 {
761         unsigned long oldest_jif;
762         unsigned long start_jif;
763         unsigned long next_jif;
764         long nr_to_write;
765         struct writeback_control wbc = {
766                 .bdi            = NULL,
767                 .sync_mode      = WB_SYNC_NONE,
768                 .older_than_this = &oldest_jif,
769                 .nr_to_write    = 0,
770                 .nonblocking    = 1,
771                 .for_kupdate    = 1,
772                 .range_cyclic   = 1,
773         };
774
775         sync_supers();
776
777         oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval * 10);
778         start_jif = jiffies;
779         next_jif = start_jif + msecs_to_jiffies(dirty_writeback_interval * 10);
780         nr_to_write = global_page_state(NR_FILE_DIRTY) +
781                         global_page_state(NR_UNSTABLE_NFS) +
782                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
783         while (nr_to_write > 0) {
784                 wbc.more_io = 0;
785                 wbc.encountered_congestion = 0;
786                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
787                 writeback_inodes(&wbc);
788                 if (wbc.nr_to_write > 0) {
789                         if (wbc.encountered_congestion || wbc.more_io)
790                                 congestion_wait(WRITE, HZ/10);
791                         else
792                                 break;  /* All the old data is written */
793                 }
794                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
795         }
796         if (time_before(next_jif, jiffies + HZ))
797                 next_jif = jiffies + HZ;
798         if (dirty_writeback_interval)
799                 mod_timer(&wb_timer, next_jif);
800 }
801
802 /*
803  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
804  */
805 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
806         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
807 {
808         proc_dointvec(table, write, file, buffer, length, ppos);
809         if (dirty_writeback_interval)
810                 mod_timer(&wb_timer, jiffies +
811                         msecs_to_jiffies(dirty_writeback_interval * 10));
812         else
813                 del_timer(&wb_timer);
814         return 0;
815 }
816
817 static void wb_timer_fn(unsigned long unused)
818 {
819         if (pdflush_operation(wb_kupdate, 0) < 0)
820                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
821 }
822
823 static void laptop_flush(unsigned long unused)
824 {
825         sys_sync();
826 }
827
828 static void laptop_timer_fn(unsigned long unused)
829 {
830         pdflush_operation(laptop_flush, 0);
831 }
832
833 /*
834  * We've spun up the disk and we're in laptop mode: schedule writeback
835  * of all dirty data a few seconds from now.  If the flush is already scheduled
836  * then push it back - the user is still using the disk.
837  */
838 void laptop_io_completion(void)
839 {
840         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
841 }
842
843 /*
844  * We're in laptop mode and we've just synced. The sync's writes will have
845  * caused another writeback to be scheduled by laptop_io_completion.
846  * Nothing needs to be written back anymore, so we unschedule the writeback.
847  */
848 void laptop_sync_completion(void)
849 {
850         del_timer(&laptop_mode_wb_timer);
851 }
852
853 /*
854  * If ratelimit_pages is too high then we can get into dirty-data overload
855  * if a large number of processes all perform writes at the same time.
856  * If it is too low then SMP machines will call the (expensive)
857  * get_writeback_state too often.
858  *
859  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
860  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
861  * thresholds before writeback cuts in.
862  *
863  * But the limit should not be set too high.  Because it also controls the
864  * amount of memory which the balance_dirty_pages() caller has to write back.
865  * If this is too large then the caller will block on the IO queue all the
866  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
867  * will write six megabyte chunks, max.
868  */
869
870 void writeback_set_ratelimit(void)
871 {
872         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
873         if (ratelimit_pages < 16)
874                 ratelimit_pages = 16;
875         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
876                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
877 }
878
879 static int __cpuinit
880 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
881 {
882         writeback_set_ratelimit();
883         return NOTIFY_DONE;
884 }
885
886 static struct notifier_block __cpuinitdata ratelimit_nb = {
887         .notifier_call  = ratelimit_handler,
888         .next           = NULL,
889 };
890
891 /*
892  * Called early on to tune the page writeback dirty limits.
893  *
894  * We used to scale dirty pages according to how total memory
895  * related to pages that could be allocated for buffers (by
896  * comparing nr_free_buffer_pages() to vm_total_pages.
897  *
898  * However, that was when we used "dirty_ratio" to scale with
899  * all memory, and we don't do that any more. "dirty_ratio"
900  * is now applied to total non-HIGHPAGE memory (by subtracting
901  * totalhigh_pages from vm_total_pages), and as such we can't
902  * get into the old insane situation any more where we had
903  * large amounts of dirty pages compared to a small amount of
904  * non-HIGHMEM memory.
905  *
906  * But we might still want to scale the dirty_ratio by how
907  * much memory the box has..
908  */
909 void __init page_writeback_init(void)
910 {
911         int shift;
912
913         mod_timer(&wb_timer,
914                   jiffies + msecs_to_jiffies(dirty_writeback_interval * 10));
915         writeback_set_ratelimit();
916         register_cpu_notifier(&ratelimit_nb);
917
918         shift = calc_period_shift();
919         prop_descriptor_init(&vm_completions, shift);
920         prop_descriptor_init(&vm_dirties, shift);
921 }
922
923 /**
924  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
925  * @mapping: address space structure to write
926  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
927  * @writepage: function called for each page
928  * @data: data passed to writepage function
929  *
930  * If a page is already under I/O, write_cache_pages() skips it, even
931  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
932  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
933  * and msync() need to guarantee that all the data which was dirty at the time
934  * the call was made get new I/O started against them.  If wbc->sync_mode is
935  * WB_SYNC_ALL then we were called for data integrity and we must wait for
936  * existing IO to complete.
937  */
938 int write_cache_pages(struct address_space *mapping,
939                       struct writeback_control *wbc, writepage_t writepage,
940                       void *data)
941 {
942         struct backing_dev_info *bdi = mapping->backing_dev_info;
943         int ret = 0;
944         int done = 0;
945         struct pagevec pvec;
946         int nr_pages;
947         pgoff_t uninitialized_var(writeback_index);
948         pgoff_t index;
949         pgoff_t end;            /* Inclusive */
950         pgoff_t done_index;
951         int cycled;
952         int range_whole = 0;
953         long nr_to_write = wbc->nr_to_write;
954
955         if (wbc->nonblocking && bdi_write_congested(bdi)) {
956                 wbc->encountered_congestion = 1;
957                 return 0;
958         }
959
960         pagevec_init(&pvec, 0);
961         if (wbc->range_cyclic) {
962                 writeback_index = mapping->writeback_index; /* prev offset */
963                 index = writeback_index;
964                 if (index == 0)
965                         cycled = 1;
966                 else
967                         cycled = 0;
968                 end = -1;
969         } else {
970                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
971                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
972                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
973                         range_whole = 1;
974                 cycled = 1; /* ignore range_cyclic tests */
975         }
976 retry:
977         done_index = index;
978         while (!done && (index <= end)) {
979                 int i;
980
981                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
982                               PAGECACHE_TAG_DIRTY,
983                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
984                 if (nr_pages == 0)
985                         break;
986
987                 for (i = 0; i < nr_pages; i++) {
988                         struct page *page = pvec.pages[i];
989
990                         /*
991                          * At this point, the page may be truncated or
992                          * invalidated (changing page->mapping to NULL), or
993                          * even swizzled back from swapper_space to tmpfs file
994                          * mapping. However, page->index will not change
995                          * because we have a reference on the page.
996                          */
997                         if (page->index > end) {
998                                 /*
999                                  * can't be range_cyclic (1st pass) because
1000                                  * end == -1 in that case.
1001                                  */
1002                                 done = 1;
1003                                 break;
1004                         }
1005
1006                         done_index = page->index + 1;
1007
1008                         lock_page(page);
1009
1010                         /*
1011                          * Page truncated or invalidated. We can freely skip it
1012                          * then, even for data integrity operations: the page
1013                          * has disappeared concurrently, so there could be no
1014                          * real expectation of this data interity operation
1015                          * even if there is now a new, dirty page at the same
1016                          * pagecache address.
1017                          */
1018                         if (unlikely(page->mapping != mapping)) {
1019 continue_unlock:
1020                                 unlock_page(page);
1021                                 continue;
1022                         }
1023
1024                         if (!PageDirty(page)) {
1025                                 /* someone wrote it for us */
1026                                 goto continue_unlock;
1027                         }
1028
1029                         if (PageWriteback(page)) {
1030                                 if (wbc->sync_mode != WB_SYNC_NONE)
1031                                         wait_on_page_writeback(page);
1032                                 else
1033                                         goto continue_unlock;
1034                         }
1035
1036                         BUG_ON(PageWriteback(page));
1037                         if (!clear_page_dirty_for_io(page))
1038                                 goto continue_unlock;
1039
1040                         ret = (*writepage)(page, wbc, data);
1041                         if (unlikely(ret)) {
1042                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1043                                         unlock_page(page);
1044                                         ret = 0;
1045                                 } else {
1046                                         /*
1047                                          * done_index is set past this page,
1048                                          * so media errors will not choke
1049                                          * background writeout for the entire
1050                                          * file. This has consequences for
1051                                          * range_cyclic semantics (ie. it may
1052                                          * not be suitable for data integrity
1053                                          * writeout).
1054                                          */
1055                                         done = 1;
1056                                         break;
1057                                 }
1058                         }
1059
1060                         if (nr_to_write > 0) {
1061                                 nr_to_write--;
1062                                 if (nr_to_write == 0 &&
1063                                     wbc->sync_mode == WB_SYNC_NONE) {
1064                                         /*
1065                                          * We stop writing back only if we are
1066                                          * not doing integrity sync. In case of
1067                                          * integrity sync we have to keep going
1068                                          * because someone may be concurrently
1069                                          * dirtying pages, and we might have
1070                                          * synced a lot of newly appeared dirty
1071                                          * pages, but have not synced all of the
1072                                          * old dirty pages.
1073                                          */
1074                                         done = 1;
1075                                         break;
1076                                 }
1077                         }
1078
1079                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
1080                                 wbc->encountered_congestion = 1;
1081                                 done = 1;
1082                                 break;
1083                         }
1084                 }
1085                 pagevec_release(&pvec);
1086                 cond_resched();
1087         }
1088         if (!cycled && !done) {
1089                 /*
1090                  * range_cyclic:
1091                  * We hit the last page and there is more work to be done: wrap
1092                  * back to the start of the file
1093                  */
1094                 cycled = 1;
1095                 index = 0;
1096                 end = writeback_index - 1;
1097                 goto retry;
1098         }
1099         if (!wbc->no_nrwrite_index_update) {
1100                 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
1101                         mapping->writeback_index = done_index;
1102                 wbc->nr_to_write = nr_to_write;
1103         }
1104
1105         return ret;
1106 }
1107 EXPORT_SYMBOL(write_cache_pages);
1108
1109 /*
1110  * Function used by generic_writepages to call the real writepage
1111  * function and set the mapping flags on error
1112  */
1113 static int __writepage(struct page *page, struct writeback_control *wbc,
1114                        void *data)
1115 {
1116         struct address_space *mapping = data;
1117         int ret = mapping->a_ops->writepage(page, wbc);
1118         mapping_set_error(mapping, ret);
1119         return ret;
1120 }
1121
1122 /**
1123  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1124  * @mapping: address space structure to write
1125  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1126  *
1127  * This is a library function, which implements the writepages()
1128  * address_space_operation.
1129  */
1130 int generic_writepages(struct address_space *mapping,
1131                        struct writeback_control *wbc)
1132 {
1133         /* deal with chardevs and other special file */
1134         if (!mapping->a_ops->writepage)
1135                 return 0;
1136
1137         return write_cache_pages(mapping, wbc, __writepage, mapping);
1138 }
1139
1140 EXPORT_SYMBOL(generic_writepages);
1141
1142 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1143 {
1144         int ret;
1145
1146         if (wbc->nr_to_write <= 0)
1147                 return 0;
1148         wbc->for_writepages = 1;
1149         if (mapping->a_ops->writepages)
1150                 ret = mapping->a_ops->writepages(mapping, wbc);
1151         else
1152                 ret = generic_writepages(mapping, wbc);
1153         wbc->for_writepages = 0;
1154         return ret;
1155 }
1156
1157 /**
1158  * write_one_page - write out a single page and optionally wait on I/O
1159  * @page: the page to write
1160  * @wait: if true, wait on writeout
1161  *
1162  * The page must be locked by the caller and will be unlocked upon return.
1163  *
1164  * write_one_page() returns a negative error code if I/O failed.
1165  */
1166 int write_one_page(struct page *page, int wait)
1167 {
1168         struct address_space *mapping = page->mapping;
1169         int ret = 0;
1170         struct writeback_control wbc = {
1171                 .sync_mode = WB_SYNC_ALL,
1172                 .nr_to_write = 1,
1173         };
1174
1175         BUG_ON(!PageLocked(page));
1176
1177         if (wait)
1178                 wait_on_page_writeback(page);
1179
1180         if (clear_page_dirty_for_io(page)) {
1181                 page_cache_get(page);
1182                 ret = mapping->a_ops->writepage(page, &wbc);
1183                 if (ret == 0 && wait) {
1184                         wait_on_page_writeback(page);
1185                         if (PageError(page))
1186                                 ret = -EIO;
1187                 }
1188                 page_cache_release(page);
1189         } else {
1190                 unlock_page(page);
1191         }
1192         return ret;
1193 }
1194 EXPORT_SYMBOL(write_one_page);
1195
1196 /*
1197  * For address_spaces which do not use buffers nor write back.
1198  */
1199 int __set_page_dirty_no_writeback(struct page *page)
1200 {
1201         if (!PageDirty(page))
1202                 SetPageDirty(page);
1203         return 0;
1204 }
1205
1206 /*
1207  * Helper function for set_page_dirty family.
1208  * NOTE: This relies on being atomic wrt interrupts.
1209  */
1210 void account_page_dirtied(struct page *page, struct address_space *mapping)
1211 {
1212         if (mapping_cap_account_dirty(mapping)) {
1213                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1214                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1215                 task_dirty_inc(current);
1216                 task_io_account_write(PAGE_CACHE_SIZE);
1217         }
1218 }
1219
1220 /*
1221  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1222  * its radix tree.
1223  *
1224  * This is also used when a single buffer is being dirtied: we want to set the
1225  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1226  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1227  *
1228  * Most callers have locked the page, which pins the address_space in memory.
1229  * But zap_pte_range() does not lock the page, however in that case the
1230  * mapping is pinned by the vma's ->vm_file reference.
1231  *
1232  * We take care to handle the case where the page was truncated from the
1233  * mapping by re-checking page_mapping() inside tree_lock.
1234  */
1235 int __set_page_dirty_nobuffers(struct page *page)
1236 {
1237         if (!TestSetPageDirty(page)) {
1238                 struct address_space *mapping = page_mapping(page);
1239                 struct address_space *mapping2;
1240
1241                 if (!mapping)
1242                         return 1;
1243
1244                 spin_lock_irq(&mapping->tree_lock);
1245                 mapping2 = page_mapping(page);
1246                 if (mapping2) { /* Race with truncate? */
1247                         BUG_ON(mapping2 != mapping);
1248                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1249                         account_page_dirtied(page, mapping);
1250                         radix_tree_tag_set(&mapping->page_tree,
1251                                 page_index(page), PAGECACHE_TAG_DIRTY);
1252                 }
1253                 spin_unlock_irq(&mapping->tree_lock);
1254                 if (mapping->host) {
1255                         /* !PageAnon && !swapper_space */
1256                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1257                 }
1258                 return 1;
1259         }
1260         return 0;
1261 }
1262 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1263
1264 /*
1265  * When a writepage implementation decides that it doesn't want to write this
1266  * page for some reason, it should redirty the locked page via
1267  * redirty_page_for_writepage() and it should then unlock the page and return 0
1268  */
1269 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1270 {
1271         wbc->pages_skipped++;
1272         return __set_page_dirty_nobuffers(page);
1273 }
1274 EXPORT_SYMBOL(redirty_page_for_writepage);
1275
1276 /*
1277  * If the mapping doesn't provide a set_page_dirty a_op, then
1278  * just fall through and assume that it wants buffer_heads.
1279  */
1280 int set_page_dirty(struct page *page)
1281 {
1282         struct address_space *mapping = page_mapping(page);
1283
1284         if (likely(mapping)) {
1285                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1286 #ifdef CONFIG_BLOCK
1287                 if (!spd)
1288                         spd = __set_page_dirty_buffers;
1289 #endif
1290                 return (*spd)(page);
1291         }
1292         if (!PageDirty(page)) {
1293                 if (!TestSetPageDirty(page))
1294                         return 1;
1295         }
1296         return 0;
1297 }
1298 EXPORT_SYMBOL(set_page_dirty);
1299
1300 /*
1301  * set_page_dirty() is racy if the caller has no reference against
1302  * page->mapping->host, and if the page is unlocked.  This is because another
1303  * CPU could truncate the page off the mapping and then free the mapping.
1304  *
1305  * Usually, the page _is_ locked, or the caller is a user-space process which
1306  * holds a reference on the inode by having an open file.
1307  *
1308  * In other cases, the page should be locked before running set_page_dirty().
1309  */
1310 int set_page_dirty_lock(struct page *page)
1311 {
1312         int ret;
1313
1314         lock_page_nosync(page);
1315         ret = set_page_dirty(page);
1316         unlock_page(page);
1317         return ret;
1318 }
1319 EXPORT_SYMBOL(set_page_dirty_lock);
1320
1321 /*
1322  * Clear a page's dirty flag, while caring for dirty memory accounting.
1323  * Returns true if the page was previously dirty.
1324  *
1325  * This is for preparing to put the page under writeout.  We leave the page
1326  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1327  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1328  * implementation will run either set_page_writeback() or set_page_dirty(),
1329  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1330  * back into sync.
1331  *
1332  * This incoherency between the page's dirty flag and radix-tree tag is
1333  * unfortunate, but it only exists while the page is locked.
1334  */
1335 int clear_page_dirty_for_io(struct page *page)
1336 {
1337         struct address_space *mapping = page_mapping(page);
1338
1339         BUG_ON(!PageLocked(page));
1340
1341         ClearPageReclaim(page);
1342         if (mapping && mapping_cap_account_dirty(mapping)) {
1343                 /*
1344                  * Yes, Virginia, this is indeed insane.
1345                  *
1346                  * We use this sequence to make sure that
1347                  *  (a) we account for dirty stats properly
1348                  *  (b) we tell the low-level filesystem to
1349                  *      mark the whole page dirty if it was
1350                  *      dirty in a pagetable. Only to then
1351                  *  (c) clean the page again and return 1 to
1352                  *      cause the writeback.
1353                  *
1354                  * This way we avoid all nasty races with the
1355                  * dirty bit in multiple places and clearing
1356                  * them concurrently from different threads.
1357                  *
1358                  * Note! Normally the "set_page_dirty(page)"
1359                  * has no effect on the actual dirty bit - since
1360                  * that will already usually be set. But we
1361                  * need the side effects, and it can help us
1362                  * avoid races.
1363                  *
1364                  * We basically use the page "master dirty bit"
1365                  * as a serialization point for all the different
1366                  * threads doing their things.
1367                  */
1368                 if (page_mkclean(page))
1369                         set_page_dirty(page);
1370                 /*
1371                  * We carefully synchronise fault handlers against
1372                  * installing a dirty pte and marking the page dirty
1373                  * at this point. We do this by having them hold the
1374                  * page lock at some point after installing their
1375                  * pte, but before marking the page dirty.
1376                  * Pages are always locked coming in here, so we get
1377                  * the desired exclusion. See mm/memory.c:do_wp_page()
1378                  * for more comments.
1379                  */
1380                 if (TestClearPageDirty(page)) {
1381                         dec_zone_page_state(page, NR_FILE_DIRTY);
1382                         dec_bdi_stat(mapping->backing_dev_info,
1383                                         BDI_RECLAIMABLE);
1384                         return 1;
1385                 }
1386                 return 0;
1387         }
1388         return TestClearPageDirty(page);
1389 }
1390 EXPORT_SYMBOL(clear_page_dirty_for_io);
1391
1392 int test_clear_page_writeback(struct page *page)
1393 {
1394         struct address_space *mapping = page_mapping(page);
1395         int ret;
1396
1397         if (mapping) {
1398                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1399                 unsigned long flags;
1400
1401                 spin_lock_irqsave(&mapping->tree_lock, flags);
1402                 ret = TestClearPageWriteback(page);
1403                 if (ret) {
1404                         radix_tree_tag_clear(&mapping->page_tree,
1405                                                 page_index(page),
1406                                                 PAGECACHE_TAG_WRITEBACK);
1407                         if (bdi_cap_account_writeback(bdi)) {
1408                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1409                                 __bdi_writeout_inc(bdi);
1410                         }
1411                 }
1412                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1413         } else {
1414                 ret = TestClearPageWriteback(page);
1415         }
1416         if (ret)
1417                 dec_zone_page_state(page, NR_WRITEBACK);
1418         return ret;
1419 }
1420
1421 int test_set_page_writeback(struct page *page)
1422 {
1423         struct address_space *mapping = page_mapping(page);
1424         int ret;
1425
1426         if (mapping) {
1427                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1428                 unsigned long flags;
1429
1430                 spin_lock_irqsave(&mapping->tree_lock, flags);
1431                 ret = TestSetPageWriteback(page);
1432                 if (!ret) {
1433                         radix_tree_tag_set(&mapping->page_tree,
1434                                                 page_index(page),
1435                                                 PAGECACHE_TAG_WRITEBACK);
1436                         if (bdi_cap_account_writeback(bdi))
1437                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1438                 }
1439                 if (!PageDirty(page))
1440                         radix_tree_tag_clear(&mapping->page_tree,
1441                                                 page_index(page),
1442                                                 PAGECACHE_TAG_DIRTY);
1443                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1444         } else {
1445                 ret = TestSetPageWriteback(page);
1446         }
1447         if (!ret)
1448                 inc_zone_page_state(page, NR_WRITEBACK);
1449         return ret;
1450
1451 }
1452 EXPORT_SYMBOL(test_set_page_writeback);
1453
1454 /*
1455  * Return true if any of the pages in the mapping are marked with the
1456  * passed tag.
1457  */
1458 int mapping_tagged(struct address_space *mapping, int tag)
1459 {
1460         int ret;
1461         rcu_read_lock();
1462         ret = radix_tree_tagged(&mapping->page_tree, tag);
1463         rcu_read_unlock();
1464         return ret;
1465 }
1466 EXPORT_SYMBOL(mapping_tagged);