writeback: skip tmpfs early in balance_dirty_pages_ratelimited_nr()
[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 #include <trace/events/writeback.h>
38
39 /*
40  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
41  * will look to see if it needs to force writeback or throttling.
42  */
43 static long ratelimit_pages = 32;
44
45 /*
46  * When balance_dirty_pages decides that the caller needs to perform some
47  * non-background writeback, this is how many pages it will attempt to write.
48  * It should be somewhat larger than dirtied pages to ensure that reasonably
49  * large amounts of I/O are submitted.
50  */
51 static inline long sync_writeback_pages(unsigned long dirtied)
52 {
53         if (dirtied < ratelimit_pages)
54                 dirtied = ratelimit_pages;
55
56         return dirtied + dirtied / 2;
57 }
58
59 /* The following parameters are exported via /proc/sys/vm */
60
61 /*
62  * Start background writeback (via writeback threads) at this percentage
63  */
64 int dirty_background_ratio = 10;
65
66 /*
67  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
68  * dirty_background_ratio * the amount of dirtyable memory
69  */
70 unsigned long dirty_background_bytes;
71
72 /*
73  * free highmem will not be subtracted from the total free memory
74  * for calculating free ratios if vm_highmem_is_dirtyable is true
75  */
76 int vm_highmem_is_dirtyable;
77
78 /*
79  * The generator of dirty data starts writeback at this percentage
80  */
81 int vm_dirty_ratio = 20;
82
83 /*
84  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
85  * vm_dirty_ratio * the amount of dirtyable memory
86  */
87 unsigned long vm_dirty_bytes;
88
89 /*
90  * The interval between `kupdate'-style writebacks
91  */
92 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
93
94 /*
95  * The longest time for which data is allowed to remain dirty
96  */
97 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
98
99 /*
100  * Flag that makes the machine dump writes/reads and block dirtyings.
101  */
102 int block_dump;
103
104 /*
105  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
106  * a full sync is triggered after this time elapses without any disk activity.
107  */
108 int laptop_mode;
109
110 EXPORT_SYMBOL(laptop_mode);
111
112 /* End of sysctl-exported parameters */
113
114
115 /*
116  * Scale the writeback cache size proportional to the relative writeout speeds.
117  *
118  * We do this by keeping a floating proportion between BDIs, based on page
119  * writeback completions [end_page_writeback()]. Those devices that write out
120  * pages fastest will get the larger share, while the slower will get a smaller
121  * share.
122  *
123  * We use page writeout completions because we are interested in getting rid of
124  * dirty pages. Having them written out is the primary goal.
125  *
126  * We introduce a concept of time, a period over which we measure these events,
127  * because demand can/will vary over time. The length of this period itself is
128  * measured in page writeback completions.
129  *
130  */
131 static struct prop_descriptor vm_completions;
132 static struct prop_descriptor vm_dirties;
133
134 /*
135  * couple the period to the dirty_ratio:
136  *
137  *   period/2 ~ roundup_pow_of_two(dirty limit)
138  */
139 static int calc_period_shift(void)
140 {
141         unsigned long dirty_total;
142
143         if (vm_dirty_bytes)
144                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
145         else
146                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
147                                 100;
148         return 2 + ilog2(dirty_total - 1);
149 }
150
151 /*
152  * update the period when the dirty threshold changes.
153  */
154 static void update_completion_period(void)
155 {
156         int shift = calc_period_shift();
157         prop_change_shift(&vm_completions, shift);
158         prop_change_shift(&vm_dirties, shift);
159 }
160
161 int dirty_background_ratio_handler(struct ctl_table *table, int write,
162                 void __user *buffer, size_t *lenp,
163                 loff_t *ppos)
164 {
165         int ret;
166
167         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
168         if (ret == 0 && write)
169                 dirty_background_bytes = 0;
170         return ret;
171 }
172
173 int dirty_background_bytes_handler(struct ctl_table *table, int write,
174                 void __user *buffer, size_t *lenp,
175                 loff_t *ppos)
176 {
177         int ret;
178
179         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
180         if (ret == 0 && write)
181                 dirty_background_ratio = 0;
182         return ret;
183 }
184
185 int dirty_ratio_handler(struct ctl_table *table, int write,
186                 void __user *buffer, size_t *lenp,
187                 loff_t *ppos)
188 {
189         int old_ratio = vm_dirty_ratio;
190         int ret;
191
192         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
193         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
194                 update_completion_period();
195                 vm_dirty_bytes = 0;
196         }
197         return ret;
198 }
199
200
201 int dirty_bytes_handler(struct ctl_table *table, int write,
202                 void __user *buffer, size_t *lenp,
203                 loff_t *ppos)
204 {
205         unsigned long old_bytes = vm_dirty_bytes;
206         int ret;
207
208         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
209         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
210                 update_completion_period();
211                 vm_dirty_ratio = 0;
212         }
213         return ret;
214 }
215
216 /*
217  * Increment the BDI's writeout completion count and the global writeout
218  * completion count. Called from test_clear_page_writeback().
219  */
220 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
221 {
222         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
223                               bdi->max_prop_frac);
224 }
225
226 void bdi_writeout_inc(struct backing_dev_info *bdi)
227 {
228         unsigned long flags;
229
230         local_irq_save(flags);
231         __bdi_writeout_inc(bdi);
232         local_irq_restore(flags);
233 }
234 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
235
236 void task_dirty_inc(struct task_struct *tsk)
237 {
238         prop_inc_single(&vm_dirties, &tsk->dirties);
239 }
240
241 /*
242  * Obtain an accurate fraction of the BDI's portion.
243  */
244 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
245                 long *numerator, long *denominator)
246 {
247         prop_fraction_percpu(&vm_completions, &bdi->completions,
248                                 numerator, denominator);
249 }
250
251 static inline void task_dirties_fraction(struct task_struct *tsk,
252                 long *numerator, long *denominator)
253 {
254         prop_fraction_single(&vm_dirties, &tsk->dirties,
255                                 numerator, denominator);
256 }
257
258 /*
259  * task_dirty_limit - scale down dirty throttling threshold for one task
260  *
261  * task specific dirty limit:
262  *
263  *   dirty -= (dirty/8) * p_{t}
264  *
265  * To protect light/slow dirtying tasks from heavier/fast ones, we start
266  * throttling individual tasks before reaching the bdi dirty limit.
267  * Relatively low thresholds will be allocated to heavy dirtiers. So when
268  * dirty pages grow large, heavy dirtiers will be throttled first, which will
269  * effectively curb the growth of dirty pages. Light dirtiers with high enough
270  * dirty threshold may never get throttled.
271  */
272 static unsigned long task_dirty_limit(struct task_struct *tsk,
273                                        unsigned long bdi_dirty)
274 {
275         long numerator, denominator;
276         unsigned long dirty = bdi_dirty;
277         u64 inv = dirty >> 3;
278
279         task_dirties_fraction(tsk, &numerator, &denominator);
280         inv *= numerator;
281         do_div(inv, denominator);
282
283         dirty -= inv;
284
285         return max(dirty, bdi_dirty/2);
286 }
287
288 /*
289  *
290  */
291 static unsigned int bdi_min_ratio;
292
293 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
294 {
295         int ret = 0;
296
297         spin_lock_bh(&bdi_lock);
298         if (min_ratio > bdi->max_ratio) {
299                 ret = -EINVAL;
300         } else {
301                 min_ratio -= bdi->min_ratio;
302                 if (bdi_min_ratio + min_ratio < 100) {
303                         bdi_min_ratio += min_ratio;
304                         bdi->min_ratio += min_ratio;
305                 } else {
306                         ret = -EINVAL;
307                 }
308         }
309         spin_unlock_bh(&bdi_lock);
310
311         return ret;
312 }
313
314 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
315 {
316         int ret = 0;
317
318         if (max_ratio > 100)
319                 return -EINVAL;
320
321         spin_lock_bh(&bdi_lock);
322         if (bdi->min_ratio > max_ratio) {
323                 ret = -EINVAL;
324         } else {
325                 bdi->max_ratio = max_ratio;
326                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
327         }
328         spin_unlock_bh(&bdi_lock);
329
330         return ret;
331 }
332 EXPORT_SYMBOL(bdi_set_max_ratio);
333
334 /*
335  * Work out the current dirty-memory clamping and background writeout
336  * thresholds.
337  *
338  * The main aim here is to lower them aggressively if there is a lot of mapped
339  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
340  * pages.  It is better to clamp down on writers than to start swapping, and
341  * performing lots of scanning.
342  *
343  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
344  *
345  * We don't permit the clamping level to fall below 5% - that is getting rather
346  * excessive.
347  *
348  * We make sure that the background writeout level is below the adjusted
349  * clamping level.
350  */
351
352 static unsigned long highmem_dirtyable_memory(unsigned long total)
353 {
354 #ifdef CONFIG_HIGHMEM
355         int node;
356         unsigned long x = 0;
357
358         for_each_node_state(node, N_HIGH_MEMORY) {
359                 struct zone *z =
360                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
361
362                 x += zone_page_state(z, NR_FREE_PAGES) +
363                      zone_reclaimable_pages(z);
364         }
365         /*
366          * Make sure that the number of highmem pages is never larger
367          * than the number of the total dirtyable memory. This can only
368          * occur in very strange VM situations but we want to make sure
369          * that this does not occur.
370          */
371         return min(x, total);
372 #else
373         return 0;
374 #endif
375 }
376
377 /**
378  * determine_dirtyable_memory - amount of memory that may be used
379  *
380  * Returns the numebr of pages that can currently be freed and used
381  * by the kernel for direct mappings.
382  */
383 unsigned long determine_dirtyable_memory(void)
384 {
385         unsigned long x;
386
387         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
388
389         if (!vm_highmem_is_dirtyable)
390                 x -= highmem_dirtyable_memory(x);
391
392         return x + 1;   /* Ensure that we never return 0 */
393 }
394
395 /*
396  * global_dirty_limits - background-writeback and dirty-throttling thresholds
397  *
398  * Calculate the dirty thresholds based on sysctl parameters
399  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
400  * - vm.dirty_ratio             or  vm.dirty_bytes
401  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
402  * real-time tasks.
403  */
404 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
405 {
406         unsigned long background;
407         unsigned long dirty;
408         unsigned long uninitialized_var(available_memory);
409         struct task_struct *tsk;
410
411         if (!vm_dirty_bytes || !dirty_background_bytes)
412                 available_memory = determine_dirtyable_memory();
413
414         if (vm_dirty_bytes)
415                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
416         else
417                 dirty = (vm_dirty_ratio * available_memory) / 100;
418
419         if (dirty_background_bytes)
420                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
421         else
422                 background = (dirty_background_ratio * available_memory) / 100;
423
424         if (background >= dirty)
425                 background = dirty / 2;
426         tsk = current;
427         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
428                 background += background / 4;
429                 dirty += dirty / 4;
430         }
431         *pbackground = background;
432         *pdirty = dirty;
433 }
434
435 /**
436  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
437  * @bdi: the backing_dev_info to query
438  * @dirty: global dirty limit in pages
439  *
440  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
441  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
442  * And the "limit" in the name is not seriously taken as hard limit in
443  * balance_dirty_pages().
444  *
445  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
446  * - starving fast devices
447  * - piling up dirty pages (that will take long time to sync) on slow devices
448  *
449  * The bdi's share of dirty limit will be adapting to its throughput and
450  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
451  */
452 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
453 {
454         u64 bdi_dirty;
455         long numerator, denominator;
456
457         /*
458          * Calculate this BDI's share of the dirty ratio.
459          */
460         bdi_writeout_fraction(bdi, &numerator, &denominator);
461
462         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
463         bdi_dirty *= numerator;
464         do_div(bdi_dirty, denominator);
465
466         bdi_dirty += (dirty * bdi->min_ratio) / 100;
467         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
468                 bdi_dirty = dirty * bdi->max_ratio / 100;
469
470         return bdi_dirty;
471 }
472
473 /*
474  * balance_dirty_pages() must be called by processes which are generating dirty
475  * data.  It looks at the number of dirty pages in the machine and will force
476  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
477  * If we're over `background_thresh' then the writeback threads are woken to
478  * perform some writeout.
479  */
480 static void balance_dirty_pages(struct address_space *mapping,
481                                 unsigned long write_chunk)
482 {
483         long nr_reclaimable, bdi_nr_reclaimable;
484         long nr_writeback, bdi_nr_writeback;
485         unsigned long background_thresh;
486         unsigned long dirty_thresh;
487         unsigned long bdi_thresh;
488         unsigned long pages_written = 0;
489         unsigned long pause = 1;
490         bool dirty_exceeded = false;
491         struct backing_dev_info *bdi = mapping->backing_dev_info;
492
493         for (;;) {
494                 struct writeback_control wbc = {
495                         .sync_mode      = WB_SYNC_NONE,
496                         .older_than_this = NULL,
497                         .nr_to_write    = write_chunk,
498                         .range_cyclic   = 1,
499                 };
500
501                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
502                                         global_page_state(NR_UNSTABLE_NFS);
503                 nr_writeback = global_page_state(NR_WRITEBACK);
504
505                 global_dirty_limits(&background_thresh, &dirty_thresh);
506
507                 /*
508                  * Throttle it only when the background writeback cannot
509                  * catch-up. This avoids (excessively) small writeouts
510                  * when the bdi limits are ramping up.
511                  */
512                 if (nr_reclaimable + nr_writeback <=
513                                 (background_thresh + dirty_thresh) / 2)
514                         break;
515
516                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
517                 bdi_thresh = task_dirty_limit(current, bdi_thresh);
518
519                 /*
520                  * In order to avoid the stacked BDI deadlock we need
521                  * to ensure we accurately count the 'dirty' pages when
522                  * the threshold is low.
523                  *
524                  * Otherwise it would be possible to get thresh+n pages
525                  * reported dirty, even though there are thresh-m pages
526                  * actually dirty; with m+n sitting in the percpu
527                  * deltas.
528                  */
529                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
530                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
531                         bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
532                 } else {
533                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
534                         bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
535                 }
536
537                 /*
538                  * The bdi thresh is somehow "soft" limit derived from the
539                  * global "hard" limit. The former helps to prevent heavy IO
540                  * bdi or process from holding back light ones; The latter is
541                  * the last resort safeguard.
542                  */
543                 dirty_exceeded =
544                         (bdi_nr_reclaimable + bdi_nr_writeback > bdi_thresh)
545                         || (nr_reclaimable + nr_writeback > dirty_thresh);
546
547                 if (!dirty_exceeded)
548                         break;
549
550                 if (!bdi->dirty_exceeded)
551                         bdi->dirty_exceeded = 1;
552
553                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
554                  * Unstable writes are a feature of certain networked
555                  * filesystems (i.e. NFS) in which data may have been
556                  * written to the server's write cache, but has not yet
557                  * been flushed to permanent storage.
558                  * Only move pages to writeback if this bdi is over its
559                  * threshold otherwise wait until the disk writes catch
560                  * up.
561                  */
562                 trace_wbc_balance_dirty_start(&wbc, bdi);
563                 if (bdi_nr_reclaimable > bdi_thresh) {
564                         writeback_inodes_wb(&bdi->wb, &wbc);
565                         pages_written += write_chunk - wbc.nr_to_write;
566                         trace_wbc_balance_dirty_written(&wbc, bdi);
567                         if (pages_written >= write_chunk)
568                                 break;          /* We've done our duty */
569                 }
570                 trace_wbc_balance_dirty_wait(&wbc, bdi);
571                 __set_current_state(TASK_UNINTERRUPTIBLE);
572                 io_schedule_timeout(pause);
573
574                 /*
575                  * Increase the delay for each loop, up to our previous
576                  * default of taking a 100ms nap.
577                  */
578                 pause <<= 1;
579                 if (pause > HZ / 10)
580                         pause = HZ / 10;
581         }
582
583         if (!dirty_exceeded && bdi->dirty_exceeded)
584                 bdi->dirty_exceeded = 0;
585
586         if (writeback_in_progress(bdi))
587                 return;
588
589         /*
590          * In laptop mode, we wait until hitting the higher threshold before
591          * starting background writeout, and then write out all the way down
592          * to the lower threshold.  So slow writers cause minimal disk activity.
593          *
594          * In normal mode, we start background writeout at the lower
595          * background_thresh, to keep the amount of dirty memory low.
596          */
597         if ((laptop_mode && pages_written) ||
598             (!laptop_mode && (nr_reclaimable > background_thresh)))
599                 bdi_start_background_writeback(bdi);
600 }
601
602 void set_page_dirty_balance(struct page *page, int page_mkwrite)
603 {
604         if (set_page_dirty(page) || page_mkwrite) {
605                 struct address_space *mapping = page_mapping(page);
606
607                 if (mapping)
608                         balance_dirty_pages_ratelimited(mapping);
609         }
610 }
611
612 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
613
614 /**
615  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
616  * @mapping: address_space which was dirtied
617  * @nr_pages_dirtied: number of pages which the caller has just dirtied
618  *
619  * Processes which are dirtying memory should call in here once for each page
620  * which was newly dirtied.  The function will periodically check the system's
621  * dirty state and will initiate writeback if needed.
622  *
623  * On really big machines, get_writeback_state is expensive, so try to avoid
624  * calling it too often (ratelimiting).  But once we're over the dirty memory
625  * limit we decrease the ratelimiting by a lot, to prevent individual processes
626  * from overshooting the limit by (ratelimit_pages) each.
627  */
628 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
629                                         unsigned long nr_pages_dirtied)
630 {
631         struct backing_dev_info *bdi = mapping->backing_dev_info;
632         unsigned long ratelimit;
633         unsigned long *p;
634
635         if (!bdi_cap_account_dirty(bdi))
636                 return;
637
638         ratelimit = ratelimit_pages;
639         if (mapping->backing_dev_info->dirty_exceeded)
640                 ratelimit = 8;
641
642         /*
643          * Check the rate limiting. Also, we do not want to throttle real-time
644          * tasks in balance_dirty_pages(). Period.
645          */
646         preempt_disable();
647         p =  &__get_cpu_var(bdp_ratelimits);
648         *p += nr_pages_dirtied;
649         if (unlikely(*p >= ratelimit)) {
650                 ratelimit = sync_writeback_pages(*p);
651                 *p = 0;
652                 preempt_enable();
653                 balance_dirty_pages(mapping, ratelimit);
654                 return;
655         }
656         preempt_enable();
657 }
658 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
659
660 void throttle_vm_writeout(gfp_t gfp_mask)
661 {
662         unsigned long background_thresh;
663         unsigned long dirty_thresh;
664
665         for ( ; ; ) {
666                 global_dirty_limits(&background_thresh, &dirty_thresh);
667
668                 /*
669                  * Boost the allowable dirty threshold a bit for page
670                  * allocators so they don't get DoS'ed by heavy writers
671                  */
672                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
673
674                 if (global_page_state(NR_UNSTABLE_NFS) +
675                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
676                                 break;
677                 congestion_wait(BLK_RW_ASYNC, HZ/10);
678
679                 /*
680                  * The caller might hold locks which can prevent IO completion
681                  * or progress in the filesystem.  So we cannot just sit here
682                  * waiting for IO to complete.
683                  */
684                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
685                         break;
686         }
687 }
688
689 /*
690  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
691  */
692 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
693         void __user *buffer, size_t *length, loff_t *ppos)
694 {
695         proc_dointvec(table, write, buffer, length, ppos);
696         bdi_arm_supers_timer();
697         return 0;
698 }
699
700 #ifdef CONFIG_BLOCK
701 void laptop_mode_timer_fn(unsigned long data)
702 {
703         struct request_queue *q = (struct request_queue *)data;
704         int nr_pages = global_page_state(NR_FILE_DIRTY) +
705                 global_page_state(NR_UNSTABLE_NFS);
706
707         /*
708          * We want to write everything out, not just down to the dirty
709          * threshold
710          */
711         if (bdi_has_dirty_io(&q->backing_dev_info))
712                 bdi_start_writeback(&q->backing_dev_info, nr_pages);
713 }
714
715 /*
716  * We've spun up the disk and we're in laptop mode: schedule writeback
717  * of all dirty data a few seconds from now.  If the flush is already scheduled
718  * then push it back - the user is still using the disk.
719  */
720 void laptop_io_completion(struct backing_dev_info *info)
721 {
722         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
723 }
724
725 /*
726  * We're in laptop mode and we've just synced. The sync's writes will have
727  * caused another writeback to be scheduled by laptop_io_completion.
728  * Nothing needs to be written back anymore, so we unschedule the writeback.
729  */
730 void laptop_sync_completion(void)
731 {
732         struct backing_dev_info *bdi;
733
734         rcu_read_lock();
735
736         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
737                 del_timer(&bdi->laptop_mode_wb_timer);
738
739         rcu_read_unlock();
740 }
741 #endif
742
743 /*
744  * If ratelimit_pages is too high then we can get into dirty-data overload
745  * if a large number of processes all perform writes at the same time.
746  * If it is too low then SMP machines will call the (expensive)
747  * get_writeback_state too often.
748  *
749  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
750  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
751  * thresholds before writeback cuts in.
752  *
753  * But the limit should not be set too high.  Because it also controls the
754  * amount of memory which the balance_dirty_pages() caller has to write back.
755  * If this is too large then the caller will block on the IO queue all the
756  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
757  * will write six megabyte chunks, max.
758  */
759
760 void writeback_set_ratelimit(void)
761 {
762         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
763         if (ratelimit_pages < 16)
764                 ratelimit_pages = 16;
765         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
766                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
767 }
768
769 static int __cpuinit
770 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
771 {
772         writeback_set_ratelimit();
773         return NOTIFY_DONE;
774 }
775
776 static struct notifier_block __cpuinitdata ratelimit_nb = {
777         .notifier_call  = ratelimit_handler,
778         .next           = NULL,
779 };
780
781 /*
782  * Called early on to tune the page writeback dirty limits.
783  *
784  * We used to scale dirty pages according to how total memory
785  * related to pages that could be allocated for buffers (by
786  * comparing nr_free_buffer_pages() to vm_total_pages.
787  *
788  * However, that was when we used "dirty_ratio" to scale with
789  * all memory, and we don't do that any more. "dirty_ratio"
790  * is now applied to total non-HIGHPAGE memory (by subtracting
791  * totalhigh_pages from vm_total_pages), and as such we can't
792  * get into the old insane situation any more where we had
793  * large amounts of dirty pages compared to a small amount of
794  * non-HIGHMEM memory.
795  *
796  * But we might still want to scale the dirty_ratio by how
797  * much memory the box has..
798  */
799 void __init page_writeback_init(void)
800 {
801         int shift;
802
803         writeback_set_ratelimit();
804         register_cpu_notifier(&ratelimit_nb);
805
806         shift = calc_period_shift();
807         prop_descriptor_init(&vm_completions, shift);
808         prop_descriptor_init(&vm_dirties, shift);
809 }
810
811 /**
812  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
813  * @mapping: address space structure to write
814  * @start: starting page index
815  * @end: ending page index (inclusive)
816  *
817  * This function scans the page range from @start to @end (inclusive) and tags
818  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
819  * that write_cache_pages (or whoever calls this function) will then use
820  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
821  * used to avoid livelocking of writeback by a process steadily creating new
822  * dirty pages in the file (thus it is important for this function to be quick
823  * so that it can tag pages faster than a dirtying process can create them).
824  */
825 /*
826  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
827  */
828 void tag_pages_for_writeback(struct address_space *mapping,
829                              pgoff_t start, pgoff_t end)
830 {
831 #define WRITEBACK_TAG_BATCH 4096
832         unsigned long tagged;
833
834         do {
835                 spin_lock_irq(&mapping->tree_lock);
836                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
837                                 &start, end, WRITEBACK_TAG_BATCH,
838                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
839                 spin_unlock_irq(&mapping->tree_lock);
840                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
841                 cond_resched();
842                 /* We check 'start' to handle wrapping when end == ~0UL */
843         } while (tagged >= WRITEBACK_TAG_BATCH && start);
844 }
845 EXPORT_SYMBOL(tag_pages_for_writeback);
846
847 /**
848  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
849  * @mapping: address space structure to write
850  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
851  * @writepage: function called for each page
852  * @data: data passed to writepage function
853  *
854  * If a page is already under I/O, write_cache_pages() skips it, even
855  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
856  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
857  * and msync() need to guarantee that all the data which was dirty at the time
858  * the call was made get new I/O started against them.  If wbc->sync_mode is
859  * WB_SYNC_ALL then we were called for data integrity and we must wait for
860  * existing IO to complete.
861  *
862  * To avoid livelocks (when other process dirties new pages), we first tag
863  * pages which should be written back with TOWRITE tag and only then start
864  * writing them. For data-integrity sync we have to be careful so that we do
865  * not miss some pages (e.g., because some other process has cleared TOWRITE
866  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
867  * by the process clearing the DIRTY tag (and submitting the page for IO).
868  */
869 int write_cache_pages(struct address_space *mapping,
870                       struct writeback_control *wbc, writepage_t writepage,
871                       void *data)
872 {
873         int ret = 0;
874         int done = 0;
875         struct pagevec pvec;
876         int nr_pages;
877         pgoff_t uninitialized_var(writeback_index);
878         pgoff_t index;
879         pgoff_t end;            /* Inclusive */
880         pgoff_t done_index;
881         int cycled;
882         int range_whole = 0;
883         int tag;
884
885         pagevec_init(&pvec, 0);
886         if (wbc->range_cyclic) {
887                 writeback_index = mapping->writeback_index; /* prev offset */
888                 index = writeback_index;
889                 if (index == 0)
890                         cycled = 1;
891                 else
892                         cycled = 0;
893                 end = -1;
894         } else {
895                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
896                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
897                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
898                         range_whole = 1;
899                 cycled = 1; /* ignore range_cyclic tests */
900         }
901         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
902                 tag = PAGECACHE_TAG_TOWRITE;
903         else
904                 tag = PAGECACHE_TAG_DIRTY;
905 retry:
906         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
907                 tag_pages_for_writeback(mapping, index, end);
908         done_index = index;
909         while (!done && (index <= end)) {
910                 int i;
911
912                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
913                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
914                 if (nr_pages == 0)
915                         break;
916
917                 for (i = 0; i < nr_pages; i++) {
918                         struct page *page = pvec.pages[i];
919
920                         /*
921                          * At this point, the page may be truncated or
922                          * invalidated (changing page->mapping to NULL), or
923                          * even swizzled back from swapper_space to tmpfs file
924                          * mapping. However, page->index will not change
925                          * because we have a reference on the page.
926                          */
927                         if (page->index > end) {
928                                 /*
929                                  * can't be range_cyclic (1st pass) because
930                                  * end == -1 in that case.
931                                  */
932                                 done = 1;
933                                 break;
934                         }
935
936                         done_index = page->index;
937
938                         lock_page(page);
939
940                         /*
941                          * Page truncated or invalidated. We can freely skip it
942                          * then, even for data integrity operations: the page
943                          * has disappeared concurrently, so there could be no
944                          * real expectation of this data interity operation
945                          * even if there is now a new, dirty page at the same
946                          * pagecache address.
947                          */
948                         if (unlikely(page->mapping != mapping)) {
949 continue_unlock:
950                                 unlock_page(page);
951                                 continue;
952                         }
953
954                         if (!PageDirty(page)) {
955                                 /* someone wrote it for us */
956                                 goto continue_unlock;
957                         }
958
959                         if (PageWriteback(page)) {
960                                 if (wbc->sync_mode != WB_SYNC_NONE)
961                                         wait_on_page_writeback(page);
962                                 else
963                                         goto continue_unlock;
964                         }
965
966                         BUG_ON(PageWriteback(page));
967                         if (!clear_page_dirty_for_io(page))
968                                 goto continue_unlock;
969
970                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
971                         ret = (*writepage)(page, wbc, data);
972                         if (unlikely(ret)) {
973                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
974                                         unlock_page(page);
975                                         ret = 0;
976                                 } else {
977                                         /*
978                                          * done_index is set past this page,
979                                          * so media errors will not choke
980                                          * background writeout for the entire
981                                          * file. This has consequences for
982                                          * range_cyclic semantics (ie. it may
983                                          * not be suitable for data integrity
984                                          * writeout).
985                                          */
986                                         done_index = page->index + 1;
987                                         done = 1;
988                                         break;
989                                 }
990                         }
991
992                         /*
993                          * We stop writing back only if we are not doing
994                          * integrity sync. In case of integrity sync we have to
995                          * keep going until we have written all the pages
996                          * we tagged for writeback prior to entering this loop.
997                          */
998                         if (--wbc->nr_to_write <= 0 &&
999                             wbc->sync_mode == WB_SYNC_NONE) {
1000                                 done = 1;
1001                                 break;
1002                         }
1003                 }
1004                 pagevec_release(&pvec);
1005                 cond_resched();
1006         }
1007         if (!cycled && !done) {
1008                 /*
1009                  * range_cyclic:
1010                  * We hit the last page and there is more work to be done: wrap
1011                  * back to the start of the file
1012                  */
1013                 cycled = 1;
1014                 index = 0;
1015                 end = writeback_index - 1;
1016                 goto retry;
1017         }
1018         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1019                 mapping->writeback_index = done_index;
1020
1021         return ret;
1022 }
1023 EXPORT_SYMBOL(write_cache_pages);
1024
1025 /*
1026  * Function used by generic_writepages to call the real writepage
1027  * function and set the mapping flags on error
1028  */
1029 static int __writepage(struct page *page, struct writeback_control *wbc,
1030                        void *data)
1031 {
1032         struct address_space *mapping = data;
1033         int ret = mapping->a_ops->writepage(page, wbc);
1034         mapping_set_error(mapping, ret);
1035         return ret;
1036 }
1037
1038 /**
1039  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1040  * @mapping: address space structure to write
1041  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1042  *
1043  * This is a library function, which implements the writepages()
1044  * address_space_operation.
1045  */
1046 int generic_writepages(struct address_space *mapping,
1047                        struct writeback_control *wbc)
1048 {
1049         struct blk_plug plug;
1050         int ret;
1051
1052         /* deal with chardevs and other special file */
1053         if (!mapping->a_ops->writepage)
1054                 return 0;
1055
1056         blk_start_plug(&plug);
1057         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1058         blk_finish_plug(&plug);
1059         return ret;
1060 }
1061
1062 EXPORT_SYMBOL(generic_writepages);
1063
1064 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1065 {
1066         int ret;
1067
1068         if (wbc->nr_to_write <= 0)
1069                 return 0;
1070         if (mapping->a_ops->writepages)
1071                 ret = mapping->a_ops->writepages(mapping, wbc);
1072         else
1073                 ret = generic_writepages(mapping, wbc);
1074         return ret;
1075 }
1076
1077 /**
1078  * write_one_page - write out a single page and optionally wait on I/O
1079  * @page: the page to write
1080  * @wait: if true, wait on writeout
1081  *
1082  * The page must be locked by the caller and will be unlocked upon return.
1083  *
1084  * write_one_page() returns a negative error code if I/O failed.
1085  */
1086 int write_one_page(struct page *page, int wait)
1087 {
1088         struct address_space *mapping = page->mapping;
1089         int ret = 0;
1090         struct writeback_control wbc = {
1091                 .sync_mode = WB_SYNC_ALL,
1092                 .nr_to_write = 1,
1093         };
1094
1095         BUG_ON(!PageLocked(page));
1096
1097         if (wait)
1098                 wait_on_page_writeback(page);
1099
1100         if (clear_page_dirty_for_io(page)) {
1101                 page_cache_get(page);
1102                 ret = mapping->a_ops->writepage(page, &wbc);
1103                 if (ret == 0 && wait) {
1104                         wait_on_page_writeback(page);
1105                         if (PageError(page))
1106                                 ret = -EIO;
1107                 }
1108                 page_cache_release(page);
1109         } else {
1110                 unlock_page(page);
1111         }
1112         return ret;
1113 }
1114 EXPORT_SYMBOL(write_one_page);
1115
1116 /*
1117  * For address_spaces which do not use buffers nor write back.
1118  */
1119 int __set_page_dirty_no_writeback(struct page *page)
1120 {
1121         if (!PageDirty(page))
1122                 return !TestSetPageDirty(page);
1123         return 0;
1124 }
1125
1126 /*
1127  * Helper function for set_page_dirty family.
1128  * NOTE: This relies on being atomic wrt interrupts.
1129  */
1130 void account_page_dirtied(struct page *page, struct address_space *mapping)
1131 {
1132         if (mapping_cap_account_dirty(mapping)) {
1133                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1134                 __inc_zone_page_state(page, NR_DIRTIED);
1135                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1136                 task_dirty_inc(current);
1137                 task_io_account_write(PAGE_CACHE_SIZE);
1138         }
1139 }
1140 EXPORT_SYMBOL(account_page_dirtied);
1141
1142 /*
1143  * Helper function for set_page_writeback family.
1144  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1145  * wrt interrupts.
1146  */
1147 void account_page_writeback(struct page *page)
1148 {
1149         inc_zone_page_state(page, NR_WRITEBACK);
1150         inc_zone_page_state(page, NR_WRITTEN);
1151 }
1152 EXPORT_SYMBOL(account_page_writeback);
1153
1154 /*
1155  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1156  * its radix tree.
1157  *
1158  * This is also used when a single buffer is being dirtied: we want to set the
1159  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1160  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1161  *
1162  * Most callers have locked the page, which pins the address_space in memory.
1163  * But zap_pte_range() does not lock the page, however in that case the
1164  * mapping is pinned by the vma's ->vm_file reference.
1165  *
1166  * We take care to handle the case where the page was truncated from the
1167  * mapping by re-checking page_mapping() inside tree_lock.
1168  */
1169 int __set_page_dirty_nobuffers(struct page *page)
1170 {
1171         if (!TestSetPageDirty(page)) {
1172                 struct address_space *mapping = page_mapping(page);
1173                 struct address_space *mapping2;
1174
1175                 if (!mapping)
1176                         return 1;
1177
1178                 spin_lock_irq(&mapping->tree_lock);
1179                 mapping2 = page_mapping(page);
1180                 if (mapping2) { /* Race with truncate? */
1181                         BUG_ON(mapping2 != mapping);
1182                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1183                         account_page_dirtied(page, mapping);
1184                         radix_tree_tag_set(&mapping->page_tree,
1185                                 page_index(page), PAGECACHE_TAG_DIRTY);
1186                 }
1187                 spin_unlock_irq(&mapping->tree_lock);
1188                 if (mapping->host) {
1189                         /* !PageAnon && !swapper_space */
1190                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1191                 }
1192                 return 1;
1193         }
1194         return 0;
1195 }
1196 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1197
1198 /*
1199  * When a writepage implementation decides that it doesn't want to write this
1200  * page for some reason, it should redirty the locked page via
1201  * redirty_page_for_writepage() and it should then unlock the page and return 0
1202  */
1203 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1204 {
1205         wbc->pages_skipped++;
1206         return __set_page_dirty_nobuffers(page);
1207 }
1208 EXPORT_SYMBOL(redirty_page_for_writepage);
1209
1210 /*
1211  * Dirty a page.
1212  *
1213  * For pages with a mapping this should be done under the page lock
1214  * for the benefit of asynchronous memory errors who prefer a consistent
1215  * dirty state. This rule can be broken in some special cases,
1216  * but should be better not to.
1217  *
1218  * If the mapping doesn't provide a set_page_dirty a_op, then
1219  * just fall through and assume that it wants buffer_heads.
1220  */
1221 int set_page_dirty(struct page *page)
1222 {
1223         struct address_space *mapping = page_mapping(page);
1224
1225         if (likely(mapping)) {
1226                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1227                 /*
1228                  * readahead/lru_deactivate_page could remain
1229                  * PG_readahead/PG_reclaim due to race with end_page_writeback
1230                  * About readahead, if the page is written, the flags would be
1231                  * reset. So no problem.
1232                  * About lru_deactivate_page, if the page is redirty, the flag
1233                  * will be reset. So no problem. but if the page is used by readahead
1234                  * it will confuse readahead and make it restart the size rampup
1235                  * process. But it's a trivial problem.
1236                  */
1237                 ClearPageReclaim(page);
1238 #ifdef CONFIG_BLOCK
1239                 if (!spd)
1240                         spd = __set_page_dirty_buffers;
1241 #endif
1242                 return (*spd)(page);
1243         }
1244         if (!PageDirty(page)) {
1245                 if (!TestSetPageDirty(page))
1246                         return 1;
1247         }
1248         return 0;
1249 }
1250 EXPORT_SYMBOL(set_page_dirty);
1251
1252 /*
1253  * set_page_dirty() is racy if the caller has no reference against
1254  * page->mapping->host, and if the page is unlocked.  This is because another
1255  * CPU could truncate the page off the mapping and then free the mapping.
1256  *
1257  * Usually, the page _is_ locked, or the caller is a user-space process which
1258  * holds a reference on the inode by having an open file.
1259  *
1260  * In other cases, the page should be locked before running set_page_dirty().
1261  */
1262 int set_page_dirty_lock(struct page *page)
1263 {
1264         int ret;
1265
1266         lock_page(page);
1267         ret = set_page_dirty(page);
1268         unlock_page(page);
1269         return ret;
1270 }
1271 EXPORT_SYMBOL(set_page_dirty_lock);
1272
1273 /*
1274  * Clear a page's dirty flag, while caring for dirty memory accounting.
1275  * Returns true if the page was previously dirty.
1276  *
1277  * This is for preparing to put the page under writeout.  We leave the page
1278  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1279  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1280  * implementation will run either set_page_writeback() or set_page_dirty(),
1281  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1282  * back into sync.
1283  *
1284  * This incoherency between the page's dirty flag and radix-tree tag is
1285  * unfortunate, but it only exists while the page is locked.
1286  */
1287 int clear_page_dirty_for_io(struct page *page)
1288 {
1289         struct address_space *mapping = page_mapping(page);
1290
1291         BUG_ON(!PageLocked(page));
1292
1293         if (mapping && mapping_cap_account_dirty(mapping)) {
1294                 /*
1295                  * Yes, Virginia, this is indeed insane.
1296                  *
1297                  * We use this sequence to make sure that
1298                  *  (a) we account for dirty stats properly
1299                  *  (b) we tell the low-level filesystem to
1300                  *      mark the whole page dirty if it was
1301                  *      dirty in a pagetable. Only to then
1302                  *  (c) clean the page again and return 1 to
1303                  *      cause the writeback.
1304                  *
1305                  * This way we avoid all nasty races with the
1306                  * dirty bit in multiple places and clearing
1307                  * them concurrently from different threads.
1308                  *
1309                  * Note! Normally the "set_page_dirty(page)"
1310                  * has no effect on the actual dirty bit - since
1311                  * that will already usually be set. But we
1312                  * need the side effects, and it can help us
1313                  * avoid races.
1314                  *
1315                  * We basically use the page "master dirty bit"
1316                  * as a serialization point for all the different
1317                  * threads doing their things.
1318                  */
1319                 if (page_mkclean(page))
1320                         set_page_dirty(page);
1321                 /*
1322                  * We carefully synchronise fault handlers against
1323                  * installing a dirty pte and marking the page dirty
1324                  * at this point. We do this by having them hold the
1325                  * page lock at some point after installing their
1326                  * pte, but before marking the page dirty.
1327                  * Pages are always locked coming in here, so we get
1328                  * the desired exclusion. See mm/memory.c:do_wp_page()
1329                  * for more comments.
1330                  */
1331                 if (TestClearPageDirty(page)) {
1332                         dec_zone_page_state(page, NR_FILE_DIRTY);
1333                         dec_bdi_stat(mapping->backing_dev_info,
1334                                         BDI_RECLAIMABLE);
1335                         return 1;
1336                 }
1337                 return 0;
1338         }
1339         return TestClearPageDirty(page);
1340 }
1341 EXPORT_SYMBOL(clear_page_dirty_for_io);
1342
1343 int test_clear_page_writeback(struct page *page)
1344 {
1345         struct address_space *mapping = page_mapping(page);
1346         int ret;
1347
1348         if (mapping) {
1349                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1350                 unsigned long flags;
1351
1352                 spin_lock_irqsave(&mapping->tree_lock, flags);
1353                 ret = TestClearPageWriteback(page);
1354                 if (ret) {
1355                         radix_tree_tag_clear(&mapping->page_tree,
1356                                                 page_index(page),
1357                                                 PAGECACHE_TAG_WRITEBACK);
1358                         if (bdi_cap_account_writeback(bdi)) {
1359                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1360                                 __bdi_writeout_inc(bdi);
1361                         }
1362                 }
1363                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1364         } else {
1365                 ret = TestClearPageWriteback(page);
1366         }
1367         if (ret)
1368                 dec_zone_page_state(page, NR_WRITEBACK);
1369         return ret;
1370 }
1371
1372 int test_set_page_writeback(struct page *page)
1373 {
1374         struct address_space *mapping = page_mapping(page);
1375         int ret;
1376
1377         if (mapping) {
1378                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1379                 unsigned long flags;
1380
1381                 spin_lock_irqsave(&mapping->tree_lock, flags);
1382                 ret = TestSetPageWriteback(page);
1383                 if (!ret) {
1384                         radix_tree_tag_set(&mapping->page_tree,
1385                                                 page_index(page),
1386                                                 PAGECACHE_TAG_WRITEBACK);
1387                         if (bdi_cap_account_writeback(bdi))
1388                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1389                 }
1390                 if (!PageDirty(page))
1391                         radix_tree_tag_clear(&mapping->page_tree,
1392                                                 page_index(page),
1393                                                 PAGECACHE_TAG_DIRTY);
1394                 radix_tree_tag_clear(&mapping->page_tree,
1395                                      page_index(page),
1396                                      PAGECACHE_TAG_TOWRITE);
1397                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1398         } else {
1399                 ret = TestSetPageWriteback(page);
1400         }
1401         if (!ret)
1402                 account_page_writeback(page);
1403         return ret;
1404
1405 }
1406 EXPORT_SYMBOL(test_set_page_writeback);
1407
1408 /*
1409  * Return true if any of the pages in the mapping are marked with the
1410  * passed tag.
1411  */
1412 int mapping_tagged(struct address_space *mapping, int tag)
1413 {
1414         int ret;
1415         rcu_read_lock();
1416         ret = radix_tree_tagged(&mapping->page_tree, tag);
1417         rcu_read_unlock();
1418         return ret;
1419 }
1420 EXPORT_SYMBOL(mapping_tagged);