4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Contains functions related to writing back dirty pages at the
10 * 10Apr2002 Andrew Morton
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/spinlock.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>
40 * Sleep at most 200ms at a time in balance_dirty_pages().
42 #define MAX_PAUSE max(HZ/5, 1)
45 * Estimate write bandwidth at 200ms intervals.
47 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
49 #define RATELIMIT_CALC_SHIFT 10
52 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
53 * will look to see if it needs to force writeback or throttling.
55 static long ratelimit_pages = 32;
58 * When balance_dirty_pages decides that the caller needs to perform some
59 * non-background writeback, this is how many pages it will attempt to write.
60 * It should be somewhat larger than dirtied pages to ensure that reasonably
61 * large amounts of I/O are submitted.
63 static inline long sync_writeback_pages(unsigned long dirtied)
65 if (dirtied < ratelimit_pages)
66 dirtied = ratelimit_pages;
68 return dirtied + dirtied / 2;
71 /* The following parameters are exported via /proc/sys/vm */
74 * Start background writeback (via writeback threads) at this percentage
76 int dirty_background_ratio = 10;
79 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
80 * dirty_background_ratio * the amount of dirtyable memory
82 unsigned long dirty_background_bytes;
85 * free highmem will not be subtracted from the total free memory
86 * for calculating free ratios if vm_highmem_is_dirtyable is true
88 int vm_highmem_is_dirtyable;
91 * The generator of dirty data starts writeback at this percentage
93 int vm_dirty_ratio = 20;
96 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
97 * vm_dirty_ratio * the amount of dirtyable memory
99 unsigned long vm_dirty_bytes;
102 * The interval between `kupdate'-style writebacks
104 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
107 * The longest time for which data is allowed to remain dirty
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
112 * Flag that makes the machine dump writes/reads and block dirtyings.
117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 * a full sync is triggered after this time elapses without any disk activity.
122 EXPORT_SYMBOL(laptop_mode);
124 /* End of sysctl-exported parameters */
126 unsigned long global_dirty_limit;
129 * Scale the writeback cache size proportional to the relative writeout speeds.
131 * We do this by keeping a floating proportion between BDIs, based on page
132 * writeback completions [end_page_writeback()]. Those devices that write out
133 * pages fastest will get the larger share, while the slower will get a smaller
136 * We use page writeout completions because we are interested in getting rid of
137 * dirty pages. Having them written out is the primary goal.
139 * We introduce a concept of time, a period over which we measure these events,
140 * because demand can/will vary over time. The length of this period itself is
141 * measured in page writeback completions.
144 static struct prop_descriptor vm_completions;
145 static struct prop_descriptor vm_dirties;
148 * couple the period to the dirty_ratio:
150 * period/2 ~ roundup_pow_of_two(dirty limit)
152 static int calc_period_shift(void)
154 unsigned long dirty_total;
157 dirty_total = vm_dirty_bytes / PAGE_SIZE;
159 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
161 return 2 + ilog2(dirty_total - 1);
165 * update the period when the dirty threshold changes.
167 static void update_completion_period(void)
169 int shift = calc_period_shift();
170 prop_change_shift(&vm_completions, shift);
171 prop_change_shift(&vm_dirties, shift);
174 int dirty_background_ratio_handler(struct ctl_table *table, int write,
175 void __user *buffer, size_t *lenp,
180 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
181 if (ret == 0 && write)
182 dirty_background_bytes = 0;
186 int dirty_background_bytes_handler(struct ctl_table *table, int write,
187 void __user *buffer, size_t *lenp,
192 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
193 if (ret == 0 && write)
194 dirty_background_ratio = 0;
198 int dirty_ratio_handler(struct ctl_table *table, int write,
199 void __user *buffer, size_t *lenp,
202 int old_ratio = vm_dirty_ratio;
205 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
206 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
207 update_completion_period();
214 int dirty_bytes_handler(struct ctl_table *table, int write,
215 void __user *buffer, size_t *lenp,
218 unsigned long old_bytes = vm_dirty_bytes;
221 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
222 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
223 update_completion_period();
230 * Increment the BDI's writeout completion count and the global writeout
231 * completion count. Called from test_clear_page_writeback().
233 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
235 __inc_bdi_stat(bdi, BDI_WRITTEN);
236 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
240 void bdi_writeout_inc(struct backing_dev_info *bdi)
244 local_irq_save(flags);
245 __bdi_writeout_inc(bdi);
246 local_irq_restore(flags);
248 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
250 void task_dirty_inc(struct task_struct *tsk)
252 prop_inc_single(&vm_dirties, &tsk->dirties);
256 * Obtain an accurate fraction of the BDI's portion.
258 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
259 long *numerator, long *denominator)
261 prop_fraction_percpu(&vm_completions, &bdi->completions,
262 numerator, denominator);
265 static inline void task_dirties_fraction(struct task_struct *tsk,
266 long *numerator, long *denominator)
268 prop_fraction_single(&vm_dirties, &tsk->dirties,
269 numerator, denominator);
273 * task_dirty_limit - scale down dirty throttling threshold for one task
275 * task specific dirty limit:
277 * dirty -= (dirty/8) * p_{t}
279 * To protect light/slow dirtying tasks from heavier/fast ones, we start
280 * throttling individual tasks before reaching the bdi dirty limit.
281 * Relatively low thresholds will be allocated to heavy dirtiers. So when
282 * dirty pages grow large, heavy dirtiers will be throttled first, which will
283 * effectively curb the growth of dirty pages. Light dirtiers with high enough
284 * dirty threshold may never get throttled.
286 #define TASK_LIMIT_FRACTION 8
287 static unsigned long task_dirty_limit(struct task_struct *tsk,
288 unsigned long bdi_dirty)
290 long numerator, denominator;
291 unsigned long dirty = bdi_dirty;
292 u64 inv = dirty / TASK_LIMIT_FRACTION;
294 task_dirties_fraction(tsk, &numerator, &denominator);
296 do_div(inv, denominator);
300 return max(dirty, bdi_dirty/2);
303 /* Minimum limit for any task */
304 static unsigned long task_min_dirty_limit(unsigned long bdi_dirty)
306 return bdi_dirty - bdi_dirty / TASK_LIMIT_FRACTION;
312 static unsigned int bdi_min_ratio;
314 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
318 spin_lock_bh(&bdi_lock);
319 if (min_ratio > bdi->max_ratio) {
322 min_ratio -= bdi->min_ratio;
323 if (bdi_min_ratio + min_ratio < 100) {
324 bdi_min_ratio += min_ratio;
325 bdi->min_ratio += min_ratio;
330 spin_unlock_bh(&bdi_lock);
335 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
342 spin_lock_bh(&bdi_lock);
343 if (bdi->min_ratio > max_ratio) {
346 bdi->max_ratio = max_ratio;
347 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
349 spin_unlock_bh(&bdi_lock);
353 EXPORT_SYMBOL(bdi_set_max_ratio);
356 * Work out the current dirty-memory clamping and background writeout
359 * The main aim here is to lower them aggressively if there is a lot of mapped
360 * memory around. To avoid stressing page reclaim with lots of unreclaimable
361 * pages. It is better to clamp down on writers than to start swapping, and
362 * performing lots of scanning.
364 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
366 * We don't permit the clamping level to fall below 5% - that is getting rather
369 * We make sure that the background writeout level is below the adjusted
373 static unsigned long highmem_dirtyable_memory(unsigned long total)
375 #ifdef CONFIG_HIGHMEM
379 for_each_node_state(node, N_HIGH_MEMORY) {
381 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
383 x += zone_page_state(z, NR_FREE_PAGES) +
384 zone_reclaimable_pages(z);
387 * Make sure that the number of highmem pages is never larger
388 * than the number of the total dirtyable memory. This can only
389 * occur in very strange VM situations but we want to make sure
390 * that this does not occur.
392 return min(x, total);
399 * determine_dirtyable_memory - amount of memory that may be used
401 * Returns the numebr of pages that can currently be freed and used
402 * by the kernel for direct mappings.
404 unsigned long determine_dirtyable_memory(void)
408 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
410 if (!vm_highmem_is_dirtyable)
411 x -= highmem_dirtyable_memory(x);
413 return x + 1; /* Ensure that we never return 0 */
416 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
417 unsigned long bg_thresh)
419 return (thresh + bg_thresh) / 2;
422 static unsigned long hard_dirty_limit(unsigned long thresh)
424 return max(thresh, global_dirty_limit);
428 * global_dirty_limits - background-writeback and dirty-throttling thresholds
430 * Calculate the dirty thresholds based on sysctl parameters
431 * - vm.dirty_background_ratio or vm.dirty_background_bytes
432 * - vm.dirty_ratio or vm.dirty_bytes
433 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
436 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
438 unsigned long background;
440 unsigned long uninitialized_var(available_memory);
441 struct task_struct *tsk;
443 if (!vm_dirty_bytes || !dirty_background_bytes)
444 available_memory = determine_dirtyable_memory();
447 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
449 dirty = (vm_dirty_ratio * available_memory) / 100;
451 if (dirty_background_bytes)
452 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
454 background = (dirty_background_ratio * available_memory) / 100;
456 if (background >= dirty)
457 background = dirty / 2;
459 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
460 background += background / 4;
463 *pbackground = background;
465 trace_global_dirty_state(background, dirty);
469 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
470 * @bdi: the backing_dev_info to query
471 * @dirty: global dirty limit in pages
473 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
474 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
475 * And the "limit" in the name is not seriously taken as hard limit in
476 * balance_dirty_pages().
478 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
479 * - starving fast devices
480 * - piling up dirty pages (that will take long time to sync) on slow devices
482 * The bdi's share of dirty limit will be adapting to its throughput and
483 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
485 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
488 long numerator, denominator;
491 * Calculate this BDI's share of the dirty ratio.
493 bdi_writeout_fraction(bdi, &numerator, &denominator);
495 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
496 bdi_dirty *= numerator;
497 do_div(bdi_dirty, denominator);
499 bdi_dirty += (dirty * bdi->min_ratio) / 100;
500 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
501 bdi_dirty = dirty * bdi->max_ratio / 100;
507 * Dirty position control.
509 * (o) global/bdi setpoints
511 * We want the dirty pages be balanced around the global/bdi setpoints.
512 * When the number of dirty pages is higher/lower than the setpoint, the
513 * dirty position control ratio (and hence task dirty ratelimit) will be
514 * decreased/increased to bring the dirty pages back to the setpoint.
516 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
518 * if (dirty < setpoint) scale up pos_ratio
519 * if (dirty > setpoint) scale down pos_ratio
521 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
522 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
524 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
526 * (o) global control line
530 * | |<===== global dirty control scope ======>|
538 * 1.0 ................................*
544 * 0 +------------.------------------.----------------------*------------->
545 * freerun^ setpoint^ limit^ dirty pages
547 * (o) bdi control line
555 * | * |<=========== span ============>|
556 * 1.0 .......................*
568 * 1/4 ...............................................* * * * * * * * * * * *
572 * 0 +----------------------.-------------------------------.------------->
573 * bdi_setpoint^ x_intercept^
575 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
576 * be smoothly throttled down to normal if it starts high in situations like
577 * - start writing to a slow SD card and a fast disk at the same time. The SD
578 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
579 * - the bdi dirty thresh drops quickly due to change of JBOD workload
581 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
582 unsigned long thresh,
583 unsigned long bg_thresh,
585 unsigned long bdi_thresh,
586 unsigned long bdi_dirty)
588 unsigned long write_bw = bdi->avg_write_bandwidth;
589 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
590 unsigned long limit = hard_dirty_limit(thresh);
591 unsigned long x_intercept;
592 unsigned long setpoint; /* dirty pages' target balance point */
593 unsigned long bdi_setpoint;
595 long long pos_ratio; /* for scaling up/down the rate limit */
598 if (unlikely(dirty >= limit))
605 * f(dirty) := 1.0 + (----------------)
608 * it's a 3rd order polynomial that subjects to
610 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
611 * (2) f(setpoint) = 1.0 => the balance point
612 * (3) f(limit) = 0 => the hard limit
613 * (4) df/dx <= 0 => negative feedback control
614 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
615 * => fast response on large errors; small oscillation near setpoint
617 setpoint = (freerun + limit) / 2;
618 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
619 limit - setpoint + 1);
621 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
622 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
623 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
626 * We have computed basic pos_ratio above based on global situation. If
627 * the bdi is over/under its share of dirty pages, we want to scale
628 * pos_ratio further down/up. That is done by the following mechanism.
634 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
636 * x_intercept - bdi_dirty
637 * := --------------------------
638 * x_intercept - bdi_setpoint
640 * The main bdi control line is a linear function that subjects to
642 * (1) f(bdi_setpoint) = 1.0
643 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
644 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
646 * For single bdi case, the dirty pages are observed to fluctuate
647 * regularly within range
648 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
649 * for various filesystems, where (2) can yield in a reasonable 12.5%
650 * fluctuation range for pos_ratio.
652 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
653 * own size, so move the slope over accordingly and choose a slope that
654 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
656 if (unlikely(bdi_thresh > thresh))
659 * scale global setpoint to bdi's:
660 * bdi_setpoint = setpoint * bdi_thresh / thresh
662 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
663 bdi_setpoint = setpoint * (u64)x >> 16;
665 * Use span=(8*write_bw) in single bdi case as indicated by
666 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
668 * bdi_thresh thresh - bdi_thresh
669 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
672 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
673 x_intercept = bdi_setpoint + span;
675 if (bdi_dirty < x_intercept - span / 4) {
676 pos_ratio *= x_intercept - bdi_dirty;
677 do_div(pos_ratio, x_intercept - bdi_setpoint + 1);
684 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
685 unsigned long elapsed,
686 unsigned long written)
688 const unsigned long period = roundup_pow_of_two(3 * HZ);
689 unsigned long avg = bdi->avg_write_bandwidth;
690 unsigned long old = bdi->write_bandwidth;
694 * bw = written * HZ / elapsed
696 * bw * elapsed + write_bandwidth * (period - elapsed)
697 * write_bandwidth = ---------------------------------------------------
700 bw = written - bdi->written_stamp;
702 if (unlikely(elapsed > period)) {
707 bw += (u64)bdi->write_bandwidth * (period - elapsed);
708 bw >>= ilog2(period);
711 * one more level of smoothing, for filtering out sudden spikes
713 if (avg > old && old >= (unsigned long)bw)
714 avg -= (avg - old) >> 3;
716 if (avg < old && old <= (unsigned long)bw)
717 avg += (old - avg) >> 3;
720 bdi->write_bandwidth = bw;
721 bdi->avg_write_bandwidth = avg;
725 * The global dirtyable memory and dirty threshold could be suddenly knocked
726 * down by a large amount (eg. on the startup of KVM in a swapless system).
727 * This may throw the system into deep dirty exceeded state and throttle
728 * heavy/light dirtiers alike. To retain good responsiveness, maintain
729 * global_dirty_limit for tracking slowly down to the knocked down dirty
732 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
734 unsigned long limit = global_dirty_limit;
737 * Follow up in one step.
739 if (limit < thresh) {
745 * Follow down slowly. Use the higher one as the target, because thresh
746 * may drop below dirty. This is exactly the reason to introduce
747 * global_dirty_limit which is guaranteed to lie above the dirty pages.
749 thresh = max(thresh, dirty);
750 if (limit > thresh) {
751 limit -= (limit - thresh) >> 5;
756 global_dirty_limit = limit;
759 static void global_update_bandwidth(unsigned long thresh,
763 static DEFINE_SPINLOCK(dirty_lock);
764 static unsigned long update_time;
767 * check locklessly first to optimize away locking for the most time
769 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
772 spin_lock(&dirty_lock);
773 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
774 update_dirty_limit(thresh, dirty);
777 spin_unlock(&dirty_lock);
780 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
781 unsigned long thresh,
783 unsigned long bdi_thresh,
784 unsigned long bdi_dirty,
785 unsigned long start_time)
787 unsigned long now = jiffies;
788 unsigned long elapsed = now - bdi->bw_time_stamp;
789 unsigned long written;
792 * rate-limit, only update once every 200ms.
794 if (elapsed < BANDWIDTH_INTERVAL)
797 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
800 * Skip quiet periods when disk bandwidth is under-utilized.
801 * (at least 1s idle time between two flusher runs)
803 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
807 global_update_bandwidth(thresh, dirty, now);
809 bdi_update_write_bandwidth(bdi, elapsed, written);
812 bdi->written_stamp = written;
813 bdi->bw_time_stamp = now;
816 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
817 unsigned long thresh,
819 unsigned long bdi_thresh,
820 unsigned long bdi_dirty,
821 unsigned long start_time)
823 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
825 spin_lock(&bdi->wb.list_lock);
826 __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty,
828 spin_unlock(&bdi->wb.list_lock);
832 * balance_dirty_pages() must be called by processes which are generating dirty
833 * data. It looks at the number of dirty pages in the machine and will force
834 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
835 * If we're over `background_thresh' then the writeback threads are woken to
836 * perform some writeout.
838 static void balance_dirty_pages(struct address_space *mapping,
839 unsigned long write_chunk)
841 unsigned long nr_reclaimable, bdi_nr_reclaimable;
842 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
843 unsigned long bdi_dirty;
844 unsigned long freerun;
845 unsigned long background_thresh;
846 unsigned long dirty_thresh;
847 unsigned long bdi_thresh;
848 unsigned long task_bdi_thresh;
849 unsigned long min_task_bdi_thresh;
850 unsigned long pages_written = 0;
851 unsigned long pause = 1;
852 bool dirty_exceeded = false;
853 bool clear_dirty_exceeded = true;
854 struct backing_dev_info *bdi = mapping->backing_dev_info;
855 unsigned long start_time = jiffies;
858 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
859 global_page_state(NR_UNSTABLE_NFS);
860 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
862 global_dirty_limits(&background_thresh, &dirty_thresh);
865 * Throttle it only when the background writeback cannot
866 * catch-up. This avoids (excessively) small writeouts
867 * when the bdi limits are ramping up.
869 freerun = dirty_freerun_ceiling(dirty_thresh,
871 if (nr_dirty <= freerun)
874 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
875 min_task_bdi_thresh = task_min_dirty_limit(bdi_thresh);
876 task_bdi_thresh = task_dirty_limit(current, bdi_thresh);
879 * In order to avoid the stacked BDI deadlock we need
880 * to ensure we accurately count the 'dirty' pages when
881 * the threshold is low.
883 * Otherwise it would be possible to get thresh+n pages
884 * reported dirty, even though there are thresh-m pages
885 * actually dirty; with m+n sitting in the percpu
888 if (task_bdi_thresh < 2 * bdi_stat_error(bdi)) {
889 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
890 bdi_dirty = bdi_nr_reclaimable +
891 bdi_stat_sum(bdi, BDI_WRITEBACK);
893 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
894 bdi_dirty = bdi_nr_reclaimable +
895 bdi_stat(bdi, BDI_WRITEBACK);
899 * The bdi thresh is somehow "soft" limit derived from the
900 * global "hard" limit. The former helps to prevent heavy IO
901 * bdi or process from holding back light ones; The latter is
902 * the last resort safeguard.
904 dirty_exceeded = (bdi_dirty > task_bdi_thresh) ||
905 (nr_dirty > dirty_thresh);
906 clear_dirty_exceeded = (bdi_dirty <= min_task_bdi_thresh) &&
907 (nr_dirty <= dirty_thresh);
912 if (!bdi->dirty_exceeded)
913 bdi->dirty_exceeded = 1;
915 bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty,
916 bdi_thresh, bdi_dirty, start_time);
918 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
919 * Unstable writes are a feature of certain networked
920 * filesystems (i.e. NFS) in which data may have been
921 * written to the server's write cache, but has not yet
922 * been flushed to permanent storage.
923 * Only move pages to writeback if this bdi is over its
924 * threshold otherwise wait until the disk writes catch
927 trace_balance_dirty_start(bdi);
928 if (bdi_nr_reclaimable > task_bdi_thresh) {
929 pages_written += writeback_inodes_wb(&bdi->wb,
931 trace_balance_dirty_written(bdi, pages_written);
932 if (pages_written >= write_chunk)
933 break; /* We've done our duty */
935 __set_current_state(TASK_UNINTERRUPTIBLE);
936 io_schedule_timeout(pause);
937 trace_balance_dirty_wait(bdi);
939 dirty_thresh = hard_dirty_limit(dirty_thresh);
941 * max-pause area. If dirty exceeded but still within this
942 * area, no need to sleep for more than 200ms: (a) 8 pages per
943 * 200ms is typically more than enough to curb heavy dirtiers;
944 * (b) the pause time limit makes the dirtiers more responsive.
946 if (nr_dirty < dirty_thresh &&
947 bdi_dirty < (task_bdi_thresh + bdi_thresh) / 2 &&
948 time_after(jiffies, start_time + MAX_PAUSE))
952 * Increase the delay for each loop, up to our previous
953 * default of taking a 100ms nap.
960 /* Clear dirty_exceeded flag only when no task can exceed the limit */
961 if (clear_dirty_exceeded && bdi->dirty_exceeded)
962 bdi->dirty_exceeded = 0;
964 if (writeback_in_progress(bdi))
968 * In laptop mode, we wait until hitting the higher threshold before
969 * starting background writeout, and then write out all the way down
970 * to the lower threshold. So slow writers cause minimal disk activity.
972 * In normal mode, we start background writeout at the lower
973 * background_thresh, to keep the amount of dirty memory low.
975 if ((laptop_mode && pages_written) ||
976 (!laptop_mode && (nr_reclaimable > background_thresh)))
977 bdi_start_background_writeback(bdi);
980 void set_page_dirty_balance(struct page *page, int page_mkwrite)
982 if (set_page_dirty(page) || page_mkwrite) {
983 struct address_space *mapping = page_mapping(page);
986 balance_dirty_pages_ratelimited(mapping);
990 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
993 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
994 * @mapping: address_space which was dirtied
995 * @nr_pages_dirtied: number of pages which the caller has just dirtied
997 * Processes which are dirtying memory should call in here once for each page
998 * which was newly dirtied. The function will periodically check the system's
999 * dirty state and will initiate writeback if needed.
1001 * On really big machines, get_writeback_state is expensive, so try to avoid
1002 * calling it too often (ratelimiting). But once we're over the dirty memory
1003 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1004 * from overshooting the limit by (ratelimit_pages) each.
1006 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1007 unsigned long nr_pages_dirtied)
1009 struct backing_dev_info *bdi = mapping->backing_dev_info;
1010 unsigned long ratelimit;
1013 if (!bdi_cap_account_dirty(bdi))
1016 ratelimit = ratelimit_pages;
1017 if (mapping->backing_dev_info->dirty_exceeded)
1021 * Check the rate limiting. Also, we do not want to throttle real-time
1022 * tasks in balance_dirty_pages(). Period.
1025 p = &__get_cpu_var(bdp_ratelimits);
1026 *p += nr_pages_dirtied;
1027 if (unlikely(*p >= ratelimit)) {
1028 ratelimit = sync_writeback_pages(*p);
1031 balance_dirty_pages(mapping, ratelimit);
1036 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1038 void throttle_vm_writeout(gfp_t gfp_mask)
1040 unsigned long background_thresh;
1041 unsigned long dirty_thresh;
1044 global_dirty_limits(&background_thresh, &dirty_thresh);
1047 * Boost the allowable dirty threshold a bit for page
1048 * allocators so they don't get DoS'ed by heavy writers
1050 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1052 if (global_page_state(NR_UNSTABLE_NFS) +
1053 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1055 congestion_wait(BLK_RW_ASYNC, HZ/10);
1058 * The caller might hold locks which can prevent IO completion
1059 * or progress in the filesystem. So we cannot just sit here
1060 * waiting for IO to complete.
1062 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1068 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1070 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1071 void __user *buffer, size_t *length, loff_t *ppos)
1073 proc_dointvec(table, write, buffer, length, ppos);
1074 bdi_arm_supers_timer();
1079 void laptop_mode_timer_fn(unsigned long data)
1081 struct request_queue *q = (struct request_queue *)data;
1082 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1083 global_page_state(NR_UNSTABLE_NFS);
1086 * We want to write everything out, not just down to the dirty
1089 if (bdi_has_dirty_io(&q->backing_dev_info))
1090 bdi_start_writeback(&q->backing_dev_info, nr_pages);
1094 * We've spun up the disk and we're in laptop mode: schedule writeback
1095 * of all dirty data a few seconds from now. If the flush is already scheduled
1096 * then push it back - the user is still using the disk.
1098 void laptop_io_completion(struct backing_dev_info *info)
1100 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1104 * We're in laptop mode and we've just synced. The sync's writes will have
1105 * caused another writeback to be scheduled by laptop_io_completion.
1106 * Nothing needs to be written back anymore, so we unschedule the writeback.
1108 void laptop_sync_completion(void)
1110 struct backing_dev_info *bdi;
1114 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1115 del_timer(&bdi->laptop_mode_wb_timer);
1122 * If ratelimit_pages is too high then we can get into dirty-data overload
1123 * if a large number of processes all perform writes at the same time.
1124 * If it is too low then SMP machines will call the (expensive)
1125 * get_writeback_state too often.
1127 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1128 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1129 * thresholds before writeback cuts in.
1131 * But the limit should not be set too high. Because it also controls the
1132 * amount of memory which the balance_dirty_pages() caller has to write back.
1133 * If this is too large then the caller will block on the IO queue all the
1134 * time. So limit it to four megabytes - the balance_dirty_pages() caller
1135 * will write six megabyte chunks, max.
1138 void writeback_set_ratelimit(void)
1140 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
1141 if (ratelimit_pages < 16)
1142 ratelimit_pages = 16;
1143 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
1144 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
1147 static int __cpuinit
1148 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1150 writeback_set_ratelimit();
1154 static struct notifier_block __cpuinitdata ratelimit_nb = {
1155 .notifier_call = ratelimit_handler,
1160 * Called early on to tune the page writeback dirty limits.
1162 * We used to scale dirty pages according to how total memory
1163 * related to pages that could be allocated for buffers (by
1164 * comparing nr_free_buffer_pages() to vm_total_pages.
1166 * However, that was when we used "dirty_ratio" to scale with
1167 * all memory, and we don't do that any more. "dirty_ratio"
1168 * is now applied to total non-HIGHPAGE memory (by subtracting
1169 * totalhigh_pages from vm_total_pages), and as such we can't
1170 * get into the old insane situation any more where we had
1171 * large amounts of dirty pages compared to a small amount of
1172 * non-HIGHMEM memory.
1174 * But we might still want to scale the dirty_ratio by how
1175 * much memory the box has..
1177 void __init page_writeback_init(void)
1181 writeback_set_ratelimit();
1182 register_cpu_notifier(&ratelimit_nb);
1184 shift = calc_period_shift();
1185 prop_descriptor_init(&vm_completions, shift);
1186 prop_descriptor_init(&vm_dirties, shift);
1190 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1191 * @mapping: address space structure to write
1192 * @start: starting page index
1193 * @end: ending page index (inclusive)
1195 * This function scans the page range from @start to @end (inclusive) and tags
1196 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1197 * that write_cache_pages (or whoever calls this function) will then use
1198 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1199 * used to avoid livelocking of writeback by a process steadily creating new
1200 * dirty pages in the file (thus it is important for this function to be quick
1201 * so that it can tag pages faster than a dirtying process can create them).
1204 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1206 void tag_pages_for_writeback(struct address_space *mapping,
1207 pgoff_t start, pgoff_t end)
1209 #define WRITEBACK_TAG_BATCH 4096
1210 unsigned long tagged;
1213 spin_lock_irq(&mapping->tree_lock);
1214 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1215 &start, end, WRITEBACK_TAG_BATCH,
1216 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1217 spin_unlock_irq(&mapping->tree_lock);
1218 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1220 /* We check 'start' to handle wrapping when end == ~0UL */
1221 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1223 EXPORT_SYMBOL(tag_pages_for_writeback);
1226 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1227 * @mapping: address space structure to write
1228 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1229 * @writepage: function called for each page
1230 * @data: data passed to writepage function
1232 * If a page is already under I/O, write_cache_pages() skips it, even
1233 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1234 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1235 * and msync() need to guarantee that all the data which was dirty at the time
1236 * the call was made get new I/O started against them. If wbc->sync_mode is
1237 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1238 * existing IO to complete.
1240 * To avoid livelocks (when other process dirties new pages), we first tag
1241 * pages which should be written back with TOWRITE tag and only then start
1242 * writing them. For data-integrity sync we have to be careful so that we do
1243 * not miss some pages (e.g., because some other process has cleared TOWRITE
1244 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1245 * by the process clearing the DIRTY tag (and submitting the page for IO).
1247 int write_cache_pages(struct address_space *mapping,
1248 struct writeback_control *wbc, writepage_t writepage,
1253 struct pagevec pvec;
1255 pgoff_t uninitialized_var(writeback_index);
1257 pgoff_t end; /* Inclusive */
1260 int range_whole = 0;
1263 pagevec_init(&pvec, 0);
1264 if (wbc->range_cyclic) {
1265 writeback_index = mapping->writeback_index; /* prev offset */
1266 index = writeback_index;
1273 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1274 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1275 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1277 cycled = 1; /* ignore range_cyclic tests */
1279 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1280 tag = PAGECACHE_TAG_TOWRITE;
1282 tag = PAGECACHE_TAG_DIRTY;
1284 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1285 tag_pages_for_writeback(mapping, index, end);
1287 while (!done && (index <= end)) {
1290 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1291 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1295 for (i = 0; i < nr_pages; i++) {
1296 struct page *page = pvec.pages[i];
1299 * At this point, the page may be truncated or
1300 * invalidated (changing page->mapping to NULL), or
1301 * even swizzled back from swapper_space to tmpfs file
1302 * mapping. However, page->index will not change
1303 * because we have a reference on the page.
1305 if (page->index > end) {
1307 * can't be range_cyclic (1st pass) because
1308 * end == -1 in that case.
1314 done_index = page->index;
1319 * Page truncated or invalidated. We can freely skip it
1320 * then, even for data integrity operations: the page
1321 * has disappeared concurrently, so there could be no
1322 * real expectation of this data interity operation
1323 * even if there is now a new, dirty page at the same
1324 * pagecache address.
1326 if (unlikely(page->mapping != mapping)) {
1332 if (!PageDirty(page)) {
1333 /* someone wrote it for us */
1334 goto continue_unlock;
1337 if (PageWriteback(page)) {
1338 if (wbc->sync_mode != WB_SYNC_NONE)
1339 wait_on_page_writeback(page);
1341 goto continue_unlock;
1344 BUG_ON(PageWriteback(page));
1345 if (!clear_page_dirty_for_io(page))
1346 goto continue_unlock;
1348 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1349 ret = (*writepage)(page, wbc, data);
1350 if (unlikely(ret)) {
1351 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1356 * done_index is set past this page,
1357 * so media errors will not choke
1358 * background writeout for the entire
1359 * file. This has consequences for
1360 * range_cyclic semantics (ie. it may
1361 * not be suitable for data integrity
1364 done_index = page->index + 1;
1371 * We stop writing back only if we are not doing
1372 * integrity sync. In case of integrity sync we have to
1373 * keep going until we have written all the pages
1374 * we tagged for writeback prior to entering this loop.
1376 if (--wbc->nr_to_write <= 0 &&
1377 wbc->sync_mode == WB_SYNC_NONE) {
1382 pagevec_release(&pvec);
1385 if (!cycled && !done) {
1388 * We hit the last page and there is more work to be done: wrap
1389 * back to the start of the file
1393 end = writeback_index - 1;
1396 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1397 mapping->writeback_index = done_index;
1401 EXPORT_SYMBOL(write_cache_pages);
1404 * Function used by generic_writepages to call the real writepage
1405 * function and set the mapping flags on error
1407 static int __writepage(struct page *page, struct writeback_control *wbc,
1410 struct address_space *mapping = data;
1411 int ret = mapping->a_ops->writepage(page, wbc);
1412 mapping_set_error(mapping, ret);
1417 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1418 * @mapping: address space structure to write
1419 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1421 * This is a library function, which implements the writepages()
1422 * address_space_operation.
1424 int generic_writepages(struct address_space *mapping,
1425 struct writeback_control *wbc)
1427 struct blk_plug plug;
1430 /* deal with chardevs and other special file */
1431 if (!mapping->a_ops->writepage)
1434 blk_start_plug(&plug);
1435 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1436 blk_finish_plug(&plug);
1440 EXPORT_SYMBOL(generic_writepages);
1442 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1446 if (wbc->nr_to_write <= 0)
1448 if (mapping->a_ops->writepages)
1449 ret = mapping->a_ops->writepages(mapping, wbc);
1451 ret = generic_writepages(mapping, wbc);
1456 * write_one_page - write out a single page and optionally wait on I/O
1457 * @page: the page to write
1458 * @wait: if true, wait on writeout
1460 * The page must be locked by the caller and will be unlocked upon return.
1462 * write_one_page() returns a negative error code if I/O failed.
1464 int write_one_page(struct page *page, int wait)
1466 struct address_space *mapping = page->mapping;
1468 struct writeback_control wbc = {
1469 .sync_mode = WB_SYNC_ALL,
1473 BUG_ON(!PageLocked(page));
1476 wait_on_page_writeback(page);
1478 if (clear_page_dirty_for_io(page)) {
1479 page_cache_get(page);
1480 ret = mapping->a_ops->writepage(page, &wbc);
1481 if (ret == 0 && wait) {
1482 wait_on_page_writeback(page);
1483 if (PageError(page))
1486 page_cache_release(page);
1492 EXPORT_SYMBOL(write_one_page);
1495 * For address_spaces which do not use buffers nor write back.
1497 int __set_page_dirty_no_writeback(struct page *page)
1499 if (!PageDirty(page))
1500 return !TestSetPageDirty(page);
1505 * Helper function for set_page_dirty family.
1506 * NOTE: This relies on being atomic wrt interrupts.
1508 void account_page_dirtied(struct page *page, struct address_space *mapping)
1510 if (mapping_cap_account_dirty(mapping)) {
1511 __inc_zone_page_state(page, NR_FILE_DIRTY);
1512 __inc_zone_page_state(page, NR_DIRTIED);
1513 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1514 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1515 task_dirty_inc(current);
1516 task_io_account_write(PAGE_CACHE_SIZE);
1519 EXPORT_SYMBOL(account_page_dirtied);
1522 * Helper function for set_page_writeback family.
1523 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1526 void account_page_writeback(struct page *page)
1528 inc_zone_page_state(page, NR_WRITEBACK);
1530 EXPORT_SYMBOL(account_page_writeback);
1533 * For address_spaces which do not use buffers. Just tag the page as dirty in
1536 * This is also used when a single buffer is being dirtied: we want to set the
1537 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1538 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1540 * Most callers have locked the page, which pins the address_space in memory.
1541 * But zap_pte_range() does not lock the page, however in that case the
1542 * mapping is pinned by the vma's ->vm_file reference.
1544 * We take care to handle the case where the page was truncated from the
1545 * mapping by re-checking page_mapping() inside tree_lock.
1547 int __set_page_dirty_nobuffers(struct page *page)
1549 if (!TestSetPageDirty(page)) {
1550 struct address_space *mapping = page_mapping(page);
1551 struct address_space *mapping2;
1556 spin_lock_irq(&mapping->tree_lock);
1557 mapping2 = page_mapping(page);
1558 if (mapping2) { /* Race with truncate? */
1559 BUG_ON(mapping2 != mapping);
1560 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1561 account_page_dirtied(page, mapping);
1562 radix_tree_tag_set(&mapping->page_tree,
1563 page_index(page), PAGECACHE_TAG_DIRTY);
1565 spin_unlock_irq(&mapping->tree_lock);
1566 if (mapping->host) {
1567 /* !PageAnon && !swapper_space */
1568 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1574 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1577 * When a writepage implementation decides that it doesn't want to write this
1578 * page for some reason, it should redirty the locked page via
1579 * redirty_page_for_writepage() and it should then unlock the page and return 0
1581 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1583 wbc->pages_skipped++;
1584 return __set_page_dirty_nobuffers(page);
1586 EXPORT_SYMBOL(redirty_page_for_writepage);
1591 * For pages with a mapping this should be done under the page lock
1592 * for the benefit of asynchronous memory errors who prefer a consistent
1593 * dirty state. This rule can be broken in some special cases,
1594 * but should be better not to.
1596 * If the mapping doesn't provide a set_page_dirty a_op, then
1597 * just fall through and assume that it wants buffer_heads.
1599 int set_page_dirty(struct page *page)
1601 struct address_space *mapping = page_mapping(page);
1603 if (likely(mapping)) {
1604 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1606 * readahead/lru_deactivate_page could remain
1607 * PG_readahead/PG_reclaim due to race with end_page_writeback
1608 * About readahead, if the page is written, the flags would be
1609 * reset. So no problem.
1610 * About lru_deactivate_page, if the page is redirty, the flag
1611 * will be reset. So no problem. but if the page is used by readahead
1612 * it will confuse readahead and make it restart the size rampup
1613 * process. But it's a trivial problem.
1615 ClearPageReclaim(page);
1618 spd = __set_page_dirty_buffers;
1620 return (*spd)(page);
1622 if (!PageDirty(page)) {
1623 if (!TestSetPageDirty(page))
1628 EXPORT_SYMBOL(set_page_dirty);
1631 * set_page_dirty() is racy if the caller has no reference against
1632 * page->mapping->host, and if the page is unlocked. This is because another
1633 * CPU could truncate the page off the mapping and then free the mapping.
1635 * Usually, the page _is_ locked, or the caller is a user-space process which
1636 * holds a reference on the inode by having an open file.
1638 * In other cases, the page should be locked before running set_page_dirty().
1640 int set_page_dirty_lock(struct page *page)
1645 ret = set_page_dirty(page);
1649 EXPORT_SYMBOL(set_page_dirty_lock);
1652 * Clear a page's dirty flag, while caring for dirty memory accounting.
1653 * Returns true if the page was previously dirty.
1655 * This is for preparing to put the page under writeout. We leave the page
1656 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1657 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1658 * implementation will run either set_page_writeback() or set_page_dirty(),
1659 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1662 * This incoherency between the page's dirty flag and radix-tree tag is
1663 * unfortunate, but it only exists while the page is locked.
1665 int clear_page_dirty_for_io(struct page *page)
1667 struct address_space *mapping = page_mapping(page);
1669 BUG_ON(!PageLocked(page));
1671 if (mapping && mapping_cap_account_dirty(mapping)) {
1673 * Yes, Virginia, this is indeed insane.
1675 * We use this sequence to make sure that
1676 * (a) we account for dirty stats properly
1677 * (b) we tell the low-level filesystem to
1678 * mark the whole page dirty if it was
1679 * dirty in a pagetable. Only to then
1680 * (c) clean the page again and return 1 to
1681 * cause the writeback.
1683 * This way we avoid all nasty races with the
1684 * dirty bit in multiple places and clearing
1685 * them concurrently from different threads.
1687 * Note! Normally the "set_page_dirty(page)"
1688 * has no effect on the actual dirty bit - since
1689 * that will already usually be set. But we
1690 * need the side effects, and it can help us
1693 * We basically use the page "master dirty bit"
1694 * as a serialization point for all the different
1695 * threads doing their things.
1697 if (page_mkclean(page))
1698 set_page_dirty(page);
1700 * We carefully synchronise fault handlers against
1701 * installing a dirty pte and marking the page dirty
1702 * at this point. We do this by having them hold the
1703 * page lock at some point after installing their
1704 * pte, but before marking the page dirty.
1705 * Pages are always locked coming in here, so we get
1706 * the desired exclusion. See mm/memory.c:do_wp_page()
1707 * for more comments.
1709 if (TestClearPageDirty(page)) {
1710 dec_zone_page_state(page, NR_FILE_DIRTY);
1711 dec_bdi_stat(mapping->backing_dev_info,
1717 return TestClearPageDirty(page);
1719 EXPORT_SYMBOL(clear_page_dirty_for_io);
1721 int test_clear_page_writeback(struct page *page)
1723 struct address_space *mapping = page_mapping(page);
1727 struct backing_dev_info *bdi = mapping->backing_dev_info;
1728 unsigned long flags;
1730 spin_lock_irqsave(&mapping->tree_lock, flags);
1731 ret = TestClearPageWriteback(page);
1733 radix_tree_tag_clear(&mapping->page_tree,
1735 PAGECACHE_TAG_WRITEBACK);
1736 if (bdi_cap_account_writeback(bdi)) {
1737 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1738 __bdi_writeout_inc(bdi);
1741 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1743 ret = TestClearPageWriteback(page);
1746 dec_zone_page_state(page, NR_WRITEBACK);
1747 inc_zone_page_state(page, NR_WRITTEN);
1752 int test_set_page_writeback(struct page *page)
1754 struct address_space *mapping = page_mapping(page);
1758 struct backing_dev_info *bdi = mapping->backing_dev_info;
1759 unsigned long flags;
1761 spin_lock_irqsave(&mapping->tree_lock, flags);
1762 ret = TestSetPageWriteback(page);
1764 radix_tree_tag_set(&mapping->page_tree,
1766 PAGECACHE_TAG_WRITEBACK);
1767 if (bdi_cap_account_writeback(bdi))
1768 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1770 if (!PageDirty(page))
1771 radix_tree_tag_clear(&mapping->page_tree,
1773 PAGECACHE_TAG_DIRTY);
1774 radix_tree_tag_clear(&mapping->page_tree,
1776 PAGECACHE_TAG_TOWRITE);
1777 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1779 ret = TestSetPageWriteback(page);
1782 account_page_writeback(page);
1786 EXPORT_SYMBOL(test_set_page_writeback);
1789 * Return true if any of the pages in the mapping are marked with the
1792 int mapping_tagged(struct address_space *mapping, int tag)
1794 return radix_tree_tagged(&mapping->page_tree, tag);
1796 EXPORT_SYMBOL(mapping_tagged);