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)
50 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
51 * will look to see if it needs to force writeback or throttling.
53 static long ratelimit_pages = 32;
56 * When balance_dirty_pages decides that the caller needs to perform some
57 * non-background writeback, this is how many pages it will attempt to write.
58 * It should be somewhat larger than dirtied pages to ensure that reasonably
59 * large amounts of I/O are submitted.
61 static inline long sync_writeback_pages(unsigned long dirtied)
63 if (dirtied < ratelimit_pages)
64 dirtied = ratelimit_pages;
66 return dirtied + dirtied / 2;
69 /* The following parameters are exported via /proc/sys/vm */
72 * Start background writeback (via writeback threads) at this percentage
74 int dirty_background_ratio = 10;
77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
80 unsigned long dirty_background_bytes;
83 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
86 int vm_highmem_is_dirtyable;
89 * The generator of dirty data starts writeback at this percentage
91 int vm_dirty_ratio = 20;
94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
97 unsigned long vm_dirty_bytes;
100 * The interval between `kupdate'-style writebacks
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
105 * The longest time for which data is allowed to remain dirty
107 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110 * Flag that makes the machine dump writes/reads and block dirtyings.
115 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
116 * a full sync is triggered after this time elapses without any disk activity.
120 EXPORT_SYMBOL(laptop_mode);
122 /* End of sysctl-exported parameters */
124 unsigned long global_dirty_limit;
127 * Scale the writeback cache size proportional to the relative writeout speeds.
129 * We do this by keeping a floating proportion between BDIs, based on page
130 * writeback completions [end_page_writeback()]. Those devices that write out
131 * pages fastest will get the larger share, while the slower will get a smaller
134 * We use page writeout completions because we are interested in getting rid of
135 * dirty pages. Having them written out is the primary goal.
137 * We introduce a concept of time, a period over which we measure these events,
138 * because demand can/will vary over time. The length of this period itself is
139 * measured in page writeback completions.
142 static struct prop_descriptor vm_completions;
143 static struct prop_descriptor vm_dirties;
146 * couple the period to the dirty_ratio:
148 * period/2 ~ roundup_pow_of_two(dirty limit)
150 static int calc_period_shift(void)
152 unsigned long dirty_total;
155 dirty_total = vm_dirty_bytes / PAGE_SIZE;
157 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
159 return 2 + ilog2(dirty_total - 1);
163 * update the period when the dirty threshold changes.
165 static void update_completion_period(void)
167 int shift = calc_period_shift();
168 prop_change_shift(&vm_completions, shift);
169 prop_change_shift(&vm_dirties, shift);
172 int dirty_background_ratio_handler(struct ctl_table *table, int write,
173 void __user *buffer, size_t *lenp,
178 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
179 if (ret == 0 && write)
180 dirty_background_bytes = 0;
184 int dirty_background_bytes_handler(struct ctl_table *table, int write,
185 void __user *buffer, size_t *lenp,
190 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
191 if (ret == 0 && write)
192 dirty_background_ratio = 0;
196 int dirty_ratio_handler(struct ctl_table *table, int write,
197 void __user *buffer, size_t *lenp,
200 int old_ratio = vm_dirty_ratio;
203 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
204 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
205 update_completion_period();
212 int dirty_bytes_handler(struct ctl_table *table, int write,
213 void __user *buffer, size_t *lenp,
216 unsigned long old_bytes = vm_dirty_bytes;
219 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
220 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
221 update_completion_period();
228 * Increment the BDI's writeout completion count and the global writeout
229 * completion count. Called from test_clear_page_writeback().
231 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
233 __inc_bdi_stat(bdi, BDI_WRITTEN);
234 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
238 void bdi_writeout_inc(struct backing_dev_info *bdi)
242 local_irq_save(flags);
243 __bdi_writeout_inc(bdi);
244 local_irq_restore(flags);
246 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
248 void task_dirty_inc(struct task_struct *tsk)
250 prop_inc_single(&vm_dirties, &tsk->dirties);
254 * Obtain an accurate fraction of the BDI's portion.
256 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
257 long *numerator, long *denominator)
259 prop_fraction_percpu(&vm_completions, &bdi->completions,
260 numerator, denominator);
263 static inline void task_dirties_fraction(struct task_struct *tsk,
264 long *numerator, long *denominator)
266 prop_fraction_single(&vm_dirties, &tsk->dirties,
267 numerator, denominator);
271 * task_dirty_limit - scale down dirty throttling threshold for one task
273 * task specific dirty limit:
275 * dirty -= (dirty/8) * p_{t}
277 * To protect light/slow dirtying tasks from heavier/fast ones, we start
278 * throttling individual tasks before reaching the bdi dirty limit.
279 * Relatively low thresholds will be allocated to heavy dirtiers. So when
280 * dirty pages grow large, heavy dirtiers will be throttled first, which will
281 * effectively curb the growth of dirty pages. Light dirtiers with high enough
282 * dirty threshold may never get throttled.
284 static unsigned long task_dirty_limit(struct task_struct *tsk,
285 unsigned long bdi_dirty)
287 long numerator, denominator;
288 unsigned long dirty = bdi_dirty;
289 u64 inv = dirty >> 3;
291 task_dirties_fraction(tsk, &numerator, &denominator);
293 do_div(inv, denominator);
297 return max(dirty, bdi_dirty/2);
303 static unsigned int bdi_min_ratio;
305 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
309 spin_lock_bh(&bdi_lock);
310 if (min_ratio > bdi->max_ratio) {
313 min_ratio -= bdi->min_ratio;
314 if (bdi_min_ratio + min_ratio < 100) {
315 bdi_min_ratio += min_ratio;
316 bdi->min_ratio += min_ratio;
321 spin_unlock_bh(&bdi_lock);
326 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
333 spin_lock_bh(&bdi_lock);
334 if (bdi->min_ratio > max_ratio) {
337 bdi->max_ratio = max_ratio;
338 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
340 spin_unlock_bh(&bdi_lock);
344 EXPORT_SYMBOL(bdi_set_max_ratio);
347 * Work out the current dirty-memory clamping and background writeout
350 * The main aim here is to lower them aggressively if there is a lot of mapped
351 * memory around. To avoid stressing page reclaim with lots of unreclaimable
352 * pages. It is better to clamp down on writers than to start swapping, and
353 * performing lots of scanning.
355 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
357 * We don't permit the clamping level to fall below 5% - that is getting rather
360 * We make sure that the background writeout level is below the adjusted
364 static unsigned long highmem_dirtyable_memory(unsigned long total)
366 #ifdef CONFIG_HIGHMEM
370 for_each_node_state(node, N_HIGH_MEMORY) {
372 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
374 x += zone_page_state(z, NR_FREE_PAGES) +
375 zone_reclaimable_pages(z);
378 * Make sure that the number of highmem pages is never larger
379 * than the number of the total dirtyable memory. This can only
380 * occur in very strange VM situations but we want to make sure
381 * that this does not occur.
383 return min(x, total);
390 * determine_dirtyable_memory - amount of memory that may be used
392 * Returns the numebr of pages that can currently be freed and used
393 * by the kernel for direct mappings.
395 unsigned long determine_dirtyable_memory(void)
399 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
401 if (!vm_highmem_is_dirtyable)
402 x -= highmem_dirtyable_memory(x);
404 return x + 1; /* Ensure that we never return 0 */
407 static unsigned long hard_dirty_limit(unsigned long thresh)
409 return max(thresh, global_dirty_limit);
413 * global_dirty_limits - background-writeback and dirty-throttling thresholds
415 * Calculate the dirty thresholds based on sysctl parameters
416 * - vm.dirty_background_ratio or vm.dirty_background_bytes
417 * - vm.dirty_ratio or vm.dirty_bytes
418 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
421 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
423 unsigned long background;
425 unsigned long uninitialized_var(available_memory);
426 struct task_struct *tsk;
428 if (!vm_dirty_bytes || !dirty_background_bytes)
429 available_memory = determine_dirtyable_memory();
432 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
434 dirty = (vm_dirty_ratio * available_memory) / 100;
436 if (dirty_background_bytes)
437 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
439 background = (dirty_background_ratio * available_memory) / 100;
441 if (background >= dirty)
442 background = dirty / 2;
444 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
445 background += background / 4;
448 *pbackground = background;
453 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
454 * @bdi: the backing_dev_info to query
455 * @dirty: global dirty limit in pages
457 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
458 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
459 * And the "limit" in the name is not seriously taken as hard limit in
460 * balance_dirty_pages().
462 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
463 * - starving fast devices
464 * - piling up dirty pages (that will take long time to sync) on slow devices
466 * The bdi's share of dirty limit will be adapting to its throughput and
467 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
469 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
472 long numerator, denominator;
475 * Calculate this BDI's share of the dirty ratio.
477 bdi_writeout_fraction(bdi, &numerator, &denominator);
479 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
480 bdi_dirty *= numerator;
481 do_div(bdi_dirty, denominator);
483 bdi_dirty += (dirty * bdi->min_ratio) / 100;
484 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
485 bdi_dirty = dirty * bdi->max_ratio / 100;
490 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
491 unsigned long elapsed,
492 unsigned long written)
494 const unsigned long period = roundup_pow_of_two(3 * HZ);
495 unsigned long avg = bdi->avg_write_bandwidth;
496 unsigned long old = bdi->write_bandwidth;
500 * bw = written * HZ / elapsed
502 * bw * elapsed + write_bandwidth * (period - elapsed)
503 * write_bandwidth = ---------------------------------------------------
506 bw = written - bdi->written_stamp;
508 if (unlikely(elapsed > period)) {
513 bw += (u64)bdi->write_bandwidth * (period - elapsed);
514 bw >>= ilog2(period);
517 * one more level of smoothing, for filtering out sudden spikes
519 if (avg > old && old >= (unsigned long)bw)
520 avg -= (avg - old) >> 3;
522 if (avg < old && old <= (unsigned long)bw)
523 avg += (old - avg) >> 3;
526 bdi->write_bandwidth = bw;
527 bdi->avg_write_bandwidth = avg;
531 * The global dirtyable memory and dirty threshold could be suddenly knocked
532 * down by a large amount (eg. on the startup of KVM in a swapless system).
533 * This may throw the system into deep dirty exceeded state and throttle
534 * heavy/light dirtiers alike. To retain good responsiveness, maintain
535 * global_dirty_limit for tracking slowly down to the knocked down dirty
538 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
540 unsigned long limit = global_dirty_limit;
543 * Follow up in one step.
545 if (limit < thresh) {
551 * Follow down slowly. Use the higher one as the target, because thresh
552 * may drop below dirty. This is exactly the reason to introduce
553 * global_dirty_limit which is guaranteed to lie above the dirty pages.
555 thresh = max(thresh, dirty);
556 if (limit > thresh) {
557 limit -= (limit - thresh) >> 5;
562 global_dirty_limit = limit;
565 static void global_update_bandwidth(unsigned long thresh,
569 static DEFINE_SPINLOCK(dirty_lock);
570 static unsigned long update_time;
573 * check locklessly first to optimize away locking for the most time
575 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
578 spin_lock(&dirty_lock);
579 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
580 update_dirty_limit(thresh, dirty);
583 spin_unlock(&dirty_lock);
586 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
587 unsigned long thresh,
589 unsigned long bdi_thresh,
590 unsigned long bdi_dirty,
591 unsigned long start_time)
593 unsigned long now = jiffies;
594 unsigned long elapsed = now - bdi->bw_time_stamp;
595 unsigned long written;
598 * rate-limit, only update once every 200ms.
600 if (elapsed < BANDWIDTH_INTERVAL)
603 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
606 * Skip quiet periods when disk bandwidth is under-utilized.
607 * (at least 1s idle time between two flusher runs)
609 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
613 global_update_bandwidth(thresh, dirty, now);
615 bdi_update_write_bandwidth(bdi, elapsed, written);
618 bdi->written_stamp = written;
619 bdi->bw_time_stamp = now;
622 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
623 unsigned long thresh,
625 unsigned long bdi_thresh,
626 unsigned long bdi_dirty,
627 unsigned long start_time)
629 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
631 spin_lock(&bdi->wb.list_lock);
632 __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty,
634 spin_unlock(&bdi->wb.list_lock);
638 * balance_dirty_pages() must be called by processes which are generating dirty
639 * data. It looks at the number of dirty pages in the machine and will force
640 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
641 * If we're over `background_thresh' then the writeback threads are woken to
642 * perform some writeout.
644 static void balance_dirty_pages(struct address_space *mapping,
645 unsigned long write_chunk)
647 unsigned long nr_reclaimable, bdi_nr_reclaimable;
648 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
649 unsigned long bdi_dirty;
650 unsigned long background_thresh;
651 unsigned long dirty_thresh;
652 unsigned long bdi_thresh;
653 unsigned long pages_written = 0;
654 unsigned long pause = 1;
655 bool dirty_exceeded = false;
656 struct backing_dev_info *bdi = mapping->backing_dev_info;
657 unsigned long start_time = jiffies;
660 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
661 global_page_state(NR_UNSTABLE_NFS);
662 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
664 global_dirty_limits(&background_thresh, &dirty_thresh);
667 * Throttle it only when the background writeback cannot
668 * catch-up. This avoids (excessively) small writeouts
669 * when the bdi limits are ramping up.
671 if (nr_dirty <= (background_thresh + dirty_thresh) / 2)
674 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
675 bdi_thresh = task_dirty_limit(current, bdi_thresh);
678 * In order to avoid the stacked BDI deadlock we need
679 * to ensure we accurately count the 'dirty' pages when
680 * the threshold is low.
682 * Otherwise it would be possible to get thresh+n pages
683 * reported dirty, even though there are thresh-m pages
684 * actually dirty; with m+n sitting in the percpu
687 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
688 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
689 bdi_dirty = bdi_nr_reclaimable +
690 bdi_stat_sum(bdi, BDI_WRITEBACK);
692 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
693 bdi_dirty = bdi_nr_reclaimable +
694 bdi_stat(bdi, BDI_WRITEBACK);
698 * The bdi thresh is somehow "soft" limit derived from the
699 * global "hard" limit. The former helps to prevent heavy IO
700 * bdi or process from holding back light ones; The latter is
701 * the last resort safeguard.
703 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
704 (nr_dirty > dirty_thresh);
709 if (!bdi->dirty_exceeded)
710 bdi->dirty_exceeded = 1;
712 bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty,
713 bdi_thresh, bdi_dirty, start_time);
715 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
716 * Unstable writes are a feature of certain networked
717 * filesystems (i.e. NFS) in which data may have been
718 * written to the server's write cache, but has not yet
719 * been flushed to permanent storage.
720 * Only move pages to writeback if this bdi is over its
721 * threshold otherwise wait until the disk writes catch
724 trace_balance_dirty_start(bdi);
725 if (bdi_nr_reclaimable > bdi_thresh) {
726 pages_written += writeback_inodes_wb(&bdi->wb,
728 trace_balance_dirty_written(bdi, pages_written);
729 if (pages_written >= write_chunk)
730 break; /* We've done our duty */
732 __set_current_state(TASK_UNINTERRUPTIBLE);
733 io_schedule_timeout(pause);
734 trace_balance_dirty_wait(bdi);
736 dirty_thresh = hard_dirty_limit(dirty_thresh);
738 * max-pause area. If dirty exceeded but still within this
739 * area, no need to sleep for more than 200ms: (a) 8 pages per
740 * 200ms is typically more than enough to curb heavy dirtiers;
741 * (b) the pause time limit makes the dirtiers more responsive.
743 if (nr_dirty < dirty_thresh +
744 dirty_thresh / DIRTY_MAXPAUSE_AREA &&
745 time_after(jiffies, start_time + MAX_PAUSE))
748 * pass-good area. When some bdi gets blocked (eg. NFS server
749 * not responding), or write bandwidth dropped dramatically due
750 * to concurrent reads, or dirty threshold suddenly dropped and
751 * the dirty pages cannot be brought down anytime soon (eg. on
752 * slow USB stick), at least let go of the good bdi's.
754 if (nr_dirty < dirty_thresh +
755 dirty_thresh / DIRTY_PASSGOOD_AREA &&
756 bdi_dirty < bdi_thresh)
760 * Increase the delay for each loop, up to our previous
761 * default of taking a 100ms nap.
768 if (!dirty_exceeded && bdi->dirty_exceeded)
769 bdi->dirty_exceeded = 0;
771 if (writeback_in_progress(bdi))
775 * In laptop mode, we wait until hitting the higher threshold before
776 * starting background writeout, and then write out all the way down
777 * to the lower threshold. So slow writers cause minimal disk activity.
779 * In normal mode, we start background writeout at the lower
780 * background_thresh, to keep the amount of dirty memory low.
782 if ((laptop_mode && pages_written) ||
783 (!laptop_mode && (nr_reclaimable > background_thresh)))
784 bdi_start_background_writeback(bdi);
787 void set_page_dirty_balance(struct page *page, int page_mkwrite)
789 if (set_page_dirty(page) || page_mkwrite) {
790 struct address_space *mapping = page_mapping(page);
793 balance_dirty_pages_ratelimited(mapping);
797 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
800 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
801 * @mapping: address_space which was dirtied
802 * @nr_pages_dirtied: number of pages which the caller has just dirtied
804 * Processes which are dirtying memory should call in here once for each page
805 * which was newly dirtied. The function will periodically check the system's
806 * dirty state and will initiate writeback if needed.
808 * On really big machines, get_writeback_state is expensive, so try to avoid
809 * calling it too often (ratelimiting). But once we're over the dirty memory
810 * limit we decrease the ratelimiting by a lot, to prevent individual processes
811 * from overshooting the limit by (ratelimit_pages) each.
813 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
814 unsigned long nr_pages_dirtied)
816 struct backing_dev_info *bdi = mapping->backing_dev_info;
817 unsigned long ratelimit;
820 if (!bdi_cap_account_dirty(bdi))
823 ratelimit = ratelimit_pages;
824 if (mapping->backing_dev_info->dirty_exceeded)
828 * Check the rate limiting. Also, we do not want to throttle real-time
829 * tasks in balance_dirty_pages(). Period.
832 p = &__get_cpu_var(bdp_ratelimits);
833 *p += nr_pages_dirtied;
834 if (unlikely(*p >= ratelimit)) {
835 ratelimit = sync_writeback_pages(*p);
838 balance_dirty_pages(mapping, ratelimit);
843 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
845 void throttle_vm_writeout(gfp_t gfp_mask)
847 unsigned long background_thresh;
848 unsigned long dirty_thresh;
851 global_dirty_limits(&background_thresh, &dirty_thresh);
854 * Boost the allowable dirty threshold a bit for page
855 * allocators so they don't get DoS'ed by heavy writers
857 dirty_thresh += dirty_thresh / 10; /* wheeee... */
859 if (global_page_state(NR_UNSTABLE_NFS) +
860 global_page_state(NR_WRITEBACK) <= dirty_thresh)
862 congestion_wait(BLK_RW_ASYNC, HZ/10);
865 * The caller might hold locks which can prevent IO completion
866 * or progress in the filesystem. So we cannot just sit here
867 * waiting for IO to complete.
869 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
875 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
877 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
878 void __user *buffer, size_t *length, loff_t *ppos)
880 proc_dointvec(table, write, buffer, length, ppos);
881 bdi_arm_supers_timer();
886 void laptop_mode_timer_fn(unsigned long data)
888 struct request_queue *q = (struct request_queue *)data;
889 int nr_pages = global_page_state(NR_FILE_DIRTY) +
890 global_page_state(NR_UNSTABLE_NFS);
893 * We want to write everything out, not just down to the dirty
896 if (bdi_has_dirty_io(&q->backing_dev_info))
897 bdi_start_writeback(&q->backing_dev_info, nr_pages);
901 * We've spun up the disk and we're in laptop mode: schedule writeback
902 * of all dirty data a few seconds from now. If the flush is already scheduled
903 * then push it back - the user is still using the disk.
905 void laptop_io_completion(struct backing_dev_info *info)
907 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
911 * We're in laptop mode and we've just synced. The sync's writes will have
912 * caused another writeback to be scheduled by laptop_io_completion.
913 * Nothing needs to be written back anymore, so we unschedule the writeback.
915 void laptop_sync_completion(void)
917 struct backing_dev_info *bdi;
921 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
922 del_timer(&bdi->laptop_mode_wb_timer);
929 * If ratelimit_pages is too high then we can get into dirty-data overload
930 * if a large number of processes all perform writes at the same time.
931 * If it is too low then SMP machines will call the (expensive)
932 * get_writeback_state too often.
934 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
935 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
936 * thresholds before writeback cuts in.
938 * But the limit should not be set too high. Because it also controls the
939 * amount of memory which the balance_dirty_pages() caller has to write back.
940 * If this is too large then the caller will block on the IO queue all the
941 * time. So limit it to four megabytes - the balance_dirty_pages() caller
942 * will write six megabyte chunks, max.
945 void writeback_set_ratelimit(void)
947 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
948 if (ratelimit_pages < 16)
949 ratelimit_pages = 16;
950 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
951 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
955 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
957 writeback_set_ratelimit();
961 static struct notifier_block __cpuinitdata ratelimit_nb = {
962 .notifier_call = ratelimit_handler,
967 * Called early on to tune the page writeback dirty limits.
969 * We used to scale dirty pages according to how total memory
970 * related to pages that could be allocated for buffers (by
971 * comparing nr_free_buffer_pages() to vm_total_pages.
973 * However, that was when we used "dirty_ratio" to scale with
974 * all memory, and we don't do that any more. "dirty_ratio"
975 * is now applied to total non-HIGHPAGE memory (by subtracting
976 * totalhigh_pages from vm_total_pages), and as such we can't
977 * get into the old insane situation any more where we had
978 * large amounts of dirty pages compared to a small amount of
979 * non-HIGHMEM memory.
981 * But we might still want to scale the dirty_ratio by how
982 * much memory the box has..
984 void __init page_writeback_init(void)
988 writeback_set_ratelimit();
989 register_cpu_notifier(&ratelimit_nb);
991 shift = calc_period_shift();
992 prop_descriptor_init(&vm_completions, shift);
993 prop_descriptor_init(&vm_dirties, shift);
997 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
998 * @mapping: address space structure to write
999 * @start: starting page index
1000 * @end: ending page index (inclusive)
1002 * This function scans the page range from @start to @end (inclusive) and tags
1003 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1004 * that write_cache_pages (or whoever calls this function) will then use
1005 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1006 * used to avoid livelocking of writeback by a process steadily creating new
1007 * dirty pages in the file (thus it is important for this function to be quick
1008 * so that it can tag pages faster than a dirtying process can create them).
1011 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1013 void tag_pages_for_writeback(struct address_space *mapping,
1014 pgoff_t start, pgoff_t end)
1016 #define WRITEBACK_TAG_BATCH 4096
1017 unsigned long tagged;
1020 spin_lock_irq(&mapping->tree_lock);
1021 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1022 &start, end, WRITEBACK_TAG_BATCH,
1023 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1024 spin_unlock_irq(&mapping->tree_lock);
1025 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1027 /* We check 'start' to handle wrapping when end == ~0UL */
1028 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1030 EXPORT_SYMBOL(tag_pages_for_writeback);
1033 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1034 * @mapping: address space structure to write
1035 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1036 * @writepage: function called for each page
1037 * @data: data passed to writepage function
1039 * If a page is already under I/O, write_cache_pages() skips it, even
1040 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1041 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1042 * and msync() need to guarantee that all the data which was dirty at the time
1043 * the call was made get new I/O started against them. If wbc->sync_mode is
1044 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1045 * existing IO to complete.
1047 * To avoid livelocks (when other process dirties new pages), we first tag
1048 * pages which should be written back with TOWRITE tag and only then start
1049 * writing them. For data-integrity sync we have to be careful so that we do
1050 * not miss some pages (e.g., because some other process has cleared TOWRITE
1051 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1052 * by the process clearing the DIRTY tag (and submitting the page for IO).
1054 int write_cache_pages(struct address_space *mapping,
1055 struct writeback_control *wbc, writepage_t writepage,
1060 struct pagevec pvec;
1062 pgoff_t uninitialized_var(writeback_index);
1064 pgoff_t end; /* Inclusive */
1067 int range_whole = 0;
1070 pagevec_init(&pvec, 0);
1071 if (wbc->range_cyclic) {
1072 writeback_index = mapping->writeback_index; /* prev offset */
1073 index = writeback_index;
1080 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1081 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1082 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1084 cycled = 1; /* ignore range_cyclic tests */
1086 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1087 tag = PAGECACHE_TAG_TOWRITE;
1089 tag = PAGECACHE_TAG_DIRTY;
1091 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1092 tag_pages_for_writeback(mapping, index, end);
1094 while (!done && (index <= end)) {
1097 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1098 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1102 for (i = 0; i < nr_pages; i++) {
1103 struct page *page = pvec.pages[i];
1106 * At this point, the page may be truncated or
1107 * invalidated (changing page->mapping to NULL), or
1108 * even swizzled back from swapper_space to tmpfs file
1109 * mapping. However, page->index will not change
1110 * because we have a reference on the page.
1112 if (page->index > end) {
1114 * can't be range_cyclic (1st pass) because
1115 * end == -1 in that case.
1121 done_index = page->index;
1126 * Page truncated or invalidated. We can freely skip it
1127 * then, even for data integrity operations: the page
1128 * has disappeared concurrently, so there could be no
1129 * real expectation of this data interity operation
1130 * even if there is now a new, dirty page at the same
1131 * pagecache address.
1133 if (unlikely(page->mapping != mapping)) {
1139 if (!PageDirty(page)) {
1140 /* someone wrote it for us */
1141 goto continue_unlock;
1144 if (PageWriteback(page)) {
1145 if (wbc->sync_mode != WB_SYNC_NONE)
1146 wait_on_page_writeback(page);
1148 goto continue_unlock;
1151 BUG_ON(PageWriteback(page));
1152 if (!clear_page_dirty_for_io(page))
1153 goto continue_unlock;
1155 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1156 ret = (*writepage)(page, wbc, data);
1157 if (unlikely(ret)) {
1158 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1163 * done_index is set past this page,
1164 * so media errors will not choke
1165 * background writeout for the entire
1166 * file. This has consequences for
1167 * range_cyclic semantics (ie. it may
1168 * not be suitable for data integrity
1171 done_index = page->index + 1;
1178 * We stop writing back only if we are not doing
1179 * integrity sync. In case of integrity sync we have to
1180 * keep going until we have written all the pages
1181 * we tagged for writeback prior to entering this loop.
1183 if (--wbc->nr_to_write <= 0 &&
1184 wbc->sync_mode == WB_SYNC_NONE) {
1189 pagevec_release(&pvec);
1192 if (!cycled && !done) {
1195 * We hit the last page and there is more work to be done: wrap
1196 * back to the start of the file
1200 end = writeback_index - 1;
1203 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1204 mapping->writeback_index = done_index;
1208 EXPORT_SYMBOL(write_cache_pages);
1211 * Function used by generic_writepages to call the real writepage
1212 * function and set the mapping flags on error
1214 static int __writepage(struct page *page, struct writeback_control *wbc,
1217 struct address_space *mapping = data;
1218 int ret = mapping->a_ops->writepage(page, wbc);
1219 mapping_set_error(mapping, ret);
1224 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1225 * @mapping: address space structure to write
1226 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1228 * This is a library function, which implements the writepages()
1229 * address_space_operation.
1231 int generic_writepages(struct address_space *mapping,
1232 struct writeback_control *wbc)
1234 struct blk_plug plug;
1237 /* deal with chardevs and other special file */
1238 if (!mapping->a_ops->writepage)
1241 blk_start_plug(&plug);
1242 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1243 blk_finish_plug(&plug);
1247 EXPORT_SYMBOL(generic_writepages);
1249 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1253 if (wbc->nr_to_write <= 0)
1255 if (mapping->a_ops->writepages)
1256 ret = mapping->a_ops->writepages(mapping, wbc);
1258 ret = generic_writepages(mapping, wbc);
1263 * write_one_page - write out a single page and optionally wait on I/O
1264 * @page: the page to write
1265 * @wait: if true, wait on writeout
1267 * The page must be locked by the caller and will be unlocked upon return.
1269 * write_one_page() returns a negative error code if I/O failed.
1271 int write_one_page(struct page *page, int wait)
1273 struct address_space *mapping = page->mapping;
1275 struct writeback_control wbc = {
1276 .sync_mode = WB_SYNC_ALL,
1280 BUG_ON(!PageLocked(page));
1283 wait_on_page_writeback(page);
1285 if (clear_page_dirty_for_io(page)) {
1286 page_cache_get(page);
1287 ret = mapping->a_ops->writepage(page, &wbc);
1288 if (ret == 0 && wait) {
1289 wait_on_page_writeback(page);
1290 if (PageError(page))
1293 page_cache_release(page);
1299 EXPORT_SYMBOL(write_one_page);
1302 * For address_spaces which do not use buffers nor write back.
1304 int __set_page_dirty_no_writeback(struct page *page)
1306 if (!PageDirty(page))
1307 return !TestSetPageDirty(page);
1312 * Helper function for set_page_dirty family.
1313 * NOTE: This relies on being atomic wrt interrupts.
1315 void account_page_dirtied(struct page *page, struct address_space *mapping)
1317 if (mapping_cap_account_dirty(mapping)) {
1318 __inc_zone_page_state(page, NR_FILE_DIRTY);
1319 __inc_zone_page_state(page, NR_DIRTIED);
1320 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1321 task_dirty_inc(current);
1322 task_io_account_write(PAGE_CACHE_SIZE);
1325 EXPORT_SYMBOL(account_page_dirtied);
1328 * Helper function for set_page_writeback family.
1329 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1332 void account_page_writeback(struct page *page)
1334 inc_zone_page_state(page, NR_WRITEBACK);
1335 inc_zone_page_state(page, NR_WRITTEN);
1337 EXPORT_SYMBOL(account_page_writeback);
1340 * For address_spaces which do not use buffers. Just tag the page as dirty in
1343 * This is also used when a single buffer is being dirtied: we want to set the
1344 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1345 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1347 * Most callers have locked the page, which pins the address_space in memory.
1348 * But zap_pte_range() does not lock the page, however in that case the
1349 * mapping is pinned by the vma's ->vm_file reference.
1351 * We take care to handle the case where the page was truncated from the
1352 * mapping by re-checking page_mapping() inside tree_lock.
1354 int __set_page_dirty_nobuffers(struct page *page)
1356 if (!TestSetPageDirty(page)) {
1357 struct address_space *mapping = page_mapping(page);
1358 struct address_space *mapping2;
1363 spin_lock_irq(&mapping->tree_lock);
1364 mapping2 = page_mapping(page);
1365 if (mapping2) { /* Race with truncate? */
1366 BUG_ON(mapping2 != mapping);
1367 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1368 account_page_dirtied(page, mapping);
1369 radix_tree_tag_set(&mapping->page_tree,
1370 page_index(page), PAGECACHE_TAG_DIRTY);
1372 spin_unlock_irq(&mapping->tree_lock);
1373 if (mapping->host) {
1374 /* !PageAnon && !swapper_space */
1375 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1381 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1384 * When a writepage implementation decides that it doesn't want to write this
1385 * page for some reason, it should redirty the locked page via
1386 * redirty_page_for_writepage() and it should then unlock the page and return 0
1388 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1390 wbc->pages_skipped++;
1391 return __set_page_dirty_nobuffers(page);
1393 EXPORT_SYMBOL(redirty_page_for_writepage);
1398 * For pages with a mapping this should be done under the page lock
1399 * for the benefit of asynchronous memory errors who prefer a consistent
1400 * dirty state. This rule can be broken in some special cases,
1401 * but should be better not to.
1403 * If the mapping doesn't provide a set_page_dirty a_op, then
1404 * just fall through and assume that it wants buffer_heads.
1406 int set_page_dirty(struct page *page)
1408 struct address_space *mapping = page_mapping(page);
1410 if (likely(mapping)) {
1411 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1413 * readahead/lru_deactivate_page could remain
1414 * PG_readahead/PG_reclaim due to race with end_page_writeback
1415 * About readahead, if the page is written, the flags would be
1416 * reset. So no problem.
1417 * About lru_deactivate_page, if the page is redirty, the flag
1418 * will be reset. So no problem. but if the page is used by readahead
1419 * it will confuse readahead and make it restart the size rampup
1420 * process. But it's a trivial problem.
1422 ClearPageReclaim(page);
1425 spd = __set_page_dirty_buffers;
1427 return (*spd)(page);
1429 if (!PageDirty(page)) {
1430 if (!TestSetPageDirty(page))
1435 EXPORT_SYMBOL(set_page_dirty);
1438 * set_page_dirty() is racy if the caller has no reference against
1439 * page->mapping->host, and if the page is unlocked. This is because another
1440 * CPU could truncate the page off the mapping and then free the mapping.
1442 * Usually, the page _is_ locked, or the caller is a user-space process which
1443 * holds a reference on the inode by having an open file.
1445 * In other cases, the page should be locked before running set_page_dirty().
1447 int set_page_dirty_lock(struct page *page)
1452 ret = set_page_dirty(page);
1456 EXPORT_SYMBOL(set_page_dirty_lock);
1459 * Clear a page's dirty flag, while caring for dirty memory accounting.
1460 * Returns true if the page was previously dirty.
1462 * This is for preparing to put the page under writeout. We leave the page
1463 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1464 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1465 * implementation will run either set_page_writeback() or set_page_dirty(),
1466 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1469 * This incoherency between the page's dirty flag and radix-tree tag is
1470 * unfortunate, but it only exists while the page is locked.
1472 int clear_page_dirty_for_io(struct page *page)
1474 struct address_space *mapping = page_mapping(page);
1476 BUG_ON(!PageLocked(page));
1478 if (mapping && mapping_cap_account_dirty(mapping)) {
1480 * Yes, Virginia, this is indeed insane.
1482 * We use this sequence to make sure that
1483 * (a) we account for dirty stats properly
1484 * (b) we tell the low-level filesystem to
1485 * mark the whole page dirty if it was
1486 * dirty in a pagetable. Only to then
1487 * (c) clean the page again and return 1 to
1488 * cause the writeback.
1490 * This way we avoid all nasty races with the
1491 * dirty bit in multiple places and clearing
1492 * them concurrently from different threads.
1494 * Note! Normally the "set_page_dirty(page)"
1495 * has no effect on the actual dirty bit - since
1496 * that will already usually be set. But we
1497 * need the side effects, and it can help us
1500 * We basically use the page "master dirty bit"
1501 * as a serialization point for all the different
1502 * threads doing their things.
1504 if (page_mkclean(page))
1505 set_page_dirty(page);
1507 * We carefully synchronise fault handlers against
1508 * installing a dirty pte and marking the page dirty
1509 * at this point. We do this by having them hold the
1510 * page lock at some point after installing their
1511 * pte, but before marking the page dirty.
1512 * Pages are always locked coming in here, so we get
1513 * the desired exclusion. See mm/memory.c:do_wp_page()
1514 * for more comments.
1516 if (TestClearPageDirty(page)) {
1517 dec_zone_page_state(page, NR_FILE_DIRTY);
1518 dec_bdi_stat(mapping->backing_dev_info,
1524 return TestClearPageDirty(page);
1526 EXPORT_SYMBOL(clear_page_dirty_for_io);
1528 int test_clear_page_writeback(struct page *page)
1530 struct address_space *mapping = page_mapping(page);
1534 struct backing_dev_info *bdi = mapping->backing_dev_info;
1535 unsigned long flags;
1537 spin_lock_irqsave(&mapping->tree_lock, flags);
1538 ret = TestClearPageWriteback(page);
1540 radix_tree_tag_clear(&mapping->page_tree,
1542 PAGECACHE_TAG_WRITEBACK);
1543 if (bdi_cap_account_writeback(bdi)) {
1544 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1545 __bdi_writeout_inc(bdi);
1548 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1550 ret = TestClearPageWriteback(page);
1553 dec_zone_page_state(page, NR_WRITEBACK);
1557 int test_set_page_writeback(struct page *page)
1559 struct address_space *mapping = page_mapping(page);
1563 struct backing_dev_info *bdi = mapping->backing_dev_info;
1564 unsigned long flags;
1566 spin_lock_irqsave(&mapping->tree_lock, flags);
1567 ret = TestSetPageWriteback(page);
1569 radix_tree_tag_set(&mapping->page_tree,
1571 PAGECACHE_TAG_WRITEBACK);
1572 if (bdi_cap_account_writeback(bdi))
1573 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1575 if (!PageDirty(page))
1576 radix_tree_tag_clear(&mapping->page_tree,
1578 PAGECACHE_TAG_DIRTY);
1579 radix_tree_tag_clear(&mapping->page_tree,
1581 PAGECACHE_TAG_TOWRITE);
1582 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1584 ret = TestSetPageWriteback(page);
1587 account_page_writeback(page);
1591 EXPORT_SYMBOL(test_set_page_writeback);
1594 * Return true if any of the pages in the mapping are marked with the
1597 int mapping_tagged(struct address_space *mapping, int tag)
1601 ret = radix_tree_tagged(&mapping->page_tree, tag);
1605 EXPORT_SYMBOL(mapping_tagged);