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>
39 * The maximum number of pages to writeout in a single bdflush/kupdate
40 * operation. We do this so we don't hold I_SYNC against an inode for
41 * enormous amounts of time, which would block a userspace task which has
42 * been forced to throttle against that inode. Also, the code reevaluates
43 * the dirty each time it has written this many pages.
45 #define MAX_WRITEBACK_PAGES 1024
48 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
49 * will look to see if it needs to force writeback or throttling.
51 static long ratelimit_pages = 32;
54 * When balance_dirty_pages decides that the caller needs to perform some
55 * non-background writeback, this is how many pages it will attempt to write.
56 * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
57 * large amounts of I/O are submitted.
59 static inline long sync_writeback_pages(void)
61 return ratelimit_pages + ratelimit_pages / 2;
64 /* The following parameters are exported via /proc/sys/vm */
67 * Start background writeback (via pdflush) at this percentage
69 int dirty_background_ratio = 10;
72 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
73 * dirty_background_ratio * the amount of dirtyable memory
75 unsigned long dirty_background_bytes;
78 * free highmem will not be subtracted from the total free memory
79 * for calculating free ratios if vm_highmem_is_dirtyable is true
81 int vm_highmem_is_dirtyable;
84 * The generator of dirty data starts writeback at this percentage
86 int vm_dirty_ratio = 20;
89 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
90 * vm_dirty_ratio * the amount of dirtyable memory
92 unsigned long vm_dirty_bytes;
95 * The interval between `kupdate'-style writebacks
97 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
100 * The longest time for which data is allowed to remain dirty
102 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
105 * Flag that makes the machine dump writes/reads and block dirtyings.
110 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
111 * a full sync is triggered after this time elapses without any disk activity.
115 EXPORT_SYMBOL(laptop_mode);
117 /* End of sysctl-exported parameters */
120 static void background_writeout(unsigned long _min_pages);
123 * Scale the writeback cache size proportional to the relative writeout speeds.
125 * We do this by keeping a floating proportion between BDIs, based on page
126 * writeback completions [end_page_writeback()]. Those devices that write out
127 * pages fastest will get the larger share, while the slower will get a smaller
130 * We use page writeout completions because we are interested in getting rid of
131 * dirty pages. Having them written out is the primary goal.
133 * We introduce a concept of time, a period over which we measure these events,
134 * because demand can/will vary over time. The length of this period itself is
135 * measured in page writeback completions.
138 static struct prop_descriptor vm_completions;
139 static struct prop_descriptor vm_dirties;
142 * couple the period to the dirty_ratio:
144 * period/2 ~ roundup_pow_of_two(dirty limit)
146 static int calc_period_shift(void)
148 unsigned long dirty_total;
151 dirty_total = vm_dirty_bytes / PAGE_SIZE;
153 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
155 return 2 + ilog2(dirty_total - 1);
159 * update the period when the dirty threshold changes.
161 static void update_completion_period(void)
163 int shift = calc_period_shift();
164 prop_change_shift(&vm_completions, shift);
165 prop_change_shift(&vm_dirties, shift);
168 int dirty_background_ratio_handler(struct ctl_table *table, int write,
169 struct file *filp, void __user *buffer, size_t *lenp,
174 ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
175 if (ret == 0 && write)
176 dirty_background_bytes = 0;
180 int dirty_background_bytes_handler(struct ctl_table *table, int write,
181 struct file *filp, void __user *buffer, size_t *lenp,
186 ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
187 if (ret == 0 && write)
188 dirty_background_ratio = 0;
192 int dirty_ratio_handler(struct ctl_table *table, int write,
193 struct file *filp, void __user *buffer, size_t *lenp,
196 int old_ratio = vm_dirty_ratio;
199 ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
200 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
201 update_completion_period();
208 int dirty_bytes_handler(struct ctl_table *table, int write,
209 struct file *filp, void __user *buffer, size_t *lenp,
212 unsigned long old_bytes = vm_dirty_bytes;
215 ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
216 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
217 update_completion_period();
224 * Increment the BDI's writeout completion count and the global writeout
225 * completion count. Called from test_clear_page_writeback().
227 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
229 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
233 void bdi_writeout_inc(struct backing_dev_info *bdi)
237 local_irq_save(flags);
238 __bdi_writeout_inc(bdi);
239 local_irq_restore(flags);
241 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
243 void task_dirty_inc(struct task_struct *tsk)
245 prop_inc_single(&vm_dirties, &tsk->dirties);
249 * Obtain an accurate fraction of the BDI's portion.
251 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
252 long *numerator, long *denominator)
254 if (bdi_cap_writeback_dirty(bdi)) {
255 prop_fraction_percpu(&vm_completions, &bdi->completions,
256 numerator, denominator);
264 * Clip the earned share of dirty pages to that which is actually available.
265 * This avoids exceeding the total dirty_limit when the floating averages
266 * fluctuate too quickly.
268 static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
269 unsigned long dirty, unsigned long *pbdi_dirty)
271 unsigned long avail_dirty;
273 avail_dirty = global_page_state(NR_FILE_DIRTY) +
274 global_page_state(NR_WRITEBACK) +
275 global_page_state(NR_UNSTABLE_NFS) +
276 global_page_state(NR_WRITEBACK_TEMP);
278 if (avail_dirty < dirty)
279 avail_dirty = dirty - avail_dirty;
283 avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
284 bdi_stat(bdi, BDI_WRITEBACK);
286 *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
289 static inline void task_dirties_fraction(struct task_struct *tsk,
290 long *numerator, long *denominator)
292 prop_fraction_single(&vm_dirties, &tsk->dirties,
293 numerator, denominator);
297 * scale the dirty limit
299 * task specific dirty limit:
301 * dirty -= (dirty/8) * p_{t}
303 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
305 long numerator, denominator;
306 unsigned long dirty = *pdirty;
307 u64 inv = dirty >> 3;
309 task_dirties_fraction(tsk, &numerator, &denominator);
311 do_div(inv, denominator);
314 if (dirty < *pdirty/2)
323 static DEFINE_SPINLOCK(bdi_lock);
324 static unsigned int bdi_min_ratio;
326 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
331 spin_lock_irqsave(&bdi_lock, flags);
332 if (min_ratio > bdi->max_ratio) {
335 min_ratio -= bdi->min_ratio;
336 if (bdi_min_ratio + min_ratio < 100) {
337 bdi_min_ratio += min_ratio;
338 bdi->min_ratio += min_ratio;
343 spin_unlock_irqrestore(&bdi_lock, flags);
348 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
356 spin_lock_irqsave(&bdi_lock, flags);
357 if (bdi->min_ratio > max_ratio) {
360 bdi->max_ratio = max_ratio;
361 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
363 spin_unlock_irqrestore(&bdi_lock, flags);
367 EXPORT_SYMBOL(bdi_set_max_ratio);
370 * Work out the current dirty-memory clamping and background writeout
373 * The main aim here is to lower them aggressively if there is a lot of mapped
374 * memory around. To avoid stressing page reclaim with lots of unreclaimable
375 * pages. It is better to clamp down on writers than to start swapping, and
376 * performing lots of scanning.
378 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
380 * We don't permit the clamping level to fall below 5% - that is getting rather
383 * We make sure that the background writeout level is below the adjusted
387 static unsigned long highmem_dirtyable_memory(unsigned long total)
389 #ifdef CONFIG_HIGHMEM
393 for_each_node_state(node, N_HIGH_MEMORY) {
395 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
397 x += zone_page_state(z, NR_FREE_PAGES) + zone_lru_pages(z);
400 * Make sure that the number of highmem pages is never larger
401 * than the number of the total dirtyable memory. This can only
402 * occur in very strange VM situations but we want to make sure
403 * that this does not occur.
405 return min(x, total);
412 * determine_dirtyable_memory - amount of memory that may be used
414 * Returns the numebr of pages that can currently be freed and used
415 * by the kernel for direct mappings.
417 unsigned long determine_dirtyable_memory(void)
421 x = global_page_state(NR_FREE_PAGES) + global_lru_pages();
423 if (!vm_highmem_is_dirtyable)
424 x -= highmem_dirtyable_memory(x);
426 return x + 1; /* Ensure that we never return 0 */
430 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
431 unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
433 unsigned long background;
435 unsigned long available_memory = determine_dirtyable_memory();
436 struct task_struct *tsk;
439 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
443 dirty_ratio = vm_dirty_ratio;
446 dirty = (dirty_ratio * available_memory) / 100;
449 if (dirty_background_bytes)
450 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
452 background = (dirty_background_ratio * available_memory) / 100;
454 if (background >= dirty)
455 background = dirty / 2;
457 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
458 background += background / 4;
461 *pbackground = background;
466 long numerator, denominator;
469 * Calculate this BDI's share of the dirty ratio.
471 bdi_writeout_fraction(bdi, &numerator, &denominator);
473 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
474 bdi_dirty *= numerator;
475 do_div(bdi_dirty, denominator);
476 bdi_dirty += (dirty * bdi->min_ratio) / 100;
477 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
478 bdi_dirty = dirty * bdi->max_ratio / 100;
480 *pbdi_dirty = bdi_dirty;
481 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
482 task_dirty_limit(current, pbdi_dirty);
487 * balance_dirty_pages() must be called by processes which are generating dirty
488 * data. It looks at the number of dirty pages in the machine and will force
489 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
490 * If we're over `background_thresh' then pdflush is woken to perform some
493 static void balance_dirty_pages(struct address_space *mapping)
495 long nr_reclaimable, bdi_nr_reclaimable;
496 long nr_writeback, bdi_nr_writeback;
497 unsigned long background_thresh;
498 unsigned long dirty_thresh;
499 unsigned long bdi_thresh;
500 unsigned long pages_written = 0;
501 unsigned long write_chunk = sync_writeback_pages();
503 struct backing_dev_info *bdi = mapping->backing_dev_info;
506 struct writeback_control wbc = {
508 .sync_mode = WB_SYNC_NONE,
509 .older_than_this = NULL,
510 .nr_to_write = write_chunk,
514 get_dirty_limits(&background_thresh, &dirty_thresh,
517 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
518 global_page_state(NR_UNSTABLE_NFS);
519 nr_writeback = global_page_state(NR_WRITEBACK);
521 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
522 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
524 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
528 * Throttle it only when the background writeback cannot
529 * catch-up. This avoids (excessively) small writeouts
530 * when the bdi limits are ramping up.
532 if (nr_reclaimable + nr_writeback <
533 (background_thresh + dirty_thresh) / 2)
536 if (!bdi->dirty_exceeded)
537 bdi->dirty_exceeded = 1;
539 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
540 * Unstable writes are a feature of certain networked
541 * filesystems (i.e. NFS) in which data may have been
542 * written to the server's write cache, but has not yet
543 * been flushed to permanent storage.
545 if (bdi_nr_reclaimable) {
546 writeback_inodes(&wbc);
547 pages_written += write_chunk - wbc.nr_to_write;
548 get_dirty_limits(&background_thresh, &dirty_thresh,
553 * In order to avoid the stacked BDI deadlock we need
554 * to ensure we accurately count the 'dirty' pages when
555 * the threshold is low.
557 * Otherwise it would be possible to get thresh+n pages
558 * reported dirty, even though there are thresh-m pages
559 * actually dirty; with m+n sitting in the percpu
562 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
563 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
564 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
565 } else if (bdi_nr_reclaimable) {
566 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
567 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
570 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
572 if (pages_written >= write_chunk)
573 break; /* We've done our duty */
575 congestion_wait(WRITE, HZ/10);
578 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
580 bdi->dirty_exceeded = 0;
582 if (writeback_in_progress(bdi))
583 return; /* pdflush is already working this queue */
586 * In laptop mode, we wait until hitting the higher threshold before
587 * starting background writeout, and then write out all the way down
588 * to the lower threshold. So slow writers cause minimal disk activity.
590 * In normal mode, we start background writeout at the lower
591 * background_thresh, to keep the amount of dirty memory low.
593 if ((laptop_mode && pages_written) ||
594 (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
595 + global_page_state(NR_UNSTABLE_NFS)
596 > background_thresh)))
597 pdflush_operation(background_writeout, 0);
600 void set_page_dirty_balance(struct page *page, int page_mkwrite)
602 if (set_page_dirty(page) || page_mkwrite) {
603 struct address_space *mapping = page_mapping(page);
606 balance_dirty_pages_ratelimited(mapping);
610 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
613 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
614 * @mapping: address_space which was dirtied
615 * @nr_pages_dirtied: number of pages which the caller has just dirtied
617 * Processes which are dirtying memory should call in here once for each page
618 * which was newly dirtied. The function will periodically check the system's
619 * dirty state and will initiate writeback if needed.
621 * On really big machines, get_writeback_state is expensive, so try to avoid
622 * calling it too often (ratelimiting). But once we're over the dirty memory
623 * limit we decrease the ratelimiting by a lot, to prevent individual processes
624 * from overshooting the limit by (ratelimit_pages) each.
626 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
627 unsigned long nr_pages_dirtied)
629 unsigned long ratelimit;
632 ratelimit = ratelimit_pages;
633 if (mapping->backing_dev_info->dirty_exceeded)
637 * Check the rate limiting. Also, we do not want to throttle real-time
638 * tasks in balance_dirty_pages(). Period.
641 p = &__get_cpu_var(bdp_ratelimits);
642 *p += nr_pages_dirtied;
643 if (unlikely(*p >= ratelimit)) {
646 balance_dirty_pages(mapping);
651 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
653 void throttle_vm_writeout(gfp_t gfp_mask)
655 unsigned long background_thresh;
656 unsigned long dirty_thresh;
659 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
662 * Boost the allowable dirty threshold a bit for page
663 * allocators so they don't get DoS'ed by heavy writers
665 dirty_thresh += dirty_thresh / 10; /* wheeee... */
667 if (global_page_state(NR_UNSTABLE_NFS) +
668 global_page_state(NR_WRITEBACK) <= dirty_thresh)
670 congestion_wait(WRITE, HZ/10);
673 * The caller might hold locks which can prevent IO completion
674 * or progress in the filesystem. So we cannot just sit here
675 * waiting for IO to complete.
677 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
683 * writeback at least _min_pages, and keep writing until the amount of dirty
684 * memory is less than the background threshold, or until we're all clean.
686 static void background_writeout(unsigned long _min_pages)
688 long min_pages = _min_pages;
689 struct writeback_control wbc = {
691 .sync_mode = WB_SYNC_NONE,
692 .older_than_this = NULL,
699 unsigned long background_thresh;
700 unsigned long dirty_thresh;
702 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
703 if (global_page_state(NR_FILE_DIRTY) +
704 global_page_state(NR_UNSTABLE_NFS) < background_thresh
708 wbc.encountered_congestion = 0;
709 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
710 wbc.pages_skipped = 0;
711 writeback_inodes(&wbc);
712 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
713 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
714 /* Wrote less than expected */
715 if (wbc.encountered_congestion || wbc.more_io)
716 congestion_wait(WRITE, HZ/10);
724 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
725 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
726 * -1 if all pdflush threads were busy.
728 int wakeup_pdflush(long nr_pages)
731 nr_pages = global_page_state(NR_FILE_DIRTY) +
732 global_page_state(NR_UNSTABLE_NFS);
733 return pdflush_operation(background_writeout, nr_pages);
736 static void wb_timer_fn(unsigned long unused);
737 static void laptop_timer_fn(unsigned long unused);
739 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
740 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
743 * Periodic writeback of "old" data.
745 * Define "old": the first time one of an inode's pages is dirtied, we mark the
746 * dirtying-time in the inode's address_space. So this periodic writeback code
747 * just walks the superblock inode list, writing back any inodes which are
748 * older than a specific point in time.
750 * Try to run once per dirty_writeback_interval. But if a writeback event
751 * takes longer than a dirty_writeback_interval interval, then leave a
754 * older_than_this takes precedence over nr_to_write. So we'll only write back
755 * all dirty pages if they are all attached to "old" mappings.
757 static void wb_kupdate(unsigned long arg)
759 unsigned long oldest_jif;
760 unsigned long start_jif;
761 unsigned long next_jif;
763 struct writeback_control wbc = {
765 .sync_mode = WB_SYNC_NONE,
766 .older_than_this = &oldest_jif,
775 oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval * 10);
777 next_jif = start_jif + msecs_to_jiffies(dirty_writeback_interval * 10);
778 nr_to_write = global_page_state(NR_FILE_DIRTY) +
779 global_page_state(NR_UNSTABLE_NFS) +
780 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
781 while (nr_to_write > 0) {
783 wbc.encountered_congestion = 0;
784 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
785 writeback_inodes(&wbc);
786 if (wbc.nr_to_write > 0) {
787 if (wbc.encountered_congestion || wbc.more_io)
788 congestion_wait(WRITE, HZ/10);
790 break; /* All the old data is written */
792 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
794 if (time_before(next_jif, jiffies + HZ))
795 next_jif = jiffies + HZ;
796 if (dirty_writeback_interval)
797 mod_timer(&wb_timer, next_jif);
801 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
803 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
804 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
806 proc_dointvec(table, write, file, buffer, length, ppos);
807 if (dirty_writeback_interval)
808 mod_timer(&wb_timer, jiffies +
809 msecs_to_jiffies(dirty_writeback_interval * 10));
811 del_timer(&wb_timer);
815 static void wb_timer_fn(unsigned long unused)
817 if (pdflush_operation(wb_kupdate, 0) < 0)
818 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
821 static void laptop_flush(unsigned long unused)
826 static void laptop_timer_fn(unsigned long unused)
828 pdflush_operation(laptop_flush, 0);
832 * We've spun up the disk and we're in laptop mode: schedule writeback
833 * of all dirty data a few seconds from now. If the flush is already scheduled
834 * then push it back - the user is still using the disk.
836 void laptop_io_completion(void)
838 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
842 * We're in laptop mode and we've just synced. The sync's writes will have
843 * caused another writeback to be scheduled by laptop_io_completion.
844 * Nothing needs to be written back anymore, so we unschedule the writeback.
846 void laptop_sync_completion(void)
848 del_timer(&laptop_mode_wb_timer);
852 * If ratelimit_pages is too high then we can get into dirty-data overload
853 * if a large number of processes all perform writes at the same time.
854 * If it is too low then SMP machines will call the (expensive)
855 * get_writeback_state too often.
857 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
858 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
859 * thresholds before writeback cuts in.
861 * But the limit should not be set too high. Because it also controls the
862 * amount of memory which the balance_dirty_pages() caller has to write back.
863 * If this is too large then the caller will block on the IO queue all the
864 * time. So limit it to four megabytes - the balance_dirty_pages() caller
865 * will write six megabyte chunks, max.
868 void writeback_set_ratelimit(void)
870 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
871 if (ratelimit_pages < 16)
872 ratelimit_pages = 16;
873 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
874 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
878 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
880 writeback_set_ratelimit();
884 static struct notifier_block __cpuinitdata ratelimit_nb = {
885 .notifier_call = ratelimit_handler,
890 * Called early on to tune the page writeback dirty limits.
892 * We used to scale dirty pages according to how total memory
893 * related to pages that could be allocated for buffers (by
894 * comparing nr_free_buffer_pages() to vm_total_pages.
896 * However, that was when we used "dirty_ratio" to scale with
897 * all memory, and we don't do that any more. "dirty_ratio"
898 * is now applied to total non-HIGHPAGE memory (by subtracting
899 * totalhigh_pages from vm_total_pages), and as such we can't
900 * get into the old insane situation any more where we had
901 * large amounts of dirty pages compared to a small amount of
902 * non-HIGHMEM memory.
904 * But we might still want to scale the dirty_ratio by how
905 * much memory the box has..
907 void __init page_writeback_init(void)
912 jiffies + msecs_to_jiffies(dirty_writeback_interval * 10));
913 writeback_set_ratelimit();
914 register_cpu_notifier(&ratelimit_nb);
916 shift = calc_period_shift();
917 prop_descriptor_init(&vm_completions, shift);
918 prop_descriptor_init(&vm_dirties, shift);
922 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
923 * @mapping: address space structure to write
924 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
925 * @writepage: function called for each page
926 * @data: data passed to writepage function
928 * If a page is already under I/O, write_cache_pages() skips it, even
929 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
930 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
931 * and msync() need to guarantee that all the data which was dirty at the time
932 * the call was made get new I/O started against them. If wbc->sync_mode is
933 * WB_SYNC_ALL then we were called for data integrity and we must wait for
934 * existing IO to complete.
936 int write_cache_pages(struct address_space *mapping,
937 struct writeback_control *wbc, writepage_t writepage,
940 struct backing_dev_info *bdi = mapping->backing_dev_info;
945 pgoff_t uninitialized_var(writeback_index);
947 pgoff_t end; /* Inclusive */
951 long nr_to_write = wbc->nr_to_write;
953 if (wbc->nonblocking && bdi_write_congested(bdi)) {
954 wbc->encountered_congestion = 1;
958 pagevec_init(&pvec, 0);
959 if (wbc->range_cyclic) {
960 writeback_index = mapping->writeback_index; /* prev offset */
961 index = writeback_index;
968 index = wbc->range_start >> PAGE_CACHE_SHIFT;
969 end = wbc->range_end >> PAGE_CACHE_SHIFT;
970 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
972 cycled = 1; /* ignore range_cyclic tests */
976 while (!done && (index <= end)) {
979 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
981 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
985 for (i = 0; i < nr_pages; i++) {
986 struct page *page = pvec.pages[i];
989 * At this point, the page may be truncated or
990 * invalidated (changing page->mapping to NULL), or
991 * even swizzled back from swapper_space to tmpfs file
992 * mapping. However, page->index will not change
993 * because we have a reference on the page.
995 if (page->index > end) {
997 * can't be range_cyclic (1st pass) because
998 * end == -1 in that case.
1004 done_index = page->index + 1;
1009 * Page truncated or invalidated. We can freely skip it
1010 * then, even for data integrity operations: the page
1011 * has disappeared concurrently, so there could be no
1012 * real expectation of this data interity operation
1013 * even if there is now a new, dirty page at the same
1014 * pagecache address.
1016 if (unlikely(page->mapping != mapping)) {
1022 if (!PageDirty(page)) {
1023 /* someone wrote it for us */
1024 goto continue_unlock;
1027 if (PageWriteback(page)) {
1028 if (wbc->sync_mode != WB_SYNC_NONE)
1029 wait_on_page_writeback(page);
1031 goto continue_unlock;
1034 BUG_ON(PageWriteback(page));
1035 if (!clear_page_dirty_for_io(page))
1036 goto continue_unlock;
1038 ret = (*writepage)(page, wbc, data);
1039 if (unlikely(ret)) {
1040 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1045 * done_index is set past this page,
1046 * so media errors will not choke
1047 * background writeout for the entire
1048 * file. This has consequences for
1049 * range_cyclic semantics (ie. it may
1050 * not be suitable for data integrity
1058 if (nr_to_write > 0) {
1060 if (nr_to_write == 0 &&
1061 wbc->sync_mode == WB_SYNC_NONE) {
1063 * We stop writing back only if we are
1064 * not doing integrity sync. In case of
1065 * integrity sync we have to keep going
1066 * because someone may be concurrently
1067 * dirtying pages, and we might have
1068 * synced a lot of newly appeared dirty
1069 * pages, but have not synced all of the
1077 if (wbc->nonblocking && bdi_write_congested(bdi)) {
1078 wbc->encountered_congestion = 1;
1083 pagevec_release(&pvec);
1086 if (!cycled && !done) {
1089 * We hit the last page and there is more work to be done: wrap
1090 * back to the start of the file
1094 end = writeback_index - 1;
1097 if (!wbc->no_nrwrite_index_update) {
1098 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
1099 mapping->writeback_index = done_index;
1100 wbc->nr_to_write = nr_to_write;
1105 EXPORT_SYMBOL(write_cache_pages);
1108 * Function used by generic_writepages to call the real writepage
1109 * function and set the mapping flags on error
1111 static int __writepage(struct page *page, struct writeback_control *wbc,
1114 struct address_space *mapping = data;
1115 int ret = mapping->a_ops->writepage(page, wbc);
1116 mapping_set_error(mapping, ret);
1121 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1122 * @mapping: address space structure to write
1123 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1125 * This is a library function, which implements the writepages()
1126 * address_space_operation.
1128 int generic_writepages(struct address_space *mapping,
1129 struct writeback_control *wbc)
1131 /* deal with chardevs and other special file */
1132 if (!mapping->a_ops->writepage)
1135 return write_cache_pages(mapping, wbc, __writepage, mapping);
1138 EXPORT_SYMBOL(generic_writepages);
1140 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1144 if (wbc->nr_to_write <= 0)
1146 wbc->for_writepages = 1;
1147 if (mapping->a_ops->writepages)
1148 ret = mapping->a_ops->writepages(mapping, wbc);
1150 ret = generic_writepages(mapping, wbc);
1151 wbc->for_writepages = 0;
1156 * write_one_page - write out a single page and optionally wait on I/O
1157 * @page: the page to write
1158 * @wait: if true, wait on writeout
1160 * The page must be locked by the caller and will be unlocked upon return.
1162 * write_one_page() returns a negative error code if I/O failed.
1164 int write_one_page(struct page *page, int wait)
1166 struct address_space *mapping = page->mapping;
1168 struct writeback_control wbc = {
1169 .sync_mode = WB_SYNC_ALL,
1173 BUG_ON(!PageLocked(page));
1176 wait_on_page_writeback(page);
1178 if (clear_page_dirty_for_io(page)) {
1179 page_cache_get(page);
1180 ret = mapping->a_ops->writepage(page, &wbc);
1181 if (ret == 0 && wait) {
1182 wait_on_page_writeback(page);
1183 if (PageError(page))
1186 page_cache_release(page);
1192 EXPORT_SYMBOL(write_one_page);
1195 * For address_spaces which do not use buffers nor write back.
1197 int __set_page_dirty_no_writeback(struct page *page)
1199 if (!PageDirty(page))
1205 * Helper function for set_page_dirty family.
1206 * NOTE: This relies on being atomic wrt interrupts.
1208 void account_page_dirtied(struct page *page, struct address_space *mapping)
1210 if (mapping_cap_account_dirty(mapping)) {
1211 __inc_zone_page_state(page, NR_FILE_DIRTY);
1212 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1213 task_dirty_inc(current);
1214 task_io_account_write(PAGE_CACHE_SIZE);
1219 * For address_spaces which do not use buffers. Just tag the page as dirty in
1222 * This is also used when a single buffer is being dirtied: we want to set the
1223 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1224 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1226 * Most callers have locked the page, which pins the address_space in memory.
1227 * But zap_pte_range() does not lock the page, however in that case the
1228 * mapping is pinned by the vma's ->vm_file reference.
1230 * We take care to handle the case where the page was truncated from the
1231 * mapping by re-checking page_mapping() inside tree_lock.
1233 int __set_page_dirty_nobuffers(struct page *page)
1235 if (!TestSetPageDirty(page)) {
1236 struct address_space *mapping = page_mapping(page);
1237 struct address_space *mapping2;
1242 spin_lock_irq(&mapping->tree_lock);
1243 mapping2 = page_mapping(page);
1244 if (mapping2) { /* Race with truncate? */
1245 BUG_ON(mapping2 != mapping);
1246 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1247 account_page_dirtied(page, mapping);
1248 radix_tree_tag_set(&mapping->page_tree,
1249 page_index(page), PAGECACHE_TAG_DIRTY);
1251 spin_unlock_irq(&mapping->tree_lock);
1252 if (mapping->host) {
1253 /* !PageAnon && !swapper_space */
1254 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1260 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1263 * When a writepage implementation decides that it doesn't want to write this
1264 * page for some reason, it should redirty the locked page via
1265 * redirty_page_for_writepage() and it should then unlock the page and return 0
1267 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1269 wbc->pages_skipped++;
1270 return __set_page_dirty_nobuffers(page);
1272 EXPORT_SYMBOL(redirty_page_for_writepage);
1275 * If the mapping doesn't provide a set_page_dirty a_op, then
1276 * just fall through and assume that it wants buffer_heads.
1278 int set_page_dirty(struct page *page)
1280 struct address_space *mapping = page_mapping(page);
1282 if (likely(mapping)) {
1283 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1286 spd = __set_page_dirty_buffers;
1288 return (*spd)(page);
1290 if (!PageDirty(page)) {
1291 if (!TestSetPageDirty(page))
1296 EXPORT_SYMBOL(set_page_dirty);
1299 * set_page_dirty() is racy if the caller has no reference against
1300 * page->mapping->host, and if the page is unlocked. This is because another
1301 * CPU could truncate the page off the mapping and then free the mapping.
1303 * Usually, the page _is_ locked, or the caller is a user-space process which
1304 * holds a reference on the inode by having an open file.
1306 * In other cases, the page should be locked before running set_page_dirty().
1308 int set_page_dirty_lock(struct page *page)
1312 lock_page_nosync(page);
1313 ret = set_page_dirty(page);
1317 EXPORT_SYMBOL(set_page_dirty_lock);
1320 * Clear a page's dirty flag, while caring for dirty memory accounting.
1321 * Returns true if the page was previously dirty.
1323 * This is for preparing to put the page under writeout. We leave the page
1324 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1325 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1326 * implementation will run either set_page_writeback() or set_page_dirty(),
1327 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1330 * This incoherency between the page's dirty flag and radix-tree tag is
1331 * unfortunate, but it only exists while the page is locked.
1333 int clear_page_dirty_for_io(struct page *page)
1335 struct address_space *mapping = page_mapping(page);
1337 BUG_ON(!PageLocked(page));
1339 ClearPageReclaim(page);
1340 if (mapping && mapping_cap_account_dirty(mapping)) {
1342 * Yes, Virginia, this is indeed insane.
1344 * We use this sequence to make sure that
1345 * (a) we account for dirty stats properly
1346 * (b) we tell the low-level filesystem to
1347 * mark the whole page dirty if it was
1348 * dirty in a pagetable. Only to then
1349 * (c) clean the page again and return 1 to
1350 * cause the writeback.
1352 * This way we avoid all nasty races with the
1353 * dirty bit in multiple places and clearing
1354 * them concurrently from different threads.
1356 * Note! Normally the "set_page_dirty(page)"
1357 * has no effect on the actual dirty bit - since
1358 * that will already usually be set. But we
1359 * need the side effects, and it can help us
1362 * We basically use the page "master dirty bit"
1363 * as a serialization point for all the different
1364 * threads doing their things.
1366 if (page_mkclean(page))
1367 set_page_dirty(page);
1369 * We carefully synchronise fault handlers against
1370 * installing a dirty pte and marking the page dirty
1371 * at this point. We do this by having them hold the
1372 * page lock at some point after installing their
1373 * pte, but before marking the page dirty.
1374 * Pages are always locked coming in here, so we get
1375 * the desired exclusion. See mm/memory.c:do_wp_page()
1376 * for more comments.
1378 if (TestClearPageDirty(page)) {
1379 dec_zone_page_state(page, NR_FILE_DIRTY);
1380 dec_bdi_stat(mapping->backing_dev_info,
1386 return TestClearPageDirty(page);
1388 EXPORT_SYMBOL(clear_page_dirty_for_io);
1390 int test_clear_page_writeback(struct page *page)
1392 struct address_space *mapping = page_mapping(page);
1396 struct backing_dev_info *bdi = mapping->backing_dev_info;
1397 unsigned long flags;
1399 spin_lock_irqsave(&mapping->tree_lock, flags);
1400 ret = TestClearPageWriteback(page);
1402 radix_tree_tag_clear(&mapping->page_tree,
1404 PAGECACHE_TAG_WRITEBACK);
1405 if (bdi_cap_account_writeback(bdi)) {
1406 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1407 __bdi_writeout_inc(bdi);
1410 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1412 ret = TestClearPageWriteback(page);
1415 dec_zone_page_state(page, NR_WRITEBACK);
1419 int test_set_page_writeback(struct page *page)
1421 struct address_space *mapping = page_mapping(page);
1425 struct backing_dev_info *bdi = mapping->backing_dev_info;
1426 unsigned long flags;
1428 spin_lock_irqsave(&mapping->tree_lock, flags);
1429 ret = TestSetPageWriteback(page);
1431 radix_tree_tag_set(&mapping->page_tree,
1433 PAGECACHE_TAG_WRITEBACK);
1434 if (bdi_cap_account_writeback(bdi))
1435 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1437 if (!PageDirty(page))
1438 radix_tree_tag_clear(&mapping->page_tree,
1440 PAGECACHE_TAG_DIRTY);
1441 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1443 ret = TestSetPageWriteback(page);
1446 inc_zone_page_state(page, NR_WRITEBACK);
1450 EXPORT_SYMBOL(test_set_page_writeback);
1453 * Return true if any of the pages in the mapping are marked with the
1456 int mapping_tagged(struct address_space *mapping, int tag)
1460 ret = radix_tree_tagged(&mapping->page_tree, tag);
1464 EXPORT_SYMBOL(mapping_tagged);