4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains functions related to writing back dirty pages at the
9 * 10Apr2002 akpm@zip.com.au
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/spinlock.h>
18 #include <linux/swap.h>
19 #include <linux/slab.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/init.h>
23 #include <linux/backing-dev.h>
24 #include <linux/blkdev.h>
25 #include <linux/mpage.h>
26 #include <linux/percpu.h>
27 #include <linux/notifier.h>
28 #include <linux/smp.h>
29 #include <linux/sysctl.h>
30 #include <linux/cpu.h>
31 #include <linux/syscalls.h>
34 * The maximum number of pages to writeout in a single bdflush/kupdate
35 * operation. We do this so we don't hold I_LOCK against an inode for
36 * enormous amounts of time, which would block a userspace task which has
37 * been forced to throttle against that inode. Also, the code reevaluates
38 * the dirty each time it has written this many pages.
40 #define MAX_WRITEBACK_PAGES 1024
43 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
44 * will look to see if it needs to force writeback or throttling.
46 static long ratelimit_pages
= 32;
48 static long total_pages
; /* The total number of pages in the machine. */
49 static int dirty_exceeded __cacheline_aligned_in_smp
; /* Dirty mem may be over limit */
52 * When balance_dirty_pages decides that the caller needs to perform some
53 * non-background writeback, this is how many pages it will attempt to write.
54 * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
55 * large amounts of I/O are submitted.
57 static inline long sync_writeback_pages(void)
59 return ratelimit_pages
+ ratelimit_pages
/ 2;
62 /* The following parameters are exported via /proc/sys/vm */
65 * Start background writeback (via pdflush) at this percentage
67 int dirty_background_ratio
= 10;
70 * The generator of dirty data starts writeback at this percentage
72 int vm_dirty_ratio
= 40;
75 * The interval between `kupdate'-style writebacks, in centiseconds
76 * (hundredths of a second)
78 int dirty_writeback_interval
= 5 * HZ
;
81 * The longest number of centiseconds for which data is allowed to remain dirty
83 int dirty_expire_interval
= 30 * HZ
;
86 * Flag that makes the machine dump writes/reads and block dirtyings.
91 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
92 * a full sync is triggered after this time elapses without any disk activity.
96 EXPORT_SYMBOL(laptop_mode
);
98 /* End of sysctl-exported parameters */
101 static void background_writeout(unsigned long _min_pages
);
103 struct writeback_state
105 unsigned long nr_dirty
;
106 unsigned long nr_unstable
;
107 unsigned long nr_mapped
;
108 unsigned long nr_writeback
;
111 static void get_writeback_state(struct writeback_state
*wbs
)
113 wbs
->nr_dirty
= read_page_state(nr_dirty
);
114 wbs
->nr_unstable
= read_page_state(nr_unstable
);
115 wbs
->nr_mapped
= read_page_state(nr_mapped
);
116 wbs
->nr_writeback
= read_page_state(nr_writeback
);
120 * Work out the current dirty-memory clamping and background writeout
123 * The main aim here is to lower them aggressively if there is a lot of mapped
124 * memory around. To avoid stressing page reclaim with lots of unreclaimable
125 * pages. It is better to clamp down on writers than to start swapping, and
126 * performing lots of scanning.
128 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
130 * We don't permit the clamping level to fall below 5% - that is getting rather
133 * We make sure that the background writeout level is below the adjusted
137 get_dirty_limits(struct writeback_state
*wbs
, long *pbackground
, long *pdirty
,
138 struct address_space
*mapping
)
140 int background_ratio
; /* Percentages */
145 unsigned long available_memory
= total_pages
;
146 struct task_struct
*tsk
;
148 get_writeback_state(wbs
);
150 #ifdef CONFIG_HIGHMEM
152 * If this mapping can only allocate from low memory,
153 * we exclude high memory from our count.
155 if (mapping
&& !(mapping_gfp_mask(mapping
) & __GFP_HIGHMEM
))
156 available_memory
-= totalhigh_pages
;
160 unmapped_ratio
= 100 - (wbs
->nr_mapped
* 100) / total_pages
;
162 dirty_ratio
= vm_dirty_ratio
;
163 if (dirty_ratio
> unmapped_ratio
/ 2)
164 dirty_ratio
= unmapped_ratio
/ 2;
169 background_ratio
= dirty_background_ratio
;
170 if (background_ratio
>= dirty_ratio
)
171 background_ratio
= dirty_ratio
/ 2;
173 background
= (background_ratio
* available_memory
) / 100;
174 dirty
= (dirty_ratio
* available_memory
) / 100;
176 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
)) {
177 background
+= background
/ 4;
180 *pbackground
= background
;
185 * balance_dirty_pages() must be called by processes which are generating dirty
186 * data. It looks at the number of dirty pages in the machine and will force
187 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
188 * If we're over `background_thresh' then pdflush is woken to perform some
191 static void balance_dirty_pages(struct address_space
*mapping
)
193 struct writeback_state wbs
;
195 long background_thresh
;
197 unsigned long pages_written
= 0;
198 unsigned long write_chunk
= sync_writeback_pages();
200 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
203 struct writeback_control wbc
= {
205 .sync_mode
= WB_SYNC_NONE
,
206 .older_than_this
= NULL
,
207 .nr_to_write
= write_chunk
,
210 get_dirty_limits(&wbs
, &background_thresh
,
211 &dirty_thresh
, mapping
);
212 nr_reclaimable
= wbs
.nr_dirty
+ wbs
.nr_unstable
;
213 if (nr_reclaimable
+ wbs
.nr_writeback
<= dirty_thresh
)
219 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
220 * Unstable writes are a feature of certain networked
221 * filesystems (i.e. NFS) in which data may have been
222 * written to the server's write cache, but has not yet
223 * been flushed to permanent storage.
225 if (nr_reclaimable
) {
226 writeback_inodes(&wbc
);
227 get_dirty_limits(&wbs
, &background_thresh
,
228 &dirty_thresh
, mapping
);
229 nr_reclaimable
= wbs
.nr_dirty
+ wbs
.nr_unstable
;
230 if (nr_reclaimable
+ wbs
.nr_writeback
<= dirty_thresh
)
232 pages_written
+= write_chunk
- wbc
.nr_to_write
;
233 if (pages_written
>= write_chunk
)
234 break; /* We've done our duty */
236 blk_congestion_wait(WRITE
, HZ
/10);
239 if (nr_reclaimable
+ wbs
.nr_writeback
<= dirty_thresh
&& dirty_exceeded
)
242 if (writeback_in_progress(bdi
))
243 return; /* pdflush is already working this queue */
246 * In laptop mode, we wait until hitting the higher threshold before
247 * starting background writeout, and then write out all the way down
248 * to the lower threshold. So slow writers cause minimal disk activity.
250 * In normal mode, we start background writeout at the lower
251 * background_thresh, to keep the amount of dirty memory low.
253 if ((laptop_mode
&& pages_written
) ||
254 (!laptop_mode
&& (nr_reclaimable
> background_thresh
)))
255 pdflush_operation(background_writeout
, 0);
259 * balance_dirty_pages_ratelimited - balance dirty memory state
260 * @mapping: address_space which was dirtied
262 * Processes which are dirtying memory should call in here once for each page
263 * which was newly dirtied. The function will periodically check the system's
264 * dirty state and will initiate writeback if needed.
266 * On really big machines, get_writeback_state is expensive, so try to avoid
267 * calling it too often (ratelimiting). But once we're over the dirty memory
268 * limit we decrease the ratelimiting by a lot, to prevent individual processes
269 * from overshooting the limit by (ratelimit_pages) each.
271 void balance_dirty_pages_ratelimited(struct address_space
*mapping
)
273 static DEFINE_PER_CPU(int, ratelimits
) = 0;
276 ratelimit
= ratelimit_pages
;
281 * Check the rate limiting. Also, we do not want to throttle real-time
282 * tasks in balance_dirty_pages(). Period.
284 if (get_cpu_var(ratelimits
)++ >= ratelimit
) {
285 __get_cpu_var(ratelimits
) = 0;
286 put_cpu_var(ratelimits
);
287 balance_dirty_pages(mapping
);
290 put_cpu_var(ratelimits
);
292 EXPORT_SYMBOL(balance_dirty_pages_ratelimited
);
294 void throttle_vm_writeout(void)
296 struct writeback_state wbs
;
297 long background_thresh
;
301 get_dirty_limits(&wbs
, &background_thresh
, &dirty_thresh
, NULL
);
304 * Boost the allowable dirty threshold a bit for page
305 * allocators so they don't get DoS'ed by heavy writers
307 dirty_thresh
+= dirty_thresh
/ 10; /* wheeee... */
309 if (wbs
.nr_unstable
+ wbs
.nr_writeback
<= dirty_thresh
)
311 blk_congestion_wait(WRITE
, HZ
/10);
317 * writeback at least _min_pages, and keep writing until the amount of dirty
318 * memory is less than the background threshold, or until we're all clean.
320 static void background_writeout(unsigned long _min_pages
)
322 long min_pages
= _min_pages
;
323 struct writeback_control wbc
= {
325 .sync_mode
= WB_SYNC_NONE
,
326 .older_than_this
= NULL
,
332 struct writeback_state wbs
;
333 long background_thresh
;
336 get_dirty_limits(&wbs
, &background_thresh
, &dirty_thresh
, NULL
);
337 if (wbs
.nr_dirty
+ wbs
.nr_unstable
< background_thresh
340 wbc
.encountered_congestion
= 0;
341 wbc
.nr_to_write
= MAX_WRITEBACK_PAGES
;
342 wbc
.pages_skipped
= 0;
343 writeback_inodes(&wbc
);
344 min_pages
-= MAX_WRITEBACK_PAGES
- wbc
.nr_to_write
;
345 if (wbc
.nr_to_write
> 0 || wbc
.pages_skipped
> 0) {
346 /* Wrote less than expected */
347 blk_congestion_wait(WRITE
, HZ
/10);
348 if (!wbc
.encountered_congestion
)
355 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
356 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
357 * -1 if all pdflush threads were busy.
359 int wakeup_pdflush(long nr_pages
)
362 struct writeback_state wbs
;
364 get_writeback_state(&wbs
);
365 nr_pages
= wbs
.nr_dirty
+ wbs
.nr_unstable
;
367 return pdflush_operation(background_writeout
, nr_pages
);
370 static void wb_timer_fn(unsigned long unused
);
371 static void laptop_timer_fn(unsigned long unused
);
373 static DEFINE_TIMER(wb_timer
, wb_timer_fn
, 0, 0);
374 static DEFINE_TIMER(laptop_mode_wb_timer
, laptop_timer_fn
, 0, 0);
377 * Periodic writeback of "old" data.
379 * Define "old": the first time one of an inode's pages is dirtied, we mark the
380 * dirtying-time in the inode's address_space. So this periodic writeback code
381 * just walks the superblock inode list, writing back any inodes which are
382 * older than a specific point in time.
384 * Try to run once per dirty_writeback_interval. But if a writeback event
385 * takes longer than a dirty_writeback_interval interval, then leave a
388 * older_than_this takes precedence over nr_to_write. So we'll only write back
389 * all dirty pages if they are all attached to "old" mappings.
391 static void wb_kupdate(unsigned long arg
)
393 unsigned long oldest_jif
;
394 unsigned long start_jif
;
395 unsigned long next_jif
;
397 struct writeback_state wbs
;
398 struct writeback_control wbc
= {
400 .sync_mode
= WB_SYNC_NONE
,
401 .older_than_this
= &oldest_jif
,
409 get_writeback_state(&wbs
);
410 oldest_jif
= jiffies
- dirty_expire_interval
;
412 next_jif
= start_jif
+ dirty_writeback_interval
;
413 nr_to_write
= wbs
.nr_dirty
+ wbs
.nr_unstable
+
414 (inodes_stat
.nr_inodes
- inodes_stat
.nr_unused
);
415 while (nr_to_write
> 0) {
416 wbc
.encountered_congestion
= 0;
417 wbc
.nr_to_write
= MAX_WRITEBACK_PAGES
;
418 writeback_inodes(&wbc
);
419 if (wbc
.nr_to_write
> 0) {
420 if (wbc
.encountered_congestion
)
421 blk_congestion_wait(WRITE
, HZ
/10);
423 break; /* All the old data is written */
425 nr_to_write
-= MAX_WRITEBACK_PAGES
- wbc
.nr_to_write
;
427 if (time_before(next_jif
, jiffies
+ HZ
))
428 next_jif
= jiffies
+ HZ
;
429 if (dirty_writeback_interval
)
430 mod_timer(&wb_timer
, next_jif
);
434 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
436 int dirty_writeback_centisecs_handler(ctl_table
*table
, int write
,
437 struct file
*file
, void __user
*buffer
, size_t *length
, loff_t
*ppos
)
439 proc_dointvec_userhz_jiffies(table
, write
, file
, buffer
, length
, ppos
);
440 if (dirty_writeback_interval
) {
442 jiffies
+ dirty_writeback_interval
);
444 del_timer(&wb_timer
);
449 static void wb_timer_fn(unsigned long unused
)
451 if (pdflush_operation(wb_kupdate
, 0) < 0)
452 mod_timer(&wb_timer
, jiffies
+ HZ
); /* delay 1 second */
455 static void laptop_flush(unsigned long unused
)
460 static void laptop_timer_fn(unsigned long unused
)
462 pdflush_operation(laptop_flush
, 0);
466 * We've spun up the disk and we're in laptop mode: schedule writeback
467 * of all dirty data a few seconds from now. If the flush is already scheduled
468 * then push it back - the user is still using the disk.
470 void laptop_io_completion(void)
472 mod_timer(&laptop_mode_wb_timer
, jiffies
+ laptop_mode
);
476 * We're in laptop mode and we've just synced. The sync's writes will have
477 * caused another writeback to be scheduled by laptop_io_completion.
478 * Nothing needs to be written back anymore, so we unschedule the writeback.
480 void laptop_sync_completion(void)
482 del_timer(&laptop_mode_wb_timer
);
486 * If ratelimit_pages is too high then we can get into dirty-data overload
487 * if a large number of processes all perform writes at the same time.
488 * If it is too low then SMP machines will call the (expensive)
489 * get_writeback_state too often.
491 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
492 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
493 * thresholds before writeback cuts in.
495 * But the limit should not be set too high. Because it also controls the
496 * amount of memory which the balance_dirty_pages() caller has to write back.
497 * If this is too large then the caller will block on the IO queue all the
498 * time. So limit it to four megabytes - the balance_dirty_pages() caller
499 * will write six megabyte chunks, max.
502 static void set_ratelimit(void)
504 ratelimit_pages
= total_pages
/ (num_online_cpus() * 32);
505 if (ratelimit_pages
< 16)
506 ratelimit_pages
= 16;
507 if (ratelimit_pages
* PAGE_CACHE_SIZE
> 4096 * 1024)
508 ratelimit_pages
= (4096 * 1024) / PAGE_CACHE_SIZE
;
512 ratelimit_handler(struct notifier_block
*self
, unsigned long u
, void *v
)
518 static struct notifier_block ratelimit_nb
= {
519 .notifier_call
= ratelimit_handler
,
524 * If the machine has a large highmem:lowmem ratio then scale back the default
525 * dirty memory thresholds: allowing too much dirty highmem pins an excessive
526 * number of buffer_heads.
528 void __init
page_writeback_init(void)
530 long buffer_pages
= nr_free_buffer_pages();
533 total_pages
= nr_free_pagecache_pages();
535 correction
= (100 * 4 * buffer_pages
) / total_pages
;
537 if (correction
< 100) {
538 dirty_background_ratio
*= correction
;
539 dirty_background_ratio
/= 100;
540 vm_dirty_ratio
*= correction
;
541 vm_dirty_ratio
/= 100;
543 if (dirty_background_ratio
<= 0)
544 dirty_background_ratio
= 1;
545 if (vm_dirty_ratio
<= 0)
548 mod_timer(&wb_timer
, jiffies
+ dirty_writeback_interval
);
550 register_cpu_notifier(&ratelimit_nb
);
553 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
557 if (wbc
->nr_to_write
<= 0)
559 wbc
->for_writepages
= 1;
560 if (mapping
->a_ops
->writepages
)
561 ret
= mapping
->a_ops
->writepages(mapping
, wbc
);
563 ret
= generic_writepages(mapping
, wbc
);
564 wbc
->for_writepages
= 0;
569 * write_one_page - write out a single page and optionally wait on I/O
571 * @page: the page to write
572 * @wait: if true, wait on writeout
574 * The page must be locked by the caller and will be unlocked upon return.
576 * write_one_page() returns a negative error code if I/O failed.
578 int write_one_page(struct page
*page
, int wait
)
580 struct address_space
*mapping
= page
->mapping
;
582 struct writeback_control wbc
= {
583 .sync_mode
= WB_SYNC_ALL
,
587 BUG_ON(!PageLocked(page
));
590 wait_on_page_writeback(page
);
592 if (clear_page_dirty_for_io(page
)) {
593 page_cache_get(page
);
594 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
595 if (ret
== 0 && wait
) {
596 wait_on_page_writeback(page
);
600 page_cache_release(page
);
606 EXPORT_SYMBOL(write_one_page
);
609 * For address_spaces which do not use buffers. Just tag the page as dirty in
612 * This is also used when a single buffer is being dirtied: we want to set the
613 * page dirty in that case, but not all the buffers. This is a "bottom-up"
614 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
616 * Most callers have locked the page, which pins the address_space in memory.
617 * But zap_pte_range() does not lock the page, however in that case the
618 * mapping is pinned by the vma's ->vm_file reference.
620 * We take care to handle the case where the page was truncated from the
621 * mapping by re-checking page_mapping() insode tree_lock.
623 int __set_page_dirty_nobuffers(struct page
*page
)
627 if (!TestSetPageDirty(page
)) {
628 struct address_space
*mapping
= page_mapping(page
);
629 struct address_space
*mapping2
;
632 write_lock_irq(&mapping
->tree_lock
);
633 mapping2
= page_mapping(page
);
634 if (mapping2
) { /* Race with truncate? */
635 BUG_ON(mapping2
!= mapping
);
636 if (mapping_cap_account_dirty(mapping
))
637 inc_page_state(nr_dirty
);
638 radix_tree_tag_set(&mapping
->page_tree
,
639 page_index(page
), PAGECACHE_TAG_DIRTY
);
641 write_unlock_irq(&mapping
->tree_lock
);
643 /* !PageAnon && !swapper_space */
644 __mark_inode_dirty(mapping
->host
,
651 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
654 * When a writepage implementation decides that it doesn't want to write this
655 * page for some reason, it should redirty the locked page via
656 * redirty_page_for_writepage() and it should then unlock the page and return 0
658 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
660 wbc
->pages_skipped
++;
661 return __set_page_dirty_nobuffers(page
);
663 EXPORT_SYMBOL(redirty_page_for_writepage
);
666 * If the mapping doesn't provide a set_page_dirty a_op, then
667 * just fall through and assume that it wants buffer_heads.
669 int fastcall
set_page_dirty(struct page
*page
)
671 struct address_space
*mapping
= page_mapping(page
);
673 if (likely(mapping
)) {
674 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
677 return __set_page_dirty_buffers(page
);
679 if (!PageDirty(page
))
683 EXPORT_SYMBOL(set_page_dirty
);
686 * set_page_dirty() is racy if the caller has no reference against
687 * page->mapping->host, and if the page is unlocked. This is because another
688 * CPU could truncate the page off the mapping and then free the mapping.
690 * Usually, the page _is_ locked, or the caller is a user-space process which
691 * holds a reference on the inode by having an open file.
693 * In other cases, the page should be locked before running set_page_dirty().
695 int set_page_dirty_lock(struct page
*page
)
700 ret
= set_page_dirty(page
);
704 EXPORT_SYMBOL(set_page_dirty_lock
);
707 * Clear a page's dirty flag, while caring for dirty memory accounting.
708 * Returns true if the page was previously dirty.
710 int test_clear_page_dirty(struct page
*page
)
712 struct address_space
*mapping
= page_mapping(page
);
716 write_lock_irqsave(&mapping
->tree_lock
, flags
);
717 if (TestClearPageDirty(page
)) {
718 radix_tree_tag_clear(&mapping
->page_tree
,
720 PAGECACHE_TAG_DIRTY
);
721 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
722 if (mapping_cap_account_dirty(mapping
))
723 dec_page_state(nr_dirty
);
726 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
729 return TestClearPageDirty(page
);
731 EXPORT_SYMBOL(test_clear_page_dirty
);
734 * Clear a page's dirty flag, while caring for dirty memory accounting.
735 * Returns true if the page was previously dirty.
737 * This is for preparing to put the page under writeout. We leave the page
738 * tagged as dirty in the radix tree so that a concurrent write-for-sync
739 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
740 * implementation will run either set_page_writeback() or set_page_dirty(),
741 * at which stage we bring the page's dirty flag and radix-tree dirty tag
744 * This incoherency between the page's dirty flag and radix-tree tag is
745 * unfortunate, but it only exists while the page is locked.
747 int clear_page_dirty_for_io(struct page
*page
)
749 struct address_space
*mapping
= page_mapping(page
);
752 if (TestClearPageDirty(page
)) {
753 if (mapping_cap_account_dirty(mapping
))
754 dec_page_state(nr_dirty
);
759 return TestClearPageDirty(page
);
761 EXPORT_SYMBOL(clear_page_dirty_for_io
);
763 int test_clear_page_writeback(struct page
*page
)
765 struct address_space
*mapping
= page_mapping(page
);
771 write_lock_irqsave(&mapping
->tree_lock
, flags
);
772 ret
= TestClearPageWriteback(page
);
774 radix_tree_tag_clear(&mapping
->page_tree
,
776 PAGECACHE_TAG_WRITEBACK
);
777 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
779 ret
= TestClearPageWriteback(page
);
784 int test_set_page_writeback(struct page
*page
)
786 struct address_space
*mapping
= page_mapping(page
);
792 write_lock_irqsave(&mapping
->tree_lock
, flags
);
793 ret
= TestSetPageWriteback(page
);
795 radix_tree_tag_set(&mapping
->page_tree
,
797 PAGECACHE_TAG_WRITEBACK
);
798 if (!PageDirty(page
))
799 radix_tree_tag_clear(&mapping
->page_tree
,
801 PAGECACHE_TAG_DIRTY
);
802 write_unlock_irqrestore(&mapping
->tree_lock
, flags
);
804 ret
= TestSetPageWriteback(page
);
809 EXPORT_SYMBOL(test_set_page_writeback
);
812 * Return true if any of the pages in the mapping are marged with the
815 int mapping_tagged(struct address_space
*mapping
, int tag
)
820 read_lock_irqsave(&mapping
->tree_lock
, flags
);
821 ret
= radix_tree_tagged(&mapping
->page_tree
, tag
);
822 read_unlock_irqrestore(&mapping
->tree_lock
, flags
);
825 EXPORT_SYMBOL(mapping_tagged
);