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; /* 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_centisecs = 5 * 100;
81 * The longest number of centiseconds for which data is allowed to remain dirty
83 int dirty_expire_centisecs = 30 * 100;
86 * Flag that makes the machine dump writes/reads and block dirtyings.
91 * Flag that puts the machine in "laptop mode".
95 /* End of sysctl-exported parameters */
98 static void background_writeout(unsigned long _min_pages);
101 * Work out the current dirty-memory clamping and background writeout
104 * The main aim here is to lower them aggressively if there is a lot of mapped
105 * memory around. To avoid stressing page reclaim with lots of unreclaimable
106 * pages. It is better to clamp down on writers than to start swapping, and
107 * performing lots of scanning.
109 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
111 * We don't permit the clamping level to fall below 5% - that is getting rather
114 * We make sure that the background writeout level is below the adjusted
118 get_dirty_limits(struct page_state *ps, long *pbackground, long *pdirty)
120 int background_ratio; /* Percentages */
125 struct task_struct *tsk;
129 unmapped_ratio = 100 - (ps->nr_mapped * 100) / total_pages;
131 dirty_ratio = vm_dirty_ratio;
132 if (dirty_ratio > unmapped_ratio / 2)
133 dirty_ratio = unmapped_ratio / 2;
138 background_ratio = dirty_background_ratio;
139 if (background_ratio >= dirty_ratio)
140 background_ratio = dirty_ratio / 2;
142 background = (background_ratio * total_pages) / 100;
143 dirty = (dirty_ratio * total_pages) / 100;
145 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
146 background += background / 4;
149 *pbackground = background;
154 * balance_dirty_pages() must be called by processes which are generating dirty
155 * data. It looks at the number of dirty pages in the machine and will force
156 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
157 * If we're over `background_thresh' then pdflush is woken to perform some
160 static void balance_dirty_pages(struct address_space *mapping)
162 struct page_state ps;
164 long background_thresh;
166 unsigned long pages_written = 0;
167 unsigned long write_chunk = sync_writeback_pages();
169 struct backing_dev_info *bdi = mapping->backing_dev_info;
172 struct writeback_control wbc = {
174 .sync_mode = WB_SYNC_NONE,
175 .older_than_this = NULL,
176 .nr_to_write = write_chunk,
179 get_dirty_limits(&ps, &background_thresh, &dirty_thresh);
180 nr_reclaimable = ps.nr_dirty + ps.nr_unstable;
181 if (nr_reclaimable + ps.nr_writeback <= dirty_thresh)
186 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
187 * Unstable writes are a feature of certain networked
188 * filesystems (i.e. NFS) in which data may have been
189 * written to the server's write cache, but has not yet
190 * been flushed to permanent storage.
192 if (nr_reclaimable) {
193 writeback_inodes(&wbc);
194 get_dirty_limits(&ps, &background_thresh,
196 nr_reclaimable = ps.nr_dirty + ps.nr_unstable;
197 if (nr_reclaimable + ps.nr_writeback <= dirty_thresh)
199 pages_written += write_chunk - wbc.nr_to_write;
200 if (pages_written >= write_chunk)
201 break; /* We've done our duty */
203 blk_congestion_wait(WRITE, HZ/10);
206 if (nr_reclaimable + ps.nr_writeback <= dirty_thresh)
209 if (writeback_in_progress(bdi))
210 return; /* pdflush is already working this queue */
213 * In laptop mode, we wait until hitting the higher threshold before
214 * starting background writeout, and then write out all the way down
215 * to the lower threshold. So slow writers cause minimal disk activity.
217 * In normal mode, we start background writeout at the lower
218 * background_thresh, to keep the amount of dirty memory low.
220 if ((laptop_mode && pages_written) ||
221 (!laptop_mode && (nr_reclaimable > background_thresh)))
222 pdflush_operation(background_writeout, 0);
226 * balance_dirty_pages_ratelimited - balance dirty memory state
227 * @mapping - address_space which was dirtied
229 * Processes which are dirtying memory should call in here once for each page
230 * which was newly dirtied. The function will periodically check the system's
231 * dirty state and will initiate writeback if needed.
233 * On really big machines, get_page_state is expensive, so try to avoid calling
234 * it too often (ratelimiting). But once we're over the dirty memory limit we
235 * decrease the ratelimiting by a lot, to prevent individual processes from
236 * overshooting the limit by (ratelimit_pages) each.
238 void balance_dirty_pages_ratelimited(struct address_space *mapping)
240 static DEFINE_PER_CPU(int, ratelimits) = 0;
243 ratelimit = ratelimit_pages;
248 * Check the rate limiting. Also, we do not want to throttle real-time
249 * tasks in balance_dirty_pages(). Period.
251 if (get_cpu_var(ratelimits)++ >= ratelimit) {
252 __get_cpu_var(ratelimits) = 0;
253 put_cpu_var(ratelimits);
254 balance_dirty_pages(mapping);
257 put_cpu_var(ratelimits);
259 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
262 * writeback at least _min_pages, and keep writing until the amount of dirty
263 * memory is less than the background threshold, or until we're all clean.
265 static void background_writeout(unsigned long _min_pages)
267 long min_pages = _min_pages;
268 struct writeback_control wbc = {
270 .sync_mode = WB_SYNC_NONE,
271 .older_than_this = NULL,
277 struct page_state ps;
278 long background_thresh;
281 get_dirty_limits(&ps, &background_thresh, &dirty_thresh);
282 if (ps.nr_dirty + ps.nr_unstable < background_thresh
285 wbc.encountered_congestion = 0;
286 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
287 wbc.pages_skipped = 0;
288 writeback_inodes(&wbc);
289 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
290 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
291 /* Wrote less than expected */
292 blk_congestion_wait(WRITE, HZ/10);
293 if (!wbc.encountered_congestion)
300 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
301 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
302 * -1 if all pdflush threads were busy.
304 int wakeup_bdflush(long nr_pages)
307 struct page_state ps;
310 nr_pages = ps.nr_dirty + ps.nr_unstable;
312 return pdflush_operation(background_writeout, nr_pages);
315 static void wb_timer_fn(unsigned long unused);
316 static void laptop_timer_fn(unsigned long unused);
318 static struct timer_list wb_timer =
319 TIMER_INITIALIZER(wb_timer_fn, 0, 0);
320 static struct timer_list laptop_mode_wb_timer =
321 TIMER_INITIALIZER(laptop_timer_fn, 0, 0);
324 * Periodic writeback of "old" data.
326 * Define "old": the first time one of an inode's pages is dirtied, we mark the
327 * dirtying-time in the inode's address_space. So this periodic writeback code
328 * just walks the superblock inode list, writing back any inodes which are
329 * older than a specific point in time.
331 * Try to run once per dirty_writeback_centisecs. But if a writeback event
332 * takes longer than a dirty_writeback_centisecs interval, then leave a
335 * older_than_this takes precedence over nr_to_write. So we'll only write back
336 * all dirty pages if they are all attached to "old" mappings.
338 static void wb_kupdate(unsigned long arg)
340 unsigned long oldest_jif;
341 unsigned long start_jif;
342 unsigned long next_jif;
344 struct page_state ps;
345 struct writeback_control wbc = {
347 .sync_mode = WB_SYNC_NONE,
348 .older_than_this = &oldest_jif,
357 oldest_jif = jiffies - (dirty_expire_centisecs * HZ) / 100;
359 next_jif = start_jif + (dirty_writeback_centisecs * HZ) / 100;
360 nr_to_write = ps.nr_dirty + ps.nr_unstable +
361 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
362 while (nr_to_write > 0) {
363 wbc.encountered_congestion = 0;
364 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
365 writeback_inodes(&wbc);
366 if (wbc.nr_to_write > 0) {
367 if (wbc.encountered_congestion)
368 blk_congestion_wait(WRITE, HZ/10);
370 break; /* All the old data is written */
372 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
374 if (time_before(next_jif, jiffies + HZ))
375 next_jif = jiffies + HZ;
376 if (dirty_writeback_centisecs)
377 mod_timer(&wb_timer, next_jif);
381 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
383 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
384 struct file *file, void __user *buffer, size_t *length)
386 proc_dointvec(table, write, file, buffer, length);
387 if (dirty_writeback_centisecs) {
389 jiffies + (dirty_writeback_centisecs * HZ) / 100);
391 del_timer(&wb_timer);
396 static void wb_timer_fn(unsigned long unused)
398 if (pdflush_operation(wb_kupdate, 0) < 0)
399 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
402 static void laptop_flush(unsigned long unused)
407 static void laptop_timer_fn(unsigned long unused)
409 pdflush_operation(laptop_flush, 0);
413 * We've spun up the disk and we're in laptop mode: schedule writeback
414 * of all dirty data a few seconds from now. If the flush is already scheduled
415 * then push it back - the user is still using the disk.
417 void laptop_io_completion(void)
419 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode * HZ);
423 * We're in laptop mode and we've just synced. The sync's writes will have
424 * caused another writeback to be scheduled by laptop_io_completion.
425 * Nothing needs to be written back anymore, so we unschedule the writeback.
427 void laptop_sync_completion(void)
429 del_timer(&laptop_mode_wb_timer);
433 * If ratelimit_pages is too high then we can get into dirty-data overload
434 * if a large number of processes all perform writes at the same time.
435 * If it is too low then SMP machines will call the (expensive) get_page_state
438 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
439 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
440 * thresholds before writeback cuts in.
442 * But the limit should not be set too high. Because it also controls the
443 * amount of memory which the balance_dirty_pages() caller has to write back.
444 * If this is too large then the caller will block on the IO queue all the
445 * time. So limit it to four megabytes - the balance_dirty_pages() caller
446 * will write six megabyte chunks, max.
449 static void set_ratelimit(void)
451 ratelimit_pages = total_pages / (num_online_cpus() * 32);
452 if (ratelimit_pages < 16)
453 ratelimit_pages = 16;
454 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
455 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
459 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
465 static struct notifier_block ratelimit_nb = {
466 .notifier_call = ratelimit_handler,
471 * If the machine has a large highmem:lowmem ratio then scale back the default
472 * dirty memory thresholds: allowing too much dirty highmem pins an excessive
473 * number of buffer_heads.
475 void __init page_writeback_init(void)
477 long buffer_pages = nr_free_buffer_pages();
480 total_pages = nr_free_pagecache_pages();
482 correction = (100 * 4 * buffer_pages) / total_pages;
484 if (correction < 100) {
485 dirty_background_ratio *= correction;
486 dirty_background_ratio /= 100;
487 vm_dirty_ratio *= correction;
488 vm_dirty_ratio /= 100;
490 mod_timer(&wb_timer, jiffies + (dirty_writeback_centisecs * HZ) / 100);
492 register_cpu_notifier(&ratelimit_nb);
495 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
497 if (wbc->nr_to_write <= 0)
499 if (mapping->a_ops->writepages)
500 return mapping->a_ops->writepages(mapping, wbc);
501 return generic_writepages(mapping, wbc);
505 * write_one_page - write out a single page and optionally wait on I/O
507 * @page - the page to write
508 * @wait - if true, wait on writeout
510 * The page must be locked by the caller and will be unlocked upon return.
512 * write_one_page() returns a negative error code if I/O failed.
514 int write_one_page(struct page *page, int wait)
516 struct address_space *mapping = page->mapping;
518 struct writeback_control wbc = {
519 .sync_mode = WB_SYNC_ALL,
523 BUG_ON(!PageLocked(page));
526 wait_on_page_writeback(page);
528 if (clear_page_dirty_for_io(page)) {
529 page_cache_get(page);
530 ret = mapping->a_ops->writepage(page, &wbc);
531 if (ret == 0 && wait) {
532 wait_on_page_writeback(page);
536 page_cache_release(page);
542 EXPORT_SYMBOL(write_one_page);
545 * For address_spaces which do not use buffers. Just tag the page as dirty in
548 * __set_page_dirty_nobuffers() may return -ENOSPC. But if it does, the page
549 * is still safe, as long as it actually manages to find some blocks at
552 * This is also used when a single buffer is being dirtied: we want to set the
553 * page dirty in that case, but not all the buffers. This is a "bottom-up"
554 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
556 int __set_page_dirty_nobuffers(struct page *page)
560 if (!TestSetPageDirty(page)) {
561 struct address_space *mapping = page->mapping;
564 spin_lock_irq(&mapping->tree_lock);
565 if (page->mapping) { /* Race with truncate? */
566 BUG_ON(page->mapping != mapping);
567 if (!mapping->backing_dev_info->memory_backed)
568 inc_page_state(nr_dirty);
569 radix_tree_tag_set(&mapping->page_tree,
570 page->index, PAGECACHE_TAG_DIRTY);
572 spin_unlock_irq(&mapping->tree_lock);
573 if (!PageSwapCache(page))
574 __mark_inode_dirty(mapping->host,
580 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
583 * When a writepage implementation decides that it doesn't want to write this
584 * page for some reason, it should redirty the locked page via
585 * redirty_page_for_writepage() and it should then unlock the page and return 0
587 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
589 wbc->pages_skipped++;
590 return __set_page_dirty_nobuffers(page);
592 EXPORT_SYMBOL(redirty_page_for_writepage);
595 * If the mapping doesn't provide a set_page_dirty a_op, then
596 * just fall through and assume that it wants buffer_heads.
598 int fastcall set_page_dirty(struct page *page)
600 struct address_space *mapping = page_mapping(page);
601 int (*spd)(struct page *);
607 spd = mapping->a_ops->set_page_dirty;
608 return spd? (*spd)(page): __set_page_dirty_buffers(page);
610 EXPORT_SYMBOL(set_page_dirty);
613 * set_page_dirty() is racy if the caller has no reference against
614 * page->mapping->host, and if the page is unlocked. This is because another
615 * CPU could truncate the page off the mapping and then free the mapping.
617 * Usually, the page _is_ locked, or the caller is a user-space process which
618 * holds a reference on the inode by having an open file.
620 * In other cases, the page should be locked before running set_page_dirty().
622 int set_page_dirty_lock(struct page *page)
627 ret = set_page_dirty(page);
631 EXPORT_SYMBOL(set_page_dirty_lock);
634 * Clear a page's dirty flag, while caring for dirty memory accounting.
635 * Returns true if the page was previously dirty.
637 int test_clear_page_dirty(struct page *page)
639 struct address_space *mapping = page_mapping(page);
643 spin_lock_irqsave(&mapping->tree_lock, flags);
644 if (TestClearPageDirty(page)) {
645 radix_tree_tag_clear(&mapping->page_tree, page->index,
646 PAGECACHE_TAG_DIRTY);
647 spin_unlock_irqrestore(&mapping->tree_lock, flags);
648 if (!mapping->backing_dev_info->memory_backed)
649 dec_page_state(nr_dirty);
652 spin_unlock_irqrestore(&mapping->tree_lock, flags);
655 return TestClearPageDirty(page);
657 EXPORT_SYMBOL(test_clear_page_dirty);
660 * Clear a page's dirty flag, while caring for dirty memory accounting.
661 * Returns true if the page was previously dirty.
663 * This is for preparing to put the page under writeout. We leave the page
664 * tagged as dirty in the radix tree so that a concurrent write-for-sync
665 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
666 * implementation will run either set_page_writeback() or set_page_dirty(),
667 * at which stage we bring the page's dirty flag and radix-tree dirty tag
670 * This incoherency between the page's dirty flag and radix-tree tag is
671 * unfortunate, but it only exists while the page is locked.
673 int clear_page_dirty_for_io(struct page *page)
675 struct address_space *mapping = page_mapping(page);
678 if (TestClearPageDirty(page)) {
679 if (!mapping->backing_dev_info->memory_backed)
680 dec_page_state(nr_dirty);
685 return TestClearPageDirty(page);
687 EXPORT_SYMBOL(clear_page_dirty_for_io);
690 * Clear a page's dirty flag while ignoring dirty memory accounting
692 int __clear_page_dirty(struct page *page)
694 struct address_space *mapping = page_mapping(page);
699 spin_lock_irqsave(&mapping->tree_lock, flags);
700 if (TestClearPageDirty(page)) {
701 radix_tree_tag_clear(&mapping->page_tree, page->index,
702 PAGECACHE_TAG_DIRTY);
703 spin_unlock_irqrestore(&mapping->tree_lock, flags);
706 spin_unlock_irqrestore(&mapping->tree_lock, flags);
709 return TestClearPageDirty(page);
712 int test_clear_page_writeback(struct page *page)
714 struct address_space *mapping = page_mapping(page);
720 spin_lock_irqsave(&mapping->tree_lock, flags);
721 ret = TestClearPageWriteback(page);
723 radix_tree_tag_clear(&mapping->page_tree, page->index,
724 PAGECACHE_TAG_WRITEBACK);
725 spin_unlock_irqrestore(&mapping->tree_lock, flags);
727 ret = TestClearPageWriteback(page);
732 int test_set_page_writeback(struct page *page)
734 struct address_space *mapping = page_mapping(page);
740 spin_lock_irqsave(&mapping->tree_lock, flags);
741 ret = TestSetPageWriteback(page);
743 radix_tree_tag_set(&mapping->page_tree, page->index,
744 PAGECACHE_TAG_WRITEBACK);
745 if (!PageDirty(page))
746 radix_tree_tag_clear(&mapping->page_tree, page->index,
747 PAGECACHE_TAG_DIRTY);
748 spin_unlock_irqrestore(&mapping->tree_lock, flags);
750 ret = TestSetPageWriteback(page);
755 EXPORT_SYMBOL(test_set_page_writeback);
758 * Return true if any of the pages in the mapping are marged with the
761 int mapping_tagged(struct address_space *mapping, int tag)
766 spin_lock_irqsave(&mapping->tree_lock, flags);
767 ret = radix_tree_tagged(&mapping->page_tree, tag);
768 spin_unlock_irqrestore(&mapping->tree_lock, flags);
771 EXPORT_SYMBOL(mapping_tagged);