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 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
41 * will look to see if it needs to force writeback or throttling.
43 static long ratelimit_pages = 32;
46 * When balance_dirty_pages decides that the caller needs to perform some
47 * non-background writeback, this is how many pages it will attempt to write.
48 * It should be somewhat larger than dirtied pages to ensure that reasonably
49 * large amounts of I/O are submitted.
51 static inline long sync_writeback_pages(unsigned long dirtied)
53 if (dirtied < ratelimit_pages)
54 dirtied = ratelimit_pages;
56 return dirtied + dirtied / 2;
59 /* The following parameters are exported via /proc/sys/vm */
62 * Start background writeback (via writeback threads) at this percentage
64 int dirty_background_ratio = 10;
67 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
68 * dirty_background_ratio * the amount of dirtyable memory
70 unsigned long dirty_background_bytes;
73 * free highmem will not be subtracted from the total free memory
74 * for calculating free ratios if vm_highmem_is_dirtyable is true
76 int vm_highmem_is_dirtyable;
79 * The generator of dirty data starts writeback at this percentage
81 int vm_dirty_ratio = 20;
84 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
85 * vm_dirty_ratio * the amount of dirtyable memory
87 unsigned long vm_dirty_bytes;
90 * The interval between `kupdate'-style writebacks
92 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
95 * The longest time for which data is allowed to remain dirty
97 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
100 * Flag that makes the machine dump writes/reads and block dirtyings.
105 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
106 * a full sync is triggered after this time elapses without any disk activity.
110 EXPORT_SYMBOL(laptop_mode);
112 /* End of sysctl-exported parameters */
116 * Scale the writeback cache size proportional to the relative writeout speeds.
118 * We do this by keeping a floating proportion between BDIs, based on page
119 * writeback completions [end_page_writeback()]. Those devices that write out
120 * pages fastest will get the larger share, while the slower will get a smaller
123 * We use page writeout completions because we are interested in getting rid of
124 * dirty pages. Having them written out is the primary goal.
126 * We introduce a concept of time, a period over which we measure these events,
127 * because demand can/will vary over time. The length of this period itself is
128 * measured in page writeback completions.
131 static struct prop_descriptor vm_completions;
132 static struct prop_descriptor vm_dirties;
135 * couple the period to the dirty_ratio:
137 * period/2 ~ roundup_pow_of_two(dirty limit)
139 static int calc_period_shift(void)
141 unsigned long dirty_total;
144 dirty_total = vm_dirty_bytes / PAGE_SIZE;
146 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
148 return 2 + ilog2(dirty_total - 1);
152 * update the period when the dirty threshold changes.
154 static void update_completion_period(void)
156 int shift = calc_period_shift();
157 prop_change_shift(&vm_completions, shift);
158 prop_change_shift(&vm_dirties, shift);
161 int dirty_background_ratio_handler(struct ctl_table *table, int write,
162 void __user *buffer, size_t *lenp,
167 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
168 if (ret == 0 && write)
169 dirty_background_bytes = 0;
173 int dirty_background_bytes_handler(struct ctl_table *table, int write,
174 void __user *buffer, size_t *lenp,
179 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
180 if (ret == 0 && write)
181 dirty_background_ratio = 0;
185 int dirty_ratio_handler(struct ctl_table *table, int write,
186 void __user *buffer, size_t *lenp,
189 int old_ratio = vm_dirty_ratio;
192 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
193 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
194 update_completion_period();
201 int dirty_bytes_handler(struct ctl_table *table, int write,
202 void __user *buffer, size_t *lenp,
205 unsigned long old_bytes = vm_dirty_bytes;
208 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
209 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
210 update_completion_period();
217 * Increment the BDI's writeout completion count and the global writeout
218 * completion count. Called from test_clear_page_writeback().
220 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
222 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
226 void bdi_writeout_inc(struct backing_dev_info *bdi)
230 local_irq_save(flags);
231 __bdi_writeout_inc(bdi);
232 local_irq_restore(flags);
234 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
236 void task_dirty_inc(struct task_struct *tsk)
238 prop_inc_single(&vm_dirties, &tsk->dirties);
242 * Obtain an accurate fraction of the BDI's portion.
244 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
245 long *numerator, long *denominator)
247 if (bdi_cap_writeback_dirty(bdi)) {
248 prop_fraction_percpu(&vm_completions, &bdi->completions,
249 numerator, denominator);
256 static inline void task_dirties_fraction(struct task_struct *tsk,
257 long *numerator, long *denominator)
259 prop_fraction_single(&vm_dirties, &tsk->dirties,
260 numerator, denominator);
264 * scale the dirty limit
266 * task specific dirty limit:
268 * dirty -= (dirty/8) * p_{t}
270 static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
272 long numerator, denominator;
273 unsigned long dirty = *pdirty;
274 u64 inv = dirty >> 3;
276 task_dirties_fraction(tsk, &numerator, &denominator);
278 do_div(inv, denominator);
281 if (dirty < *pdirty/2)
290 static unsigned int bdi_min_ratio;
292 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
296 spin_lock_bh(&bdi_lock);
297 if (min_ratio > bdi->max_ratio) {
300 min_ratio -= bdi->min_ratio;
301 if (bdi_min_ratio + min_ratio < 100) {
302 bdi_min_ratio += min_ratio;
303 bdi->min_ratio += min_ratio;
308 spin_unlock_bh(&bdi_lock);
313 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
320 spin_lock_bh(&bdi_lock);
321 if (bdi->min_ratio > max_ratio) {
324 bdi->max_ratio = max_ratio;
325 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
327 spin_unlock_bh(&bdi_lock);
331 EXPORT_SYMBOL(bdi_set_max_ratio);
334 * Work out the current dirty-memory clamping and background writeout
337 * The main aim here is to lower them aggressively if there is a lot of mapped
338 * memory around. To avoid stressing page reclaim with lots of unreclaimable
339 * pages. It is better to clamp down on writers than to start swapping, and
340 * performing lots of scanning.
342 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
344 * We don't permit the clamping level to fall below 5% - that is getting rather
347 * We make sure that the background writeout level is below the adjusted
351 static unsigned long highmem_dirtyable_memory(unsigned long total)
353 #ifdef CONFIG_HIGHMEM
357 for_each_node_state(node, N_HIGH_MEMORY) {
359 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
361 x += zone_page_state(z, NR_FREE_PAGES) +
362 zone_reclaimable_pages(z);
365 * Make sure that the number of highmem pages is never larger
366 * than the number of the total dirtyable memory. This can only
367 * occur in very strange VM situations but we want to make sure
368 * that this does not occur.
370 return min(x, total);
377 * determine_dirtyable_memory - amount of memory that may be used
379 * Returns the numebr of pages that can currently be freed and used
380 * by the kernel for direct mappings.
382 unsigned long determine_dirtyable_memory(void)
386 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
388 if (!vm_highmem_is_dirtyable)
389 x -= highmem_dirtyable_memory(x);
391 return x + 1; /* Ensure that we never return 0 */
395 get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
396 unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
398 unsigned long background;
400 unsigned long available_memory = determine_dirtyable_memory();
401 struct task_struct *tsk;
404 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
408 dirty_ratio = vm_dirty_ratio;
411 dirty = (dirty_ratio * available_memory) / 100;
414 if (dirty_background_bytes)
415 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
417 background = (dirty_background_ratio * available_memory) / 100;
419 if (background >= dirty)
420 background = dirty / 2;
422 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
423 background += background / 4;
426 *pbackground = background;
431 long numerator, denominator;
434 * Calculate this BDI's share of the dirty ratio.
436 bdi_writeout_fraction(bdi, &numerator, &denominator);
438 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
439 bdi_dirty *= numerator;
440 do_div(bdi_dirty, denominator);
441 bdi_dirty += (dirty * bdi->min_ratio) / 100;
442 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
443 bdi_dirty = dirty * bdi->max_ratio / 100;
445 *pbdi_dirty = bdi_dirty;
446 task_dirty_limit(current, pbdi_dirty);
451 * balance_dirty_pages() must be called by processes which are generating dirty
452 * data. It looks at the number of dirty pages in the machine and will force
453 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
454 * If we're over `background_thresh' then the writeback threads are woken to
455 * perform some writeout.
457 static void balance_dirty_pages(struct address_space *mapping,
458 unsigned long write_chunk)
460 long nr_reclaimable, bdi_nr_reclaimable;
461 long nr_writeback, bdi_nr_writeback;
462 unsigned long background_thresh;
463 unsigned long dirty_thresh;
464 unsigned long bdi_thresh;
465 unsigned long pages_written = 0;
466 unsigned long pause = 1;
467 bool dirty_exceeded = false;
468 struct backing_dev_info *bdi = mapping->backing_dev_info;
471 struct writeback_control wbc = {
472 .sync_mode = WB_SYNC_NONE,
473 .older_than_this = NULL,
474 .nr_to_write = write_chunk,
478 get_dirty_limits(&background_thresh, &dirty_thresh,
481 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
482 global_page_state(NR_UNSTABLE_NFS);
483 nr_writeback = global_page_state(NR_WRITEBACK);
486 * In order to avoid the stacked BDI deadlock we need
487 * to ensure we accurately count the 'dirty' pages when
488 * the threshold is low.
490 * Otherwise it would be possible to get thresh+n pages
491 * reported dirty, even though there are thresh-m pages
492 * actually dirty; with m+n sitting in the percpu
495 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
496 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
497 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
499 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
500 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
504 * The bdi thresh is somehow "soft" limit derived from the
505 * global "hard" limit. The former helps to prevent heavy IO
506 * bdi or process from holding back light ones; The latter is
507 * the last resort safeguard.
510 (bdi_nr_reclaimable + bdi_nr_writeback >= bdi_thresh)
511 || (nr_reclaimable + nr_writeback >= dirty_thresh);
517 * Throttle it only when the background writeback cannot
518 * catch-up. This avoids (excessively) small writeouts
519 * when the bdi limits are ramping up.
521 if (nr_reclaimable + nr_writeback <
522 (background_thresh + dirty_thresh) / 2)
525 if (!bdi->dirty_exceeded)
526 bdi->dirty_exceeded = 1;
528 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
529 * Unstable writes are a feature of certain networked
530 * filesystems (i.e. NFS) in which data may have been
531 * written to the server's write cache, but has not yet
532 * been flushed to permanent storage.
533 * Only move pages to writeback if this bdi is over its
534 * threshold otherwise wait until the disk writes catch
537 trace_wbc_balance_dirty_start(&wbc, bdi);
538 if (bdi_nr_reclaimable > bdi_thresh) {
539 writeback_inodes_wb(&bdi->wb, &wbc);
540 pages_written += write_chunk - wbc.nr_to_write;
541 trace_wbc_balance_dirty_written(&wbc, bdi);
542 if (pages_written >= write_chunk)
543 break; /* We've done our duty */
545 trace_wbc_balance_dirty_wait(&wbc, bdi);
546 __set_current_state(TASK_INTERRUPTIBLE);
547 io_schedule_timeout(pause);
550 * Increase the delay for each loop, up to our previous
551 * default of taking a 100ms nap.
558 if (!dirty_exceeded && bdi->dirty_exceeded)
559 bdi->dirty_exceeded = 0;
561 if (writeback_in_progress(bdi))
565 * In laptop mode, we wait until hitting the higher threshold before
566 * starting background writeout, and then write out all the way down
567 * to the lower threshold. So slow writers cause minimal disk activity.
569 * In normal mode, we start background writeout at the lower
570 * background_thresh, to keep the amount of dirty memory low.
572 if ((laptop_mode && pages_written) ||
573 (!laptop_mode && (nr_reclaimable > background_thresh)))
574 bdi_start_background_writeback(bdi);
577 void set_page_dirty_balance(struct page *page, int page_mkwrite)
579 if (set_page_dirty(page) || page_mkwrite) {
580 struct address_space *mapping = page_mapping(page);
583 balance_dirty_pages_ratelimited(mapping);
587 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
590 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
591 * @mapping: address_space which was dirtied
592 * @nr_pages_dirtied: number of pages which the caller has just dirtied
594 * Processes which are dirtying memory should call in here once for each page
595 * which was newly dirtied. The function will periodically check the system's
596 * dirty state and will initiate writeback if needed.
598 * On really big machines, get_writeback_state is expensive, so try to avoid
599 * calling it too often (ratelimiting). But once we're over the dirty memory
600 * limit we decrease the ratelimiting by a lot, to prevent individual processes
601 * from overshooting the limit by (ratelimit_pages) each.
603 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
604 unsigned long nr_pages_dirtied)
606 unsigned long ratelimit;
609 ratelimit = ratelimit_pages;
610 if (mapping->backing_dev_info->dirty_exceeded)
614 * Check the rate limiting. Also, we do not want to throttle real-time
615 * tasks in balance_dirty_pages(). Period.
618 p = &__get_cpu_var(bdp_ratelimits);
619 *p += nr_pages_dirtied;
620 if (unlikely(*p >= ratelimit)) {
621 ratelimit = sync_writeback_pages(*p);
624 balance_dirty_pages(mapping, ratelimit);
629 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
631 void throttle_vm_writeout(gfp_t gfp_mask)
633 unsigned long background_thresh;
634 unsigned long dirty_thresh;
637 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
640 * Boost the allowable dirty threshold a bit for page
641 * allocators so they don't get DoS'ed by heavy writers
643 dirty_thresh += dirty_thresh / 10; /* wheeee... */
645 if (global_page_state(NR_UNSTABLE_NFS) +
646 global_page_state(NR_WRITEBACK) <= dirty_thresh)
648 congestion_wait(BLK_RW_ASYNC, HZ/10);
651 * The caller might hold locks which can prevent IO completion
652 * or progress in the filesystem. So we cannot just sit here
653 * waiting for IO to complete.
655 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
661 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
663 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
664 void __user *buffer, size_t *length, loff_t *ppos)
666 proc_dointvec(table, write, buffer, length, ppos);
667 bdi_arm_supers_timer();
672 void laptop_mode_timer_fn(unsigned long data)
674 struct request_queue *q = (struct request_queue *)data;
675 int nr_pages = global_page_state(NR_FILE_DIRTY) +
676 global_page_state(NR_UNSTABLE_NFS);
679 * We want to write everything out, not just down to the dirty
682 if (bdi_has_dirty_io(&q->backing_dev_info))
683 bdi_start_writeback(&q->backing_dev_info, nr_pages);
687 * We've spun up the disk and we're in laptop mode: schedule writeback
688 * of all dirty data a few seconds from now. If the flush is already scheduled
689 * then push it back - the user is still using the disk.
691 void laptop_io_completion(struct backing_dev_info *info)
693 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
697 * We're in laptop mode and we've just synced. The sync's writes will have
698 * caused another writeback to be scheduled by laptop_io_completion.
699 * Nothing needs to be written back anymore, so we unschedule the writeback.
701 void laptop_sync_completion(void)
703 struct backing_dev_info *bdi;
707 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
708 del_timer(&bdi->laptop_mode_wb_timer);
715 * If ratelimit_pages is too high then we can get into dirty-data overload
716 * if a large number of processes all perform writes at the same time.
717 * If it is too low then SMP machines will call the (expensive)
718 * get_writeback_state too often.
720 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
721 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
722 * thresholds before writeback cuts in.
724 * But the limit should not be set too high. Because it also controls the
725 * amount of memory which the balance_dirty_pages() caller has to write back.
726 * If this is too large then the caller will block on the IO queue all the
727 * time. So limit it to four megabytes - the balance_dirty_pages() caller
728 * will write six megabyte chunks, max.
731 void writeback_set_ratelimit(void)
733 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
734 if (ratelimit_pages < 16)
735 ratelimit_pages = 16;
736 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
737 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
741 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
743 writeback_set_ratelimit();
747 static struct notifier_block __cpuinitdata ratelimit_nb = {
748 .notifier_call = ratelimit_handler,
753 * Called early on to tune the page writeback dirty limits.
755 * We used to scale dirty pages according to how total memory
756 * related to pages that could be allocated for buffers (by
757 * comparing nr_free_buffer_pages() to vm_total_pages.
759 * However, that was when we used "dirty_ratio" to scale with
760 * all memory, and we don't do that any more. "dirty_ratio"
761 * is now applied to total non-HIGHPAGE memory (by subtracting
762 * totalhigh_pages from vm_total_pages), and as such we can't
763 * get into the old insane situation any more where we had
764 * large amounts of dirty pages compared to a small amount of
765 * non-HIGHMEM memory.
767 * But we might still want to scale the dirty_ratio by how
768 * much memory the box has..
770 void __init page_writeback_init(void)
774 writeback_set_ratelimit();
775 register_cpu_notifier(&ratelimit_nb);
777 shift = calc_period_shift();
778 prop_descriptor_init(&vm_completions, shift);
779 prop_descriptor_init(&vm_dirties, shift);
783 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
784 * @mapping: address space structure to write
785 * @start: starting page index
786 * @end: ending page index (inclusive)
788 * This function scans the page range from @start to @end (inclusive) and tags
789 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
790 * that write_cache_pages (or whoever calls this function) will then use
791 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
792 * used to avoid livelocking of writeback by a process steadily creating new
793 * dirty pages in the file (thus it is important for this function to be quick
794 * so that it can tag pages faster than a dirtying process can create them).
797 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
799 void tag_pages_for_writeback(struct address_space *mapping,
800 pgoff_t start, pgoff_t end)
802 #define WRITEBACK_TAG_BATCH 4096
803 unsigned long tagged;
806 spin_lock_irq(&mapping->tree_lock);
807 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
808 &start, end, WRITEBACK_TAG_BATCH,
809 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
810 spin_unlock_irq(&mapping->tree_lock);
811 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
813 } while (tagged >= WRITEBACK_TAG_BATCH);
815 EXPORT_SYMBOL(tag_pages_for_writeback);
818 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
819 * @mapping: address space structure to write
820 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
821 * @writepage: function called for each page
822 * @data: data passed to writepage function
824 * If a page is already under I/O, write_cache_pages() skips it, even
825 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
826 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
827 * and msync() need to guarantee that all the data which was dirty at the time
828 * the call was made get new I/O started against them. If wbc->sync_mode is
829 * WB_SYNC_ALL then we were called for data integrity and we must wait for
830 * existing IO to complete.
832 * To avoid livelocks (when other process dirties new pages), we first tag
833 * pages which should be written back with TOWRITE tag and only then start
834 * writing them. For data-integrity sync we have to be careful so that we do
835 * not miss some pages (e.g., because some other process has cleared TOWRITE
836 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
837 * by the process clearing the DIRTY tag (and submitting the page for IO).
839 int write_cache_pages(struct address_space *mapping,
840 struct writeback_control *wbc, writepage_t writepage,
847 pgoff_t uninitialized_var(writeback_index);
849 pgoff_t end; /* Inclusive */
855 pagevec_init(&pvec, 0);
856 if (wbc->range_cyclic) {
857 writeback_index = mapping->writeback_index; /* prev offset */
858 index = writeback_index;
865 index = wbc->range_start >> PAGE_CACHE_SHIFT;
866 end = wbc->range_end >> PAGE_CACHE_SHIFT;
867 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
869 cycled = 1; /* ignore range_cyclic tests */
871 if (wbc->sync_mode == WB_SYNC_ALL)
872 tag = PAGECACHE_TAG_TOWRITE;
874 tag = PAGECACHE_TAG_DIRTY;
876 if (wbc->sync_mode == WB_SYNC_ALL)
877 tag_pages_for_writeback(mapping, index, end);
879 while (!done && (index <= end)) {
882 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
883 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
887 for (i = 0; i < nr_pages; i++) {
888 struct page *page = pvec.pages[i];
891 * At this point, the page may be truncated or
892 * invalidated (changing page->mapping to NULL), or
893 * even swizzled back from swapper_space to tmpfs file
894 * mapping. However, page->index will not change
895 * because we have a reference on the page.
897 if (page->index > end) {
899 * can't be range_cyclic (1st pass) because
900 * end == -1 in that case.
906 done_index = page->index + 1;
911 * Page truncated or invalidated. We can freely skip it
912 * then, even for data integrity operations: the page
913 * has disappeared concurrently, so there could be no
914 * real expectation of this data interity operation
915 * even if there is now a new, dirty page at the same
918 if (unlikely(page->mapping != mapping)) {
924 if (!PageDirty(page)) {
925 /* someone wrote it for us */
926 goto continue_unlock;
929 if (PageWriteback(page)) {
930 if (wbc->sync_mode != WB_SYNC_NONE)
931 wait_on_page_writeback(page);
933 goto continue_unlock;
936 BUG_ON(PageWriteback(page));
937 if (!clear_page_dirty_for_io(page))
938 goto continue_unlock;
940 trace_wbc_writepage(wbc, mapping->backing_dev_info);
941 ret = (*writepage)(page, wbc, data);
943 if (ret == AOP_WRITEPAGE_ACTIVATE) {
948 * done_index is set past this page,
949 * so media errors will not choke
950 * background writeout for the entire
951 * file. This has consequences for
952 * range_cyclic semantics (ie. it may
953 * not be suitable for data integrity
961 if (wbc->nr_to_write > 0) {
962 if (--wbc->nr_to_write == 0 &&
963 wbc->sync_mode == WB_SYNC_NONE) {
965 * We stop writing back only if we are
966 * not doing integrity sync. In case of
967 * integrity sync we have to keep going
968 * because someone may be concurrently
969 * dirtying pages, and we might have
970 * synced a lot of newly appeared dirty
971 * pages, but have not synced all of the
979 pagevec_release(&pvec);
982 if (!cycled && !done) {
985 * We hit the last page and there is more work to be done: wrap
986 * back to the start of the file
990 end = writeback_index - 1;
993 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
994 mapping->writeback_index = done_index;
998 EXPORT_SYMBOL(write_cache_pages);
1001 * Function used by generic_writepages to call the real writepage
1002 * function and set the mapping flags on error
1004 static int __writepage(struct page *page, struct writeback_control *wbc,
1007 struct address_space *mapping = data;
1008 int ret = mapping->a_ops->writepage(page, wbc);
1009 mapping_set_error(mapping, ret);
1014 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1015 * @mapping: address space structure to write
1016 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1018 * This is a library function, which implements the writepages()
1019 * address_space_operation.
1021 int generic_writepages(struct address_space *mapping,
1022 struct writeback_control *wbc)
1024 /* deal with chardevs and other special file */
1025 if (!mapping->a_ops->writepage)
1028 return write_cache_pages(mapping, wbc, __writepage, mapping);
1031 EXPORT_SYMBOL(generic_writepages);
1033 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1037 if (wbc->nr_to_write <= 0)
1039 if (mapping->a_ops->writepages)
1040 ret = mapping->a_ops->writepages(mapping, wbc);
1042 ret = generic_writepages(mapping, wbc);
1047 * write_one_page - write out a single page and optionally wait on I/O
1048 * @page: the page to write
1049 * @wait: if true, wait on writeout
1051 * The page must be locked by the caller and will be unlocked upon return.
1053 * write_one_page() returns a negative error code if I/O failed.
1055 int write_one_page(struct page *page, int wait)
1057 struct address_space *mapping = page->mapping;
1059 struct writeback_control wbc = {
1060 .sync_mode = WB_SYNC_ALL,
1064 BUG_ON(!PageLocked(page));
1067 wait_on_page_writeback(page);
1069 if (clear_page_dirty_for_io(page)) {
1070 page_cache_get(page);
1071 ret = mapping->a_ops->writepage(page, &wbc);
1072 if (ret == 0 && wait) {
1073 wait_on_page_writeback(page);
1074 if (PageError(page))
1077 page_cache_release(page);
1083 EXPORT_SYMBOL(write_one_page);
1086 * For address_spaces which do not use buffers nor write back.
1088 int __set_page_dirty_no_writeback(struct page *page)
1090 if (!PageDirty(page))
1096 * Helper function for set_page_dirty family.
1097 * NOTE: This relies on being atomic wrt interrupts.
1099 void account_page_dirtied(struct page *page, struct address_space *mapping)
1101 if (mapping_cap_account_dirty(mapping)) {
1102 __inc_zone_page_state(page, NR_FILE_DIRTY);
1103 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1104 task_dirty_inc(current);
1105 task_io_account_write(PAGE_CACHE_SIZE);
1110 * For address_spaces which do not use buffers. Just tag the page as dirty in
1113 * This is also used when a single buffer is being dirtied: we want to set the
1114 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1115 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1117 * Most callers have locked the page, which pins the address_space in memory.
1118 * But zap_pte_range() does not lock the page, however in that case the
1119 * mapping is pinned by the vma's ->vm_file reference.
1121 * We take care to handle the case where the page was truncated from the
1122 * mapping by re-checking page_mapping() inside tree_lock.
1124 int __set_page_dirty_nobuffers(struct page *page)
1126 if (!TestSetPageDirty(page)) {
1127 struct address_space *mapping = page_mapping(page);
1128 struct address_space *mapping2;
1133 spin_lock_irq(&mapping->tree_lock);
1134 mapping2 = page_mapping(page);
1135 if (mapping2) { /* Race with truncate? */
1136 BUG_ON(mapping2 != mapping);
1137 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1138 account_page_dirtied(page, mapping);
1139 radix_tree_tag_set(&mapping->page_tree,
1140 page_index(page), PAGECACHE_TAG_DIRTY);
1142 spin_unlock_irq(&mapping->tree_lock);
1143 if (mapping->host) {
1144 /* !PageAnon && !swapper_space */
1145 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1151 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1154 * When a writepage implementation decides that it doesn't want to write this
1155 * page for some reason, it should redirty the locked page via
1156 * redirty_page_for_writepage() and it should then unlock the page and return 0
1158 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1160 wbc->pages_skipped++;
1161 return __set_page_dirty_nobuffers(page);
1163 EXPORT_SYMBOL(redirty_page_for_writepage);
1168 * For pages with a mapping this should be done under the page lock
1169 * for the benefit of asynchronous memory errors who prefer a consistent
1170 * dirty state. This rule can be broken in some special cases,
1171 * but should be better not to.
1173 * If the mapping doesn't provide a set_page_dirty a_op, then
1174 * just fall through and assume that it wants buffer_heads.
1176 int set_page_dirty(struct page *page)
1178 struct address_space *mapping = page_mapping(page);
1180 if (likely(mapping)) {
1181 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1184 spd = __set_page_dirty_buffers;
1186 return (*spd)(page);
1188 if (!PageDirty(page)) {
1189 if (!TestSetPageDirty(page))
1194 EXPORT_SYMBOL(set_page_dirty);
1197 * set_page_dirty() is racy if the caller has no reference against
1198 * page->mapping->host, and if the page is unlocked. This is because another
1199 * CPU could truncate the page off the mapping and then free the mapping.
1201 * Usually, the page _is_ locked, or the caller is a user-space process which
1202 * holds a reference on the inode by having an open file.
1204 * In other cases, the page should be locked before running set_page_dirty().
1206 int set_page_dirty_lock(struct page *page)
1210 lock_page_nosync(page);
1211 ret = set_page_dirty(page);
1215 EXPORT_SYMBOL(set_page_dirty_lock);
1218 * Clear a page's dirty flag, while caring for dirty memory accounting.
1219 * Returns true if the page was previously dirty.
1221 * This is for preparing to put the page under writeout. We leave the page
1222 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1223 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1224 * implementation will run either set_page_writeback() or set_page_dirty(),
1225 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1228 * This incoherency between the page's dirty flag and radix-tree tag is
1229 * unfortunate, but it only exists while the page is locked.
1231 int clear_page_dirty_for_io(struct page *page)
1233 struct address_space *mapping = page_mapping(page);
1235 BUG_ON(!PageLocked(page));
1237 ClearPageReclaim(page);
1238 if (mapping && mapping_cap_account_dirty(mapping)) {
1240 * Yes, Virginia, this is indeed insane.
1242 * We use this sequence to make sure that
1243 * (a) we account for dirty stats properly
1244 * (b) we tell the low-level filesystem to
1245 * mark the whole page dirty if it was
1246 * dirty in a pagetable. Only to then
1247 * (c) clean the page again and return 1 to
1248 * cause the writeback.
1250 * This way we avoid all nasty races with the
1251 * dirty bit in multiple places and clearing
1252 * them concurrently from different threads.
1254 * Note! Normally the "set_page_dirty(page)"
1255 * has no effect on the actual dirty bit - since
1256 * that will already usually be set. But we
1257 * need the side effects, and it can help us
1260 * We basically use the page "master dirty bit"
1261 * as a serialization point for all the different
1262 * threads doing their things.
1264 if (page_mkclean(page))
1265 set_page_dirty(page);
1267 * We carefully synchronise fault handlers against
1268 * installing a dirty pte and marking the page dirty
1269 * at this point. We do this by having them hold the
1270 * page lock at some point after installing their
1271 * pte, but before marking the page dirty.
1272 * Pages are always locked coming in here, so we get
1273 * the desired exclusion. See mm/memory.c:do_wp_page()
1274 * for more comments.
1276 if (TestClearPageDirty(page)) {
1277 dec_zone_page_state(page, NR_FILE_DIRTY);
1278 dec_bdi_stat(mapping->backing_dev_info,
1284 return TestClearPageDirty(page);
1286 EXPORT_SYMBOL(clear_page_dirty_for_io);
1288 int test_clear_page_writeback(struct page *page)
1290 struct address_space *mapping = page_mapping(page);
1294 struct backing_dev_info *bdi = mapping->backing_dev_info;
1295 unsigned long flags;
1297 spin_lock_irqsave(&mapping->tree_lock, flags);
1298 ret = TestClearPageWriteback(page);
1300 radix_tree_tag_clear(&mapping->page_tree,
1302 PAGECACHE_TAG_WRITEBACK);
1303 if (bdi_cap_account_writeback(bdi)) {
1304 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1305 __bdi_writeout_inc(bdi);
1308 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1310 ret = TestClearPageWriteback(page);
1313 dec_zone_page_state(page, NR_WRITEBACK);
1317 int test_set_page_writeback(struct page *page)
1319 struct address_space *mapping = page_mapping(page);
1323 struct backing_dev_info *bdi = mapping->backing_dev_info;
1324 unsigned long flags;
1326 spin_lock_irqsave(&mapping->tree_lock, flags);
1327 ret = TestSetPageWriteback(page);
1329 radix_tree_tag_set(&mapping->page_tree,
1331 PAGECACHE_TAG_WRITEBACK);
1332 if (bdi_cap_account_writeback(bdi))
1333 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1335 if (!PageDirty(page))
1336 radix_tree_tag_clear(&mapping->page_tree,
1338 PAGECACHE_TAG_DIRTY);
1339 radix_tree_tag_clear(&mapping->page_tree,
1341 PAGECACHE_TAG_TOWRITE);
1342 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1344 ret = TestSetPageWriteback(page);
1347 inc_zone_page_state(page, NR_WRITEBACK);
1351 EXPORT_SYMBOL(test_set_page_writeback);
1354 * Return true if any of the pages in the mapping are marked with the
1357 int mapping_tagged(struct address_space *mapping, int tag)
1361 ret = radix_tree_tagged(&mapping->page_tree, tag);
1365 EXPORT_SYMBOL(mapping_tagged);