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 unsigned long task_dirty_limit(struct task_struct *tsk,
271 unsigned long bdi_dirty)
273 long numerator, denominator;
274 unsigned long dirty = bdi_dirty;
275 u64 inv = dirty >> 3;
277 task_dirties_fraction(tsk, &numerator, &denominator);
279 do_div(inv, denominator);
283 return max(dirty, bdi_dirty/2);
289 static unsigned int bdi_min_ratio;
291 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
295 spin_lock_bh(&bdi_lock);
296 if (min_ratio > bdi->max_ratio) {
299 min_ratio -= bdi->min_ratio;
300 if (bdi_min_ratio + min_ratio < 100) {
301 bdi_min_ratio += min_ratio;
302 bdi->min_ratio += min_ratio;
307 spin_unlock_bh(&bdi_lock);
312 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
319 spin_lock_bh(&bdi_lock);
320 if (bdi->min_ratio > max_ratio) {
323 bdi->max_ratio = max_ratio;
324 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
326 spin_unlock_bh(&bdi_lock);
330 EXPORT_SYMBOL(bdi_set_max_ratio);
333 * Work out the current dirty-memory clamping and background writeout
336 * The main aim here is to lower them aggressively if there is a lot of mapped
337 * memory around. To avoid stressing page reclaim with lots of unreclaimable
338 * pages. It is better to clamp down on writers than to start swapping, and
339 * performing lots of scanning.
341 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
343 * We don't permit the clamping level to fall below 5% - that is getting rather
346 * We make sure that the background writeout level is below the adjusted
350 static unsigned long highmem_dirtyable_memory(unsigned long total)
352 #ifdef CONFIG_HIGHMEM
356 for_each_node_state(node, N_HIGH_MEMORY) {
358 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
360 x += zone_page_state(z, NR_FREE_PAGES) +
361 zone_reclaimable_pages(z);
364 * Make sure that the number of highmem pages is never larger
365 * than the number of the total dirtyable memory. This can only
366 * occur in very strange VM situations but we want to make sure
367 * that this does not occur.
369 return min(x, total);
376 * determine_dirtyable_memory - amount of memory that may be used
378 * Returns the numebr of pages that can currently be freed and used
379 * by the kernel for direct mappings.
381 unsigned long determine_dirtyable_memory(void)
385 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
387 if (!vm_highmem_is_dirtyable)
388 x -= highmem_dirtyable_memory(x);
390 return x + 1; /* Ensure that we never return 0 */
393 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
395 unsigned long background;
397 unsigned long available_memory = determine_dirtyable_memory();
398 struct task_struct *tsk;
401 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
405 dirty_ratio = vm_dirty_ratio;
408 dirty = (dirty_ratio * available_memory) / 100;
411 if (dirty_background_bytes)
412 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
414 background = (dirty_background_ratio * available_memory) / 100;
416 if (background >= dirty)
417 background = dirty / 2;
419 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
420 background += background / 4;
423 *pbackground = background;
427 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi,
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);
442 bdi_dirty += (dirty * bdi->min_ratio) / 100;
443 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
444 bdi_dirty = dirty * bdi->max_ratio / 100;
450 * balance_dirty_pages() must be called by processes which are generating dirty
451 * data. It looks at the number of dirty pages in the machine and will force
452 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
453 * If we're over `background_thresh' then the writeback threads are woken to
454 * perform some writeout.
456 static void balance_dirty_pages(struct address_space *mapping,
457 unsigned long write_chunk)
459 long nr_reclaimable, bdi_nr_reclaimable;
460 long nr_writeback, bdi_nr_writeback;
461 unsigned long background_thresh;
462 unsigned long dirty_thresh;
463 unsigned long bdi_thresh;
464 unsigned long pages_written = 0;
465 unsigned long pause = 1;
466 bool dirty_exceeded = false;
467 struct backing_dev_info *bdi = mapping->backing_dev_info;
470 struct writeback_control wbc = {
471 .sync_mode = WB_SYNC_NONE,
472 .older_than_this = NULL,
473 .nr_to_write = write_chunk,
477 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
478 global_page_state(NR_UNSTABLE_NFS);
479 nr_writeback = global_page_state(NR_WRITEBACK);
481 global_dirty_limits(&background_thresh, &dirty_thresh);
484 * Throttle it only when the background writeback cannot
485 * catch-up. This avoids (excessively) small writeouts
486 * when the bdi limits are ramping up.
488 if (nr_reclaimable + nr_writeback <
489 (background_thresh + dirty_thresh) / 2)
492 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
493 bdi_thresh = task_dirty_limit(current, bdi_thresh);
496 * In order to avoid the stacked BDI deadlock we need
497 * to ensure we accurately count the 'dirty' pages when
498 * the threshold is low.
500 * Otherwise it would be possible to get thresh+n pages
501 * reported dirty, even though there are thresh-m pages
502 * actually dirty; with m+n sitting in the percpu
505 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
506 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
507 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
509 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
510 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
514 * The bdi thresh is somehow "soft" limit derived from the
515 * global "hard" limit. The former helps to prevent heavy IO
516 * bdi or process from holding back light ones; The latter is
517 * the last resort safeguard.
520 (bdi_nr_reclaimable + bdi_nr_writeback >= bdi_thresh)
521 || (nr_reclaimable + nr_writeback >= dirty_thresh);
526 if (!bdi->dirty_exceeded)
527 bdi->dirty_exceeded = 1;
529 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
530 * Unstable writes are a feature of certain networked
531 * filesystems (i.e. NFS) in which data may have been
532 * written to the server's write cache, but has not yet
533 * been flushed to permanent storage.
534 * Only move pages to writeback if this bdi is over its
535 * threshold otherwise wait until the disk writes catch
538 trace_wbc_balance_dirty_start(&wbc, bdi);
539 if (bdi_nr_reclaimable > bdi_thresh) {
540 writeback_inodes_wb(&bdi->wb, &wbc);
541 pages_written += write_chunk - wbc.nr_to_write;
542 trace_wbc_balance_dirty_written(&wbc, bdi);
543 if (pages_written >= write_chunk)
544 break; /* We've done our duty */
546 trace_wbc_balance_dirty_wait(&wbc, bdi);
547 __set_current_state(TASK_INTERRUPTIBLE);
548 io_schedule_timeout(pause);
551 * Increase the delay for each loop, up to our previous
552 * default of taking a 100ms nap.
559 if (!dirty_exceeded && bdi->dirty_exceeded)
560 bdi->dirty_exceeded = 0;
562 if (writeback_in_progress(bdi))
566 * In laptop mode, we wait until hitting the higher threshold before
567 * starting background writeout, and then write out all the way down
568 * to the lower threshold. So slow writers cause minimal disk activity.
570 * In normal mode, we start background writeout at the lower
571 * background_thresh, to keep the amount of dirty memory low.
573 if ((laptop_mode && pages_written) ||
574 (!laptop_mode && (nr_reclaimable > background_thresh)))
575 bdi_start_background_writeback(bdi);
578 void set_page_dirty_balance(struct page *page, int page_mkwrite)
580 if (set_page_dirty(page) || page_mkwrite) {
581 struct address_space *mapping = page_mapping(page);
584 balance_dirty_pages_ratelimited(mapping);
588 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
591 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
592 * @mapping: address_space which was dirtied
593 * @nr_pages_dirtied: number of pages which the caller has just dirtied
595 * Processes which are dirtying memory should call in here once for each page
596 * which was newly dirtied. The function will periodically check the system's
597 * dirty state and will initiate writeback if needed.
599 * On really big machines, get_writeback_state is expensive, so try to avoid
600 * calling it too often (ratelimiting). But once we're over the dirty memory
601 * limit we decrease the ratelimiting by a lot, to prevent individual processes
602 * from overshooting the limit by (ratelimit_pages) each.
604 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
605 unsigned long nr_pages_dirtied)
607 unsigned long ratelimit;
610 ratelimit = ratelimit_pages;
611 if (mapping->backing_dev_info->dirty_exceeded)
615 * Check the rate limiting. Also, we do not want to throttle real-time
616 * tasks in balance_dirty_pages(). Period.
619 p = &__get_cpu_var(bdp_ratelimits);
620 *p += nr_pages_dirtied;
621 if (unlikely(*p >= ratelimit)) {
622 ratelimit = sync_writeback_pages(*p);
625 balance_dirty_pages(mapping, ratelimit);
630 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
632 void throttle_vm_writeout(gfp_t gfp_mask)
634 unsigned long background_thresh;
635 unsigned long dirty_thresh;
638 global_dirty_limits(&background_thresh, &dirty_thresh);
641 * Boost the allowable dirty threshold a bit for page
642 * allocators so they don't get DoS'ed by heavy writers
644 dirty_thresh += dirty_thresh / 10; /* wheeee... */
646 if (global_page_state(NR_UNSTABLE_NFS) +
647 global_page_state(NR_WRITEBACK) <= dirty_thresh)
649 congestion_wait(BLK_RW_ASYNC, HZ/10);
652 * The caller might hold locks which can prevent IO completion
653 * or progress in the filesystem. So we cannot just sit here
654 * waiting for IO to complete.
656 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
662 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
664 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
665 void __user *buffer, size_t *length, loff_t *ppos)
667 proc_dointvec(table, write, buffer, length, ppos);
668 bdi_arm_supers_timer();
673 void laptop_mode_timer_fn(unsigned long data)
675 struct request_queue *q = (struct request_queue *)data;
676 int nr_pages = global_page_state(NR_FILE_DIRTY) +
677 global_page_state(NR_UNSTABLE_NFS);
680 * We want to write everything out, not just down to the dirty
683 if (bdi_has_dirty_io(&q->backing_dev_info))
684 bdi_start_writeback(&q->backing_dev_info, nr_pages);
688 * We've spun up the disk and we're in laptop mode: schedule writeback
689 * of all dirty data a few seconds from now. If the flush is already scheduled
690 * then push it back - the user is still using the disk.
692 void laptop_io_completion(struct backing_dev_info *info)
694 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
698 * We're in laptop mode and we've just synced. The sync's writes will have
699 * caused another writeback to be scheduled by laptop_io_completion.
700 * Nothing needs to be written back anymore, so we unschedule the writeback.
702 void laptop_sync_completion(void)
704 struct backing_dev_info *bdi;
708 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
709 del_timer(&bdi->laptop_mode_wb_timer);
716 * If ratelimit_pages is too high then we can get into dirty-data overload
717 * if a large number of processes all perform writes at the same time.
718 * If it is too low then SMP machines will call the (expensive)
719 * get_writeback_state too often.
721 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
722 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
723 * thresholds before writeback cuts in.
725 * But the limit should not be set too high. Because it also controls the
726 * amount of memory which the balance_dirty_pages() caller has to write back.
727 * If this is too large then the caller will block on the IO queue all the
728 * time. So limit it to four megabytes - the balance_dirty_pages() caller
729 * will write six megabyte chunks, max.
732 void writeback_set_ratelimit(void)
734 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
735 if (ratelimit_pages < 16)
736 ratelimit_pages = 16;
737 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
738 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
742 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
744 writeback_set_ratelimit();
748 static struct notifier_block __cpuinitdata ratelimit_nb = {
749 .notifier_call = ratelimit_handler,
754 * Called early on to tune the page writeback dirty limits.
756 * We used to scale dirty pages according to how total memory
757 * related to pages that could be allocated for buffers (by
758 * comparing nr_free_buffer_pages() to vm_total_pages.
760 * However, that was when we used "dirty_ratio" to scale with
761 * all memory, and we don't do that any more. "dirty_ratio"
762 * is now applied to total non-HIGHPAGE memory (by subtracting
763 * totalhigh_pages from vm_total_pages), and as such we can't
764 * get into the old insane situation any more where we had
765 * large amounts of dirty pages compared to a small amount of
766 * non-HIGHMEM memory.
768 * But we might still want to scale the dirty_ratio by how
769 * much memory the box has..
771 void __init page_writeback_init(void)
775 writeback_set_ratelimit();
776 register_cpu_notifier(&ratelimit_nb);
778 shift = calc_period_shift();
779 prop_descriptor_init(&vm_completions, shift);
780 prop_descriptor_init(&vm_dirties, shift);
784 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
785 * @mapping: address space structure to write
786 * @start: starting page index
787 * @end: ending page index (inclusive)
789 * This function scans the page range from @start to @end (inclusive) and tags
790 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
791 * that write_cache_pages (or whoever calls this function) will then use
792 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
793 * used to avoid livelocking of writeback by a process steadily creating new
794 * dirty pages in the file (thus it is important for this function to be quick
795 * so that it can tag pages faster than a dirtying process can create them).
798 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
800 void tag_pages_for_writeback(struct address_space *mapping,
801 pgoff_t start, pgoff_t end)
803 #define WRITEBACK_TAG_BATCH 4096
804 unsigned long tagged;
807 spin_lock_irq(&mapping->tree_lock);
808 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
809 &start, end, WRITEBACK_TAG_BATCH,
810 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
811 spin_unlock_irq(&mapping->tree_lock);
812 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
814 } while (tagged >= WRITEBACK_TAG_BATCH);
816 EXPORT_SYMBOL(tag_pages_for_writeback);
819 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
820 * @mapping: address space structure to write
821 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
822 * @writepage: function called for each page
823 * @data: data passed to writepage function
825 * If a page is already under I/O, write_cache_pages() skips it, even
826 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
827 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
828 * and msync() need to guarantee that all the data which was dirty at the time
829 * the call was made get new I/O started against them. If wbc->sync_mode is
830 * WB_SYNC_ALL then we were called for data integrity and we must wait for
831 * existing IO to complete.
833 * To avoid livelocks (when other process dirties new pages), we first tag
834 * pages which should be written back with TOWRITE tag and only then start
835 * writing them. For data-integrity sync we have to be careful so that we do
836 * not miss some pages (e.g., because some other process has cleared TOWRITE
837 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
838 * by the process clearing the DIRTY tag (and submitting the page for IO).
840 int write_cache_pages(struct address_space *mapping,
841 struct writeback_control *wbc, writepage_t writepage,
848 pgoff_t uninitialized_var(writeback_index);
850 pgoff_t end; /* Inclusive */
856 pagevec_init(&pvec, 0);
857 if (wbc->range_cyclic) {
858 writeback_index = mapping->writeback_index; /* prev offset */
859 index = writeback_index;
866 index = wbc->range_start >> PAGE_CACHE_SHIFT;
867 end = wbc->range_end >> PAGE_CACHE_SHIFT;
868 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
870 cycled = 1; /* ignore range_cyclic tests */
872 if (wbc->sync_mode == WB_SYNC_ALL)
873 tag = PAGECACHE_TAG_TOWRITE;
875 tag = PAGECACHE_TAG_DIRTY;
877 if (wbc->sync_mode == WB_SYNC_ALL)
878 tag_pages_for_writeback(mapping, index, end);
880 while (!done && (index <= end)) {
883 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
884 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
888 for (i = 0; i < nr_pages; i++) {
889 struct page *page = pvec.pages[i];
892 * At this point, the page may be truncated or
893 * invalidated (changing page->mapping to NULL), or
894 * even swizzled back from swapper_space to tmpfs file
895 * mapping. However, page->index will not change
896 * because we have a reference on the page.
898 if (page->index > end) {
900 * can't be range_cyclic (1st pass) because
901 * end == -1 in that case.
907 done_index = page->index + 1;
912 * Page truncated or invalidated. We can freely skip it
913 * then, even for data integrity operations: the page
914 * has disappeared concurrently, so there could be no
915 * real expectation of this data interity operation
916 * even if there is now a new, dirty page at the same
919 if (unlikely(page->mapping != mapping)) {
925 if (!PageDirty(page)) {
926 /* someone wrote it for us */
927 goto continue_unlock;
930 if (PageWriteback(page)) {
931 if (wbc->sync_mode != WB_SYNC_NONE)
932 wait_on_page_writeback(page);
934 goto continue_unlock;
937 BUG_ON(PageWriteback(page));
938 if (!clear_page_dirty_for_io(page))
939 goto continue_unlock;
941 trace_wbc_writepage(wbc, mapping->backing_dev_info);
942 ret = (*writepage)(page, wbc, data);
944 if (ret == AOP_WRITEPAGE_ACTIVATE) {
949 * done_index is set past this page,
950 * so media errors will not choke
951 * background writeout for the entire
952 * file. This has consequences for
953 * range_cyclic semantics (ie. it may
954 * not be suitable for data integrity
962 if (wbc->nr_to_write > 0) {
963 if (--wbc->nr_to_write == 0 &&
964 wbc->sync_mode == WB_SYNC_NONE) {
966 * We stop writing back only if we are
967 * not doing integrity sync. In case of
968 * integrity sync we have to keep going
969 * because someone may be concurrently
970 * dirtying pages, and we might have
971 * synced a lot of newly appeared dirty
972 * pages, but have not synced all of the
980 pagevec_release(&pvec);
983 if (!cycled && !done) {
986 * We hit the last page and there is more work to be done: wrap
987 * back to the start of the file
991 end = writeback_index - 1;
994 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
995 mapping->writeback_index = done_index;
999 EXPORT_SYMBOL(write_cache_pages);
1002 * Function used by generic_writepages to call the real writepage
1003 * function and set the mapping flags on error
1005 static int __writepage(struct page *page, struct writeback_control *wbc,
1008 struct address_space *mapping = data;
1009 int ret = mapping->a_ops->writepage(page, wbc);
1010 mapping_set_error(mapping, ret);
1015 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1016 * @mapping: address space structure to write
1017 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1019 * This is a library function, which implements the writepages()
1020 * address_space_operation.
1022 int generic_writepages(struct address_space *mapping,
1023 struct writeback_control *wbc)
1025 /* deal with chardevs and other special file */
1026 if (!mapping->a_ops->writepage)
1029 return write_cache_pages(mapping, wbc, __writepage, mapping);
1032 EXPORT_SYMBOL(generic_writepages);
1034 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1038 if (wbc->nr_to_write <= 0)
1040 if (mapping->a_ops->writepages)
1041 ret = mapping->a_ops->writepages(mapping, wbc);
1043 ret = generic_writepages(mapping, wbc);
1048 * write_one_page - write out a single page and optionally wait on I/O
1049 * @page: the page to write
1050 * @wait: if true, wait on writeout
1052 * The page must be locked by the caller and will be unlocked upon return.
1054 * write_one_page() returns a negative error code if I/O failed.
1056 int write_one_page(struct page *page, int wait)
1058 struct address_space *mapping = page->mapping;
1060 struct writeback_control wbc = {
1061 .sync_mode = WB_SYNC_ALL,
1065 BUG_ON(!PageLocked(page));
1068 wait_on_page_writeback(page);
1070 if (clear_page_dirty_for_io(page)) {
1071 page_cache_get(page);
1072 ret = mapping->a_ops->writepage(page, &wbc);
1073 if (ret == 0 && wait) {
1074 wait_on_page_writeback(page);
1075 if (PageError(page))
1078 page_cache_release(page);
1084 EXPORT_SYMBOL(write_one_page);
1087 * For address_spaces which do not use buffers nor write back.
1089 int __set_page_dirty_no_writeback(struct page *page)
1091 if (!PageDirty(page))
1097 * Helper function for set_page_dirty family.
1098 * NOTE: This relies on being atomic wrt interrupts.
1100 void account_page_dirtied(struct page *page, struct address_space *mapping)
1102 if (mapping_cap_account_dirty(mapping)) {
1103 __inc_zone_page_state(page, NR_FILE_DIRTY);
1104 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1105 task_dirty_inc(current);
1106 task_io_account_write(PAGE_CACHE_SIZE);
1111 * For address_spaces which do not use buffers. Just tag the page as dirty in
1114 * This is also used when a single buffer is being dirtied: we want to set the
1115 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1116 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1118 * Most callers have locked the page, which pins the address_space in memory.
1119 * But zap_pte_range() does not lock the page, however in that case the
1120 * mapping is pinned by the vma's ->vm_file reference.
1122 * We take care to handle the case where the page was truncated from the
1123 * mapping by re-checking page_mapping() inside tree_lock.
1125 int __set_page_dirty_nobuffers(struct page *page)
1127 if (!TestSetPageDirty(page)) {
1128 struct address_space *mapping = page_mapping(page);
1129 struct address_space *mapping2;
1134 spin_lock_irq(&mapping->tree_lock);
1135 mapping2 = page_mapping(page);
1136 if (mapping2) { /* Race with truncate? */
1137 BUG_ON(mapping2 != mapping);
1138 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1139 account_page_dirtied(page, mapping);
1140 radix_tree_tag_set(&mapping->page_tree,
1141 page_index(page), PAGECACHE_TAG_DIRTY);
1143 spin_unlock_irq(&mapping->tree_lock);
1144 if (mapping->host) {
1145 /* !PageAnon && !swapper_space */
1146 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1152 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1155 * When a writepage implementation decides that it doesn't want to write this
1156 * page for some reason, it should redirty the locked page via
1157 * redirty_page_for_writepage() and it should then unlock the page and return 0
1159 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1161 wbc->pages_skipped++;
1162 return __set_page_dirty_nobuffers(page);
1164 EXPORT_SYMBOL(redirty_page_for_writepage);
1169 * For pages with a mapping this should be done under the page lock
1170 * for the benefit of asynchronous memory errors who prefer a consistent
1171 * dirty state. This rule can be broken in some special cases,
1172 * but should be better not to.
1174 * If the mapping doesn't provide a set_page_dirty a_op, then
1175 * just fall through and assume that it wants buffer_heads.
1177 int set_page_dirty(struct page *page)
1179 struct address_space *mapping = page_mapping(page);
1181 if (likely(mapping)) {
1182 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1185 spd = __set_page_dirty_buffers;
1187 return (*spd)(page);
1189 if (!PageDirty(page)) {
1190 if (!TestSetPageDirty(page))
1195 EXPORT_SYMBOL(set_page_dirty);
1198 * set_page_dirty() is racy if the caller has no reference against
1199 * page->mapping->host, and if the page is unlocked. This is because another
1200 * CPU could truncate the page off the mapping and then free the mapping.
1202 * Usually, the page _is_ locked, or the caller is a user-space process which
1203 * holds a reference on the inode by having an open file.
1205 * In other cases, the page should be locked before running set_page_dirty().
1207 int set_page_dirty_lock(struct page *page)
1211 lock_page_nosync(page);
1212 ret = set_page_dirty(page);
1216 EXPORT_SYMBOL(set_page_dirty_lock);
1219 * Clear a page's dirty flag, while caring for dirty memory accounting.
1220 * Returns true if the page was previously dirty.
1222 * This is for preparing to put the page under writeout. We leave the page
1223 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1224 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1225 * implementation will run either set_page_writeback() or set_page_dirty(),
1226 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1229 * This incoherency between the page's dirty flag and radix-tree tag is
1230 * unfortunate, but it only exists while the page is locked.
1232 int clear_page_dirty_for_io(struct page *page)
1234 struct address_space *mapping = page_mapping(page);
1236 BUG_ON(!PageLocked(page));
1238 ClearPageReclaim(page);
1239 if (mapping && mapping_cap_account_dirty(mapping)) {
1241 * Yes, Virginia, this is indeed insane.
1243 * We use this sequence to make sure that
1244 * (a) we account for dirty stats properly
1245 * (b) we tell the low-level filesystem to
1246 * mark the whole page dirty if it was
1247 * dirty in a pagetable. Only to then
1248 * (c) clean the page again and return 1 to
1249 * cause the writeback.
1251 * This way we avoid all nasty races with the
1252 * dirty bit in multiple places and clearing
1253 * them concurrently from different threads.
1255 * Note! Normally the "set_page_dirty(page)"
1256 * has no effect on the actual dirty bit - since
1257 * that will already usually be set. But we
1258 * need the side effects, and it can help us
1261 * We basically use the page "master dirty bit"
1262 * as a serialization point for all the different
1263 * threads doing their things.
1265 if (page_mkclean(page))
1266 set_page_dirty(page);
1268 * We carefully synchronise fault handlers against
1269 * installing a dirty pte and marking the page dirty
1270 * at this point. We do this by having them hold the
1271 * page lock at some point after installing their
1272 * pte, but before marking the page dirty.
1273 * Pages are always locked coming in here, so we get
1274 * the desired exclusion. See mm/memory.c:do_wp_page()
1275 * for more comments.
1277 if (TestClearPageDirty(page)) {
1278 dec_zone_page_state(page, NR_FILE_DIRTY);
1279 dec_bdi_stat(mapping->backing_dev_info,
1285 return TestClearPageDirty(page);
1287 EXPORT_SYMBOL(clear_page_dirty_for_io);
1289 int test_clear_page_writeback(struct page *page)
1291 struct address_space *mapping = page_mapping(page);
1295 struct backing_dev_info *bdi = mapping->backing_dev_info;
1296 unsigned long flags;
1298 spin_lock_irqsave(&mapping->tree_lock, flags);
1299 ret = TestClearPageWriteback(page);
1301 radix_tree_tag_clear(&mapping->page_tree,
1303 PAGECACHE_TAG_WRITEBACK);
1304 if (bdi_cap_account_writeback(bdi)) {
1305 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1306 __bdi_writeout_inc(bdi);
1309 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1311 ret = TestClearPageWriteback(page);
1314 dec_zone_page_state(page, NR_WRITEBACK);
1318 int test_set_page_writeback(struct page *page)
1320 struct address_space *mapping = page_mapping(page);
1324 struct backing_dev_info *bdi = mapping->backing_dev_info;
1325 unsigned long flags;
1327 spin_lock_irqsave(&mapping->tree_lock, flags);
1328 ret = TestSetPageWriteback(page);
1330 radix_tree_tag_set(&mapping->page_tree,
1332 PAGECACHE_TAG_WRITEBACK);
1333 if (bdi_cap_account_writeback(bdi))
1334 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1336 if (!PageDirty(page))
1337 radix_tree_tag_clear(&mapping->page_tree,
1339 PAGECACHE_TAG_DIRTY);
1340 radix_tree_tag_clear(&mapping->page_tree,
1342 PAGECACHE_TAG_TOWRITE);
1343 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1345 ret = TestSetPageWriteback(page);
1348 inc_zone_page_state(page, NR_WRITEBACK);
1352 EXPORT_SYMBOL(test_set_page_writeback);
1355 * Return true if any of the pages in the mapping are marked with the
1358 int mapping_tagged(struct address_space *mapping, int tag)
1362 ret = radix_tree_tagged(&mapping->page_tree, tag);
1366 EXPORT_SYMBOL(mapping_tagged);