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 * task_dirty_limit - scale down dirty throttling threshold for one task
266 * task specific dirty limit:
268 * dirty -= (dirty/8) * p_{t}
270 * To protect light/slow dirtying tasks from heavier/fast ones, we start
271 * throttling individual tasks before reaching the bdi dirty limit.
272 * Relatively low thresholds will be allocated to heavy dirtiers. So when
273 * dirty pages grow large, heavy dirtiers will be throttled first, which will
274 * effectively curb the growth of dirty pages. Light dirtiers with high enough
275 * dirty threshold may never get throttled.
277 static unsigned long task_dirty_limit(struct task_struct *tsk,
278 unsigned long bdi_dirty)
280 long numerator, denominator;
281 unsigned long dirty = bdi_dirty;
282 u64 inv = dirty >> 3;
284 task_dirties_fraction(tsk, &numerator, &denominator);
286 do_div(inv, denominator);
290 return max(dirty, bdi_dirty/2);
296 static unsigned int bdi_min_ratio;
298 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
302 spin_lock_bh(&bdi_lock);
303 if (min_ratio > bdi->max_ratio) {
306 min_ratio -= bdi->min_ratio;
307 if (bdi_min_ratio + min_ratio < 100) {
308 bdi_min_ratio += min_ratio;
309 bdi->min_ratio += min_ratio;
314 spin_unlock_bh(&bdi_lock);
319 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
326 spin_lock_bh(&bdi_lock);
327 if (bdi->min_ratio > max_ratio) {
330 bdi->max_ratio = max_ratio;
331 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
333 spin_unlock_bh(&bdi_lock);
337 EXPORT_SYMBOL(bdi_set_max_ratio);
340 * Work out the current dirty-memory clamping and background writeout
343 * The main aim here is to lower them aggressively if there is a lot of mapped
344 * memory around. To avoid stressing page reclaim with lots of unreclaimable
345 * pages. It is better to clamp down on writers than to start swapping, and
346 * performing lots of scanning.
348 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
350 * We don't permit the clamping level to fall below 5% - that is getting rather
353 * We make sure that the background writeout level is below the adjusted
357 static unsigned long highmem_dirtyable_memory(unsigned long total)
359 #ifdef CONFIG_HIGHMEM
363 for_each_node_state(node, N_HIGH_MEMORY) {
365 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
367 x += zone_page_state(z, NR_FREE_PAGES) +
368 zone_reclaimable_pages(z);
371 * Make sure that the number of highmem pages is never larger
372 * than the number of the total dirtyable memory. This can only
373 * occur in very strange VM situations but we want to make sure
374 * that this does not occur.
376 return min(x, total);
383 * determine_dirtyable_memory - amount of memory that may be used
385 * Returns the numebr of pages that can currently be freed and used
386 * by the kernel for direct mappings.
388 unsigned long determine_dirtyable_memory(void)
392 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
394 if (!vm_highmem_is_dirtyable)
395 x -= highmem_dirtyable_memory(x);
397 return x + 1; /* Ensure that we never return 0 */
401 * global_dirty_limits - background-writeback and dirty-throttling thresholds
403 * Calculate the dirty thresholds based on sysctl parameters
404 * - vm.dirty_background_ratio or vm.dirty_background_bytes
405 * - vm.dirty_ratio or vm.dirty_bytes
406 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
409 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
411 unsigned long background;
413 unsigned long uninitialized_var(available_memory);
414 struct task_struct *tsk;
416 if (!vm_dirty_bytes || !dirty_background_bytes)
417 available_memory = determine_dirtyable_memory();
420 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
422 dirty = (vm_dirty_ratio * available_memory) / 100;
424 if (dirty_background_bytes)
425 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
427 background = (dirty_background_ratio * available_memory) / 100;
429 if (background >= dirty)
430 background = dirty / 2;
432 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
433 background += background / 4;
436 *pbackground = background;
441 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
442 * @bdi: the backing_dev_info to query
443 * @dirty: global dirty limit in pages
445 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
446 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
447 * And the "limit" in the name is not seriously taken as hard limit in
448 * balance_dirty_pages().
450 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
451 * - starving fast devices
452 * - piling up dirty pages (that will take long time to sync) on slow devices
454 * The bdi's share of dirty limit will be adapting to its throughput and
455 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
457 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
460 long numerator, denominator;
463 * Calculate this BDI's share of the dirty ratio.
465 bdi_writeout_fraction(bdi, &numerator, &denominator);
467 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
468 bdi_dirty *= numerator;
469 do_div(bdi_dirty, denominator);
471 bdi_dirty += (dirty * bdi->min_ratio) / 100;
472 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
473 bdi_dirty = dirty * bdi->max_ratio / 100;
479 * balance_dirty_pages() must be called by processes which are generating dirty
480 * data. It looks at the number of dirty pages in the machine and will force
481 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
482 * If we're over `background_thresh' then the writeback threads are woken to
483 * perform some writeout.
485 static void balance_dirty_pages(struct address_space *mapping,
486 unsigned long write_chunk)
488 long nr_reclaimable, bdi_nr_reclaimable;
489 long nr_writeback, bdi_nr_writeback;
490 unsigned long background_thresh;
491 unsigned long dirty_thresh;
492 unsigned long bdi_thresh;
493 unsigned long pages_written = 0;
494 unsigned long pause = 1;
495 bool dirty_exceeded = false;
496 struct backing_dev_info *bdi = mapping->backing_dev_info;
499 struct writeback_control wbc = {
500 .sync_mode = WB_SYNC_NONE,
501 .older_than_this = NULL,
502 .nr_to_write = write_chunk,
506 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
507 global_page_state(NR_UNSTABLE_NFS);
508 nr_writeback = global_page_state(NR_WRITEBACK);
510 global_dirty_limits(&background_thresh, &dirty_thresh);
513 * Throttle it only when the background writeback cannot
514 * catch-up. This avoids (excessively) small writeouts
515 * when the bdi limits are ramping up.
517 if (nr_reclaimable + nr_writeback <=
518 (background_thresh + dirty_thresh) / 2)
521 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
522 bdi_thresh = task_dirty_limit(current, bdi_thresh);
525 * In order to avoid the stacked BDI deadlock we need
526 * to ensure we accurately count the 'dirty' pages when
527 * the threshold is low.
529 * Otherwise it would be possible to get thresh+n pages
530 * reported dirty, even though there are thresh-m pages
531 * actually dirty; with m+n sitting in the percpu
534 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
535 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
536 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
538 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
539 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
543 * The bdi thresh is somehow "soft" limit derived from the
544 * global "hard" limit. The former helps to prevent heavy IO
545 * bdi or process from holding back light ones; The latter is
546 * the last resort safeguard.
549 (bdi_nr_reclaimable + bdi_nr_writeback > bdi_thresh)
550 || (nr_reclaimable + nr_writeback > dirty_thresh);
555 if (!bdi->dirty_exceeded)
556 bdi->dirty_exceeded = 1;
558 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
559 * Unstable writes are a feature of certain networked
560 * filesystems (i.e. NFS) in which data may have been
561 * written to the server's write cache, but has not yet
562 * been flushed to permanent storage.
563 * Only move pages to writeback if this bdi is over its
564 * threshold otherwise wait until the disk writes catch
567 trace_wbc_balance_dirty_start(&wbc, bdi);
568 if (bdi_nr_reclaimable > bdi_thresh) {
569 writeback_inodes_wb(&bdi->wb, &wbc);
570 pages_written += write_chunk - wbc.nr_to_write;
571 trace_wbc_balance_dirty_written(&wbc, bdi);
572 if (pages_written >= write_chunk)
573 break; /* We've done our duty */
575 trace_wbc_balance_dirty_wait(&wbc, bdi);
576 __set_current_state(TASK_UNINTERRUPTIBLE);
577 io_schedule_timeout(pause);
580 * Increase the delay for each loop, up to our previous
581 * default of taking a 100ms nap.
588 if (!dirty_exceeded && bdi->dirty_exceeded)
589 bdi->dirty_exceeded = 0;
591 if (writeback_in_progress(bdi))
595 * In laptop mode, we wait until hitting the higher threshold before
596 * starting background writeout, and then write out all the way down
597 * to the lower threshold. So slow writers cause minimal disk activity.
599 * In normal mode, we start background writeout at the lower
600 * background_thresh, to keep the amount of dirty memory low.
602 if ((laptop_mode && pages_written) ||
603 (!laptop_mode && (nr_reclaimable > background_thresh)))
604 bdi_start_background_writeback(bdi);
607 void set_page_dirty_balance(struct page *page, int page_mkwrite)
609 if (set_page_dirty(page) || page_mkwrite) {
610 struct address_space *mapping = page_mapping(page);
613 balance_dirty_pages_ratelimited(mapping);
617 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
620 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
621 * @mapping: address_space which was dirtied
622 * @nr_pages_dirtied: number of pages which the caller has just dirtied
624 * Processes which are dirtying memory should call in here once for each page
625 * which was newly dirtied. The function will periodically check the system's
626 * dirty state and will initiate writeback if needed.
628 * On really big machines, get_writeback_state is expensive, so try to avoid
629 * calling it too often (ratelimiting). But once we're over the dirty memory
630 * limit we decrease the ratelimiting by a lot, to prevent individual processes
631 * from overshooting the limit by (ratelimit_pages) each.
633 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
634 unsigned long nr_pages_dirtied)
636 unsigned long ratelimit;
639 ratelimit = ratelimit_pages;
640 if (mapping->backing_dev_info->dirty_exceeded)
644 * Check the rate limiting. Also, we do not want to throttle real-time
645 * tasks in balance_dirty_pages(). Period.
648 p = &__get_cpu_var(bdp_ratelimits);
649 *p += nr_pages_dirtied;
650 if (unlikely(*p >= ratelimit)) {
651 ratelimit = sync_writeback_pages(*p);
654 balance_dirty_pages(mapping, ratelimit);
659 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
661 void throttle_vm_writeout(gfp_t gfp_mask)
663 unsigned long background_thresh;
664 unsigned long dirty_thresh;
667 global_dirty_limits(&background_thresh, &dirty_thresh);
670 * Boost the allowable dirty threshold a bit for page
671 * allocators so they don't get DoS'ed by heavy writers
673 dirty_thresh += dirty_thresh / 10; /* wheeee... */
675 if (global_page_state(NR_UNSTABLE_NFS) +
676 global_page_state(NR_WRITEBACK) <= dirty_thresh)
678 congestion_wait(BLK_RW_ASYNC, HZ/10);
681 * The caller might hold locks which can prevent IO completion
682 * or progress in the filesystem. So we cannot just sit here
683 * waiting for IO to complete.
685 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
691 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
693 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
694 void __user *buffer, size_t *length, loff_t *ppos)
696 proc_dointvec(table, write, buffer, length, ppos);
697 bdi_arm_supers_timer();
702 void laptop_mode_timer_fn(unsigned long data)
704 struct request_queue *q = (struct request_queue *)data;
705 int nr_pages = global_page_state(NR_FILE_DIRTY) +
706 global_page_state(NR_UNSTABLE_NFS);
709 * We want to write everything out, not just down to the dirty
712 if (bdi_has_dirty_io(&q->backing_dev_info))
713 bdi_start_writeback(&q->backing_dev_info, nr_pages);
717 * We've spun up the disk and we're in laptop mode: schedule writeback
718 * of all dirty data a few seconds from now. If the flush is already scheduled
719 * then push it back - the user is still using the disk.
721 void laptop_io_completion(struct backing_dev_info *info)
723 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
727 * We're in laptop mode and we've just synced. The sync's writes will have
728 * caused another writeback to be scheduled by laptop_io_completion.
729 * Nothing needs to be written back anymore, so we unschedule the writeback.
731 void laptop_sync_completion(void)
733 struct backing_dev_info *bdi;
737 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
738 del_timer(&bdi->laptop_mode_wb_timer);
745 * If ratelimit_pages is too high then we can get into dirty-data overload
746 * if a large number of processes all perform writes at the same time.
747 * If it is too low then SMP machines will call the (expensive)
748 * get_writeback_state too often.
750 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
751 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
752 * thresholds before writeback cuts in.
754 * But the limit should not be set too high. Because it also controls the
755 * amount of memory which the balance_dirty_pages() caller has to write back.
756 * If this is too large then the caller will block on the IO queue all the
757 * time. So limit it to four megabytes - the balance_dirty_pages() caller
758 * will write six megabyte chunks, max.
761 void writeback_set_ratelimit(void)
763 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
764 if (ratelimit_pages < 16)
765 ratelimit_pages = 16;
766 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
767 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
771 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
773 writeback_set_ratelimit();
777 static struct notifier_block __cpuinitdata ratelimit_nb = {
778 .notifier_call = ratelimit_handler,
783 * Called early on to tune the page writeback dirty limits.
785 * We used to scale dirty pages according to how total memory
786 * related to pages that could be allocated for buffers (by
787 * comparing nr_free_buffer_pages() to vm_total_pages.
789 * However, that was when we used "dirty_ratio" to scale with
790 * all memory, and we don't do that any more. "dirty_ratio"
791 * is now applied to total non-HIGHPAGE memory (by subtracting
792 * totalhigh_pages from vm_total_pages), and as such we can't
793 * get into the old insane situation any more where we had
794 * large amounts of dirty pages compared to a small amount of
795 * non-HIGHMEM memory.
797 * But we might still want to scale the dirty_ratio by how
798 * much memory the box has..
800 void __init page_writeback_init(void)
804 writeback_set_ratelimit();
805 register_cpu_notifier(&ratelimit_nb);
807 shift = calc_period_shift();
808 prop_descriptor_init(&vm_completions, shift);
809 prop_descriptor_init(&vm_dirties, shift);
813 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
814 * @mapping: address space structure to write
815 * @start: starting page index
816 * @end: ending page index (inclusive)
818 * This function scans the page range from @start to @end (inclusive) and tags
819 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
820 * that write_cache_pages (or whoever calls this function) will then use
821 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
822 * used to avoid livelocking of writeback by a process steadily creating new
823 * dirty pages in the file (thus it is important for this function to be quick
824 * so that it can tag pages faster than a dirtying process can create them).
827 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
829 void tag_pages_for_writeback(struct address_space *mapping,
830 pgoff_t start, pgoff_t end)
832 #define WRITEBACK_TAG_BATCH 4096
833 unsigned long tagged;
836 spin_lock_irq(&mapping->tree_lock);
837 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
838 &start, end, WRITEBACK_TAG_BATCH,
839 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
840 spin_unlock_irq(&mapping->tree_lock);
841 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
843 /* We check 'start' to handle wrapping when end == ~0UL */
844 } while (tagged >= WRITEBACK_TAG_BATCH && start);
846 EXPORT_SYMBOL(tag_pages_for_writeback);
849 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
850 * @mapping: address space structure to write
851 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
852 * @writepage: function called for each page
853 * @data: data passed to writepage function
855 * If a page is already under I/O, write_cache_pages() skips it, even
856 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
857 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
858 * and msync() need to guarantee that all the data which was dirty at the time
859 * the call was made get new I/O started against them. If wbc->sync_mode is
860 * WB_SYNC_ALL then we were called for data integrity and we must wait for
861 * existing IO to complete.
863 * To avoid livelocks (when other process dirties new pages), we first tag
864 * pages which should be written back with TOWRITE tag and only then start
865 * writing them. For data-integrity sync we have to be careful so that we do
866 * not miss some pages (e.g., because some other process has cleared TOWRITE
867 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
868 * by the process clearing the DIRTY tag (and submitting the page for IO).
870 int write_cache_pages(struct address_space *mapping,
871 struct writeback_control *wbc, writepage_t writepage,
878 pgoff_t uninitialized_var(writeback_index);
880 pgoff_t end; /* Inclusive */
886 pagevec_init(&pvec, 0);
887 if (wbc->range_cyclic) {
888 writeback_index = mapping->writeback_index; /* prev offset */
889 index = writeback_index;
896 index = wbc->range_start >> PAGE_CACHE_SHIFT;
897 end = wbc->range_end >> PAGE_CACHE_SHIFT;
898 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
900 cycled = 1; /* ignore range_cyclic tests */
902 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
903 tag = PAGECACHE_TAG_TOWRITE;
905 tag = PAGECACHE_TAG_DIRTY;
907 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
908 tag_pages_for_writeback(mapping, index, end);
910 while (!done && (index <= end)) {
913 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
914 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
918 for (i = 0; i < nr_pages; i++) {
919 struct page *page = pvec.pages[i];
922 * At this point, the page may be truncated or
923 * invalidated (changing page->mapping to NULL), or
924 * even swizzled back from swapper_space to tmpfs file
925 * mapping. However, page->index will not change
926 * because we have a reference on the page.
928 if (page->index > end) {
930 * can't be range_cyclic (1st pass) because
931 * end == -1 in that case.
937 done_index = page->index;
942 * Page truncated or invalidated. We can freely skip it
943 * then, even for data integrity operations: the page
944 * has disappeared concurrently, so there could be no
945 * real expectation of this data interity operation
946 * even if there is now a new, dirty page at the same
949 if (unlikely(page->mapping != mapping)) {
955 if (!PageDirty(page)) {
956 /* someone wrote it for us */
957 goto continue_unlock;
960 if (PageWriteback(page)) {
961 if (wbc->sync_mode != WB_SYNC_NONE)
962 wait_on_page_writeback(page);
964 goto continue_unlock;
967 BUG_ON(PageWriteback(page));
968 if (!clear_page_dirty_for_io(page))
969 goto continue_unlock;
971 trace_wbc_writepage(wbc, mapping->backing_dev_info);
972 ret = (*writepage)(page, wbc, data);
974 if (ret == AOP_WRITEPAGE_ACTIVATE) {
979 * done_index is set past this page,
980 * so media errors will not choke
981 * background writeout for the entire
982 * file. This has consequences for
983 * range_cyclic semantics (ie. it may
984 * not be suitable for data integrity
987 done_index = page->index + 1;
994 * We stop writing back only if we are not doing
995 * integrity sync. In case of integrity sync we have to
996 * keep going until we have written all the pages
997 * we tagged for writeback prior to entering this loop.
999 if (--wbc->nr_to_write <= 0 &&
1000 wbc->sync_mode == WB_SYNC_NONE) {
1005 pagevec_release(&pvec);
1008 if (!cycled && !done) {
1011 * We hit the last page and there is more work to be done: wrap
1012 * back to the start of the file
1016 end = writeback_index - 1;
1019 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1020 mapping->writeback_index = done_index;
1024 EXPORT_SYMBOL(write_cache_pages);
1027 * Function used by generic_writepages to call the real writepage
1028 * function and set the mapping flags on error
1030 static int __writepage(struct page *page, struct writeback_control *wbc,
1033 struct address_space *mapping = data;
1034 int ret = mapping->a_ops->writepage(page, wbc);
1035 mapping_set_error(mapping, ret);
1040 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1041 * @mapping: address space structure to write
1042 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1044 * This is a library function, which implements the writepages()
1045 * address_space_operation.
1047 int generic_writepages(struct address_space *mapping,
1048 struct writeback_control *wbc)
1050 struct blk_plug plug;
1053 /* deal with chardevs and other special file */
1054 if (!mapping->a_ops->writepage)
1057 blk_start_plug(&plug);
1058 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1059 blk_finish_plug(&plug);
1063 EXPORT_SYMBOL(generic_writepages);
1065 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1069 if (wbc->nr_to_write <= 0)
1071 if (mapping->a_ops->writepages)
1072 ret = mapping->a_ops->writepages(mapping, wbc);
1074 ret = generic_writepages(mapping, wbc);
1079 * write_one_page - write out a single page and optionally wait on I/O
1080 * @page: the page to write
1081 * @wait: if true, wait on writeout
1083 * The page must be locked by the caller and will be unlocked upon return.
1085 * write_one_page() returns a negative error code if I/O failed.
1087 int write_one_page(struct page *page, int wait)
1089 struct address_space *mapping = page->mapping;
1091 struct writeback_control wbc = {
1092 .sync_mode = WB_SYNC_ALL,
1096 BUG_ON(!PageLocked(page));
1099 wait_on_page_writeback(page);
1101 if (clear_page_dirty_for_io(page)) {
1102 page_cache_get(page);
1103 ret = mapping->a_ops->writepage(page, &wbc);
1104 if (ret == 0 && wait) {
1105 wait_on_page_writeback(page);
1106 if (PageError(page))
1109 page_cache_release(page);
1115 EXPORT_SYMBOL(write_one_page);
1118 * For address_spaces which do not use buffers nor write back.
1120 int __set_page_dirty_no_writeback(struct page *page)
1122 if (!PageDirty(page))
1123 return !TestSetPageDirty(page);
1128 * Helper function for set_page_dirty family.
1129 * NOTE: This relies on being atomic wrt interrupts.
1131 void account_page_dirtied(struct page *page, struct address_space *mapping)
1133 if (mapping_cap_account_dirty(mapping)) {
1134 __inc_zone_page_state(page, NR_FILE_DIRTY);
1135 __inc_zone_page_state(page, NR_DIRTIED);
1136 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1137 task_dirty_inc(current);
1138 task_io_account_write(PAGE_CACHE_SIZE);
1141 EXPORT_SYMBOL(account_page_dirtied);
1144 * Helper function for set_page_writeback family.
1145 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1148 void account_page_writeback(struct page *page)
1150 inc_zone_page_state(page, NR_WRITEBACK);
1151 inc_zone_page_state(page, NR_WRITTEN);
1153 EXPORT_SYMBOL(account_page_writeback);
1156 * For address_spaces which do not use buffers. Just tag the page as dirty in
1159 * This is also used when a single buffer is being dirtied: we want to set the
1160 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1161 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1163 * Most callers have locked the page, which pins the address_space in memory.
1164 * But zap_pte_range() does not lock the page, however in that case the
1165 * mapping is pinned by the vma's ->vm_file reference.
1167 * We take care to handle the case where the page was truncated from the
1168 * mapping by re-checking page_mapping() inside tree_lock.
1170 int __set_page_dirty_nobuffers(struct page *page)
1172 if (!TestSetPageDirty(page)) {
1173 struct address_space *mapping = page_mapping(page);
1174 struct address_space *mapping2;
1179 spin_lock_irq(&mapping->tree_lock);
1180 mapping2 = page_mapping(page);
1181 if (mapping2) { /* Race with truncate? */
1182 BUG_ON(mapping2 != mapping);
1183 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1184 account_page_dirtied(page, mapping);
1185 radix_tree_tag_set(&mapping->page_tree,
1186 page_index(page), PAGECACHE_TAG_DIRTY);
1188 spin_unlock_irq(&mapping->tree_lock);
1189 if (mapping->host) {
1190 /* !PageAnon && !swapper_space */
1191 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1197 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1200 * When a writepage implementation decides that it doesn't want to write this
1201 * page for some reason, it should redirty the locked page via
1202 * redirty_page_for_writepage() and it should then unlock the page and return 0
1204 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1206 wbc->pages_skipped++;
1207 return __set_page_dirty_nobuffers(page);
1209 EXPORT_SYMBOL(redirty_page_for_writepage);
1214 * For pages with a mapping this should be done under the page lock
1215 * for the benefit of asynchronous memory errors who prefer a consistent
1216 * dirty state. This rule can be broken in some special cases,
1217 * but should be better not to.
1219 * If the mapping doesn't provide a set_page_dirty a_op, then
1220 * just fall through and assume that it wants buffer_heads.
1222 int set_page_dirty(struct page *page)
1224 struct address_space *mapping = page_mapping(page);
1226 if (likely(mapping)) {
1227 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1229 * readahead/lru_deactivate_page could remain
1230 * PG_readahead/PG_reclaim due to race with end_page_writeback
1231 * About readahead, if the page is written, the flags would be
1232 * reset. So no problem.
1233 * About lru_deactivate_page, if the page is redirty, the flag
1234 * will be reset. So no problem. but if the page is used by readahead
1235 * it will confuse readahead and make it restart the size rampup
1236 * process. But it's a trivial problem.
1238 ClearPageReclaim(page);
1241 spd = __set_page_dirty_buffers;
1243 return (*spd)(page);
1245 if (!PageDirty(page)) {
1246 if (!TestSetPageDirty(page))
1251 EXPORT_SYMBOL(set_page_dirty);
1254 * set_page_dirty() is racy if the caller has no reference against
1255 * page->mapping->host, and if the page is unlocked. This is because another
1256 * CPU could truncate the page off the mapping and then free the mapping.
1258 * Usually, the page _is_ locked, or the caller is a user-space process which
1259 * holds a reference on the inode by having an open file.
1261 * In other cases, the page should be locked before running set_page_dirty().
1263 int set_page_dirty_lock(struct page *page)
1268 ret = set_page_dirty(page);
1272 EXPORT_SYMBOL(set_page_dirty_lock);
1275 * Clear a page's dirty flag, while caring for dirty memory accounting.
1276 * Returns true if the page was previously dirty.
1278 * This is for preparing to put the page under writeout. We leave the page
1279 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1280 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1281 * implementation will run either set_page_writeback() or set_page_dirty(),
1282 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1285 * This incoherency between the page's dirty flag and radix-tree tag is
1286 * unfortunate, but it only exists while the page is locked.
1288 int clear_page_dirty_for_io(struct page *page)
1290 struct address_space *mapping = page_mapping(page);
1292 BUG_ON(!PageLocked(page));
1294 if (mapping && mapping_cap_account_dirty(mapping)) {
1296 * Yes, Virginia, this is indeed insane.
1298 * We use this sequence to make sure that
1299 * (a) we account for dirty stats properly
1300 * (b) we tell the low-level filesystem to
1301 * mark the whole page dirty if it was
1302 * dirty in a pagetable. Only to then
1303 * (c) clean the page again and return 1 to
1304 * cause the writeback.
1306 * This way we avoid all nasty races with the
1307 * dirty bit in multiple places and clearing
1308 * them concurrently from different threads.
1310 * Note! Normally the "set_page_dirty(page)"
1311 * has no effect on the actual dirty bit - since
1312 * that will already usually be set. But we
1313 * need the side effects, and it can help us
1316 * We basically use the page "master dirty bit"
1317 * as a serialization point for all the different
1318 * threads doing their things.
1320 if (page_mkclean(page))
1321 set_page_dirty(page);
1323 * We carefully synchronise fault handlers against
1324 * installing a dirty pte and marking the page dirty
1325 * at this point. We do this by having them hold the
1326 * page lock at some point after installing their
1327 * pte, but before marking the page dirty.
1328 * Pages are always locked coming in here, so we get
1329 * the desired exclusion. See mm/memory.c:do_wp_page()
1330 * for more comments.
1332 if (TestClearPageDirty(page)) {
1333 dec_zone_page_state(page, NR_FILE_DIRTY);
1334 dec_bdi_stat(mapping->backing_dev_info,
1340 return TestClearPageDirty(page);
1342 EXPORT_SYMBOL(clear_page_dirty_for_io);
1344 int test_clear_page_writeback(struct page *page)
1346 struct address_space *mapping = page_mapping(page);
1350 struct backing_dev_info *bdi = mapping->backing_dev_info;
1351 unsigned long flags;
1353 spin_lock_irqsave(&mapping->tree_lock, flags);
1354 ret = TestClearPageWriteback(page);
1356 radix_tree_tag_clear(&mapping->page_tree,
1358 PAGECACHE_TAG_WRITEBACK);
1359 if (bdi_cap_account_writeback(bdi)) {
1360 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1361 __bdi_writeout_inc(bdi);
1364 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1366 ret = TestClearPageWriteback(page);
1369 dec_zone_page_state(page, NR_WRITEBACK);
1373 int test_set_page_writeback(struct page *page)
1375 struct address_space *mapping = page_mapping(page);
1379 struct backing_dev_info *bdi = mapping->backing_dev_info;
1380 unsigned long flags;
1382 spin_lock_irqsave(&mapping->tree_lock, flags);
1383 ret = TestSetPageWriteback(page);
1385 radix_tree_tag_set(&mapping->page_tree,
1387 PAGECACHE_TAG_WRITEBACK);
1388 if (bdi_cap_account_writeback(bdi))
1389 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1391 if (!PageDirty(page))
1392 radix_tree_tag_clear(&mapping->page_tree,
1394 PAGECACHE_TAG_DIRTY);
1395 radix_tree_tag_clear(&mapping->page_tree,
1397 PAGECACHE_TAG_TOWRITE);
1398 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1400 ret = TestSetPageWriteback(page);
1403 account_page_writeback(page);
1407 EXPORT_SYMBOL(test_set_page_writeback);
1410 * Return true if any of the pages in the mapping are marked with the
1413 int mapping_tagged(struct address_space *mapping, int tag)
1417 ret = radix_tree_tagged(&mapping->page_tree, tag);
1421 EXPORT_SYMBOL(mapping_tagged);