Linux 4.19.133
[linux/fpc-iii.git] / fs / fs-writeback.c
blob7bfeb1643c1f4d74dfc573a874fa60a604c44e74
1 /*
2 * fs/fs-writeback.c
4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include <linux/device.h>
30 #include <linux/memcontrol.h>
31 #include "internal.h"
34 * 4MB minimal write chunk size
36 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
38 struct wb_completion {
39 atomic_t cnt;
43 * Passed into wb_writeback(), essentially a subset of writeback_control
45 struct wb_writeback_work {
46 long nr_pages;
47 struct super_block *sb;
48 unsigned long *older_than_this;
49 enum writeback_sync_modes sync_mode;
50 unsigned int tagged_writepages:1;
51 unsigned int for_kupdate:1;
52 unsigned int range_cyclic:1;
53 unsigned int for_background:1;
54 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
55 unsigned int auto_free:1; /* free on completion */
56 enum wb_reason reason; /* why was writeback initiated? */
58 struct list_head list; /* pending work list */
59 struct wb_completion *done; /* set if the caller waits */
63 * If one wants to wait for one or more wb_writeback_works, each work's
64 * ->done should be set to a wb_completion defined using the following
65 * macro. Once all work items are issued with wb_queue_work(), the caller
66 * can wait for the completion of all using wb_wait_for_completion(). Work
67 * items which are waited upon aren't freed automatically on completion.
69 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
70 struct wb_completion cmpl = { \
71 .cnt = ATOMIC_INIT(1), \
76 * If an inode is constantly having its pages dirtied, but then the
77 * updates stop dirtytime_expire_interval seconds in the past, it's
78 * possible for the worst case time between when an inode has its
79 * timestamps updated and when they finally get written out to be two
80 * dirtytime_expire_intervals. We set the default to 12 hours (in
81 * seconds), which means most of the time inodes will have their
82 * timestamps written to disk after 12 hours, but in the worst case a
83 * few inodes might not their timestamps updated for 24 hours.
85 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
87 static inline struct inode *wb_inode(struct list_head *head)
89 return list_entry(head, struct inode, i_io_list);
93 * Include the creation of the trace points after defining the
94 * wb_writeback_work structure and inline functions so that the definition
95 * remains local to this file.
97 #define CREATE_TRACE_POINTS
98 #include <trace/events/writeback.h>
100 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
102 static bool wb_io_lists_populated(struct bdi_writeback *wb)
104 if (wb_has_dirty_io(wb)) {
105 return false;
106 } else {
107 set_bit(WB_has_dirty_io, &wb->state);
108 WARN_ON_ONCE(!wb->avg_write_bandwidth);
109 atomic_long_add(wb->avg_write_bandwidth,
110 &wb->bdi->tot_write_bandwidth);
111 return true;
115 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
117 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
118 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
119 clear_bit(WB_has_dirty_io, &wb->state);
120 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
121 &wb->bdi->tot_write_bandwidth) < 0);
126 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
127 * @inode: inode to be moved
128 * @wb: target bdi_writeback
129 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
131 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
132 * Returns %true if @inode is the first occupant of the !dirty_time IO
133 * lists; otherwise, %false.
135 static bool inode_io_list_move_locked(struct inode *inode,
136 struct bdi_writeback *wb,
137 struct list_head *head)
139 assert_spin_locked(&wb->list_lock);
141 list_move(&inode->i_io_list, head);
143 /* dirty_time doesn't count as dirty_io until expiration */
144 if (head != &wb->b_dirty_time)
145 return wb_io_lists_populated(wb);
147 wb_io_lists_depopulated(wb);
148 return false;
152 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
153 * @inode: inode to be removed
154 * @wb: bdi_writeback @inode is being removed from
156 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
157 * clear %WB_has_dirty_io if all are empty afterwards.
159 static void inode_io_list_del_locked(struct inode *inode,
160 struct bdi_writeback *wb)
162 assert_spin_locked(&wb->list_lock);
164 list_del_init(&inode->i_io_list);
165 wb_io_lists_depopulated(wb);
168 static void wb_wakeup(struct bdi_writeback *wb)
170 spin_lock_bh(&wb->work_lock);
171 if (test_bit(WB_registered, &wb->state))
172 mod_delayed_work(bdi_wq, &wb->dwork, 0);
173 spin_unlock_bh(&wb->work_lock);
176 static void finish_writeback_work(struct bdi_writeback *wb,
177 struct wb_writeback_work *work)
179 struct wb_completion *done = work->done;
181 if (work->auto_free)
182 kfree(work);
183 if (done && atomic_dec_and_test(&done->cnt))
184 wake_up_all(&wb->bdi->wb_waitq);
187 static void wb_queue_work(struct bdi_writeback *wb,
188 struct wb_writeback_work *work)
190 trace_writeback_queue(wb, work);
192 if (work->done)
193 atomic_inc(&work->done->cnt);
195 spin_lock_bh(&wb->work_lock);
197 if (test_bit(WB_registered, &wb->state)) {
198 list_add_tail(&work->list, &wb->work_list);
199 mod_delayed_work(bdi_wq, &wb->dwork, 0);
200 } else
201 finish_writeback_work(wb, work);
203 spin_unlock_bh(&wb->work_lock);
207 * wb_wait_for_completion - wait for completion of bdi_writeback_works
208 * @bdi: bdi work items were issued to
209 * @done: target wb_completion
211 * Wait for one or more work items issued to @bdi with their ->done field
212 * set to @done, which should have been defined with
213 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
214 * work items are completed. Work items which are waited upon aren't freed
215 * automatically on completion.
217 static void wb_wait_for_completion(struct backing_dev_info *bdi,
218 struct wb_completion *done)
220 atomic_dec(&done->cnt); /* put down the initial count */
221 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
224 #ifdef CONFIG_CGROUP_WRITEBACK
226 /* parameters for foreign inode detection, see wb_detach_inode() */
227 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
228 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
229 #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
230 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
232 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
233 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
234 /* each slot's duration is 2s / 16 */
235 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
236 /* if foreign slots >= 8, switch */
237 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
238 /* one round can affect upto 5 slots */
240 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
241 static struct workqueue_struct *isw_wq;
243 void __inode_attach_wb(struct inode *inode, struct page *page)
245 struct backing_dev_info *bdi = inode_to_bdi(inode);
246 struct bdi_writeback *wb = NULL;
248 if (inode_cgwb_enabled(inode)) {
249 struct cgroup_subsys_state *memcg_css;
251 if (page) {
252 memcg_css = mem_cgroup_css_from_page(page);
253 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
254 } else {
255 /* must pin memcg_css, see wb_get_create() */
256 memcg_css = task_get_css(current, memory_cgrp_id);
257 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
258 css_put(memcg_css);
262 if (!wb)
263 wb = &bdi->wb;
266 * There may be multiple instances of this function racing to
267 * update the same inode. Use cmpxchg() to tell the winner.
269 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
270 wb_put(wb);
272 EXPORT_SYMBOL_GPL(__inode_attach_wb);
275 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
276 * @inode: inode of interest with i_lock held
278 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
279 * held on entry and is released on return. The returned wb is guaranteed
280 * to stay @inode's associated wb until its list_lock is released.
282 static struct bdi_writeback *
283 locked_inode_to_wb_and_lock_list(struct inode *inode)
284 __releases(&inode->i_lock)
285 __acquires(&wb->list_lock)
287 while (true) {
288 struct bdi_writeback *wb = inode_to_wb(inode);
291 * inode_to_wb() association is protected by both
292 * @inode->i_lock and @wb->list_lock but list_lock nests
293 * outside i_lock. Drop i_lock and verify that the
294 * association hasn't changed after acquiring list_lock.
296 wb_get(wb);
297 spin_unlock(&inode->i_lock);
298 spin_lock(&wb->list_lock);
300 /* i_wb may have changed inbetween, can't use inode_to_wb() */
301 if (likely(wb == inode->i_wb)) {
302 wb_put(wb); /* @inode already has ref */
303 return wb;
306 spin_unlock(&wb->list_lock);
307 wb_put(wb);
308 cpu_relax();
309 spin_lock(&inode->i_lock);
314 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
315 * @inode: inode of interest
317 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
318 * on entry.
320 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
321 __acquires(&wb->list_lock)
323 spin_lock(&inode->i_lock);
324 return locked_inode_to_wb_and_lock_list(inode);
327 struct inode_switch_wbs_context {
328 struct inode *inode;
329 struct bdi_writeback *new_wb;
331 struct rcu_head rcu_head;
332 struct work_struct work;
335 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
337 down_write(&bdi->wb_switch_rwsem);
340 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
342 up_write(&bdi->wb_switch_rwsem);
345 static void inode_switch_wbs_work_fn(struct work_struct *work)
347 struct inode_switch_wbs_context *isw =
348 container_of(work, struct inode_switch_wbs_context, work);
349 struct inode *inode = isw->inode;
350 struct backing_dev_info *bdi = inode_to_bdi(inode);
351 struct address_space *mapping = inode->i_mapping;
352 struct bdi_writeback *old_wb = inode->i_wb;
353 struct bdi_writeback *new_wb = isw->new_wb;
354 struct radix_tree_iter iter;
355 bool switched = false;
356 void **slot;
359 * If @inode switches cgwb membership while sync_inodes_sb() is
360 * being issued, sync_inodes_sb() might miss it. Synchronize.
362 down_read(&bdi->wb_switch_rwsem);
365 * By the time control reaches here, RCU grace period has passed
366 * since I_WB_SWITCH assertion and all wb stat update transactions
367 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
368 * synchronizing against the i_pages lock.
370 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
371 * gives us exclusion against all wb related operations on @inode
372 * including IO list manipulations and stat updates.
374 if (old_wb < new_wb) {
375 spin_lock(&old_wb->list_lock);
376 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
377 } else {
378 spin_lock(&new_wb->list_lock);
379 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
381 spin_lock(&inode->i_lock);
382 xa_lock_irq(&mapping->i_pages);
385 * Once I_FREEING is visible under i_lock, the eviction path owns
386 * the inode and we shouldn't modify ->i_io_list.
388 if (unlikely(inode->i_state & I_FREEING))
389 goto skip_switch;
392 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
393 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
394 * pages actually under writeback.
396 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, 0,
397 PAGECACHE_TAG_DIRTY) {
398 struct page *page = radix_tree_deref_slot_protected(slot,
399 &mapping->i_pages.xa_lock);
400 if (likely(page) && PageDirty(page)) {
401 dec_wb_stat(old_wb, WB_RECLAIMABLE);
402 inc_wb_stat(new_wb, WB_RECLAIMABLE);
406 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, 0,
407 PAGECACHE_TAG_WRITEBACK) {
408 struct page *page = radix_tree_deref_slot_protected(slot,
409 &mapping->i_pages.xa_lock);
410 if (likely(page)) {
411 WARN_ON_ONCE(!PageWriteback(page));
412 dec_wb_stat(old_wb, WB_WRITEBACK);
413 inc_wb_stat(new_wb, WB_WRITEBACK);
417 wb_get(new_wb);
420 * Transfer to @new_wb's IO list if necessary. The specific list
421 * @inode was on is ignored and the inode is put on ->b_dirty which
422 * is always correct including from ->b_dirty_time. The transfer
423 * preserves @inode->dirtied_when ordering.
425 if (!list_empty(&inode->i_io_list)) {
426 struct inode *pos;
428 inode_io_list_del_locked(inode, old_wb);
429 inode->i_wb = new_wb;
430 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
431 if (time_after_eq(inode->dirtied_when,
432 pos->dirtied_when))
433 break;
434 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
435 } else {
436 inode->i_wb = new_wb;
439 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
440 inode->i_wb_frn_winner = 0;
441 inode->i_wb_frn_avg_time = 0;
442 inode->i_wb_frn_history = 0;
443 switched = true;
444 skip_switch:
446 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
447 * ensures that the new wb is visible if they see !I_WB_SWITCH.
449 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
451 xa_unlock_irq(&mapping->i_pages);
452 spin_unlock(&inode->i_lock);
453 spin_unlock(&new_wb->list_lock);
454 spin_unlock(&old_wb->list_lock);
456 up_read(&bdi->wb_switch_rwsem);
458 if (switched) {
459 wb_wakeup(new_wb);
460 wb_put(old_wb);
462 wb_put(new_wb);
464 iput(inode);
465 kfree(isw);
467 atomic_dec(&isw_nr_in_flight);
470 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
472 struct inode_switch_wbs_context *isw = container_of(rcu_head,
473 struct inode_switch_wbs_context, rcu_head);
475 /* needs to grab bh-unsafe locks, bounce to work item */
476 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
477 queue_work(isw_wq, &isw->work);
481 * inode_switch_wbs - change the wb association of an inode
482 * @inode: target inode
483 * @new_wb_id: ID of the new wb
485 * Switch @inode's wb association to the wb identified by @new_wb_id. The
486 * switching is performed asynchronously and may fail silently.
488 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
490 struct backing_dev_info *bdi = inode_to_bdi(inode);
491 struct cgroup_subsys_state *memcg_css;
492 struct inode_switch_wbs_context *isw;
494 /* noop if seems to be already in progress */
495 if (inode->i_state & I_WB_SWITCH)
496 return;
499 * Avoid starting new switches while sync_inodes_sb() is in
500 * progress. Otherwise, if the down_write protected issue path
501 * blocks heavily, we might end up starting a large number of
502 * switches which will block on the rwsem.
504 if (!down_read_trylock(&bdi->wb_switch_rwsem))
505 return;
507 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
508 if (!isw)
509 goto out_unlock;
511 /* find and pin the new wb */
512 rcu_read_lock();
513 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
514 if (memcg_css)
515 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
516 rcu_read_unlock();
517 if (!isw->new_wb)
518 goto out_free;
520 /* while holding I_WB_SWITCH, no one else can update the association */
521 spin_lock(&inode->i_lock);
522 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
523 inode->i_state & (I_WB_SWITCH | I_FREEING) ||
524 inode_to_wb(inode) == isw->new_wb) {
525 spin_unlock(&inode->i_lock);
526 goto out_free;
528 inode->i_state |= I_WB_SWITCH;
529 __iget(inode);
530 spin_unlock(&inode->i_lock);
532 isw->inode = inode;
535 * In addition to synchronizing among switchers, I_WB_SWITCH tells
536 * the RCU protected stat update paths to grab the i_page
537 * lock so that stat transfer can synchronize against them.
538 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
540 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
542 atomic_inc(&isw_nr_in_flight);
544 goto out_unlock;
546 out_free:
547 if (isw->new_wb)
548 wb_put(isw->new_wb);
549 kfree(isw);
550 out_unlock:
551 up_read(&bdi->wb_switch_rwsem);
555 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
556 * @wbc: writeback_control of interest
557 * @inode: target inode
559 * @inode is locked and about to be written back under the control of @wbc.
560 * Record @inode's writeback context into @wbc and unlock the i_lock. On
561 * writeback completion, wbc_detach_inode() should be called. This is used
562 * to track the cgroup writeback context.
564 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
565 struct inode *inode)
567 if (!inode_cgwb_enabled(inode)) {
568 spin_unlock(&inode->i_lock);
569 return;
572 wbc->wb = inode_to_wb(inode);
573 wbc->inode = inode;
575 wbc->wb_id = wbc->wb->memcg_css->id;
576 wbc->wb_lcand_id = inode->i_wb_frn_winner;
577 wbc->wb_tcand_id = 0;
578 wbc->wb_bytes = 0;
579 wbc->wb_lcand_bytes = 0;
580 wbc->wb_tcand_bytes = 0;
582 wb_get(wbc->wb);
583 spin_unlock(&inode->i_lock);
586 * A dying wb indicates that either the blkcg associated with the
587 * memcg changed or the associated memcg is dying. In the first
588 * case, a replacement wb should already be available and we should
589 * refresh the wb immediately. In the second case, trying to
590 * refresh will keep failing.
592 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
593 inode_switch_wbs(inode, wbc->wb_id);
597 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
598 * @wbc: writeback_control of the just finished writeback
600 * To be called after a writeback attempt of an inode finishes and undoes
601 * wbc_attach_and_unlock_inode(). Can be called under any context.
603 * As concurrent write sharing of an inode is expected to be very rare and
604 * memcg only tracks page ownership on first-use basis severely confining
605 * the usefulness of such sharing, cgroup writeback tracks ownership
606 * per-inode. While the support for concurrent write sharing of an inode
607 * is deemed unnecessary, an inode being written to by different cgroups at
608 * different points in time is a lot more common, and, more importantly,
609 * charging only by first-use can too readily lead to grossly incorrect
610 * behaviors (single foreign page can lead to gigabytes of writeback to be
611 * incorrectly attributed).
613 * To resolve this issue, cgroup writeback detects the majority dirtier of
614 * an inode and transfers the ownership to it. To avoid unnnecessary
615 * oscillation, the detection mechanism keeps track of history and gives
616 * out the switch verdict only if the foreign usage pattern is stable over
617 * a certain amount of time and/or writeback attempts.
619 * On each writeback attempt, @wbc tries to detect the majority writer
620 * using Boyer-Moore majority vote algorithm. In addition to the byte
621 * count from the majority voting, it also counts the bytes written for the
622 * current wb and the last round's winner wb (max of last round's current
623 * wb, the winner from two rounds ago, and the last round's majority
624 * candidate). Keeping track of the historical winner helps the algorithm
625 * to semi-reliably detect the most active writer even when it's not the
626 * absolute majority.
628 * Once the winner of the round is determined, whether the winner is
629 * foreign or not and how much IO time the round consumed is recorded in
630 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
631 * over a certain threshold, the switch verdict is given.
633 void wbc_detach_inode(struct writeback_control *wbc)
635 struct bdi_writeback *wb = wbc->wb;
636 struct inode *inode = wbc->inode;
637 unsigned long avg_time, max_bytes, max_time;
638 u16 history;
639 int max_id;
641 if (!wb)
642 return;
644 history = inode->i_wb_frn_history;
645 avg_time = inode->i_wb_frn_avg_time;
647 /* pick the winner of this round */
648 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
649 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
650 max_id = wbc->wb_id;
651 max_bytes = wbc->wb_bytes;
652 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
653 max_id = wbc->wb_lcand_id;
654 max_bytes = wbc->wb_lcand_bytes;
655 } else {
656 max_id = wbc->wb_tcand_id;
657 max_bytes = wbc->wb_tcand_bytes;
661 * Calculate the amount of IO time the winner consumed and fold it
662 * into the running average kept per inode. If the consumed IO
663 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
664 * deciding whether to switch or not. This is to prevent one-off
665 * small dirtiers from skewing the verdict.
667 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
668 wb->avg_write_bandwidth);
669 if (avg_time)
670 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
671 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
672 else
673 avg_time = max_time; /* immediate catch up on first run */
675 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
676 int slots;
679 * The switch verdict is reached if foreign wb's consume
680 * more than a certain proportion of IO time in a
681 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
682 * history mask where each bit represents one sixteenth of
683 * the period. Determine the number of slots to shift into
684 * history from @max_time.
686 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
687 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
688 history <<= slots;
689 if (wbc->wb_id != max_id)
690 history |= (1U << slots) - 1;
693 * Switch if the current wb isn't the consistent winner.
694 * If there are multiple closely competing dirtiers, the
695 * inode may switch across them repeatedly over time, which
696 * is okay. The main goal is avoiding keeping an inode on
697 * the wrong wb for an extended period of time.
699 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
700 inode_switch_wbs(inode, max_id);
704 * Multiple instances of this function may race to update the
705 * following fields but we don't mind occassional inaccuracies.
707 inode->i_wb_frn_winner = max_id;
708 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
709 inode->i_wb_frn_history = history;
711 wb_put(wbc->wb);
712 wbc->wb = NULL;
716 * wbc_account_io - account IO issued during writeback
717 * @wbc: writeback_control of the writeback in progress
718 * @page: page being written out
719 * @bytes: number of bytes being written out
721 * @bytes from @page are about to written out during the writeback
722 * controlled by @wbc. Keep the book for foreign inode detection. See
723 * wbc_detach_inode().
725 void wbc_account_io(struct writeback_control *wbc, struct page *page,
726 size_t bytes)
728 struct cgroup_subsys_state *css;
729 int id;
732 * pageout() path doesn't attach @wbc to the inode being written
733 * out. This is intentional as we don't want the function to block
734 * behind a slow cgroup. Ultimately, we want pageout() to kick off
735 * regular writeback instead of writing things out itself.
737 if (!wbc->wb)
738 return;
740 css = mem_cgroup_css_from_page(page);
741 /* dead cgroups shouldn't contribute to inode ownership arbitration */
742 if (!(css->flags & CSS_ONLINE))
743 return;
745 id = css->id;
747 if (id == wbc->wb_id) {
748 wbc->wb_bytes += bytes;
749 return;
752 if (id == wbc->wb_lcand_id)
753 wbc->wb_lcand_bytes += bytes;
755 /* Boyer-Moore majority vote algorithm */
756 if (!wbc->wb_tcand_bytes)
757 wbc->wb_tcand_id = id;
758 if (id == wbc->wb_tcand_id)
759 wbc->wb_tcand_bytes += bytes;
760 else
761 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
763 EXPORT_SYMBOL_GPL(wbc_account_io);
766 * inode_congested - test whether an inode is congested
767 * @inode: inode to test for congestion (may be NULL)
768 * @cong_bits: mask of WB_[a]sync_congested bits to test
770 * Tests whether @inode is congested. @cong_bits is the mask of congestion
771 * bits to test and the return value is the mask of set bits.
773 * If cgroup writeback is enabled for @inode, the congestion state is
774 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
775 * associated with @inode is congested; otherwise, the root wb's congestion
776 * state is used.
778 * @inode is allowed to be NULL as this function is often called on
779 * mapping->host which is NULL for the swapper space.
781 int inode_congested(struct inode *inode, int cong_bits)
784 * Once set, ->i_wb never becomes NULL while the inode is alive.
785 * Start transaction iff ->i_wb is visible.
787 if (inode && inode_to_wb_is_valid(inode)) {
788 struct bdi_writeback *wb;
789 struct wb_lock_cookie lock_cookie = {};
790 bool congested;
792 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
793 congested = wb_congested(wb, cong_bits);
794 unlocked_inode_to_wb_end(inode, &lock_cookie);
795 return congested;
798 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
800 EXPORT_SYMBOL_GPL(inode_congested);
803 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
804 * @wb: target bdi_writeback to split @nr_pages to
805 * @nr_pages: number of pages to write for the whole bdi
807 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
808 * relation to the total write bandwidth of all wb's w/ dirty inodes on
809 * @wb->bdi.
811 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
813 unsigned long this_bw = wb->avg_write_bandwidth;
814 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
816 if (nr_pages == LONG_MAX)
817 return LONG_MAX;
820 * This may be called on clean wb's and proportional distribution
821 * may not make sense, just use the original @nr_pages in those
822 * cases. In general, we wanna err on the side of writing more.
824 if (!tot_bw || this_bw >= tot_bw)
825 return nr_pages;
826 else
827 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
831 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
832 * @bdi: target backing_dev_info
833 * @base_work: wb_writeback_work to issue
834 * @skip_if_busy: skip wb's which already have writeback in progress
836 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
837 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
838 * distributed to the busy wbs according to each wb's proportion in the
839 * total active write bandwidth of @bdi.
841 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
842 struct wb_writeback_work *base_work,
843 bool skip_if_busy)
845 struct bdi_writeback *last_wb = NULL;
846 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
847 struct bdi_writeback, bdi_node);
849 might_sleep();
850 restart:
851 rcu_read_lock();
852 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
853 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
854 struct wb_writeback_work fallback_work;
855 struct wb_writeback_work *work;
856 long nr_pages;
858 if (last_wb) {
859 wb_put(last_wb);
860 last_wb = NULL;
863 /* SYNC_ALL writes out I_DIRTY_TIME too */
864 if (!wb_has_dirty_io(wb) &&
865 (base_work->sync_mode == WB_SYNC_NONE ||
866 list_empty(&wb->b_dirty_time)))
867 continue;
868 if (skip_if_busy && writeback_in_progress(wb))
869 continue;
871 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
873 work = kmalloc(sizeof(*work), GFP_ATOMIC);
874 if (work) {
875 *work = *base_work;
876 work->nr_pages = nr_pages;
877 work->auto_free = 1;
878 wb_queue_work(wb, work);
879 continue;
882 /* alloc failed, execute synchronously using on-stack fallback */
883 work = &fallback_work;
884 *work = *base_work;
885 work->nr_pages = nr_pages;
886 work->auto_free = 0;
887 work->done = &fallback_work_done;
889 wb_queue_work(wb, work);
892 * Pin @wb so that it stays on @bdi->wb_list. This allows
893 * continuing iteration from @wb after dropping and
894 * regrabbing rcu read lock.
896 wb_get(wb);
897 last_wb = wb;
899 rcu_read_unlock();
900 wb_wait_for_completion(bdi, &fallback_work_done);
901 goto restart;
903 rcu_read_unlock();
905 if (last_wb)
906 wb_put(last_wb);
910 * cgroup_writeback_umount - flush inode wb switches for umount
912 * This function is called when a super_block is about to be destroyed and
913 * flushes in-flight inode wb switches. An inode wb switch goes through
914 * RCU and then workqueue, so the two need to be flushed in order to ensure
915 * that all previously scheduled switches are finished. As wb switches are
916 * rare occurrences and synchronize_rcu() can take a while, perform
917 * flushing iff wb switches are in flight.
919 void cgroup_writeback_umount(void)
921 if (atomic_read(&isw_nr_in_flight)) {
923 * Use rcu_barrier() to wait for all pending callbacks to
924 * ensure that all in-flight wb switches are in the workqueue.
926 rcu_barrier();
927 flush_workqueue(isw_wq);
931 static int __init cgroup_writeback_init(void)
933 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
934 if (!isw_wq)
935 return -ENOMEM;
936 return 0;
938 fs_initcall(cgroup_writeback_init);
940 #else /* CONFIG_CGROUP_WRITEBACK */
942 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
943 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
945 static struct bdi_writeback *
946 locked_inode_to_wb_and_lock_list(struct inode *inode)
947 __releases(&inode->i_lock)
948 __acquires(&wb->list_lock)
950 struct bdi_writeback *wb = inode_to_wb(inode);
952 spin_unlock(&inode->i_lock);
953 spin_lock(&wb->list_lock);
954 return wb;
957 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
958 __acquires(&wb->list_lock)
960 struct bdi_writeback *wb = inode_to_wb(inode);
962 spin_lock(&wb->list_lock);
963 return wb;
966 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
968 return nr_pages;
971 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
972 struct wb_writeback_work *base_work,
973 bool skip_if_busy)
975 might_sleep();
977 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
978 base_work->auto_free = 0;
979 wb_queue_work(&bdi->wb, base_work);
983 #endif /* CONFIG_CGROUP_WRITEBACK */
986 * Add in the number of potentially dirty inodes, because each inode
987 * write can dirty pagecache in the underlying blockdev.
989 static unsigned long get_nr_dirty_pages(void)
991 return global_node_page_state(NR_FILE_DIRTY) +
992 global_node_page_state(NR_UNSTABLE_NFS) +
993 get_nr_dirty_inodes();
996 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
998 if (!wb_has_dirty_io(wb))
999 return;
1002 * All callers of this function want to start writeback of all
1003 * dirty pages. Places like vmscan can call this at a very
1004 * high frequency, causing pointless allocations of tons of
1005 * work items and keeping the flusher threads busy retrieving
1006 * that work. Ensure that we only allow one of them pending and
1007 * inflight at the time.
1009 if (test_bit(WB_start_all, &wb->state) ||
1010 test_and_set_bit(WB_start_all, &wb->state))
1011 return;
1013 wb->start_all_reason = reason;
1014 wb_wakeup(wb);
1018 * wb_start_background_writeback - start background writeback
1019 * @wb: bdi_writback to write from
1021 * Description:
1022 * This makes sure WB_SYNC_NONE background writeback happens. When
1023 * this function returns, it is only guaranteed that for given wb
1024 * some IO is happening if we are over background dirty threshold.
1025 * Caller need not hold sb s_umount semaphore.
1027 void wb_start_background_writeback(struct bdi_writeback *wb)
1030 * We just wake up the flusher thread. It will perform background
1031 * writeback as soon as there is no other work to do.
1033 trace_writeback_wake_background(wb);
1034 wb_wakeup(wb);
1038 * Remove the inode from the writeback list it is on.
1040 void inode_io_list_del(struct inode *inode)
1042 struct bdi_writeback *wb;
1044 wb = inode_to_wb_and_lock_list(inode);
1045 inode_io_list_del_locked(inode, wb);
1046 spin_unlock(&wb->list_lock);
1050 * mark an inode as under writeback on the sb
1052 void sb_mark_inode_writeback(struct inode *inode)
1054 struct super_block *sb = inode->i_sb;
1055 unsigned long flags;
1057 if (list_empty(&inode->i_wb_list)) {
1058 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1059 if (list_empty(&inode->i_wb_list)) {
1060 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1061 trace_sb_mark_inode_writeback(inode);
1063 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1068 * clear an inode as under writeback on the sb
1070 void sb_clear_inode_writeback(struct inode *inode)
1072 struct super_block *sb = inode->i_sb;
1073 unsigned long flags;
1075 if (!list_empty(&inode->i_wb_list)) {
1076 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1077 if (!list_empty(&inode->i_wb_list)) {
1078 list_del_init(&inode->i_wb_list);
1079 trace_sb_clear_inode_writeback(inode);
1081 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1086 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1087 * furthest end of its superblock's dirty-inode list.
1089 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1090 * already the most-recently-dirtied inode on the b_dirty list. If that is
1091 * the case then the inode must have been redirtied while it was being written
1092 * out and we don't reset its dirtied_when.
1094 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1096 if (!list_empty(&wb->b_dirty)) {
1097 struct inode *tail;
1099 tail = wb_inode(wb->b_dirty.next);
1100 if (time_before(inode->dirtied_when, tail->dirtied_when))
1101 inode->dirtied_when = jiffies;
1103 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1107 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1109 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1111 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1114 static void inode_sync_complete(struct inode *inode)
1116 inode->i_state &= ~I_SYNC;
1117 /* If inode is clean an unused, put it into LRU now... */
1118 inode_add_lru(inode);
1119 /* Waiters must see I_SYNC cleared before being woken up */
1120 smp_mb();
1121 wake_up_bit(&inode->i_state, __I_SYNC);
1124 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1126 bool ret = time_after(inode->dirtied_when, t);
1127 #ifndef CONFIG_64BIT
1129 * For inodes being constantly redirtied, dirtied_when can get stuck.
1130 * It _appears_ to be in the future, but is actually in distant past.
1131 * This test is necessary to prevent such wrapped-around relative times
1132 * from permanently stopping the whole bdi writeback.
1134 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1135 #endif
1136 return ret;
1139 #define EXPIRE_DIRTY_ATIME 0x0001
1142 * Move expired (dirtied before work->older_than_this) dirty inodes from
1143 * @delaying_queue to @dispatch_queue.
1145 static int move_expired_inodes(struct list_head *delaying_queue,
1146 struct list_head *dispatch_queue,
1147 int flags,
1148 struct wb_writeback_work *work)
1150 unsigned long *older_than_this = NULL;
1151 unsigned long expire_time;
1152 LIST_HEAD(tmp);
1153 struct list_head *pos, *node;
1154 struct super_block *sb = NULL;
1155 struct inode *inode;
1156 int do_sb_sort = 0;
1157 int moved = 0;
1159 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1160 older_than_this = work->older_than_this;
1161 else if (!work->for_sync) {
1162 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1163 older_than_this = &expire_time;
1165 while (!list_empty(delaying_queue)) {
1166 inode = wb_inode(delaying_queue->prev);
1167 if (older_than_this &&
1168 inode_dirtied_after(inode, *older_than_this))
1169 break;
1170 list_move(&inode->i_io_list, &tmp);
1171 moved++;
1172 if (flags & EXPIRE_DIRTY_ATIME)
1173 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1174 if (sb_is_blkdev_sb(inode->i_sb))
1175 continue;
1176 if (sb && sb != inode->i_sb)
1177 do_sb_sort = 1;
1178 sb = inode->i_sb;
1181 /* just one sb in list, splice to dispatch_queue and we're done */
1182 if (!do_sb_sort) {
1183 list_splice(&tmp, dispatch_queue);
1184 goto out;
1187 /* Move inodes from one superblock together */
1188 while (!list_empty(&tmp)) {
1189 sb = wb_inode(tmp.prev)->i_sb;
1190 list_for_each_prev_safe(pos, node, &tmp) {
1191 inode = wb_inode(pos);
1192 if (inode->i_sb == sb)
1193 list_move(&inode->i_io_list, dispatch_queue);
1196 out:
1197 return moved;
1201 * Queue all expired dirty inodes for io, eldest first.
1202 * Before
1203 * newly dirtied b_dirty b_io b_more_io
1204 * =============> gf edc BA
1205 * After
1206 * newly dirtied b_dirty b_io b_more_io
1207 * =============> g fBAedc
1209 * +--> dequeue for IO
1211 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1213 int moved;
1215 assert_spin_locked(&wb->list_lock);
1216 list_splice_init(&wb->b_more_io, &wb->b_io);
1217 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1218 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1219 EXPIRE_DIRTY_ATIME, work);
1220 if (moved)
1221 wb_io_lists_populated(wb);
1222 trace_writeback_queue_io(wb, work, moved);
1225 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1227 int ret;
1229 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1230 trace_writeback_write_inode_start(inode, wbc);
1231 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1232 trace_writeback_write_inode(inode, wbc);
1233 return ret;
1235 return 0;
1239 * Wait for writeback on an inode to complete. Called with i_lock held.
1240 * Caller must make sure inode cannot go away when we drop i_lock.
1242 static void __inode_wait_for_writeback(struct inode *inode)
1243 __releases(inode->i_lock)
1244 __acquires(inode->i_lock)
1246 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1247 wait_queue_head_t *wqh;
1249 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1250 while (inode->i_state & I_SYNC) {
1251 spin_unlock(&inode->i_lock);
1252 __wait_on_bit(wqh, &wq, bit_wait,
1253 TASK_UNINTERRUPTIBLE);
1254 spin_lock(&inode->i_lock);
1259 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1261 void inode_wait_for_writeback(struct inode *inode)
1263 spin_lock(&inode->i_lock);
1264 __inode_wait_for_writeback(inode);
1265 spin_unlock(&inode->i_lock);
1269 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1270 * held and drops it. It is aimed for callers not holding any inode reference
1271 * so once i_lock is dropped, inode can go away.
1273 static void inode_sleep_on_writeback(struct inode *inode)
1274 __releases(inode->i_lock)
1276 DEFINE_WAIT(wait);
1277 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1278 int sleep;
1280 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1281 sleep = inode->i_state & I_SYNC;
1282 spin_unlock(&inode->i_lock);
1283 if (sleep)
1284 schedule();
1285 finish_wait(wqh, &wait);
1289 * Find proper writeback list for the inode depending on its current state and
1290 * possibly also change of its state while we were doing writeback. Here we
1291 * handle things such as livelock prevention or fairness of writeback among
1292 * inodes. This function can be called only by flusher thread - noone else
1293 * processes all inodes in writeback lists and requeueing inodes behind flusher
1294 * thread's back can have unexpected consequences.
1296 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1297 struct writeback_control *wbc)
1299 if (inode->i_state & I_FREEING)
1300 return;
1303 * Sync livelock prevention. Each inode is tagged and synced in one
1304 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1305 * the dirty time to prevent enqueue and sync it again.
1307 if ((inode->i_state & I_DIRTY) &&
1308 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1309 inode->dirtied_when = jiffies;
1311 if (wbc->pages_skipped) {
1313 * writeback is not making progress due to locked
1314 * buffers. Skip this inode for now.
1316 redirty_tail(inode, wb);
1317 return;
1320 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1322 * We didn't write back all the pages. nfs_writepages()
1323 * sometimes bales out without doing anything.
1325 if (wbc->nr_to_write <= 0) {
1326 /* Slice used up. Queue for next turn. */
1327 requeue_io(inode, wb);
1328 } else {
1330 * Writeback blocked by something other than
1331 * congestion. Delay the inode for some time to
1332 * avoid spinning on the CPU (100% iowait)
1333 * retrying writeback of the dirty page/inode
1334 * that cannot be performed immediately.
1336 redirty_tail(inode, wb);
1338 } else if (inode->i_state & I_DIRTY) {
1340 * Filesystems can dirty the inode during writeback operations,
1341 * such as delayed allocation during submission or metadata
1342 * updates after data IO completion.
1344 redirty_tail(inode, wb);
1345 } else if (inode->i_state & I_DIRTY_TIME) {
1346 inode->dirtied_when = jiffies;
1347 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1348 } else {
1349 /* The inode is clean. Remove from writeback lists. */
1350 inode_io_list_del_locked(inode, wb);
1355 * Write out an inode and its dirty pages. Do not update the writeback list
1356 * linkage. That is left to the caller. The caller is also responsible for
1357 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1359 static int
1360 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1362 struct address_space *mapping = inode->i_mapping;
1363 long nr_to_write = wbc->nr_to_write;
1364 unsigned dirty;
1365 int ret;
1367 WARN_ON(!(inode->i_state & I_SYNC));
1369 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1371 ret = do_writepages(mapping, wbc);
1374 * Make sure to wait on the data before writing out the metadata.
1375 * This is important for filesystems that modify metadata on data
1376 * I/O completion. We don't do it for sync(2) writeback because it has a
1377 * separate, external IO completion path and ->sync_fs for guaranteeing
1378 * inode metadata is written back correctly.
1380 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1381 int err = filemap_fdatawait(mapping);
1382 if (ret == 0)
1383 ret = err;
1387 * Some filesystems may redirty the inode during the writeback
1388 * due to delalloc, clear dirty metadata flags right before
1389 * write_inode()
1391 spin_lock(&inode->i_lock);
1393 dirty = inode->i_state & I_DIRTY;
1394 if (inode->i_state & I_DIRTY_TIME) {
1395 if ((dirty & I_DIRTY_INODE) ||
1396 wbc->sync_mode == WB_SYNC_ALL ||
1397 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1398 unlikely(time_after(jiffies,
1399 (inode->dirtied_time_when +
1400 dirtytime_expire_interval * HZ)))) {
1401 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1402 trace_writeback_lazytime(inode);
1404 } else
1405 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1406 inode->i_state &= ~dirty;
1409 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1410 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1411 * either they see the I_DIRTY bits cleared or we see the dirtied
1412 * inode.
1414 * I_DIRTY_PAGES is always cleared together above even if @mapping
1415 * still has dirty pages. The flag is reinstated after smp_mb() if
1416 * necessary. This guarantees that either __mark_inode_dirty()
1417 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1419 smp_mb();
1421 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1422 inode->i_state |= I_DIRTY_PAGES;
1424 spin_unlock(&inode->i_lock);
1426 if (dirty & I_DIRTY_TIME)
1427 mark_inode_dirty_sync(inode);
1428 /* Don't write the inode if only I_DIRTY_PAGES was set */
1429 if (dirty & ~I_DIRTY_PAGES) {
1430 int err = write_inode(inode, wbc);
1431 if (ret == 0)
1432 ret = err;
1434 trace_writeback_single_inode(inode, wbc, nr_to_write);
1435 return ret;
1439 * Write out an inode's dirty pages. Either the caller has an active reference
1440 * on the inode or the inode has I_WILL_FREE set.
1442 * This function is designed to be called for writing back one inode which
1443 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1444 * and does more profound writeback list handling in writeback_sb_inodes().
1446 static int writeback_single_inode(struct inode *inode,
1447 struct writeback_control *wbc)
1449 struct bdi_writeback *wb;
1450 int ret = 0;
1452 spin_lock(&inode->i_lock);
1453 if (!atomic_read(&inode->i_count))
1454 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1455 else
1456 WARN_ON(inode->i_state & I_WILL_FREE);
1458 if (inode->i_state & I_SYNC) {
1459 if (wbc->sync_mode != WB_SYNC_ALL)
1460 goto out;
1462 * It's a data-integrity sync. We must wait. Since callers hold
1463 * inode reference or inode has I_WILL_FREE set, it cannot go
1464 * away under us.
1466 __inode_wait_for_writeback(inode);
1468 WARN_ON(inode->i_state & I_SYNC);
1470 * Skip inode if it is clean and we have no outstanding writeback in
1471 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1472 * function since flusher thread may be doing for example sync in
1473 * parallel and if we move the inode, it could get skipped. So here we
1474 * make sure inode is on some writeback list and leave it there unless
1475 * we have completely cleaned the inode.
1477 if (!(inode->i_state & I_DIRTY_ALL) &&
1478 (wbc->sync_mode != WB_SYNC_ALL ||
1479 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1480 goto out;
1481 inode->i_state |= I_SYNC;
1482 wbc_attach_and_unlock_inode(wbc, inode);
1484 ret = __writeback_single_inode(inode, wbc);
1486 wbc_detach_inode(wbc);
1488 wb = inode_to_wb_and_lock_list(inode);
1489 spin_lock(&inode->i_lock);
1491 * If inode is clean, remove it from writeback lists. Otherwise don't
1492 * touch it. See comment above for explanation.
1494 if (!(inode->i_state & I_DIRTY_ALL))
1495 inode_io_list_del_locked(inode, wb);
1496 spin_unlock(&wb->list_lock);
1497 inode_sync_complete(inode);
1498 out:
1499 spin_unlock(&inode->i_lock);
1500 return ret;
1503 static long writeback_chunk_size(struct bdi_writeback *wb,
1504 struct wb_writeback_work *work)
1506 long pages;
1509 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1510 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1511 * here avoids calling into writeback_inodes_wb() more than once.
1513 * The intended call sequence for WB_SYNC_ALL writeback is:
1515 * wb_writeback()
1516 * writeback_sb_inodes() <== called only once
1517 * write_cache_pages() <== called once for each inode
1518 * (quickly) tag currently dirty pages
1519 * (maybe slowly) sync all tagged pages
1521 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1522 pages = LONG_MAX;
1523 else {
1524 pages = min(wb->avg_write_bandwidth / 2,
1525 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1526 pages = min(pages, work->nr_pages);
1527 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1528 MIN_WRITEBACK_PAGES);
1531 return pages;
1535 * Write a portion of b_io inodes which belong to @sb.
1537 * Return the number of pages and/or inodes written.
1539 * NOTE! This is called with wb->list_lock held, and will
1540 * unlock and relock that for each inode it ends up doing
1541 * IO for.
1543 static long writeback_sb_inodes(struct super_block *sb,
1544 struct bdi_writeback *wb,
1545 struct wb_writeback_work *work)
1547 struct writeback_control wbc = {
1548 .sync_mode = work->sync_mode,
1549 .tagged_writepages = work->tagged_writepages,
1550 .for_kupdate = work->for_kupdate,
1551 .for_background = work->for_background,
1552 .for_sync = work->for_sync,
1553 .range_cyclic = work->range_cyclic,
1554 .range_start = 0,
1555 .range_end = LLONG_MAX,
1557 unsigned long start_time = jiffies;
1558 long write_chunk;
1559 long wrote = 0; /* count both pages and inodes */
1561 while (!list_empty(&wb->b_io)) {
1562 struct inode *inode = wb_inode(wb->b_io.prev);
1563 struct bdi_writeback *tmp_wb;
1565 if (inode->i_sb != sb) {
1566 if (work->sb) {
1568 * We only want to write back data for this
1569 * superblock, move all inodes not belonging
1570 * to it back onto the dirty list.
1572 redirty_tail(inode, wb);
1573 continue;
1577 * The inode belongs to a different superblock.
1578 * Bounce back to the caller to unpin this and
1579 * pin the next superblock.
1581 break;
1585 * Don't bother with new inodes or inodes being freed, first
1586 * kind does not need periodic writeout yet, and for the latter
1587 * kind writeout is handled by the freer.
1589 spin_lock(&inode->i_lock);
1590 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1591 spin_unlock(&inode->i_lock);
1592 redirty_tail(inode, wb);
1593 continue;
1595 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1597 * If this inode is locked for writeback and we are not
1598 * doing writeback-for-data-integrity, move it to
1599 * b_more_io so that writeback can proceed with the
1600 * other inodes on s_io.
1602 * We'll have another go at writing back this inode
1603 * when we completed a full scan of b_io.
1605 spin_unlock(&inode->i_lock);
1606 requeue_io(inode, wb);
1607 trace_writeback_sb_inodes_requeue(inode);
1608 continue;
1610 spin_unlock(&wb->list_lock);
1613 * We already requeued the inode if it had I_SYNC set and we
1614 * are doing WB_SYNC_NONE writeback. So this catches only the
1615 * WB_SYNC_ALL case.
1617 if (inode->i_state & I_SYNC) {
1618 /* Wait for I_SYNC. This function drops i_lock... */
1619 inode_sleep_on_writeback(inode);
1620 /* Inode may be gone, start again */
1621 spin_lock(&wb->list_lock);
1622 continue;
1624 inode->i_state |= I_SYNC;
1625 wbc_attach_and_unlock_inode(&wbc, inode);
1627 write_chunk = writeback_chunk_size(wb, work);
1628 wbc.nr_to_write = write_chunk;
1629 wbc.pages_skipped = 0;
1632 * We use I_SYNC to pin the inode in memory. While it is set
1633 * evict_inode() will wait so the inode cannot be freed.
1635 __writeback_single_inode(inode, &wbc);
1637 wbc_detach_inode(&wbc);
1638 work->nr_pages -= write_chunk - wbc.nr_to_write;
1639 wrote += write_chunk - wbc.nr_to_write;
1641 if (need_resched()) {
1643 * We're trying to balance between building up a nice
1644 * long list of IOs to improve our merge rate, and
1645 * getting those IOs out quickly for anyone throttling
1646 * in balance_dirty_pages(). cond_resched() doesn't
1647 * unplug, so get our IOs out the door before we
1648 * give up the CPU.
1650 blk_flush_plug(current);
1651 cond_resched();
1655 * Requeue @inode if still dirty. Be careful as @inode may
1656 * have been switched to another wb in the meantime.
1658 tmp_wb = inode_to_wb_and_lock_list(inode);
1659 spin_lock(&inode->i_lock);
1660 if (!(inode->i_state & I_DIRTY_ALL))
1661 wrote++;
1662 requeue_inode(inode, tmp_wb, &wbc);
1663 inode_sync_complete(inode);
1664 spin_unlock(&inode->i_lock);
1666 if (unlikely(tmp_wb != wb)) {
1667 spin_unlock(&tmp_wb->list_lock);
1668 spin_lock(&wb->list_lock);
1672 * bail out to wb_writeback() often enough to check
1673 * background threshold and other termination conditions.
1675 if (wrote) {
1676 if (time_is_before_jiffies(start_time + HZ / 10UL))
1677 break;
1678 if (work->nr_pages <= 0)
1679 break;
1682 return wrote;
1685 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1686 struct wb_writeback_work *work)
1688 unsigned long start_time = jiffies;
1689 long wrote = 0;
1691 while (!list_empty(&wb->b_io)) {
1692 struct inode *inode = wb_inode(wb->b_io.prev);
1693 struct super_block *sb = inode->i_sb;
1695 if (!trylock_super(sb)) {
1697 * trylock_super() may fail consistently due to
1698 * s_umount being grabbed by someone else. Don't use
1699 * requeue_io() to avoid busy retrying the inode/sb.
1701 redirty_tail(inode, wb);
1702 continue;
1704 wrote += writeback_sb_inodes(sb, wb, work);
1705 up_read(&sb->s_umount);
1707 /* refer to the same tests at the end of writeback_sb_inodes */
1708 if (wrote) {
1709 if (time_is_before_jiffies(start_time + HZ / 10UL))
1710 break;
1711 if (work->nr_pages <= 0)
1712 break;
1715 /* Leave any unwritten inodes on b_io */
1716 return wrote;
1719 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1720 enum wb_reason reason)
1722 struct wb_writeback_work work = {
1723 .nr_pages = nr_pages,
1724 .sync_mode = WB_SYNC_NONE,
1725 .range_cyclic = 1,
1726 .reason = reason,
1728 struct blk_plug plug;
1730 blk_start_plug(&plug);
1731 spin_lock(&wb->list_lock);
1732 if (list_empty(&wb->b_io))
1733 queue_io(wb, &work);
1734 __writeback_inodes_wb(wb, &work);
1735 spin_unlock(&wb->list_lock);
1736 blk_finish_plug(&plug);
1738 return nr_pages - work.nr_pages;
1742 * Explicit flushing or periodic writeback of "old" data.
1744 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1745 * dirtying-time in the inode's address_space. So this periodic writeback code
1746 * just walks the superblock inode list, writing back any inodes which are
1747 * older than a specific point in time.
1749 * Try to run once per dirty_writeback_interval. But if a writeback event
1750 * takes longer than a dirty_writeback_interval interval, then leave a
1751 * one-second gap.
1753 * older_than_this takes precedence over nr_to_write. So we'll only write back
1754 * all dirty pages if they are all attached to "old" mappings.
1756 static long wb_writeback(struct bdi_writeback *wb,
1757 struct wb_writeback_work *work)
1759 unsigned long wb_start = jiffies;
1760 long nr_pages = work->nr_pages;
1761 unsigned long oldest_jif;
1762 struct inode *inode;
1763 long progress;
1764 struct blk_plug plug;
1766 oldest_jif = jiffies;
1767 work->older_than_this = &oldest_jif;
1769 blk_start_plug(&plug);
1770 spin_lock(&wb->list_lock);
1771 for (;;) {
1773 * Stop writeback when nr_pages has been consumed
1775 if (work->nr_pages <= 0)
1776 break;
1779 * Background writeout and kupdate-style writeback may
1780 * run forever. Stop them if there is other work to do
1781 * so that e.g. sync can proceed. They'll be restarted
1782 * after the other works are all done.
1784 if ((work->for_background || work->for_kupdate) &&
1785 !list_empty(&wb->work_list))
1786 break;
1789 * For background writeout, stop when we are below the
1790 * background dirty threshold
1792 if (work->for_background && !wb_over_bg_thresh(wb))
1793 break;
1796 * Kupdate and background works are special and we want to
1797 * include all inodes that need writing. Livelock avoidance is
1798 * handled by these works yielding to any other work so we are
1799 * safe.
1801 if (work->for_kupdate) {
1802 oldest_jif = jiffies -
1803 msecs_to_jiffies(dirty_expire_interval * 10);
1804 } else if (work->for_background)
1805 oldest_jif = jiffies;
1807 trace_writeback_start(wb, work);
1808 if (list_empty(&wb->b_io))
1809 queue_io(wb, work);
1810 if (work->sb)
1811 progress = writeback_sb_inodes(work->sb, wb, work);
1812 else
1813 progress = __writeback_inodes_wb(wb, work);
1814 trace_writeback_written(wb, work);
1816 wb_update_bandwidth(wb, wb_start);
1819 * Did we write something? Try for more
1821 * Dirty inodes are moved to b_io for writeback in batches.
1822 * The completion of the current batch does not necessarily
1823 * mean the overall work is done. So we keep looping as long
1824 * as made some progress on cleaning pages or inodes.
1826 if (progress)
1827 continue;
1829 * No more inodes for IO, bail
1831 if (list_empty(&wb->b_more_io))
1832 break;
1834 * Nothing written. Wait for some inode to
1835 * become available for writeback. Otherwise
1836 * we'll just busyloop.
1838 trace_writeback_wait(wb, work);
1839 inode = wb_inode(wb->b_more_io.prev);
1840 spin_lock(&inode->i_lock);
1841 spin_unlock(&wb->list_lock);
1842 /* This function drops i_lock... */
1843 inode_sleep_on_writeback(inode);
1844 spin_lock(&wb->list_lock);
1846 spin_unlock(&wb->list_lock);
1847 blk_finish_plug(&plug);
1849 return nr_pages - work->nr_pages;
1853 * Return the next wb_writeback_work struct that hasn't been processed yet.
1855 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1857 struct wb_writeback_work *work = NULL;
1859 spin_lock_bh(&wb->work_lock);
1860 if (!list_empty(&wb->work_list)) {
1861 work = list_entry(wb->work_list.next,
1862 struct wb_writeback_work, list);
1863 list_del_init(&work->list);
1865 spin_unlock_bh(&wb->work_lock);
1866 return work;
1869 static long wb_check_background_flush(struct bdi_writeback *wb)
1871 if (wb_over_bg_thresh(wb)) {
1873 struct wb_writeback_work work = {
1874 .nr_pages = LONG_MAX,
1875 .sync_mode = WB_SYNC_NONE,
1876 .for_background = 1,
1877 .range_cyclic = 1,
1878 .reason = WB_REASON_BACKGROUND,
1881 return wb_writeback(wb, &work);
1884 return 0;
1887 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1889 unsigned long expired;
1890 long nr_pages;
1893 * When set to zero, disable periodic writeback
1895 if (!dirty_writeback_interval)
1896 return 0;
1898 expired = wb->last_old_flush +
1899 msecs_to_jiffies(dirty_writeback_interval * 10);
1900 if (time_before(jiffies, expired))
1901 return 0;
1903 wb->last_old_flush = jiffies;
1904 nr_pages = get_nr_dirty_pages();
1906 if (nr_pages) {
1907 struct wb_writeback_work work = {
1908 .nr_pages = nr_pages,
1909 .sync_mode = WB_SYNC_NONE,
1910 .for_kupdate = 1,
1911 .range_cyclic = 1,
1912 .reason = WB_REASON_PERIODIC,
1915 return wb_writeback(wb, &work);
1918 return 0;
1921 static long wb_check_start_all(struct bdi_writeback *wb)
1923 long nr_pages;
1925 if (!test_bit(WB_start_all, &wb->state))
1926 return 0;
1928 nr_pages = get_nr_dirty_pages();
1929 if (nr_pages) {
1930 struct wb_writeback_work work = {
1931 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
1932 .sync_mode = WB_SYNC_NONE,
1933 .range_cyclic = 1,
1934 .reason = wb->start_all_reason,
1937 nr_pages = wb_writeback(wb, &work);
1940 clear_bit(WB_start_all, &wb->state);
1941 return nr_pages;
1946 * Retrieve work items and do the writeback they describe
1948 static long wb_do_writeback(struct bdi_writeback *wb)
1950 struct wb_writeback_work *work;
1951 long wrote = 0;
1953 set_bit(WB_writeback_running, &wb->state);
1954 while ((work = get_next_work_item(wb)) != NULL) {
1955 trace_writeback_exec(wb, work);
1956 wrote += wb_writeback(wb, work);
1957 finish_writeback_work(wb, work);
1961 * Check for a flush-everything request
1963 wrote += wb_check_start_all(wb);
1966 * Check for periodic writeback, kupdated() style
1968 wrote += wb_check_old_data_flush(wb);
1969 wrote += wb_check_background_flush(wb);
1970 clear_bit(WB_writeback_running, &wb->state);
1972 return wrote;
1976 * Handle writeback of dirty data for the device backed by this bdi. Also
1977 * reschedules periodically and does kupdated style flushing.
1979 void wb_workfn(struct work_struct *work)
1981 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1982 struct bdi_writeback, dwork);
1983 long pages_written;
1985 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1986 current->flags |= PF_SWAPWRITE;
1988 if (likely(!current_is_workqueue_rescuer() ||
1989 !test_bit(WB_registered, &wb->state))) {
1991 * The normal path. Keep writing back @wb until its
1992 * work_list is empty. Note that this path is also taken
1993 * if @wb is shutting down even when we're running off the
1994 * rescuer as work_list needs to be drained.
1996 do {
1997 pages_written = wb_do_writeback(wb);
1998 trace_writeback_pages_written(pages_written);
1999 } while (!list_empty(&wb->work_list));
2000 } else {
2002 * bdi_wq can't get enough workers and we're running off
2003 * the emergency worker. Don't hog it. Hopefully, 1024 is
2004 * enough for efficient IO.
2006 pages_written = writeback_inodes_wb(wb, 1024,
2007 WB_REASON_FORKER_THREAD);
2008 trace_writeback_pages_written(pages_written);
2011 if (!list_empty(&wb->work_list))
2012 wb_wakeup(wb);
2013 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2014 wb_wakeup_delayed(wb);
2016 current->flags &= ~PF_SWAPWRITE;
2020 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2021 * write back the whole world.
2023 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2024 enum wb_reason reason)
2026 struct bdi_writeback *wb;
2028 if (!bdi_has_dirty_io(bdi))
2029 return;
2031 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2032 wb_start_writeback(wb, reason);
2035 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2036 enum wb_reason reason)
2038 rcu_read_lock();
2039 __wakeup_flusher_threads_bdi(bdi, reason);
2040 rcu_read_unlock();
2044 * Wakeup the flusher threads to start writeback of all currently dirty pages
2046 void wakeup_flusher_threads(enum wb_reason reason)
2048 struct backing_dev_info *bdi;
2051 * If we are expecting writeback progress we must submit plugged IO.
2053 if (blk_needs_flush_plug(current))
2054 blk_schedule_flush_plug(current);
2056 rcu_read_lock();
2057 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2058 __wakeup_flusher_threads_bdi(bdi, reason);
2059 rcu_read_unlock();
2063 * Wake up bdi's periodically to make sure dirtytime inodes gets
2064 * written back periodically. We deliberately do *not* check the
2065 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2066 * kernel to be constantly waking up once there are any dirtytime
2067 * inodes on the system. So instead we define a separate delayed work
2068 * function which gets called much more rarely. (By default, only
2069 * once every 12 hours.)
2071 * If there is any other write activity going on in the file system,
2072 * this function won't be necessary. But if the only thing that has
2073 * happened on the file system is a dirtytime inode caused by an atime
2074 * update, we need this infrastructure below to make sure that inode
2075 * eventually gets pushed out to disk.
2077 static void wakeup_dirtytime_writeback(struct work_struct *w);
2078 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2080 static void wakeup_dirtytime_writeback(struct work_struct *w)
2082 struct backing_dev_info *bdi;
2084 rcu_read_lock();
2085 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2086 struct bdi_writeback *wb;
2088 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2089 if (!list_empty(&wb->b_dirty_time))
2090 wb_wakeup(wb);
2092 rcu_read_unlock();
2093 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2096 static int __init start_dirtytime_writeback(void)
2098 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2099 return 0;
2101 __initcall(start_dirtytime_writeback);
2103 int dirtytime_interval_handler(struct ctl_table *table, int write,
2104 void __user *buffer, size_t *lenp, loff_t *ppos)
2106 int ret;
2108 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2109 if (ret == 0 && write)
2110 mod_delayed_work(system_wq, &dirtytime_work, 0);
2111 return ret;
2114 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2116 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2117 struct dentry *dentry;
2118 const char *name = "?";
2120 dentry = d_find_alias(inode);
2121 if (dentry) {
2122 spin_lock(&dentry->d_lock);
2123 name = (const char *) dentry->d_name.name;
2125 printk(KERN_DEBUG
2126 "%s(%d): dirtied inode %lu (%s) on %s\n",
2127 current->comm, task_pid_nr(current), inode->i_ino,
2128 name, inode->i_sb->s_id);
2129 if (dentry) {
2130 spin_unlock(&dentry->d_lock);
2131 dput(dentry);
2137 * __mark_inode_dirty - internal function
2139 * @inode: inode to mark
2140 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2142 * Mark an inode as dirty. Callers should use mark_inode_dirty or
2143 * mark_inode_dirty_sync.
2145 * Put the inode on the super block's dirty list.
2147 * CAREFUL! We mark it dirty unconditionally, but move it onto the
2148 * dirty list only if it is hashed or if it refers to a blockdev.
2149 * If it was not hashed, it will never be added to the dirty list
2150 * even if it is later hashed, as it will have been marked dirty already.
2152 * In short, make sure you hash any inodes _before_ you start marking
2153 * them dirty.
2155 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2156 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2157 * the kernel-internal blockdev inode represents the dirtying time of the
2158 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2159 * page->mapping->host, so the page-dirtying time is recorded in the internal
2160 * blockdev inode.
2162 void __mark_inode_dirty(struct inode *inode, int flags)
2164 struct super_block *sb = inode->i_sb;
2165 int dirtytime;
2167 trace_writeback_mark_inode_dirty(inode, flags);
2170 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2171 * dirty the inode itself
2173 if (flags & (I_DIRTY_INODE | I_DIRTY_TIME)) {
2174 trace_writeback_dirty_inode_start(inode, flags);
2176 if (sb->s_op->dirty_inode)
2177 sb->s_op->dirty_inode(inode, flags);
2179 trace_writeback_dirty_inode(inode, flags);
2181 if (flags & I_DIRTY_INODE)
2182 flags &= ~I_DIRTY_TIME;
2183 dirtytime = flags & I_DIRTY_TIME;
2186 * Paired with smp_mb() in __writeback_single_inode() for the
2187 * following lockless i_state test. See there for details.
2189 smp_mb();
2191 if (((inode->i_state & flags) == flags) ||
2192 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2193 return;
2195 if (unlikely(block_dump))
2196 block_dump___mark_inode_dirty(inode);
2198 spin_lock(&inode->i_lock);
2199 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2200 goto out_unlock_inode;
2201 if ((inode->i_state & flags) != flags) {
2202 const int was_dirty = inode->i_state & I_DIRTY;
2204 inode_attach_wb(inode, NULL);
2206 if (flags & I_DIRTY_INODE)
2207 inode->i_state &= ~I_DIRTY_TIME;
2208 inode->i_state |= flags;
2211 * If the inode is being synced, just update its dirty state.
2212 * The unlocker will place the inode on the appropriate
2213 * superblock list, based upon its state.
2215 if (inode->i_state & I_SYNC)
2216 goto out_unlock_inode;
2219 * Only add valid (hashed) inodes to the superblock's
2220 * dirty list. Add blockdev inodes as well.
2222 if (!S_ISBLK(inode->i_mode)) {
2223 if (inode_unhashed(inode))
2224 goto out_unlock_inode;
2226 if (inode->i_state & I_FREEING)
2227 goto out_unlock_inode;
2230 * If the inode was already on b_dirty/b_io/b_more_io, don't
2231 * reposition it (that would break b_dirty time-ordering).
2233 if (!was_dirty) {
2234 struct bdi_writeback *wb;
2235 struct list_head *dirty_list;
2236 bool wakeup_bdi = false;
2238 wb = locked_inode_to_wb_and_lock_list(inode);
2240 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2241 !test_bit(WB_registered, &wb->state),
2242 "bdi-%s not registered\n", wb->bdi->name);
2244 inode->dirtied_when = jiffies;
2245 if (dirtytime)
2246 inode->dirtied_time_when = jiffies;
2248 if (inode->i_state & I_DIRTY)
2249 dirty_list = &wb->b_dirty;
2250 else
2251 dirty_list = &wb->b_dirty_time;
2253 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2254 dirty_list);
2256 spin_unlock(&wb->list_lock);
2257 trace_writeback_dirty_inode_enqueue(inode);
2260 * If this is the first dirty inode for this bdi,
2261 * we have to wake-up the corresponding bdi thread
2262 * to make sure background write-back happens
2263 * later.
2265 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2266 wb_wakeup_delayed(wb);
2267 return;
2270 out_unlock_inode:
2271 spin_unlock(&inode->i_lock);
2273 EXPORT_SYMBOL(__mark_inode_dirty);
2276 * The @s_sync_lock is used to serialise concurrent sync operations
2277 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2278 * Concurrent callers will block on the s_sync_lock rather than doing contending
2279 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2280 * has been issued up to the time this function is enter is guaranteed to be
2281 * completed by the time we have gained the lock and waited for all IO that is
2282 * in progress regardless of the order callers are granted the lock.
2284 static void wait_sb_inodes(struct super_block *sb)
2286 LIST_HEAD(sync_list);
2289 * We need to be protected against the filesystem going from
2290 * r/o to r/w or vice versa.
2292 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2294 mutex_lock(&sb->s_sync_lock);
2297 * Splice the writeback list onto a temporary list to avoid waiting on
2298 * inodes that have started writeback after this point.
2300 * Use rcu_read_lock() to keep the inodes around until we have a
2301 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2302 * the local list because inodes can be dropped from either by writeback
2303 * completion.
2305 rcu_read_lock();
2306 spin_lock_irq(&sb->s_inode_wblist_lock);
2307 list_splice_init(&sb->s_inodes_wb, &sync_list);
2310 * Data integrity sync. Must wait for all pages under writeback, because
2311 * there may have been pages dirtied before our sync call, but which had
2312 * writeout started before we write it out. In which case, the inode
2313 * may not be on the dirty list, but we still have to wait for that
2314 * writeout.
2316 while (!list_empty(&sync_list)) {
2317 struct inode *inode = list_first_entry(&sync_list, struct inode,
2318 i_wb_list);
2319 struct address_space *mapping = inode->i_mapping;
2322 * Move each inode back to the wb list before we drop the lock
2323 * to preserve consistency between i_wb_list and the mapping
2324 * writeback tag. Writeback completion is responsible to remove
2325 * the inode from either list once the writeback tag is cleared.
2327 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2330 * The mapping can appear untagged while still on-list since we
2331 * do not have the mapping lock. Skip it here, wb completion
2332 * will remove it.
2334 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2335 continue;
2337 spin_unlock_irq(&sb->s_inode_wblist_lock);
2339 spin_lock(&inode->i_lock);
2340 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2341 spin_unlock(&inode->i_lock);
2343 spin_lock_irq(&sb->s_inode_wblist_lock);
2344 continue;
2346 __iget(inode);
2347 spin_unlock(&inode->i_lock);
2348 rcu_read_unlock();
2351 * We keep the error status of individual mapping so that
2352 * applications can catch the writeback error using fsync(2).
2353 * See filemap_fdatawait_keep_errors() for details.
2355 filemap_fdatawait_keep_errors(mapping);
2357 cond_resched();
2359 iput(inode);
2361 rcu_read_lock();
2362 spin_lock_irq(&sb->s_inode_wblist_lock);
2364 spin_unlock_irq(&sb->s_inode_wblist_lock);
2365 rcu_read_unlock();
2366 mutex_unlock(&sb->s_sync_lock);
2369 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2370 enum wb_reason reason, bool skip_if_busy)
2372 DEFINE_WB_COMPLETION_ONSTACK(done);
2373 struct wb_writeback_work work = {
2374 .sb = sb,
2375 .sync_mode = WB_SYNC_NONE,
2376 .tagged_writepages = 1,
2377 .done = &done,
2378 .nr_pages = nr,
2379 .reason = reason,
2381 struct backing_dev_info *bdi = sb->s_bdi;
2383 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2384 return;
2385 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2387 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2388 wb_wait_for_completion(bdi, &done);
2392 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2393 * @sb: the superblock
2394 * @nr: the number of pages to write
2395 * @reason: reason why some writeback work initiated
2397 * Start writeback on some inodes on this super_block. No guarantees are made
2398 * on how many (if any) will be written, and this function does not wait
2399 * for IO completion of submitted IO.
2401 void writeback_inodes_sb_nr(struct super_block *sb,
2402 unsigned long nr,
2403 enum wb_reason reason)
2405 __writeback_inodes_sb_nr(sb, nr, reason, false);
2407 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2410 * writeback_inodes_sb - writeback dirty inodes from given super_block
2411 * @sb: the superblock
2412 * @reason: reason why some writeback work was initiated
2414 * Start writeback on some inodes on this super_block. No guarantees are made
2415 * on how many (if any) will be written, and this function does not wait
2416 * for IO completion of submitted IO.
2418 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2420 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2422 EXPORT_SYMBOL(writeback_inodes_sb);
2425 * try_to_writeback_inodes_sb - try to start writeback if none underway
2426 * @sb: the superblock
2427 * @reason: reason why some writeback work was initiated
2429 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2431 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2433 if (!down_read_trylock(&sb->s_umount))
2434 return;
2436 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2437 up_read(&sb->s_umount);
2439 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2442 * sync_inodes_sb - sync sb inode pages
2443 * @sb: the superblock
2445 * This function writes and waits on any dirty inode belonging to this
2446 * super_block.
2448 void sync_inodes_sb(struct super_block *sb)
2450 DEFINE_WB_COMPLETION_ONSTACK(done);
2451 struct wb_writeback_work work = {
2452 .sb = sb,
2453 .sync_mode = WB_SYNC_ALL,
2454 .nr_pages = LONG_MAX,
2455 .range_cyclic = 0,
2456 .done = &done,
2457 .reason = WB_REASON_SYNC,
2458 .for_sync = 1,
2460 struct backing_dev_info *bdi = sb->s_bdi;
2463 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2464 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2465 * bdi_has_dirty() need to be written out too.
2467 if (bdi == &noop_backing_dev_info)
2468 return;
2469 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2471 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2472 bdi_down_write_wb_switch_rwsem(bdi);
2473 bdi_split_work_to_wbs(bdi, &work, false);
2474 wb_wait_for_completion(bdi, &done);
2475 bdi_up_write_wb_switch_rwsem(bdi);
2477 wait_sb_inodes(sb);
2479 EXPORT_SYMBOL(sync_inodes_sb);
2482 * write_inode_now - write an inode to disk
2483 * @inode: inode to write to disk
2484 * @sync: whether the write should be synchronous or not
2486 * This function commits an inode to disk immediately if it is dirty. This is
2487 * primarily needed by knfsd.
2489 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2491 int write_inode_now(struct inode *inode, int sync)
2493 struct writeback_control wbc = {
2494 .nr_to_write = LONG_MAX,
2495 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2496 .range_start = 0,
2497 .range_end = LLONG_MAX,
2500 if (!mapping_cap_writeback_dirty(inode->i_mapping))
2501 wbc.nr_to_write = 0;
2503 might_sleep();
2504 return writeback_single_inode(inode, &wbc);
2506 EXPORT_SYMBOL(write_inode_now);
2509 * sync_inode - write an inode and its pages to disk.
2510 * @inode: the inode to sync
2511 * @wbc: controls the writeback mode
2513 * sync_inode() will write an inode and its pages to disk. It will also
2514 * correctly update the inode on its superblock's dirty inode lists and will
2515 * update inode->i_state.
2517 * The caller must have a ref on the inode.
2519 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2521 return writeback_single_inode(inode, wbc);
2523 EXPORT_SYMBOL(sync_inode);
2526 * sync_inode_metadata - write an inode to disk
2527 * @inode: the inode to sync
2528 * @wait: wait for I/O to complete.
2530 * Write an inode to disk and adjust its dirty state after completion.
2532 * Note: only writes the actual inode, no associated data or other metadata.
2534 int sync_inode_metadata(struct inode *inode, int wait)
2536 struct writeback_control wbc = {
2537 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2538 .nr_to_write = 0, /* metadata-only */
2541 return sync_inode(inode, &wbc);
2543 EXPORT_SYMBOL(sync_inode_metadata);