1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2002, Linus Torvalds.
7 * Contains all the functions related to writing back and waiting
8 * upon dirty inodes against superblocks, and writing back dirty
9 * pages against inodes. ie: data writeback. Writeout of the
10 * inode itself is not handled here.
12 * 10Apr2002 Andrew Morton
13 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
17 #include <linux/kernel.h>
18 #include <linux/export.h>
19 #include <linux/spinlock.h>
20 #include <linux/slab.h>
21 #include <linux/sched.h>
24 #include <linux/pagemap.h>
25 #include <linux/kthread.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include <linux/device.h>
31 #include <linux/memcontrol.h>
35 * 4MB minimal write chunk size
37 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
40 * Passed into wb_writeback(), essentially a subset of writeback_control
42 struct wb_writeback_work {
44 struct super_block *sb;
45 enum writeback_sync_modes sync_mode;
46 unsigned int tagged_writepages:1;
47 unsigned int for_kupdate:1;
48 unsigned int range_cyclic:1;
49 unsigned int for_background:1;
50 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
51 unsigned int auto_free:1; /* free on completion */
52 enum wb_reason reason; /* why was writeback initiated? */
54 struct list_head list; /* pending work list */
55 struct wb_completion *done; /* set if the caller waits */
59 * If an inode is constantly having its pages dirtied, but then the
60 * updates stop dirtytime_expire_interval seconds in the past, it's
61 * possible for the worst case time between when an inode has its
62 * timestamps updated and when they finally get written out to be two
63 * dirtytime_expire_intervals. We set the default to 12 hours (in
64 * seconds), which means most of the time inodes will have their
65 * timestamps written to disk after 12 hours, but in the worst case a
66 * few inodes might not their timestamps updated for 24 hours.
68 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
70 static inline struct inode *wb_inode(struct list_head *head)
72 return list_entry(head, struct inode, i_io_list);
76 * Include the creation of the trace points after defining the
77 * wb_writeback_work structure and inline functions so that the definition
78 * remains local to this file.
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/writeback.h>
83 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
85 static bool wb_io_lists_populated(struct bdi_writeback *wb)
87 if (wb_has_dirty_io(wb)) {
90 set_bit(WB_has_dirty_io, &wb->state);
91 WARN_ON_ONCE(!wb->avg_write_bandwidth);
92 atomic_long_add(wb->avg_write_bandwidth,
93 &wb->bdi->tot_write_bandwidth);
98 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
100 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102 clear_bit(WB_has_dirty_io, &wb->state);
103 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104 &wb->bdi->tot_write_bandwidth) < 0);
109 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110 * @inode: inode to be moved
111 * @wb: target bdi_writeback
112 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
114 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115 * Returns %true if @inode is the first occupant of the !dirty_time IO
116 * lists; otherwise, %false.
118 static bool inode_io_list_move_locked(struct inode *inode,
119 struct bdi_writeback *wb,
120 struct list_head *head)
122 assert_spin_locked(&wb->list_lock);
124 list_move(&inode->i_io_list, head);
126 /* dirty_time doesn't count as dirty_io until expiration */
127 if (head != &wb->b_dirty_time)
128 return wb_io_lists_populated(wb);
130 wb_io_lists_depopulated(wb);
135 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
136 * @inode: inode to be removed
137 * @wb: bdi_writeback @inode is being removed from
139 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
140 * clear %WB_has_dirty_io if all are empty afterwards.
142 static void inode_io_list_del_locked(struct inode *inode,
143 struct bdi_writeback *wb)
145 assert_spin_locked(&wb->list_lock);
146 assert_spin_locked(&inode->i_lock);
148 inode->i_state &= ~I_SYNC_QUEUED;
149 list_del_init(&inode->i_io_list);
150 wb_io_lists_depopulated(wb);
153 static void wb_wakeup(struct bdi_writeback *wb)
155 spin_lock_bh(&wb->work_lock);
156 if (test_bit(WB_registered, &wb->state))
157 mod_delayed_work(bdi_wq, &wb->dwork, 0);
158 spin_unlock_bh(&wb->work_lock);
161 static void finish_writeback_work(struct bdi_writeback *wb,
162 struct wb_writeback_work *work)
164 struct wb_completion *done = work->done;
169 wait_queue_head_t *waitq = done->waitq;
171 /* @done can't be accessed after the following dec */
172 if (atomic_dec_and_test(&done->cnt))
177 static void wb_queue_work(struct bdi_writeback *wb,
178 struct wb_writeback_work *work)
180 trace_writeback_queue(wb, work);
183 atomic_inc(&work->done->cnt);
185 spin_lock_bh(&wb->work_lock);
187 if (test_bit(WB_registered, &wb->state)) {
188 list_add_tail(&work->list, &wb->work_list);
189 mod_delayed_work(bdi_wq, &wb->dwork, 0);
191 finish_writeback_work(wb, work);
193 spin_unlock_bh(&wb->work_lock);
197 * wb_wait_for_completion - wait for completion of bdi_writeback_works
198 * @done: target wb_completion
200 * Wait for one or more work items issued to @bdi with their ->done field
201 * set to @done, which should have been initialized with
202 * DEFINE_WB_COMPLETION(). This function returns after all such work items
203 * are completed. Work items which are waited upon aren't freed
204 * automatically on completion.
206 void wb_wait_for_completion(struct wb_completion *done)
208 atomic_dec(&done->cnt); /* put down the initial count */
209 wait_event(*done->waitq, !atomic_read(&done->cnt));
212 #ifdef CONFIG_CGROUP_WRITEBACK
215 * Parameters for foreign inode detection, see wbc_detach_inode() to see
218 * These paramters are inherently heuristical as the detection target
219 * itself is fuzzy. All we want to do is detaching an inode from the
220 * current owner if it's being written to by some other cgroups too much.
222 * The current cgroup writeback is built on the assumption that multiple
223 * cgroups writing to the same inode concurrently is very rare and a mode
224 * of operation which isn't well supported. As such, the goal is not
225 * taking too long when a different cgroup takes over an inode while
226 * avoiding too aggressive flip-flops from occasional foreign writes.
228 * We record, very roughly, 2s worth of IO time history and if more than
229 * half of that is foreign, trigger the switch. The recording is quantized
230 * to 16 slots. To avoid tiny writes from swinging the decision too much,
231 * writes smaller than 1/8 of avg size are ignored.
233 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
234 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
235 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
236 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
238 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
239 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
240 /* each slot's duration is 2s / 16 */
241 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
242 /* if foreign slots >= 8, switch */
243 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
244 /* one round can affect upto 5 slots */
245 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
247 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
248 static struct workqueue_struct *isw_wq;
250 void __inode_attach_wb(struct inode *inode, struct page *page)
252 struct backing_dev_info *bdi = inode_to_bdi(inode);
253 struct bdi_writeback *wb = NULL;
255 if (inode_cgwb_enabled(inode)) {
256 struct cgroup_subsys_state *memcg_css;
259 memcg_css = mem_cgroup_css_from_page(page);
260 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
262 /* must pin memcg_css, see wb_get_create() */
263 memcg_css = task_get_css(current, memory_cgrp_id);
264 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
273 * There may be multiple instances of this function racing to
274 * update the same inode. Use cmpxchg() to tell the winner.
276 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
279 EXPORT_SYMBOL_GPL(__inode_attach_wb);
282 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
283 * @inode: inode of interest with i_lock held
285 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
286 * held on entry and is released on return. The returned wb is guaranteed
287 * to stay @inode's associated wb until its list_lock is released.
289 static struct bdi_writeback *
290 locked_inode_to_wb_and_lock_list(struct inode *inode)
291 __releases(&inode->i_lock)
292 __acquires(&wb->list_lock)
295 struct bdi_writeback *wb = inode_to_wb(inode);
298 * inode_to_wb() association is protected by both
299 * @inode->i_lock and @wb->list_lock but list_lock nests
300 * outside i_lock. Drop i_lock and verify that the
301 * association hasn't changed after acquiring list_lock.
304 spin_unlock(&inode->i_lock);
305 spin_lock(&wb->list_lock);
307 /* i_wb may have changed inbetween, can't use inode_to_wb() */
308 if (likely(wb == inode->i_wb)) {
309 wb_put(wb); /* @inode already has ref */
313 spin_unlock(&wb->list_lock);
316 spin_lock(&inode->i_lock);
321 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
322 * @inode: inode of interest
324 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
327 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
328 __acquires(&wb->list_lock)
330 spin_lock(&inode->i_lock);
331 return locked_inode_to_wb_and_lock_list(inode);
334 struct inode_switch_wbs_context {
336 struct bdi_writeback *new_wb;
338 struct rcu_head rcu_head;
339 struct work_struct work;
342 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
344 down_write(&bdi->wb_switch_rwsem);
347 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
349 up_write(&bdi->wb_switch_rwsem);
352 static void inode_switch_wbs_work_fn(struct work_struct *work)
354 struct inode_switch_wbs_context *isw =
355 container_of(work, struct inode_switch_wbs_context, work);
356 struct inode *inode = isw->inode;
357 struct backing_dev_info *bdi = inode_to_bdi(inode);
358 struct address_space *mapping = inode->i_mapping;
359 struct bdi_writeback *old_wb = inode->i_wb;
360 struct bdi_writeback *new_wb = isw->new_wb;
361 XA_STATE(xas, &mapping->i_pages, 0);
363 bool switched = false;
366 * If @inode switches cgwb membership while sync_inodes_sb() is
367 * being issued, sync_inodes_sb() might miss it. Synchronize.
369 down_read(&bdi->wb_switch_rwsem);
372 * By the time control reaches here, RCU grace period has passed
373 * since I_WB_SWITCH assertion and all wb stat update transactions
374 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
375 * synchronizing against the i_pages lock.
377 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
378 * gives us exclusion against all wb related operations on @inode
379 * including IO list manipulations and stat updates.
381 if (old_wb < new_wb) {
382 spin_lock(&old_wb->list_lock);
383 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
385 spin_lock(&new_wb->list_lock);
386 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
388 spin_lock(&inode->i_lock);
389 xa_lock_irq(&mapping->i_pages);
392 * Once I_FREEING is visible under i_lock, the eviction path owns
393 * the inode and we shouldn't modify ->i_io_list.
395 if (unlikely(inode->i_state & I_FREEING))
398 trace_inode_switch_wbs(inode, old_wb, new_wb);
401 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
402 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
403 * pages actually under writeback.
405 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
406 if (PageDirty(page)) {
407 dec_wb_stat(old_wb, WB_RECLAIMABLE);
408 inc_wb_stat(new_wb, WB_RECLAIMABLE);
413 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
414 WARN_ON_ONCE(!PageWriteback(page));
415 dec_wb_stat(old_wb, WB_WRITEBACK);
416 inc_wb_stat(new_wb, WB_WRITEBACK);
422 * Transfer to @new_wb's IO list if necessary. The specific list
423 * @inode was on is ignored and the inode is put on ->b_dirty which
424 * is always correct including from ->b_dirty_time. The transfer
425 * preserves @inode->dirtied_when ordering.
427 if (!list_empty(&inode->i_io_list)) {
430 inode_io_list_del_locked(inode, old_wb);
431 inode->i_wb = new_wb;
432 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
433 if (time_after_eq(inode->dirtied_when,
436 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
438 inode->i_wb = new_wb;
441 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
442 inode->i_wb_frn_winner = 0;
443 inode->i_wb_frn_avg_time = 0;
444 inode->i_wb_frn_history = 0;
448 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
449 * ensures that the new wb is visible if they see !I_WB_SWITCH.
451 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
453 xa_unlock_irq(&mapping->i_pages);
454 spin_unlock(&inode->i_lock);
455 spin_unlock(&new_wb->list_lock);
456 spin_unlock(&old_wb->list_lock);
458 up_read(&bdi->wb_switch_rwsem);
469 atomic_dec(&isw_nr_in_flight);
472 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
474 struct inode_switch_wbs_context *isw = container_of(rcu_head,
475 struct inode_switch_wbs_context, rcu_head);
477 /* needs to grab bh-unsafe locks, bounce to work item */
478 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
479 queue_work(isw_wq, &isw->work);
483 * inode_switch_wbs - change the wb association of an inode
484 * @inode: target inode
485 * @new_wb_id: ID of the new wb
487 * Switch @inode's wb association to the wb identified by @new_wb_id. The
488 * switching is performed asynchronously and may fail silently.
490 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
492 struct backing_dev_info *bdi = inode_to_bdi(inode);
493 struct cgroup_subsys_state *memcg_css;
494 struct inode_switch_wbs_context *isw;
496 /* noop if seems to be already in progress */
497 if (inode->i_state & I_WB_SWITCH)
500 /* avoid queueing a new switch if too many are already in flight */
501 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
504 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
508 atomic_inc(&isw_nr_in_flight);
510 /* find and pin the new wb */
512 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
513 if (memcg_css && !css_tryget(memcg_css))
519 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
524 /* while holding I_WB_SWITCH, no one else can update the association */
525 spin_lock(&inode->i_lock);
526 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
527 inode->i_state & (I_WB_SWITCH | I_FREEING) ||
528 inode_to_wb(inode) == isw->new_wb) {
529 spin_unlock(&inode->i_lock);
532 inode->i_state |= I_WB_SWITCH;
534 spin_unlock(&inode->i_lock);
539 * In addition to synchronizing among switchers, I_WB_SWITCH tells
540 * the RCU protected stat update paths to grab the i_page
541 * lock so that stat transfer can synchronize against them.
542 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
544 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
548 atomic_dec(&isw_nr_in_flight);
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,
567 if (!inode_cgwb_enabled(inode)) {
568 spin_unlock(&inode->i_lock);
572 wbc->wb = inode_to_wb(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;
579 wbc->wb_lcand_bytes = 0;
580 wbc->wb_tcand_bytes = 0;
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);
595 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
598 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
599 * @wbc: writeback_control of the just finished writeback
601 * To be called after a writeback attempt of an inode finishes and undoes
602 * wbc_attach_and_unlock_inode(). Can be called under any context.
604 * As concurrent write sharing of an inode is expected to be very rare and
605 * memcg only tracks page ownership on first-use basis severely confining
606 * the usefulness of such sharing, cgroup writeback tracks ownership
607 * per-inode. While the support for concurrent write sharing of an inode
608 * is deemed unnecessary, an inode being written to by different cgroups at
609 * different points in time is a lot more common, and, more importantly,
610 * charging only by first-use can too readily lead to grossly incorrect
611 * behaviors (single foreign page can lead to gigabytes of writeback to be
612 * incorrectly attributed).
614 * To resolve this issue, cgroup writeback detects the majority dirtier of
615 * an inode and transfers the ownership to it. To avoid unnnecessary
616 * oscillation, the detection mechanism keeps track of history and gives
617 * out the switch verdict only if the foreign usage pattern is stable over
618 * a certain amount of time and/or writeback attempts.
620 * On each writeback attempt, @wbc tries to detect the majority writer
621 * using Boyer-Moore majority vote algorithm. In addition to the byte
622 * count from the majority voting, it also counts the bytes written for the
623 * current wb and the last round's winner wb (max of last round's current
624 * wb, the winner from two rounds ago, and the last round's majority
625 * candidate). Keeping track of the historical winner helps the algorithm
626 * to semi-reliably detect the most active writer even when it's not the
629 * Once the winner of the round is determined, whether the winner is
630 * foreign or not and how much IO time the round consumed is recorded in
631 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
632 * over a certain threshold, the switch verdict is given.
634 void wbc_detach_inode(struct writeback_control *wbc)
636 struct bdi_writeback *wb = wbc->wb;
637 struct inode *inode = wbc->inode;
638 unsigned long avg_time, max_bytes, max_time;
645 history = inode->i_wb_frn_history;
646 avg_time = inode->i_wb_frn_avg_time;
648 /* pick the winner of this round */
649 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
650 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
652 max_bytes = wbc->wb_bytes;
653 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
654 max_id = wbc->wb_lcand_id;
655 max_bytes = wbc->wb_lcand_bytes;
657 max_id = wbc->wb_tcand_id;
658 max_bytes = wbc->wb_tcand_bytes;
662 * Calculate the amount of IO time the winner consumed and fold it
663 * into the running average kept per inode. If the consumed IO
664 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
665 * deciding whether to switch or not. This is to prevent one-off
666 * small dirtiers from skewing the verdict.
668 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
669 wb->avg_write_bandwidth);
671 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
672 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
674 avg_time = max_time; /* immediate catch up on first run */
676 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
680 * The switch verdict is reached if foreign wb's consume
681 * more than a certain proportion of IO time in a
682 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
683 * history mask where each bit represents one sixteenth of
684 * the period. Determine the number of slots to shift into
685 * history from @max_time.
687 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
688 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
690 if (wbc->wb_id != max_id)
691 history |= (1U << slots) - 1;
694 trace_inode_foreign_history(inode, wbc, history);
697 * Switch if the current wb isn't the consistent winner.
698 * If there are multiple closely competing dirtiers, the
699 * inode may switch across them repeatedly over time, which
700 * is okay. The main goal is avoiding keeping an inode on
701 * the wrong wb for an extended period of time.
703 if (hweight16(history) > WB_FRN_HIST_THR_SLOTS)
704 inode_switch_wbs(inode, max_id);
708 * Multiple instances of this function may race to update the
709 * following fields but we don't mind occassional inaccuracies.
711 inode->i_wb_frn_winner = max_id;
712 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
713 inode->i_wb_frn_history = history;
718 EXPORT_SYMBOL_GPL(wbc_detach_inode);
721 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
722 * @wbc: writeback_control of the writeback in progress
723 * @page: page being written out
724 * @bytes: number of bytes being written out
726 * @bytes from @page are about to written out during the writeback
727 * controlled by @wbc. Keep the book for foreign inode detection. See
728 * wbc_detach_inode().
730 void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
733 struct cgroup_subsys_state *css;
737 * pageout() path doesn't attach @wbc to the inode being written
738 * out. This is intentional as we don't want the function to block
739 * behind a slow cgroup. Ultimately, we want pageout() to kick off
740 * regular writeback instead of writing things out itself.
742 if (!wbc->wb || wbc->no_cgroup_owner)
745 css = mem_cgroup_css_from_page(page);
746 /* dead cgroups shouldn't contribute to inode ownership arbitration */
747 if (!(css->flags & CSS_ONLINE))
752 if (id == wbc->wb_id) {
753 wbc->wb_bytes += bytes;
757 if (id == wbc->wb_lcand_id)
758 wbc->wb_lcand_bytes += bytes;
760 /* Boyer-Moore majority vote algorithm */
761 if (!wbc->wb_tcand_bytes)
762 wbc->wb_tcand_id = id;
763 if (id == wbc->wb_tcand_id)
764 wbc->wb_tcand_bytes += bytes;
766 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
768 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
771 * inode_congested - test whether an inode is congested
772 * @inode: inode to test for congestion (may be NULL)
773 * @cong_bits: mask of WB_[a]sync_congested bits to test
775 * Tests whether @inode is congested. @cong_bits is the mask of congestion
776 * bits to test and the return value is the mask of set bits.
778 * If cgroup writeback is enabled for @inode, the congestion state is
779 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
780 * associated with @inode is congested; otherwise, the root wb's congestion
783 * @inode is allowed to be NULL as this function is often called on
784 * mapping->host which is NULL for the swapper space.
786 int inode_congested(struct inode *inode, int cong_bits)
789 * Once set, ->i_wb never becomes NULL while the inode is alive.
790 * Start transaction iff ->i_wb is visible.
792 if (inode && inode_to_wb_is_valid(inode)) {
793 struct bdi_writeback *wb;
794 struct wb_lock_cookie lock_cookie = {};
797 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
798 congested = wb_congested(wb, cong_bits);
799 unlocked_inode_to_wb_end(inode, &lock_cookie);
803 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
805 EXPORT_SYMBOL_GPL(inode_congested);
808 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
809 * @wb: target bdi_writeback to split @nr_pages to
810 * @nr_pages: number of pages to write for the whole bdi
812 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
813 * relation to the total write bandwidth of all wb's w/ dirty inodes on
816 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
818 unsigned long this_bw = wb->avg_write_bandwidth;
819 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
821 if (nr_pages == LONG_MAX)
825 * This may be called on clean wb's and proportional distribution
826 * may not make sense, just use the original @nr_pages in those
827 * cases. In general, we wanna err on the side of writing more.
829 if (!tot_bw || this_bw >= tot_bw)
832 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
836 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
837 * @bdi: target backing_dev_info
838 * @base_work: wb_writeback_work to issue
839 * @skip_if_busy: skip wb's which already have writeback in progress
841 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
842 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
843 * distributed to the busy wbs according to each wb's proportion in the
844 * total active write bandwidth of @bdi.
846 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
847 struct wb_writeback_work *base_work,
850 struct bdi_writeback *last_wb = NULL;
851 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
852 struct bdi_writeback, bdi_node);
857 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
858 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
859 struct wb_writeback_work fallback_work;
860 struct wb_writeback_work *work;
868 /* SYNC_ALL writes out I_DIRTY_TIME too */
869 if (!wb_has_dirty_io(wb) &&
870 (base_work->sync_mode == WB_SYNC_NONE ||
871 list_empty(&wb->b_dirty_time)))
873 if (skip_if_busy && writeback_in_progress(wb))
876 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
878 work = kmalloc(sizeof(*work), GFP_ATOMIC);
881 work->nr_pages = nr_pages;
883 wb_queue_work(wb, work);
887 /* alloc failed, execute synchronously using on-stack fallback */
888 work = &fallback_work;
890 work->nr_pages = nr_pages;
892 work->done = &fallback_work_done;
894 wb_queue_work(wb, work);
897 * Pin @wb so that it stays on @bdi->wb_list. This allows
898 * continuing iteration from @wb after dropping and
899 * regrabbing rcu read lock.
905 wb_wait_for_completion(&fallback_work_done);
915 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
916 * @bdi_id: target bdi id
917 * @memcg_id: target memcg css id
918 * @nr: number of pages to write, 0 for best-effort dirty flushing
919 * @reason: reason why some writeback work initiated
920 * @done: target wb_completion
922 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
923 * with the specified parameters.
925 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
926 enum wb_reason reason, struct wb_completion *done)
928 struct backing_dev_info *bdi;
929 struct cgroup_subsys_state *memcg_css;
930 struct bdi_writeback *wb;
931 struct wb_writeback_work *work;
934 /* lookup bdi and memcg */
935 bdi = bdi_get_by_id(bdi_id);
940 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
941 if (memcg_css && !css_tryget(memcg_css))
950 * And find the associated wb. If the wb isn't there already
951 * there's nothing to flush, don't create one.
953 wb = wb_get_lookup(bdi, memcg_css);
960 * If @nr is zero, the caller is attempting to write out most of
961 * the currently dirty pages. Let's take the current dirty page
962 * count and inflate it by 25% which should be large enough to
963 * flush out most dirty pages while avoiding getting livelocked by
964 * concurrent dirtiers.
967 unsigned long filepages, headroom, dirty, writeback;
969 mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
974 /* issue the writeback work */
975 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
978 work->sync_mode = WB_SYNC_NONE;
979 work->range_cyclic = 1;
980 work->reason = reason;
983 wb_queue_work(wb, work);
998 * cgroup_writeback_umount - flush inode wb switches for umount
1000 * This function is called when a super_block is about to be destroyed and
1001 * flushes in-flight inode wb switches. An inode wb switch goes through
1002 * RCU and then workqueue, so the two need to be flushed in order to ensure
1003 * that all previously scheduled switches are finished. As wb switches are
1004 * rare occurrences and synchronize_rcu() can take a while, perform
1005 * flushing iff wb switches are in flight.
1007 void cgroup_writeback_umount(void)
1009 if (atomic_read(&isw_nr_in_flight)) {
1011 * Use rcu_barrier() to wait for all pending callbacks to
1012 * ensure that all in-flight wb switches are in the workqueue.
1015 flush_workqueue(isw_wq);
1019 static int __init cgroup_writeback_init(void)
1021 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1026 fs_initcall(cgroup_writeback_init);
1028 #else /* CONFIG_CGROUP_WRITEBACK */
1030 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1031 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1033 static struct bdi_writeback *
1034 locked_inode_to_wb_and_lock_list(struct inode *inode)
1035 __releases(&inode->i_lock)
1036 __acquires(&wb->list_lock)
1038 struct bdi_writeback *wb = inode_to_wb(inode);
1040 spin_unlock(&inode->i_lock);
1041 spin_lock(&wb->list_lock);
1045 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1046 __acquires(&wb->list_lock)
1048 struct bdi_writeback *wb = inode_to_wb(inode);
1050 spin_lock(&wb->list_lock);
1054 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1059 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1060 struct wb_writeback_work *base_work,
1065 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1066 base_work->auto_free = 0;
1067 wb_queue_work(&bdi->wb, base_work);
1071 #endif /* CONFIG_CGROUP_WRITEBACK */
1074 * Add in the number of potentially dirty inodes, because each inode
1075 * write can dirty pagecache in the underlying blockdev.
1077 static unsigned long get_nr_dirty_pages(void)
1079 return global_node_page_state(NR_FILE_DIRTY) +
1080 global_node_page_state(NR_UNSTABLE_NFS) +
1081 get_nr_dirty_inodes();
1084 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1086 if (!wb_has_dirty_io(wb))
1090 * All callers of this function want to start writeback of all
1091 * dirty pages. Places like vmscan can call this at a very
1092 * high frequency, causing pointless allocations of tons of
1093 * work items and keeping the flusher threads busy retrieving
1094 * that work. Ensure that we only allow one of them pending and
1095 * inflight at the time.
1097 if (test_bit(WB_start_all, &wb->state) ||
1098 test_and_set_bit(WB_start_all, &wb->state))
1101 wb->start_all_reason = reason;
1106 * wb_start_background_writeback - start background writeback
1107 * @wb: bdi_writback to write from
1110 * This makes sure WB_SYNC_NONE background writeback happens. When
1111 * this function returns, it is only guaranteed that for given wb
1112 * some IO is happening if we are over background dirty threshold.
1113 * Caller need not hold sb s_umount semaphore.
1115 void wb_start_background_writeback(struct bdi_writeback *wb)
1118 * We just wake up the flusher thread. It will perform background
1119 * writeback as soon as there is no other work to do.
1121 trace_writeback_wake_background(wb);
1126 * Remove the inode from the writeback list it is on.
1128 void inode_io_list_del(struct inode *inode)
1130 struct bdi_writeback *wb;
1132 wb = inode_to_wb_and_lock_list(inode);
1133 spin_lock(&inode->i_lock);
1134 inode_io_list_del_locked(inode, wb);
1135 spin_unlock(&inode->i_lock);
1136 spin_unlock(&wb->list_lock);
1140 * mark an inode as under writeback on the sb
1142 void sb_mark_inode_writeback(struct inode *inode)
1144 struct super_block *sb = inode->i_sb;
1145 unsigned long flags;
1147 if (list_empty(&inode->i_wb_list)) {
1148 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1149 if (list_empty(&inode->i_wb_list)) {
1150 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1151 trace_sb_mark_inode_writeback(inode);
1153 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1158 * clear an inode as under writeback on the sb
1160 void sb_clear_inode_writeback(struct inode *inode)
1162 struct super_block *sb = inode->i_sb;
1163 unsigned long flags;
1165 if (!list_empty(&inode->i_wb_list)) {
1166 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1167 if (!list_empty(&inode->i_wb_list)) {
1168 list_del_init(&inode->i_wb_list);
1169 trace_sb_clear_inode_writeback(inode);
1171 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1176 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1177 * furthest end of its superblock's dirty-inode list.
1179 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1180 * already the most-recently-dirtied inode on the b_dirty list. If that is
1181 * the case then the inode must have been redirtied while it was being written
1182 * out and we don't reset its dirtied_when.
1184 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1186 assert_spin_locked(&inode->i_lock);
1188 if (!list_empty(&wb->b_dirty)) {
1191 tail = wb_inode(wb->b_dirty.next);
1192 if (time_before(inode->dirtied_when, tail->dirtied_when))
1193 inode->dirtied_when = jiffies;
1195 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1196 inode->i_state &= ~I_SYNC_QUEUED;
1199 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1201 spin_lock(&inode->i_lock);
1202 redirty_tail_locked(inode, wb);
1203 spin_unlock(&inode->i_lock);
1207 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1209 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1211 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1214 static void inode_sync_complete(struct inode *inode)
1216 inode->i_state &= ~I_SYNC;
1217 /* If inode is clean an unused, put it into LRU now... */
1218 inode_add_lru(inode);
1219 /* Waiters must see I_SYNC cleared before being woken up */
1221 wake_up_bit(&inode->i_state, __I_SYNC);
1224 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1226 bool ret = time_after(inode->dirtied_when, t);
1227 #ifndef CONFIG_64BIT
1229 * For inodes being constantly redirtied, dirtied_when can get stuck.
1230 * It _appears_ to be in the future, but is actually in distant past.
1231 * This test is necessary to prevent such wrapped-around relative times
1232 * from permanently stopping the whole bdi writeback.
1234 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1239 #define EXPIRE_DIRTY_ATIME 0x0001
1242 * Move expired (dirtied before dirtied_before) dirty inodes from
1243 * @delaying_queue to @dispatch_queue.
1245 static int move_expired_inodes(struct list_head *delaying_queue,
1246 struct list_head *dispatch_queue,
1247 unsigned long dirtied_before)
1250 struct list_head *pos, *node;
1251 struct super_block *sb = NULL;
1252 struct inode *inode;
1256 while (!list_empty(delaying_queue)) {
1257 inode = wb_inode(delaying_queue->prev);
1258 if (inode_dirtied_after(inode, dirtied_before))
1260 list_move(&inode->i_io_list, &tmp);
1262 spin_lock(&inode->i_lock);
1263 inode->i_state |= I_SYNC_QUEUED;
1264 spin_unlock(&inode->i_lock);
1265 if (sb_is_blkdev_sb(inode->i_sb))
1267 if (sb && sb != inode->i_sb)
1272 /* just one sb in list, splice to dispatch_queue and we're done */
1274 list_splice(&tmp, dispatch_queue);
1278 /* Move inodes from one superblock together */
1279 while (!list_empty(&tmp)) {
1280 sb = wb_inode(tmp.prev)->i_sb;
1281 list_for_each_prev_safe(pos, node, &tmp) {
1282 inode = wb_inode(pos);
1283 if (inode->i_sb == sb)
1284 list_move(&inode->i_io_list, dispatch_queue);
1292 * Queue all expired dirty inodes for io, eldest first.
1294 * newly dirtied b_dirty b_io b_more_io
1295 * =============> gf edc BA
1297 * newly dirtied b_dirty b_io b_more_io
1298 * =============> g fBAedc
1300 * +--> dequeue for IO
1302 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1303 unsigned long dirtied_before)
1306 unsigned long time_expire_jif = dirtied_before;
1308 assert_spin_locked(&wb->list_lock);
1309 list_splice_init(&wb->b_more_io, &wb->b_io);
1310 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1311 if (!work->for_sync)
1312 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1313 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1316 wb_io_lists_populated(wb);
1317 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1320 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1324 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1325 trace_writeback_write_inode_start(inode, wbc);
1326 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1327 trace_writeback_write_inode(inode, wbc);
1334 * Wait for writeback on an inode to complete. Called with i_lock held.
1335 * Caller must make sure inode cannot go away when we drop i_lock.
1337 static void __inode_wait_for_writeback(struct inode *inode)
1338 __releases(inode->i_lock)
1339 __acquires(inode->i_lock)
1341 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1342 wait_queue_head_t *wqh;
1344 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1345 while (inode->i_state & I_SYNC) {
1346 spin_unlock(&inode->i_lock);
1347 __wait_on_bit(wqh, &wq, bit_wait,
1348 TASK_UNINTERRUPTIBLE);
1349 spin_lock(&inode->i_lock);
1354 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1356 void inode_wait_for_writeback(struct inode *inode)
1358 spin_lock(&inode->i_lock);
1359 __inode_wait_for_writeback(inode);
1360 spin_unlock(&inode->i_lock);
1364 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1365 * held and drops it. It is aimed for callers not holding any inode reference
1366 * so once i_lock is dropped, inode can go away.
1368 static void inode_sleep_on_writeback(struct inode *inode)
1369 __releases(inode->i_lock)
1372 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1375 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1376 sleep = inode->i_state & I_SYNC;
1377 spin_unlock(&inode->i_lock);
1380 finish_wait(wqh, &wait);
1384 * Find proper writeback list for the inode depending on its current state and
1385 * possibly also change of its state while we were doing writeback. Here we
1386 * handle things such as livelock prevention or fairness of writeback among
1387 * inodes. This function can be called only by flusher thread - noone else
1388 * processes all inodes in writeback lists and requeueing inodes behind flusher
1389 * thread's back can have unexpected consequences.
1391 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1392 struct writeback_control *wbc)
1394 if (inode->i_state & I_FREEING)
1398 * Sync livelock prevention. Each inode is tagged and synced in one
1399 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1400 * the dirty time to prevent enqueue and sync it again.
1402 if ((inode->i_state & I_DIRTY) &&
1403 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1404 inode->dirtied_when = jiffies;
1406 if (wbc->pages_skipped) {
1408 * writeback is not making progress due to locked
1409 * buffers. Skip this inode for now.
1411 redirty_tail_locked(inode, wb);
1415 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1417 * We didn't write back all the pages. nfs_writepages()
1418 * sometimes bales out without doing anything.
1420 if (wbc->nr_to_write <= 0) {
1421 /* Slice used up. Queue for next turn. */
1422 requeue_io(inode, wb);
1425 * Writeback blocked by something other than
1426 * congestion. Delay the inode for some time to
1427 * avoid spinning on the CPU (100% iowait)
1428 * retrying writeback of the dirty page/inode
1429 * that cannot be performed immediately.
1431 redirty_tail_locked(inode, wb);
1433 } else if (inode->i_state & I_DIRTY) {
1435 * Filesystems can dirty the inode during writeback operations,
1436 * such as delayed allocation during submission or metadata
1437 * updates after data IO completion.
1439 redirty_tail_locked(inode, wb);
1440 } else if (inode->i_state & I_DIRTY_TIME) {
1441 inode->dirtied_when = jiffies;
1442 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1443 inode->i_state &= ~I_SYNC_QUEUED;
1445 /* The inode is clean. Remove from writeback lists. */
1446 inode_io_list_del_locked(inode, wb);
1451 * Write out an inode and its dirty pages. Do not update the writeback list
1452 * linkage. That is left to the caller. The caller is also responsible for
1453 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1456 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1458 struct address_space *mapping = inode->i_mapping;
1459 long nr_to_write = wbc->nr_to_write;
1463 WARN_ON(!(inode->i_state & I_SYNC));
1465 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1467 ret = do_writepages(mapping, wbc);
1470 * Make sure to wait on the data before writing out the metadata.
1471 * This is important for filesystems that modify metadata on data
1472 * I/O completion. We don't do it for sync(2) writeback because it has a
1473 * separate, external IO completion path and ->sync_fs for guaranteeing
1474 * inode metadata is written back correctly.
1476 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1477 int err = filemap_fdatawait(mapping);
1483 * If the inode has dirty timestamps and we need to write them, call
1484 * mark_inode_dirty_sync() to notify the filesystem about it and to
1485 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1487 if ((inode->i_state & I_DIRTY_TIME) &&
1488 (wbc->sync_mode == WB_SYNC_ALL || wbc->for_sync ||
1489 time_after(jiffies, inode->dirtied_time_when +
1490 dirtytime_expire_interval * HZ))) {
1491 trace_writeback_lazytime(inode);
1492 mark_inode_dirty_sync(inode);
1496 * Some filesystems may redirty the inode during the writeback
1497 * due to delalloc, clear dirty metadata flags right before
1500 spin_lock(&inode->i_lock);
1501 dirty = inode->i_state & I_DIRTY;
1502 inode->i_state &= ~dirty;
1505 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1506 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1507 * either they see the I_DIRTY bits cleared or we see the dirtied
1510 * I_DIRTY_PAGES is always cleared together above even if @mapping
1511 * still has dirty pages. The flag is reinstated after smp_mb() if
1512 * necessary. This guarantees that either __mark_inode_dirty()
1513 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1517 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1518 inode->i_state |= I_DIRTY_PAGES;
1520 spin_unlock(&inode->i_lock);
1522 /* Don't write the inode if only I_DIRTY_PAGES was set */
1523 if (dirty & ~I_DIRTY_PAGES) {
1524 int err = write_inode(inode, wbc);
1528 trace_writeback_single_inode(inode, wbc, nr_to_write);
1533 * Write out an inode's dirty pages. Either the caller has an active reference
1534 * on the inode or the inode has I_WILL_FREE set.
1536 * This function is designed to be called for writing back one inode which
1537 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1538 * and does more profound writeback list handling in writeback_sb_inodes().
1540 static int writeback_single_inode(struct inode *inode,
1541 struct writeback_control *wbc)
1543 struct bdi_writeback *wb;
1546 spin_lock(&inode->i_lock);
1547 if (!atomic_read(&inode->i_count))
1548 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1550 WARN_ON(inode->i_state & I_WILL_FREE);
1552 if (inode->i_state & I_SYNC) {
1553 if (wbc->sync_mode != WB_SYNC_ALL)
1556 * It's a data-integrity sync. We must wait. Since callers hold
1557 * inode reference or inode has I_WILL_FREE set, it cannot go
1560 __inode_wait_for_writeback(inode);
1562 WARN_ON(inode->i_state & I_SYNC);
1564 * Skip inode if it is clean and we have no outstanding writeback in
1565 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1566 * function since flusher thread may be doing for example sync in
1567 * parallel and if we move the inode, it could get skipped. So here we
1568 * make sure inode is on some writeback list and leave it there unless
1569 * we have completely cleaned the inode.
1571 if (!(inode->i_state & I_DIRTY_ALL) &&
1572 (wbc->sync_mode != WB_SYNC_ALL ||
1573 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1575 inode->i_state |= I_SYNC;
1576 wbc_attach_and_unlock_inode(wbc, inode);
1578 ret = __writeback_single_inode(inode, wbc);
1580 wbc_detach_inode(wbc);
1582 wb = inode_to_wb_and_lock_list(inode);
1583 spin_lock(&inode->i_lock);
1585 * If inode is clean, remove it from writeback lists. Otherwise don't
1586 * touch it. See comment above for explanation.
1588 if (!(inode->i_state & I_DIRTY_ALL))
1589 inode_io_list_del_locked(inode, wb);
1590 spin_unlock(&wb->list_lock);
1591 inode_sync_complete(inode);
1593 spin_unlock(&inode->i_lock);
1597 static long writeback_chunk_size(struct bdi_writeback *wb,
1598 struct wb_writeback_work *work)
1603 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1604 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1605 * here avoids calling into writeback_inodes_wb() more than once.
1607 * The intended call sequence for WB_SYNC_ALL writeback is:
1610 * writeback_sb_inodes() <== called only once
1611 * write_cache_pages() <== called once for each inode
1612 * (quickly) tag currently dirty pages
1613 * (maybe slowly) sync all tagged pages
1615 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1618 pages = min(wb->avg_write_bandwidth / 2,
1619 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1620 pages = min(pages, work->nr_pages);
1621 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1622 MIN_WRITEBACK_PAGES);
1629 * Write a portion of b_io inodes which belong to @sb.
1631 * Return the number of pages and/or inodes written.
1633 * NOTE! This is called with wb->list_lock held, and will
1634 * unlock and relock that for each inode it ends up doing
1637 static long writeback_sb_inodes(struct super_block *sb,
1638 struct bdi_writeback *wb,
1639 struct wb_writeback_work *work)
1641 struct writeback_control wbc = {
1642 .sync_mode = work->sync_mode,
1643 .tagged_writepages = work->tagged_writepages,
1644 .for_kupdate = work->for_kupdate,
1645 .for_background = work->for_background,
1646 .for_sync = work->for_sync,
1647 .range_cyclic = work->range_cyclic,
1649 .range_end = LLONG_MAX,
1651 unsigned long start_time = jiffies;
1653 long total_wrote = 0; /* count both pages and inodes */
1655 while (!list_empty(&wb->b_io)) {
1656 struct inode *inode = wb_inode(wb->b_io.prev);
1657 struct bdi_writeback *tmp_wb;
1660 if (inode->i_sb != sb) {
1663 * We only want to write back data for this
1664 * superblock, move all inodes not belonging
1665 * to it back onto the dirty list.
1667 redirty_tail(inode, wb);
1672 * The inode belongs to a different superblock.
1673 * Bounce back to the caller to unpin this and
1674 * pin the next superblock.
1680 * Don't bother with new inodes or inodes being freed, first
1681 * kind does not need periodic writeout yet, and for the latter
1682 * kind writeout is handled by the freer.
1684 spin_lock(&inode->i_lock);
1685 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1686 redirty_tail_locked(inode, wb);
1687 spin_unlock(&inode->i_lock);
1690 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1692 * If this inode is locked for writeback and we are not
1693 * doing writeback-for-data-integrity, move it to
1694 * b_more_io so that writeback can proceed with the
1695 * other inodes on s_io.
1697 * We'll have another go at writing back this inode
1698 * when we completed a full scan of b_io.
1700 spin_unlock(&inode->i_lock);
1701 requeue_io(inode, wb);
1702 trace_writeback_sb_inodes_requeue(inode);
1705 spin_unlock(&wb->list_lock);
1708 * We already requeued the inode if it had I_SYNC set and we
1709 * are doing WB_SYNC_NONE writeback. So this catches only the
1712 if (inode->i_state & I_SYNC) {
1713 /* Wait for I_SYNC. This function drops i_lock... */
1714 inode_sleep_on_writeback(inode);
1715 /* Inode may be gone, start again */
1716 spin_lock(&wb->list_lock);
1719 inode->i_state |= I_SYNC;
1720 wbc_attach_and_unlock_inode(&wbc, inode);
1722 write_chunk = writeback_chunk_size(wb, work);
1723 wbc.nr_to_write = write_chunk;
1724 wbc.pages_skipped = 0;
1727 * We use I_SYNC to pin the inode in memory. While it is set
1728 * evict_inode() will wait so the inode cannot be freed.
1730 __writeback_single_inode(inode, &wbc);
1732 wbc_detach_inode(&wbc);
1733 work->nr_pages -= write_chunk - wbc.nr_to_write;
1734 wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1735 wrote = wrote < 0 ? 0 : wrote;
1736 total_wrote += wrote;
1738 if (need_resched()) {
1740 * We're trying to balance between building up a nice
1741 * long list of IOs to improve our merge rate, and
1742 * getting those IOs out quickly for anyone throttling
1743 * in balance_dirty_pages(). cond_resched() doesn't
1744 * unplug, so get our IOs out the door before we
1747 blk_flush_plug(current);
1752 * Requeue @inode if still dirty. Be careful as @inode may
1753 * have been switched to another wb in the meantime.
1755 tmp_wb = inode_to_wb_and_lock_list(inode);
1756 spin_lock(&inode->i_lock);
1757 if (!(inode->i_state & I_DIRTY_ALL))
1759 requeue_inode(inode, tmp_wb, &wbc);
1760 inode_sync_complete(inode);
1761 spin_unlock(&inode->i_lock);
1763 if (unlikely(tmp_wb != wb)) {
1764 spin_unlock(&tmp_wb->list_lock);
1765 spin_lock(&wb->list_lock);
1769 * bail out to wb_writeback() often enough to check
1770 * background threshold and other termination conditions.
1773 if (time_is_before_jiffies(start_time + HZ / 10UL))
1775 if (work->nr_pages <= 0)
1782 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1783 struct wb_writeback_work *work)
1785 unsigned long start_time = jiffies;
1788 while (!list_empty(&wb->b_io)) {
1789 struct inode *inode = wb_inode(wb->b_io.prev);
1790 struct super_block *sb = inode->i_sb;
1792 if (!trylock_super(sb)) {
1794 * trylock_super() may fail consistently due to
1795 * s_umount being grabbed by someone else. Don't use
1796 * requeue_io() to avoid busy retrying the inode/sb.
1798 redirty_tail(inode, wb);
1801 wrote += writeback_sb_inodes(sb, wb, work);
1802 up_read(&sb->s_umount);
1804 /* refer to the same tests at the end of writeback_sb_inodes */
1806 if (time_is_before_jiffies(start_time + HZ / 10UL))
1808 if (work->nr_pages <= 0)
1812 /* Leave any unwritten inodes on b_io */
1816 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1817 enum wb_reason reason)
1819 struct wb_writeback_work work = {
1820 .nr_pages = nr_pages,
1821 .sync_mode = WB_SYNC_NONE,
1825 struct blk_plug plug;
1827 blk_start_plug(&plug);
1828 spin_lock(&wb->list_lock);
1829 if (list_empty(&wb->b_io))
1830 queue_io(wb, &work, jiffies);
1831 __writeback_inodes_wb(wb, &work);
1832 spin_unlock(&wb->list_lock);
1833 blk_finish_plug(&plug);
1835 return nr_pages - work.nr_pages;
1839 * Explicit flushing or periodic writeback of "old" data.
1841 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1842 * dirtying-time in the inode's address_space. So this periodic writeback code
1843 * just walks the superblock inode list, writing back any inodes which are
1844 * older than a specific point in time.
1846 * Try to run once per dirty_writeback_interval. But if a writeback event
1847 * takes longer than a dirty_writeback_interval interval, then leave a
1850 * dirtied_before takes precedence over nr_to_write. So we'll only write back
1851 * all dirty pages if they are all attached to "old" mappings.
1853 static long wb_writeback(struct bdi_writeback *wb,
1854 struct wb_writeback_work *work)
1856 unsigned long wb_start = jiffies;
1857 long nr_pages = work->nr_pages;
1858 unsigned long dirtied_before = jiffies;
1859 struct inode *inode;
1861 struct blk_plug plug;
1863 blk_start_plug(&plug);
1864 spin_lock(&wb->list_lock);
1867 * Stop writeback when nr_pages has been consumed
1869 if (work->nr_pages <= 0)
1873 * Background writeout and kupdate-style writeback may
1874 * run forever. Stop them if there is other work to do
1875 * so that e.g. sync can proceed. They'll be restarted
1876 * after the other works are all done.
1878 if ((work->for_background || work->for_kupdate) &&
1879 !list_empty(&wb->work_list))
1883 * For background writeout, stop when we are below the
1884 * background dirty threshold
1886 if (work->for_background && !wb_over_bg_thresh(wb))
1890 * Kupdate and background works are special and we want to
1891 * include all inodes that need writing. Livelock avoidance is
1892 * handled by these works yielding to any other work so we are
1895 if (work->for_kupdate) {
1896 dirtied_before = jiffies -
1897 msecs_to_jiffies(dirty_expire_interval * 10);
1898 } else if (work->for_background)
1899 dirtied_before = jiffies;
1901 trace_writeback_start(wb, work);
1902 if (list_empty(&wb->b_io))
1903 queue_io(wb, work, dirtied_before);
1905 progress = writeback_sb_inodes(work->sb, wb, work);
1907 progress = __writeback_inodes_wb(wb, work);
1908 trace_writeback_written(wb, work);
1910 wb_update_bandwidth(wb, wb_start);
1913 * Did we write something? Try for more
1915 * Dirty inodes are moved to b_io for writeback in batches.
1916 * The completion of the current batch does not necessarily
1917 * mean the overall work is done. So we keep looping as long
1918 * as made some progress on cleaning pages or inodes.
1923 * No more inodes for IO, bail
1925 if (list_empty(&wb->b_more_io))
1928 * Nothing written. Wait for some inode to
1929 * become available for writeback. Otherwise
1930 * we'll just busyloop.
1932 trace_writeback_wait(wb, work);
1933 inode = wb_inode(wb->b_more_io.prev);
1934 spin_lock(&inode->i_lock);
1935 spin_unlock(&wb->list_lock);
1936 /* This function drops i_lock... */
1937 inode_sleep_on_writeback(inode);
1938 spin_lock(&wb->list_lock);
1940 spin_unlock(&wb->list_lock);
1941 blk_finish_plug(&plug);
1943 return nr_pages - work->nr_pages;
1947 * Return the next wb_writeback_work struct that hasn't been processed yet.
1949 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1951 struct wb_writeback_work *work = NULL;
1953 spin_lock_bh(&wb->work_lock);
1954 if (!list_empty(&wb->work_list)) {
1955 work = list_entry(wb->work_list.next,
1956 struct wb_writeback_work, list);
1957 list_del_init(&work->list);
1959 spin_unlock_bh(&wb->work_lock);
1963 static long wb_check_background_flush(struct bdi_writeback *wb)
1965 if (wb_over_bg_thresh(wb)) {
1967 struct wb_writeback_work work = {
1968 .nr_pages = LONG_MAX,
1969 .sync_mode = WB_SYNC_NONE,
1970 .for_background = 1,
1972 .reason = WB_REASON_BACKGROUND,
1975 return wb_writeback(wb, &work);
1981 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1983 unsigned long expired;
1987 * When set to zero, disable periodic writeback
1989 if (!dirty_writeback_interval)
1992 expired = wb->last_old_flush +
1993 msecs_to_jiffies(dirty_writeback_interval * 10);
1994 if (time_before(jiffies, expired))
1997 wb->last_old_flush = jiffies;
1998 nr_pages = get_nr_dirty_pages();
2001 struct wb_writeback_work work = {
2002 .nr_pages = nr_pages,
2003 .sync_mode = WB_SYNC_NONE,
2006 .reason = WB_REASON_PERIODIC,
2009 return wb_writeback(wb, &work);
2015 static long wb_check_start_all(struct bdi_writeback *wb)
2019 if (!test_bit(WB_start_all, &wb->state))
2022 nr_pages = get_nr_dirty_pages();
2024 struct wb_writeback_work work = {
2025 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2026 .sync_mode = WB_SYNC_NONE,
2028 .reason = wb->start_all_reason,
2031 nr_pages = wb_writeback(wb, &work);
2034 clear_bit(WB_start_all, &wb->state);
2040 * Retrieve work items and do the writeback they describe
2042 static long wb_do_writeback(struct bdi_writeback *wb)
2044 struct wb_writeback_work *work;
2047 set_bit(WB_writeback_running, &wb->state);
2048 while ((work = get_next_work_item(wb)) != NULL) {
2049 trace_writeback_exec(wb, work);
2050 wrote += wb_writeback(wb, work);
2051 finish_writeback_work(wb, work);
2055 * Check for a flush-everything request
2057 wrote += wb_check_start_all(wb);
2060 * Check for periodic writeback, kupdated() style
2062 wrote += wb_check_old_data_flush(wb);
2063 wrote += wb_check_background_flush(wb);
2064 clear_bit(WB_writeback_running, &wb->state);
2070 * Handle writeback of dirty data for the device backed by this bdi. Also
2071 * reschedules periodically and does kupdated style flushing.
2073 void wb_workfn(struct work_struct *work)
2075 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2076 struct bdi_writeback, dwork);
2079 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2080 current->flags |= PF_SWAPWRITE;
2082 if (likely(!current_is_workqueue_rescuer() ||
2083 !test_bit(WB_registered, &wb->state))) {
2085 * The normal path. Keep writing back @wb until its
2086 * work_list is empty. Note that this path is also taken
2087 * if @wb is shutting down even when we're running off the
2088 * rescuer as work_list needs to be drained.
2091 pages_written = wb_do_writeback(wb);
2092 trace_writeback_pages_written(pages_written);
2093 } while (!list_empty(&wb->work_list));
2096 * bdi_wq can't get enough workers and we're running off
2097 * the emergency worker. Don't hog it. Hopefully, 1024 is
2098 * enough for efficient IO.
2100 pages_written = writeback_inodes_wb(wb, 1024,
2101 WB_REASON_FORKER_THREAD);
2102 trace_writeback_pages_written(pages_written);
2105 if (!list_empty(&wb->work_list))
2107 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2108 wb_wakeup_delayed(wb);
2110 current->flags &= ~PF_SWAPWRITE;
2114 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2115 * write back the whole world.
2117 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2118 enum wb_reason reason)
2120 struct bdi_writeback *wb;
2122 if (!bdi_has_dirty_io(bdi))
2125 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2126 wb_start_writeback(wb, reason);
2129 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2130 enum wb_reason reason)
2133 __wakeup_flusher_threads_bdi(bdi, reason);
2138 * Wakeup the flusher threads to start writeback of all currently dirty pages
2140 void wakeup_flusher_threads(enum wb_reason reason)
2142 struct backing_dev_info *bdi;
2145 * If we are expecting writeback progress we must submit plugged IO.
2147 if (blk_needs_flush_plug(current))
2148 blk_schedule_flush_plug(current);
2151 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2152 __wakeup_flusher_threads_bdi(bdi, reason);
2157 * Wake up bdi's periodically to make sure dirtytime inodes gets
2158 * written back periodically. We deliberately do *not* check the
2159 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2160 * kernel to be constantly waking up once there are any dirtytime
2161 * inodes on the system. So instead we define a separate delayed work
2162 * function which gets called much more rarely. (By default, only
2163 * once every 12 hours.)
2165 * If there is any other write activity going on in the file system,
2166 * this function won't be necessary. But if the only thing that has
2167 * happened on the file system is a dirtytime inode caused by an atime
2168 * update, we need this infrastructure below to make sure that inode
2169 * eventually gets pushed out to disk.
2171 static void wakeup_dirtytime_writeback(struct work_struct *w);
2172 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2174 static void wakeup_dirtytime_writeback(struct work_struct *w)
2176 struct backing_dev_info *bdi;
2179 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2180 struct bdi_writeback *wb;
2182 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2183 if (!list_empty(&wb->b_dirty_time))
2187 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2190 static int __init start_dirtytime_writeback(void)
2192 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2195 __initcall(start_dirtytime_writeback);
2197 int dirtytime_interval_handler(struct ctl_table *table, int write,
2198 void __user *buffer, size_t *lenp, loff_t *ppos)
2202 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2203 if (ret == 0 && write)
2204 mod_delayed_work(system_wq, &dirtytime_work, 0);
2209 * __mark_inode_dirty - internal function
2211 * @inode: inode to mark
2212 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2214 * Mark an inode as dirty. Callers should use mark_inode_dirty or
2215 * mark_inode_dirty_sync.
2217 * Put the inode on the super block's dirty list.
2219 * CAREFUL! We mark it dirty unconditionally, but move it onto the
2220 * dirty list only if it is hashed or if it refers to a blockdev.
2221 * If it was not hashed, it will never be added to the dirty list
2222 * even if it is later hashed, as it will have been marked dirty already.
2224 * In short, make sure you hash any inodes _before_ you start marking
2227 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2228 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2229 * the kernel-internal blockdev inode represents the dirtying time of the
2230 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2231 * page->mapping->host, so the page-dirtying time is recorded in the internal
2234 void __mark_inode_dirty(struct inode *inode, int flags)
2236 struct super_block *sb = inode->i_sb;
2239 trace_writeback_mark_inode_dirty(inode, flags);
2242 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2243 * dirty the inode itself
2245 if (flags & (I_DIRTY_INODE | I_DIRTY_TIME)) {
2246 trace_writeback_dirty_inode_start(inode, flags);
2248 if (sb->s_op->dirty_inode)
2249 sb->s_op->dirty_inode(inode, flags);
2251 trace_writeback_dirty_inode(inode, flags);
2253 if (flags & I_DIRTY_INODE)
2254 flags &= ~I_DIRTY_TIME;
2255 dirtytime = flags & I_DIRTY_TIME;
2258 * Paired with smp_mb() in __writeback_single_inode() for the
2259 * following lockless i_state test. See there for details.
2263 if (((inode->i_state & flags) == flags) ||
2264 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2267 spin_lock(&inode->i_lock);
2268 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2269 goto out_unlock_inode;
2270 if ((inode->i_state & flags) != flags) {
2271 const int was_dirty = inode->i_state & I_DIRTY;
2273 inode_attach_wb(inode, NULL);
2275 if (flags & I_DIRTY_INODE)
2276 inode->i_state &= ~I_DIRTY_TIME;
2277 inode->i_state |= flags;
2280 * If the inode is queued for writeback by flush worker, just
2281 * update its dirty state. Once the flush worker is done with
2282 * the inode it will place it on the appropriate superblock
2283 * list, based upon its state.
2285 if (inode->i_state & I_SYNC_QUEUED)
2286 goto out_unlock_inode;
2289 * Only add valid (hashed) inodes to the superblock's
2290 * dirty list. Add blockdev inodes as well.
2292 if (!S_ISBLK(inode->i_mode)) {
2293 if (inode_unhashed(inode))
2294 goto out_unlock_inode;
2296 if (inode->i_state & I_FREEING)
2297 goto out_unlock_inode;
2300 * If the inode was already on b_dirty/b_io/b_more_io, don't
2301 * reposition it (that would break b_dirty time-ordering).
2304 struct bdi_writeback *wb;
2305 struct list_head *dirty_list;
2306 bool wakeup_bdi = false;
2308 wb = locked_inode_to_wb_and_lock_list(inode);
2310 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2311 !test_bit(WB_registered, &wb->state),
2312 "bdi-%s not registered\n", wb->bdi->name);
2314 inode->dirtied_when = jiffies;
2316 inode->dirtied_time_when = jiffies;
2318 if (inode->i_state & I_DIRTY)
2319 dirty_list = &wb->b_dirty;
2321 dirty_list = &wb->b_dirty_time;
2323 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2326 spin_unlock(&wb->list_lock);
2327 trace_writeback_dirty_inode_enqueue(inode);
2330 * If this is the first dirty inode for this bdi,
2331 * we have to wake-up the corresponding bdi thread
2332 * to make sure background write-back happens
2335 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2336 wb_wakeup_delayed(wb);
2341 spin_unlock(&inode->i_lock);
2343 EXPORT_SYMBOL(__mark_inode_dirty);
2346 * The @s_sync_lock is used to serialise concurrent sync operations
2347 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2348 * Concurrent callers will block on the s_sync_lock rather than doing contending
2349 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2350 * has been issued up to the time this function is enter is guaranteed to be
2351 * completed by the time we have gained the lock and waited for all IO that is
2352 * in progress regardless of the order callers are granted the lock.
2354 static void wait_sb_inodes(struct super_block *sb)
2356 LIST_HEAD(sync_list);
2359 * We need to be protected against the filesystem going from
2360 * r/o to r/w or vice versa.
2362 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2364 mutex_lock(&sb->s_sync_lock);
2367 * Splice the writeback list onto a temporary list to avoid waiting on
2368 * inodes that have started writeback after this point.
2370 * Use rcu_read_lock() to keep the inodes around until we have a
2371 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2372 * the local list because inodes can be dropped from either by writeback
2376 spin_lock_irq(&sb->s_inode_wblist_lock);
2377 list_splice_init(&sb->s_inodes_wb, &sync_list);
2380 * Data integrity sync. Must wait for all pages under writeback, because
2381 * there may have been pages dirtied before our sync call, but which had
2382 * writeout started before we write it out. In which case, the inode
2383 * may not be on the dirty list, but we still have to wait for that
2386 while (!list_empty(&sync_list)) {
2387 struct inode *inode = list_first_entry(&sync_list, struct inode,
2389 struct address_space *mapping = inode->i_mapping;
2392 * Move each inode back to the wb list before we drop the lock
2393 * to preserve consistency between i_wb_list and the mapping
2394 * writeback tag. Writeback completion is responsible to remove
2395 * the inode from either list once the writeback tag is cleared.
2397 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2400 * The mapping can appear untagged while still on-list since we
2401 * do not have the mapping lock. Skip it here, wb completion
2404 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2407 spin_unlock_irq(&sb->s_inode_wblist_lock);
2409 spin_lock(&inode->i_lock);
2410 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2411 spin_unlock(&inode->i_lock);
2413 spin_lock_irq(&sb->s_inode_wblist_lock);
2417 spin_unlock(&inode->i_lock);
2421 * We keep the error status of individual mapping so that
2422 * applications can catch the writeback error using fsync(2).
2423 * See filemap_fdatawait_keep_errors() for details.
2425 filemap_fdatawait_keep_errors(mapping);
2432 spin_lock_irq(&sb->s_inode_wblist_lock);
2434 spin_unlock_irq(&sb->s_inode_wblist_lock);
2436 mutex_unlock(&sb->s_sync_lock);
2439 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2440 enum wb_reason reason, bool skip_if_busy)
2442 struct backing_dev_info *bdi = sb->s_bdi;
2443 DEFINE_WB_COMPLETION(done, bdi);
2444 struct wb_writeback_work work = {
2446 .sync_mode = WB_SYNC_NONE,
2447 .tagged_writepages = 1,
2453 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2455 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2457 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2458 wb_wait_for_completion(&done);
2462 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2463 * @sb: the superblock
2464 * @nr: the number of pages to write
2465 * @reason: reason why some writeback work initiated
2467 * Start writeback on some inodes on this super_block. No guarantees are made
2468 * on how many (if any) will be written, and this function does not wait
2469 * for IO completion of submitted IO.
2471 void writeback_inodes_sb_nr(struct super_block *sb,
2473 enum wb_reason reason)
2475 __writeback_inodes_sb_nr(sb, nr, reason, false);
2477 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2480 * writeback_inodes_sb - writeback dirty inodes from given super_block
2481 * @sb: the superblock
2482 * @reason: reason why some writeback work was initiated
2484 * Start writeback on some inodes on this super_block. No guarantees are made
2485 * on how many (if any) will be written, and this function does not wait
2486 * for IO completion of submitted IO.
2488 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2490 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2492 EXPORT_SYMBOL(writeback_inodes_sb);
2495 * try_to_writeback_inodes_sb - try to start writeback if none underway
2496 * @sb: the superblock
2497 * @reason: reason why some writeback work was initiated
2499 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2501 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2503 if (!down_read_trylock(&sb->s_umount))
2506 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2507 up_read(&sb->s_umount);
2509 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2512 * sync_inodes_sb - sync sb inode pages
2513 * @sb: the superblock
2515 * This function writes and waits on any dirty inode belonging to this
2518 void sync_inodes_sb(struct super_block *sb)
2520 struct backing_dev_info *bdi = sb->s_bdi;
2521 DEFINE_WB_COMPLETION(done, bdi);
2522 struct wb_writeback_work work = {
2524 .sync_mode = WB_SYNC_ALL,
2525 .nr_pages = LONG_MAX,
2528 .reason = WB_REASON_SYNC,
2533 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2534 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2535 * bdi_has_dirty() need to be written out too.
2537 if (bdi == &noop_backing_dev_info)
2539 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2541 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2542 bdi_down_write_wb_switch_rwsem(bdi);
2543 bdi_split_work_to_wbs(bdi, &work, false);
2544 wb_wait_for_completion(&done);
2545 bdi_up_write_wb_switch_rwsem(bdi);
2549 EXPORT_SYMBOL(sync_inodes_sb);
2552 * write_inode_now - write an inode to disk
2553 * @inode: inode to write to disk
2554 * @sync: whether the write should be synchronous or not
2556 * This function commits an inode to disk immediately if it is dirty. This is
2557 * primarily needed by knfsd.
2559 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2561 int write_inode_now(struct inode *inode, int sync)
2563 struct writeback_control wbc = {
2564 .nr_to_write = LONG_MAX,
2565 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2567 .range_end = LLONG_MAX,
2570 if (!mapping_cap_writeback_dirty(inode->i_mapping))
2571 wbc.nr_to_write = 0;
2574 return writeback_single_inode(inode, &wbc);
2576 EXPORT_SYMBOL(write_inode_now);
2579 * sync_inode - write an inode and its pages to disk.
2580 * @inode: the inode to sync
2581 * @wbc: controls the writeback mode
2583 * sync_inode() will write an inode and its pages to disk. It will also
2584 * correctly update the inode on its superblock's dirty inode lists and will
2585 * update inode->i_state.
2587 * The caller must have a ref on the inode.
2589 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2591 return writeback_single_inode(inode, wbc);
2593 EXPORT_SYMBOL(sync_inode);
2596 * sync_inode_metadata - write an inode to disk
2597 * @inode: the inode to sync
2598 * @wait: wait for I/O to complete.
2600 * Write an inode to disk and adjust its dirty state after completion.
2602 * Note: only writes the actual inode, no associated data or other metadata.
2604 int sync_inode_metadata(struct inode *inode, int wait)
2606 struct writeback_control wbc = {
2607 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2608 .nr_to_write = 0, /* metadata-only */
2611 return sync_inode(inode, &wbc);
2613 EXPORT_SYMBOL(sync_inode_metadata);