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);
134 static void wb_wakeup(struct bdi_writeback *wb)
136 spin_lock_bh(&wb->work_lock);
137 if (test_bit(WB_registered, &wb->state))
138 mod_delayed_work(bdi_wq, &wb->dwork, 0);
139 spin_unlock_bh(&wb->work_lock);
142 static void finish_writeback_work(struct bdi_writeback *wb,
143 struct wb_writeback_work *work)
145 struct wb_completion *done = work->done;
150 wait_queue_head_t *waitq = done->waitq;
152 /* @done can't be accessed after the following dec */
153 if (atomic_dec_and_test(&done->cnt))
158 static void wb_queue_work(struct bdi_writeback *wb,
159 struct wb_writeback_work *work)
161 trace_writeback_queue(wb, work);
164 atomic_inc(&work->done->cnt);
166 spin_lock_bh(&wb->work_lock);
168 if (test_bit(WB_registered, &wb->state)) {
169 list_add_tail(&work->list, &wb->work_list);
170 mod_delayed_work(bdi_wq, &wb->dwork, 0);
172 finish_writeback_work(wb, work);
174 spin_unlock_bh(&wb->work_lock);
178 * wb_wait_for_completion - wait for completion of bdi_writeback_works
179 * @done: target wb_completion
181 * Wait for one or more work items issued to @bdi with their ->done field
182 * set to @done, which should have been initialized with
183 * DEFINE_WB_COMPLETION(). This function returns after all such work items
184 * are completed. Work items which are waited upon aren't freed
185 * automatically on completion.
187 void wb_wait_for_completion(struct wb_completion *done)
189 atomic_dec(&done->cnt); /* put down the initial count */
190 wait_event(*done->waitq, !atomic_read(&done->cnt));
193 #ifdef CONFIG_CGROUP_WRITEBACK
196 * Parameters for foreign inode detection, see wbc_detach_inode() to see
199 * These paramters are inherently heuristical as the detection target
200 * itself is fuzzy. All we want to do is detaching an inode from the
201 * current owner if it's being written to by some other cgroups too much.
203 * The current cgroup writeback is built on the assumption that multiple
204 * cgroups writing to the same inode concurrently is very rare and a mode
205 * of operation which isn't well supported. As such, the goal is not
206 * taking too long when a different cgroup takes over an inode while
207 * avoiding too aggressive flip-flops from occasional foreign writes.
209 * We record, very roughly, 2s worth of IO time history and if more than
210 * half of that is foreign, trigger the switch. The recording is quantized
211 * to 16 slots. To avoid tiny writes from swinging the decision too much,
212 * writes smaller than 1/8 of avg size are ignored.
214 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
215 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
216 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
217 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
219 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
220 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
221 /* each slot's duration is 2s / 16 */
222 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
223 /* if foreign slots >= 8, switch */
224 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
225 /* one round can affect upto 5 slots */
226 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
228 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
229 static struct workqueue_struct *isw_wq;
231 void __inode_attach_wb(struct inode *inode, struct page *page)
233 struct backing_dev_info *bdi = inode_to_bdi(inode);
234 struct bdi_writeback *wb = NULL;
236 if (inode_cgwb_enabled(inode)) {
237 struct cgroup_subsys_state *memcg_css;
240 memcg_css = mem_cgroup_css_from_page(page);
241 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
243 /* must pin memcg_css, see wb_get_create() */
244 memcg_css = task_get_css(current, memory_cgrp_id);
245 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
254 * There may be multiple instances of this function racing to
255 * update the same inode. Use cmpxchg() to tell the winner.
257 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
260 EXPORT_SYMBOL_GPL(__inode_attach_wb);
263 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
264 * @inode: inode of interest with i_lock held
265 * @wb: target bdi_writeback
267 * Remove the inode from wb's io lists and if necessarily put onto b_attached
268 * list. Only inodes attached to cgwb's are kept on this list.
270 static void inode_cgwb_move_to_attached(struct inode *inode,
271 struct bdi_writeback *wb)
273 assert_spin_locked(&wb->list_lock);
274 assert_spin_locked(&inode->i_lock);
276 inode->i_state &= ~I_SYNC_QUEUED;
277 if (wb != &wb->bdi->wb)
278 list_move(&inode->i_io_list, &wb->b_attached);
280 list_del_init(&inode->i_io_list);
281 wb_io_lists_depopulated(wb);
285 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
286 * @inode: inode of interest with i_lock held
288 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
289 * held on entry and is released on return. The returned wb is guaranteed
290 * to stay @inode's associated wb until its list_lock is released.
292 static struct bdi_writeback *
293 locked_inode_to_wb_and_lock_list(struct inode *inode)
294 __releases(&inode->i_lock)
295 __acquires(&wb->list_lock)
298 struct bdi_writeback *wb = inode_to_wb(inode);
301 * inode_to_wb() association is protected by both
302 * @inode->i_lock and @wb->list_lock but list_lock nests
303 * outside i_lock. Drop i_lock and verify that the
304 * association hasn't changed after acquiring list_lock.
307 spin_unlock(&inode->i_lock);
308 spin_lock(&wb->list_lock);
310 /* i_wb may have changed inbetween, can't use inode_to_wb() */
311 if (likely(wb == inode->i_wb)) {
312 wb_put(wb); /* @inode already has ref */
316 spin_unlock(&wb->list_lock);
319 spin_lock(&inode->i_lock);
324 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
325 * @inode: inode of interest
327 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
330 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
331 __acquires(&wb->list_lock)
333 spin_lock(&inode->i_lock);
334 return locked_inode_to_wb_and_lock_list(inode);
337 struct inode_switch_wbs_context {
339 struct bdi_writeback *new_wb;
341 struct rcu_work work;
344 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
346 down_write(&bdi->wb_switch_rwsem);
349 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
351 up_write(&bdi->wb_switch_rwsem);
354 static void inode_switch_wbs_work_fn(struct work_struct *work)
356 struct inode_switch_wbs_context *isw =
357 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
358 struct inode *inode = isw->inode;
359 struct backing_dev_info *bdi = inode_to_bdi(inode);
360 struct address_space *mapping = inode->i_mapping;
361 struct bdi_writeback *old_wb = inode->i_wb;
362 struct bdi_writeback *new_wb = isw->new_wb;
363 XA_STATE(xas, &mapping->i_pages, 0);
365 bool switched = false;
368 * If @inode switches cgwb membership while sync_inodes_sb() is
369 * being issued, sync_inodes_sb() might miss it. Synchronize.
371 down_read(&bdi->wb_switch_rwsem);
374 * By the time control reaches here, RCU grace period has passed
375 * since I_WB_SWITCH assertion and all wb stat update transactions
376 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
377 * synchronizing against the i_pages lock.
379 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
380 * gives us exclusion against all wb related operations on @inode
381 * including IO list manipulations and stat updates.
383 if (old_wb < new_wb) {
384 spin_lock(&old_wb->list_lock);
385 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
387 spin_lock(&new_wb->list_lock);
388 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
390 spin_lock(&inode->i_lock);
391 xa_lock_irq(&mapping->i_pages);
394 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
395 * path owns the inode and we shouldn't modify ->i_io_list.
397 if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
400 trace_inode_switch_wbs(inode, old_wb, new_wb);
403 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
404 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
405 * pages actually under writeback.
407 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
408 if (PageDirty(page)) {
409 dec_wb_stat(old_wb, WB_RECLAIMABLE);
410 inc_wb_stat(new_wb, WB_RECLAIMABLE);
415 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
416 WARN_ON_ONCE(!PageWriteback(page));
417 dec_wb_stat(old_wb, WB_WRITEBACK);
418 inc_wb_stat(new_wb, WB_WRITEBACK);
424 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
425 * the specific list @inode was on is ignored and the @inode is put on
426 * ->b_dirty which is always correct including from ->b_dirty_time.
427 * The transfer preserves @inode->dirtied_when ordering. If the @inode
428 * was clean, it means it was on the b_attached list, so move it onto
429 * the b_attached list of @new_wb.
431 if (!list_empty(&inode->i_io_list)) {
432 inode->i_wb = new_wb;
434 if (inode->i_state & I_DIRTY_ALL) {
437 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
438 if (time_after_eq(inode->dirtied_when,
441 inode_io_list_move_locked(inode, new_wb,
442 pos->i_io_list.prev);
444 inode_cgwb_move_to_attached(inode, new_wb);
447 inode->i_wb = new_wb;
450 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
451 inode->i_wb_frn_winner = 0;
452 inode->i_wb_frn_avg_time = 0;
453 inode->i_wb_frn_history = 0;
457 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
458 * ensures that the new wb is visible if they see !I_WB_SWITCH.
460 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
462 xa_unlock_irq(&mapping->i_pages);
463 spin_unlock(&inode->i_lock);
464 spin_unlock(&new_wb->list_lock);
465 spin_unlock(&old_wb->list_lock);
467 up_read(&bdi->wb_switch_rwsem);
478 atomic_dec(&isw_nr_in_flight);
482 * inode_switch_wbs - change the wb association of an inode
483 * @inode: target inode
484 * @new_wb_id: ID of the new wb
486 * Switch @inode's wb association to the wb identified by @new_wb_id. The
487 * switching is performed asynchronously and may fail silently.
489 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
491 struct backing_dev_info *bdi = inode_to_bdi(inode);
492 struct cgroup_subsys_state *memcg_css;
493 struct inode_switch_wbs_context *isw;
495 /* noop if seems to be already in progress */
496 if (inode->i_state & I_WB_SWITCH)
499 /* avoid queueing a new switch if too many are already in flight */
500 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
503 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
507 atomic_inc(&isw_nr_in_flight);
509 /* find and pin the new wb */
511 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
513 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
518 /* while holding I_WB_SWITCH, no one else can update the association */
519 spin_lock(&inode->i_lock);
520 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
521 inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
522 inode_to_wb(inode) == isw->new_wb) {
523 spin_unlock(&inode->i_lock);
526 inode->i_state |= I_WB_SWITCH;
528 spin_unlock(&inode->i_lock);
533 * In addition to synchronizing among switchers, I_WB_SWITCH tells
534 * the RCU protected stat update paths to grab the i_page
535 * lock so that stat transfer can synchronize against them.
536 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
538 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
539 queue_rcu_work(isw_wq, &isw->work);
543 atomic_dec(&isw_nr_in_flight);
550 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
551 * @wbc: writeback_control of interest
552 * @inode: target inode
554 * @inode is locked and about to be written back under the control of @wbc.
555 * Record @inode's writeback context into @wbc and unlock the i_lock. On
556 * writeback completion, wbc_detach_inode() should be called. This is used
557 * to track the cgroup writeback context.
559 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
562 if (!inode_cgwb_enabled(inode)) {
563 spin_unlock(&inode->i_lock);
567 wbc->wb = inode_to_wb(inode);
570 wbc->wb_id = wbc->wb->memcg_css->id;
571 wbc->wb_lcand_id = inode->i_wb_frn_winner;
572 wbc->wb_tcand_id = 0;
574 wbc->wb_lcand_bytes = 0;
575 wbc->wb_tcand_bytes = 0;
578 spin_unlock(&inode->i_lock);
581 * A dying wb indicates that either the blkcg associated with the
582 * memcg changed or the associated memcg is dying. In the first
583 * case, a replacement wb should already be available and we should
584 * refresh the wb immediately. In the second case, trying to
585 * refresh will keep failing.
587 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
588 inode_switch_wbs(inode, wbc->wb_id);
590 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
593 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
594 * @wbc: writeback_control of the just finished writeback
596 * To be called after a writeback attempt of an inode finishes and undoes
597 * wbc_attach_and_unlock_inode(). Can be called under any context.
599 * As concurrent write sharing of an inode is expected to be very rare and
600 * memcg only tracks page ownership on first-use basis severely confining
601 * the usefulness of such sharing, cgroup writeback tracks ownership
602 * per-inode. While the support for concurrent write sharing of an inode
603 * is deemed unnecessary, an inode being written to by different cgroups at
604 * different points in time is a lot more common, and, more importantly,
605 * charging only by first-use can too readily lead to grossly incorrect
606 * behaviors (single foreign page can lead to gigabytes of writeback to be
607 * incorrectly attributed).
609 * To resolve this issue, cgroup writeback detects the majority dirtier of
610 * an inode and transfers the ownership to it. To avoid unnnecessary
611 * oscillation, the detection mechanism keeps track of history and gives
612 * out the switch verdict only if the foreign usage pattern is stable over
613 * a certain amount of time and/or writeback attempts.
615 * On each writeback attempt, @wbc tries to detect the majority writer
616 * using Boyer-Moore majority vote algorithm. In addition to the byte
617 * count from the majority voting, it also counts the bytes written for the
618 * current wb and the last round's winner wb (max of last round's current
619 * wb, the winner from two rounds ago, and the last round's majority
620 * candidate). Keeping track of the historical winner helps the algorithm
621 * to semi-reliably detect the most active writer even when it's not the
624 * Once the winner of the round is determined, whether the winner is
625 * foreign or not and how much IO time the round consumed is recorded in
626 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
627 * over a certain threshold, the switch verdict is given.
629 void wbc_detach_inode(struct writeback_control *wbc)
631 struct bdi_writeback *wb = wbc->wb;
632 struct inode *inode = wbc->inode;
633 unsigned long avg_time, max_bytes, max_time;
640 history = inode->i_wb_frn_history;
641 avg_time = inode->i_wb_frn_avg_time;
643 /* pick the winner of this round */
644 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
645 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
647 max_bytes = wbc->wb_bytes;
648 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
649 max_id = wbc->wb_lcand_id;
650 max_bytes = wbc->wb_lcand_bytes;
652 max_id = wbc->wb_tcand_id;
653 max_bytes = wbc->wb_tcand_bytes;
657 * Calculate the amount of IO time the winner consumed and fold it
658 * into the running average kept per inode. If the consumed IO
659 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
660 * deciding whether to switch or not. This is to prevent one-off
661 * small dirtiers from skewing the verdict.
663 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
664 wb->avg_write_bandwidth);
666 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
667 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
669 avg_time = max_time; /* immediate catch up on first run */
671 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
675 * The switch verdict is reached if foreign wb's consume
676 * more than a certain proportion of IO time in a
677 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
678 * history mask where each bit represents one sixteenth of
679 * the period. Determine the number of slots to shift into
680 * history from @max_time.
682 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
683 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
685 if (wbc->wb_id != max_id)
686 history |= (1U << slots) - 1;
689 trace_inode_foreign_history(inode, wbc, history);
692 * Switch if the current wb isn't the consistent winner.
693 * If there are multiple closely competing dirtiers, the
694 * inode may switch across them repeatedly over time, which
695 * is okay. The main goal is avoiding keeping an inode on
696 * the wrong wb for an extended period of time.
698 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
699 inode_switch_wbs(inode, max_id);
703 * Multiple instances of this function may race to update the
704 * following fields but we don't mind occassional inaccuracies.
706 inode->i_wb_frn_winner = max_id;
707 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
708 inode->i_wb_frn_history = history;
713 EXPORT_SYMBOL_GPL(wbc_detach_inode);
716 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
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_cgroup_owner(struct writeback_control *wbc, struct page *page,
728 struct cgroup_subsys_state *css;
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 || wbc->no_cgroup_owner)
740 css = mem_cgroup_css_from_page(page);
741 /* dead cgroups shouldn't contribute to inode ownership arbitration */
742 if (!(css->flags & CSS_ONLINE))
747 if (id == wbc->wb_id) {
748 wbc->wb_bytes += bytes;
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;
761 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
763 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
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
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 = {};
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);
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
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)
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)
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,
845 struct bdi_writeback *last_wb = NULL;
846 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
847 struct bdi_writeback, bdi_node);
852 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
853 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
854 struct wb_writeback_work fallback_work;
855 struct wb_writeback_work *work;
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)))
868 if (skip_if_busy && writeback_in_progress(wb))
871 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
873 work = kmalloc(sizeof(*work), GFP_ATOMIC);
876 work->nr_pages = nr_pages;
878 wb_queue_work(wb, work);
882 /* alloc failed, execute synchronously using on-stack fallback */
883 work = &fallback_work;
885 work->nr_pages = nr_pages;
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.
900 wb_wait_for_completion(&fallback_work_done);
910 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
911 * @bdi_id: target bdi id
912 * @memcg_id: target memcg css id
913 * @nr: number of pages to write, 0 for best-effort dirty flushing
914 * @reason: reason why some writeback work initiated
915 * @done: target wb_completion
917 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
918 * with the specified parameters.
920 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
921 enum wb_reason reason, struct wb_completion *done)
923 struct backing_dev_info *bdi;
924 struct cgroup_subsys_state *memcg_css;
925 struct bdi_writeback *wb;
926 struct wb_writeback_work *work;
929 /* lookup bdi and memcg */
930 bdi = bdi_get_by_id(bdi_id);
935 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
936 if (memcg_css && !css_tryget(memcg_css))
945 * And find the associated wb. If the wb isn't there already
946 * there's nothing to flush, don't create one.
948 wb = wb_get_lookup(bdi, memcg_css);
955 * If @nr is zero, the caller is attempting to write out most of
956 * the currently dirty pages. Let's take the current dirty page
957 * count and inflate it by 25% which should be large enough to
958 * flush out most dirty pages while avoiding getting livelocked by
959 * concurrent dirtiers.
962 unsigned long filepages, headroom, dirty, writeback;
964 mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
969 /* issue the writeback work */
970 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
973 work->sync_mode = WB_SYNC_NONE;
974 work->range_cyclic = 1;
975 work->reason = reason;
978 wb_queue_work(wb, work);
993 * cgroup_writeback_umount - flush inode wb switches for umount
995 * This function is called when a super_block is about to be destroyed and
996 * flushes in-flight inode wb switches. An inode wb switch goes through
997 * RCU and then workqueue, so the two need to be flushed in order to ensure
998 * that all previously scheduled switches are finished. As wb switches are
999 * rare occurrences and synchronize_rcu() can take a while, perform
1000 * flushing iff wb switches are in flight.
1002 void cgroup_writeback_umount(void)
1005 * SB_ACTIVE should be reliably cleared before checking
1006 * isw_nr_in_flight, see generic_shutdown_super().
1010 if (atomic_read(&isw_nr_in_flight)) {
1012 * Use rcu_barrier() to wait for all pending callbacks to
1013 * ensure that all in-flight wb switches are in the workqueue.
1016 flush_workqueue(isw_wq);
1020 static int __init cgroup_writeback_init(void)
1022 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1027 fs_initcall(cgroup_writeback_init);
1029 #else /* CONFIG_CGROUP_WRITEBACK */
1031 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1032 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1034 static void inode_cgwb_move_to_attached(struct inode *inode,
1035 struct bdi_writeback *wb)
1037 assert_spin_locked(&wb->list_lock);
1038 assert_spin_locked(&inode->i_lock);
1040 inode->i_state &= ~I_SYNC_QUEUED;
1041 list_del_init(&inode->i_io_list);
1042 wb_io_lists_depopulated(wb);
1045 static struct bdi_writeback *
1046 locked_inode_to_wb_and_lock_list(struct inode *inode)
1047 __releases(&inode->i_lock)
1048 __acquires(&wb->list_lock)
1050 struct bdi_writeback *wb = inode_to_wb(inode);
1052 spin_unlock(&inode->i_lock);
1053 spin_lock(&wb->list_lock);
1057 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1058 __acquires(&wb->list_lock)
1060 struct bdi_writeback *wb = inode_to_wb(inode);
1062 spin_lock(&wb->list_lock);
1066 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1071 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1072 struct wb_writeback_work *base_work,
1077 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1078 base_work->auto_free = 0;
1079 wb_queue_work(&bdi->wb, base_work);
1083 #endif /* CONFIG_CGROUP_WRITEBACK */
1086 * Add in the number of potentially dirty inodes, because each inode
1087 * write can dirty pagecache in the underlying blockdev.
1089 static unsigned long get_nr_dirty_pages(void)
1091 return global_node_page_state(NR_FILE_DIRTY) +
1092 get_nr_dirty_inodes();
1095 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1097 if (!wb_has_dirty_io(wb))
1101 * All callers of this function want to start writeback of all
1102 * dirty pages. Places like vmscan can call this at a very
1103 * high frequency, causing pointless allocations of tons of
1104 * work items and keeping the flusher threads busy retrieving
1105 * that work. Ensure that we only allow one of them pending and
1106 * inflight at the time.
1108 if (test_bit(WB_start_all, &wb->state) ||
1109 test_and_set_bit(WB_start_all, &wb->state))
1112 wb->start_all_reason = reason;
1117 * wb_start_background_writeback - start background writeback
1118 * @wb: bdi_writback to write from
1121 * This makes sure WB_SYNC_NONE background writeback happens. When
1122 * this function returns, it is only guaranteed that for given wb
1123 * some IO is happening if we are over background dirty threshold.
1124 * Caller need not hold sb s_umount semaphore.
1126 void wb_start_background_writeback(struct bdi_writeback *wb)
1129 * We just wake up the flusher thread. It will perform background
1130 * writeback as soon as there is no other work to do.
1132 trace_writeback_wake_background(wb);
1137 * Remove the inode from the writeback list it is on.
1139 void inode_io_list_del(struct inode *inode)
1141 struct bdi_writeback *wb;
1143 wb = inode_to_wb_and_lock_list(inode);
1144 spin_lock(&inode->i_lock);
1146 inode->i_state &= ~I_SYNC_QUEUED;
1147 list_del_init(&inode->i_io_list);
1148 wb_io_lists_depopulated(wb);
1150 spin_unlock(&inode->i_lock);
1151 spin_unlock(&wb->list_lock);
1153 EXPORT_SYMBOL(inode_io_list_del);
1156 * mark an inode as under writeback on the sb
1158 void sb_mark_inode_writeback(struct inode *inode)
1160 struct super_block *sb = inode->i_sb;
1161 unsigned long flags;
1163 if (list_empty(&inode->i_wb_list)) {
1164 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1165 if (list_empty(&inode->i_wb_list)) {
1166 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1167 trace_sb_mark_inode_writeback(inode);
1169 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1174 * clear an inode as under writeback on the sb
1176 void sb_clear_inode_writeback(struct inode *inode)
1178 struct super_block *sb = inode->i_sb;
1179 unsigned long flags;
1181 if (!list_empty(&inode->i_wb_list)) {
1182 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1183 if (!list_empty(&inode->i_wb_list)) {
1184 list_del_init(&inode->i_wb_list);
1185 trace_sb_clear_inode_writeback(inode);
1187 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1192 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1193 * furthest end of its superblock's dirty-inode list.
1195 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1196 * already the most-recently-dirtied inode on the b_dirty list. If that is
1197 * the case then the inode must have been redirtied while it was being written
1198 * out and we don't reset its dirtied_when.
1200 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1202 assert_spin_locked(&inode->i_lock);
1204 if (!list_empty(&wb->b_dirty)) {
1207 tail = wb_inode(wb->b_dirty.next);
1208 if (time_before(inode->dirtied_when, tail->dirtied_when))
1209 inode->dirtied_when = jiffies;
1211 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1212 inode->i_state &= ~I_SYNC_QUEUED;
1215 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1217 spin_lock(&inode->i_lock);
1218 redirty_tail_locked(inode, wb);
1219 spin_unlock(&inode->i_lock);
1223 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1225 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1227 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1230 static void inode_sync_complete(struct inode *inode)
1232 inode->i_state &= ~I_SYNC;
1233 /* If inode is clean an unused, put it into LRU now... */
1234 inode_add_lru(inode);
1235 /* Waiters must see I_SYNC cleared before being woken up */
1237 wake_up_bit(&inode->i_state, __I_SYNC);
1240 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1242 bool ret = time_after(inode->dirtied_when, t);
1243 #ifndef CONFIG_64BIT
1245 * For inodes being constantly redirtied, dirtied_when can get stuck.
1246 * It _appears_ to be in the future, but is actually in distant past.
1247 * This test is necessary to prevent such wrapped-around relative times
1248 * from permanently stopping the whole bdi writeback.
1250 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1255 #define EXPIRE_DIRTY_ATIME 0x0001
1258 * Move expired (dirtied before dirtied_before) dirty inodes from
1259 * @delaying_queue to @dispatch_queue.
1261 static int move_expired_inodes(struct list_head *delaying_queue,
1262 struct list_head *dispatch_queue,
1263 unsigned long dirtied_before)
1266 struct list_head *pos, *node;
1267 struct super_block *sb = NULL;
1268 struct inode *inode;
1272 while (!list_empty(delaying_queue)) {
1273 inode = wb_inode(delaying_queue->prev);
1274 if (inode_dirtied_after(inode, dirtied_before))
1276 list_move(&inode->i_io_list, &tmp);
1278 spin_lock(&inode->i_lock);
1279 inode->i_state |= I_SYNC_QUEUED;
1280 spin_unlock(&inode->i_lock);
1281 if (sb_is_blkdev_sb(inode->i_sb))
1283 if (sb && sb != inode->i_sb)
1288 /* just one sb in list, splice to dispatch_queue and we're done */
1290 list_splice(&tmp, dispatch_queue);
1294 /* Move inodes from one superblock together */
1295 while (!list_empty(&tmp)) {
1296 sb = wb_inode(tmp.prev)->i_sb;
1297 list_for_each_prev_safe(pos, node, &tmp) {
1298 inode = wb_inode(pos);
1299 if (inode->i_sb == sb)
1300 list_move(&inode->i_io_list, dispatch_queue);
1308 * Queue all expired dirty inodes for io, eldest first.
1310 * newly dirtied b_dirty b_io b_more_io
1311 * =============> gf edc BA
1313 * newly dirtied b_dirty b_io b_more_io
1314 * =============> g fBAedc
1316 * +--> dequeue for IO
1318 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1319 unsigned long dirtied_before)
1322 unsigned long time_expire_jif = dirtied_before;
1324 assert_spin_locked(&wb->list_lock);
1325 list_splice_init(&wb->b_more_io, &wb->b_io);
1326 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1327 if (!work->for_sync)
1328 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1329 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1332 wb_io_lists_populated(wb);
1333 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1336 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1340 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1341 trace_writeback_write_inode_start(inode, wbc);
1342 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1343 trace_writeback_write_inode(inode, wbc);
1350 * Wait for writeback on an inode to complete. Called with i_lock held.
1351 * Caller must make sure inode cannot go away when we drop i_lock.
1353 static void __inode_wait_for_writeback(struct inode *inode)
1354 __releases(inode->i_lock)
1355 __acquires(inode->i_lock)
1357 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1358 wait_queue_head_t *wqh;
1360 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1361 while (inode->i_state & I_SYNC) {
1362 spin_unlock(&inode->i_lock);
1363 __wait_on_bit(wqh, &wq, bit_wait,
1364 TASK_UNINTERRUPTIBLE);
1365 spin_lock(&inode->i_lock);
1370 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1372 void inode_wait_for_writeback(struct inode *inode)
1374 spin_lock(&inode->i_lock);
1375 __inode_wait_for_writeback(inode);
1376 spin_unlock(&inode->i_lock);
1380 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1381 * held and drops it. It is aimed for callers not holding any inode reference
1382 * so once i_lock is dropped, inode can go away.
1384 static void inode_sleep_on_writeback(struct inode *inode)
1385 __releases(inode->i_lock)
1388 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1391 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1392 sleep = inode->i_state & I_SYNC;
1393 spin_unlock(&inode->i_lock);
1396 finish_wait(wqh, &wait);
1400 * Find proper writeback list for the inode depending on its current state and
1401 * possibly also change of its state while we were doing writeback. Here we
1402 * handle things such as livelock prevention or fairness of writeback among
1403 * inodes. This function can be called only by flusher thread - noone else
1404 * processes all inodes in writeback lists and requeueing inodes behind flusher
1405 * thread's back can have unexpected consequences.
1407 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1408 struct writeback_control *wbc)
1410 if (inode->i_state & I_FREEING)
1414 * Sync livelock prevention. Each inode is tagged and synced in one
1415 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1416 * the dirty time to prevent enqueue and sync it again.
1418 if ((inode->i_state & I_DIRTY) &&
1419 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1420 inode->dirtied_when = jiffies;
1422 if (wbc->pages_skipped) {
1424 * writeback is not making progress due to locked
1425 * buffers. Skip this inode for now.
1427 redirty_tail_locked(inode, wb);
1431 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1433 * We didn't write back all the pages. nfs_writepages()
1434 * sometimes bales out without doing anything.
1436 if (wbc->nr_to_write <= 0) {
1437 /* Slice used up. Queue for next turn. */
1438 requeue_io(inode, wb);
1441 * Writeback blocked by something other than
1442 * congestion. Delay the inode for some time to
1443 * avoid spinning on the CPU (100% iowait)
1444 * retrying writeback of the dirty page/inode
1445 * that cannot be performed immediately.
1447 redirty_tail_locked(inode, wb);
1449 } else if (inode->i_state & I_DIRTY) {
1451 * Filesystems can dirty the inode during writeback operations,
1452 * such as delayed allocation during submission or metadata
1453 * updates after data IO completion.
1455 redirty_tail_locked(inode, wb);
1456 } else if (inode->i_state & I_DIRTY_TIME) {
1457 inode->dirtied_when = jiffies;
1458 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1459 inode->i_state &= ~I_SYNC_QUEUED;
1461 /* The inode is clean. Remove from writeback lists. */
1462 inode_cgwb_move_to_attached(inode, wb);
1467 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1468 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1470 * This doesn't remove the inode from the writeback list it is on, except
1471 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1472 * expiration. The caller is otherwise responsible for writeback list handling.
1474 * The caller is also responsible for setting the I_SYNC flag beforehand and
1475 * calling inode_sync_complete() to clear it afterwards.
1478 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1480 struct address_space *mapping = inode->i_mapping;
1481 long nr_to_write = wbc->nr_to_write;
1485 WARN_ON(!(inode->i_state & I_SYNC));
1487 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1489 ret = do_writepages(mapping, wbc);
1492 * Make sure to wait on the data before writing out the metadata.
1493 * This is important for filesystems that modify metadata on data
1494 * I/O completion. We don't do it for sync(2) writeback because it has a
1495 * separate, external IO completion path and ->sync_fs for guaranteeing
1496 * inode metadata is written back correctly.
1498 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1499 int err = filemap_fdatawait(mapping);
1505 * If the inode has dirty timestamps and we need to write them, call
1506 * mark_inode_dirty_sync() to notify the filesystem about it and to
1507 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1509 if ((inode->i_state & I_DIRTY_TIME) &&
1510 (wbc->sync_mode == WB_SYNC_ALL ||
1511 time_after(jiffies, inode->dirtied_time_when +
1512 dirtytime_expire_interval * HZ))) {
1513 trace_writeback_lazytime(inode);
1514 mark_inode_dirty_sync(inode);
1518 * Get and clear the dirty flags from i_state. This needs to be done
1519 * after calling writepages because some filesystems may redirty the
1520 * inode during writepages due to delalloc. It also needs to be done
1521 * after handling timestamp expiration, as that may dirty the inode too.
1523 spin_lock(&inode->i_lock);
1524 dirty = inode->i_state & I_DIRTY;
1525 inode->i_state &= ~dirty;
1528 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1529 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1530 * either they see the I_DIRTY bits cleared or we see the dirtied
1533 * I_DIRTY_PAGES is always cleared together above even if @mapping
1534 * still has dirty pages. The flag is reinstated after smp_mb() if
1535 * necessary. This guarantees that either __mark_inode_dirty()
1536 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1540 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1541 inode->i_state |= I_DIRTY_PAGES;
1543 spin_unlock(&inode->i_lock);
1545 /* Don't write the inode if only I_DIRTY_PAGES was set */
1546 if (dirty & ~I_DIRTY_PAGES) {
1547 int err = write_inode(inode, wbc);
1551 trace_writeback_single_inode(inode, wbc, nr_to_write);
1556 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1557 * the regular batched writeback done by the flusher threads in
1558 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1559 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1561 * To prevent the inode from going away, either the caller must have a reference
1562 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1564 static int writeback_single_inode(struct inode *inode,
1565 struct writeback_control *wbc)
1567 struct bdi_writeback *wb;
1570 spin_lock(&inode->i_lock);
1571 if (!atomic_read(&inode->i_count))
1572 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1574 WARN_ON(inode->i_state & I_WILL_FREE);
1576 if (inode->i_state & I_SYNC) {
1578 * Writeback is already running on the inode. For WB_SYNC_NONE,
1579 * that's enough and we can just return. For WB_SYNC_ALL, we
1580 * must wait for the existing writeback to complete, then do
1581 * writeback again if there's anything left.
1583 if (wbc->sync_mode != WB_SYNC_ALL)
1585 __inode_wait_for_writeback(inode);
1587 WARN_ON(inode->i_state & I_SYNC);
1589 * If the inode is already fully clean, then there's nothing to do.
1591 * For data-integrity syncs we also need to check whether any pages are
1592 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1593 * there are any such pages, we'll need to wait for them.
1595 if (!(inode->i_state & I_DIRTY_ALL) &&
1596 (wbc->sync_mode != WB_SYNC_ALL ||
1597 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1599 inode->i_state |= I_SYNC;
1600 wbc_attach_and_unlock_inode(wbc, inode);
1602 ret = __writeback_single_inode(inode, wbc);
1604 wbc_detach_inode(wbc);
1606 wb = inode_to_wb_and_lock_list(inode);
1607 spin_lock(&inode->i_lock);
1609 * If the inode is now fully clean, then it can be safely removed from
1610 * its writeback list (if any). Otherwise the flusher threads are
1611 * responsible for the writeback lists.
1613 if (!(inode->i_state & I_DIRTY_ALL))
1614 inode_cgwb_move_to_attached(inode, wb);
1615 spin_unlock(&wb->list_lock);
1616 inode_sync_complete(inode);
1618 spin_unlock(&inode->i_lock);
1622 static long writeback_chunk_size(struct bdi_writeback *wb,
1623 struct wb_writeback_work *work)
1628 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1629 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1630 * here avoids calling into writeback_inodes_wb() more than once.
1632 * The intended call sequence for WB_SYNC_ALL writeback is:
1635 * writeback_sb_inodes() <== called only once
1636 * write_cache_pages() <== called once for each inode
1637 * (quickly) tag currently dirty pages
1638 * (maybe slowly) sync all tagged pages
1640 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1643 pages = min(wb->avg_write_bandwidth / 2,
1644 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1645 pages = min(pages, work->nr_pages);
1646 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1647 MIN_WRITEBACK_PAGES);
1654 * Write a portion of b_io inodes which belong to @sb.
1656 * Return the number of pages and/or inodes written.
1658 * NOTE! This is called with wb->list_lock held, and will
1659 * unlock and relock that for each inode it ends up doing
1662 static long writeback_sb_inodes(struct super_block *sb,
1663 struct bdi_writeback *wb,
1664 struct wb_writeback_work *work)
1666 struct writeback_control wbc = {
1667 .sync_mode = work->sync_mode,
1668 .tagged_writepages = work->tagged_writepages,
1669 .for_kupdate = work->for_kupdate,
1670 .for_background = work->for_background,
1671 .for_sync = work->for_sync,
1672 .range_cyclic = work->range_cyclic,
1674 .range_end = LLONG_MAX,
1676 unsigned long start_time = jiffies;
1678 long wrote = 0; /* count both pages and inodes */
1680 while (!list_empty(&wb->b_io)) {
1681 struct inode *inode = wb_inode(wb->b_io.prev);
1682 struct bdi_writeback *tmp_wb;
1684 if (inode->i_sb != sb) {
1687 * We only want to write back data for this
1688 * superblock, move all inodes not belonging
1689 * to it back onto the dirty list.
1691 redirty_tail(inode, wb);
1696 * The inode belongs to a different superblock.
1697 * Bounce back to the caller to unpin this and
1698 * pin the next superblock.
1704 * Don't bother with new inodes or inodes being freed, first
1705 * kind does not need periodic writeout yet, and for the latter
1706 * kind writeout is handled by the freer.
1708 spin_lock(&inode->i_lock);
1709 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1710 redirty_tail_locked(inode, wb);
1711 spin_unlock(&inode->i_lock);
1714 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1716 * If this inode is locked for writeback and we are not
1717 * doing writeback-for-data-integrity, move it to
1718 * b_more_io so that writeback can proceed with the
1719 * other inodes on s_io.
1721 * We'll have another go at writing back this inode
1722 * when we completed a full scan of b_io.
1724 spin_unlock(&inode->i_lock);
1725 requeue_io(inode, wb);
1726 trace_writeback_sb_inodes_requeue(inode);
1729 spin_unlock(&wb->list_lock);
1732 * We already requeued the inode if it had I_SYNC set and we
1733 * are doing WB_SYNC_NONE writeback. So this catches only the
1736 if (inode->i_state & I_SYNC) {
1737 /* Wait for I_SYNC. This function drops i_lock... */
1738 inode_sleep_on_writeback(inode);
1739 /* Inode may be gone, start again */
1740 spin_lock(&wb->list_lock);
1743 inode->i_state |= I_SYNC;
1744 wbc_attach_and_unlock_inode(&wbc, inode);
1746 write_chunk = writeback_chunk_size(wb, work);
1747 wbc.nr_to_write = write_chunk;
1748 wbc.pages_skipped = 0;
1751 * We use I_SYNC to pin the inode in memory. While it is set
1752 * evict_inode() will wait so the inode cannot be freed.
1754 __writeback_single_inode(inode, &wbc);
1756 wbc_detach_inode(&wbc);
1757 work->nr_pages -= write_chunk - wbc.nr_to_write;
1758 wrote += write_chunk - wbc.nr_to_write;
1760 if (need_resched()) {
1762 * We're trying to balance between building up a nice
1763 * long list of IOs to improve our merge rate, and
1764 * getting those IOs out quickly for anyone throttling
1765 * in balance_dirty_pages(). cond_resched() doesn't
1766 * unplug, so get our IOs out the door before we
1769 blk_flush_plug(current);
1774 * Requeue @inode if still dirty. Be careful as @inode may
1775 * have been switched to another wb in the meantime.
1777 tmp_wb = inode_to_wb_and_lock_list(inode);
1778 spin_lock(&inode->i_lock);
1779 if (!(inode->i_state & I_DIRTY_ALL))
1781 requeue_inode(inode, tmp_wb, &wbc);
1782 inode_sync_complete(inode);
1783 spin_unlock(&inode->i_lock);
1785 if (unlikely(tmp_wb != wb)) {
1786 spin_unlock(&tmp_wb->list_lock);
1787 spin_lock(&wb->list_lock);
1791 * bail out to wb_writeback() often enough to check
1792 * background threshold and other termination conditions.
1795 if (time_is_before_jiffies(start_time + HZ / 10UL))
1797 if (work->nr_pages <= 0)
1804 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1805 struct wb_writeback_work *work)
1807 unsigned long start_time = jiffies;
1810 while (!list_empty(&wb->b_io)) {
1811 struct inode *inode = wb_inode(wb->b_io.prev);
1812 struct super_block *sb = inode->i_sb;
1814 if (!trylock_super(sb)) {
1816 * trylock_super() may fail consistently due to
1817 * s_umount being grabbed by someone else. Don't use
1818 * requeue_io() to avoid busy retrying the inode/sb.
1820 redirty_tail(inode, wb);
1823 wrote += writeback_sb_inodes(sb, wb, work);
1824 up_read(&sb->s_umount);
1826 /* refer to the same tests at the end of writeback_sb_inodes */
1828 if (time_is_before_jiffies(start_time + HZ / 10UL))
1830 if (work->nr_pages <= 0)
1834 /* Leave any unwritten inodes on b_io */
1838 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1839 enum wb_reason reason)
1841 struct wb_writeback_work work = {
1842 .nr_pages = nr_pages,
1843 .sync_mode = WB_SYNC_NONE,
1847 struct blk_plug plug;
1849 blk_start_plug(&plug);
1850 spin_lock(&wb->list_lock);
1851 if (list_empty(&wb->b_io))
1852 queue_io(wb, &work, jiffies);
1853 __writeback_inodes_wb(wb, &work);
1854 spin_unlock(&wb->list_lock);
1855 blk_finish_plug(&plug);
1857 return nr_pages - work.nr_pages;
1861 * Explicit flushing or periodic writeback of "old" data.
1863 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1864 * dirtying-time in the inode's address_space. So this periodic writeback code
1865 * just walks the superblock inode list, writing back any inodes which are
1866 * older than a specific point in time.
1868 * Try to run once per dirty_writeback_interval. But if a writeback event
1869 * takes longer than a dirty_writeback_interval interval, then leave a
1872 * dirtied_before takes precedence over nr_to_write. So we'll only write back
1873 * all dirty pages if they are all attached to "old" mappings.
1875 static long wb_writeback(struct bdi_writeback *wb,
1876 struct wb_writeback_work *work)
1878 unsigned long wb_start = jiffies;
1879 long nr_pages = work->nr_pages;
1880 unsigned long dirtied_before = jiffies;
1881 struct inode *inode;
1883 struct blk_plug plug;
1885 blk_start_plug(&plug);
1886 spin_lock(&wb->list_lock);
1889 * Stop writeback when nr_pages has been consumed
1891 if (work->nr_pages <= 0)
1895 * Background writeout and kupdate-style writeback may
1896 * run forever. Stop them if there is other work to do
1897 * so that e.g. sync can proceed. They'll be restarted
1898 * after the other works are all done.
1900 if ((work->for_background || work->for_kupdate) &&
1901 !list_empty(&wb->work_list))
1905 * For background writeout, stop when we are below the
1906 * background dirty threshold
1908 if (work->for_background && !wb_over_bg_thresh(wb))
1912 * Kupdate and background works are special and we want to
1913 * include all inodes that need writing. Livelock avoidance is
1914 * handled by these works yielding to any other work so we are
1917 if (work->for_kupdate) {
1918 dirtied_before = jiffies -
1919 msecs_to_jiffies(dirty_expire_interval * 10);
1920 } else if (work->for_background)
1921 dirtied_before = jiffies;
1923 trace_writeback_start(wb, work);
1924 if (list_empty(&wb->b_io))
1925 queue_io(wb, work, dirtied_before);
1927 progress = writeback_sb_inodes(work->sb, wb, work);
1929 progress = __writeback_inodes_wb(wb, work);
1930 trace_writeback_written(wb, work);
1932 wb_update_bandwidth(wb, wb_start);
1935 * Did we write something? Try for more
1937 * Dirty inodes are moved to b_io for writeback in batches.
1938 * The completion of the current batch does not necessarily
1939 * mean the overall work is done. So we keep looping as long
1940 * as made some progress on cleaning pages or inodes.
1945 * No more inodes for IO, bail
1947 if (list_empty(&wb->b_more_io))
1950 * Nothing written. Wait for some inode to
1951 * become available for writeback. Otherwise
1952 * we'll just busyloop.
1954 trace_writeback_wait(wb, work);
1955 inode = wb_inode(wb->b_more_io.prev);
1956 spin_lock(&inode->i_lock);
1957 spin_unlock(&wb->list_lock);
1958 /* This function drops i_lock... */
1959 inode_sleep_on_writeback(inode);
1960 spin_lock(&wb->list_lock);
1962 spin_unlock(&wb->list_lock);
1963 blk_finish_plug(&plug);
1965 return nr_pages - work->nr_pages;
1969 * Return the next wb_writeback_work struct that hasn't been processed yet.
1971 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1973 struct wb_writeback_work *work = NULL;
1975 spin_lock_bh(&wb->work_lock);
1976 if (!list_empty(&wb->work_list)) {
1977 work = list_entry(wb->work_list.next,
1978 struct wb_writeback_work, list);
1979 list_del_init(&work->list);
1981 spin_unlock_bh(&wb->work_lock);
1985 static long wb_check_background_flush(struct bdi_writeback *wb)
1987 if (wb_over_bg_thresh(wb)) {
1989 struct wb_writeback_work work = {
1990 .nr_pages = LONG_MAX,
1991 .sync_mode = WB_SYNC_NONE,
1992 .for_background = 1,
1994 .reason = WB_REASON_BACKGROUND,
1997 return wb_writeback(wb, &work);
2003 static long wb_check_old_data_flush(struct bdi_writeback *wb)
2005 unsigned long expired;
2009 * When set to zero, disable periodic writeback
2011 if (!dirty_writeback_interval)
2014 expired = wb->last_old_flush +
2015 msecs_to_jiffies(dirty_writeback_interval * 10);
2016 if (time_before(jiffies, expired))
2019 wb->last_old_flush = jiffies;
2020 nr_pages = get_nr_dirty_pages();
2023 struct wb_writeback_work work = {
2024 .nr_pages = nr_pages,
2025 .sync_mode = WB_SYNC_NONE,
2028 .reason = WB_REASON_PERIODIC,
2031 return wb_writeback(wb, &work);
2037 static long wb_check_start_all(struct bdi_writeback *wb)
2041 if (!test_bit(WB_start_all, &wb->state))
2044 nr_pages = get_nr_dirty_pages();
2046 struct wb_writeback_work work = {
2047 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2048 .sync_mode = WB_SYNC_NONE,
2050 .reason = wb->start_all_reason,
2053 nr_pages = wb_writeback(wb, &work);
2056 clear_bit(WB_start_all, &wb->state);
2062 * Retrieve work items and do the writeback they describe
2064 static long wb_do_writeback(struct bdi_writeback *wb)
2066 struct wb_writeback_work *work;
2069 set_bit(WB_writeback_running, &wb->state);
2070 while ((work = get_next_work_item(wb)) != NULL) {
2071 trace_writeback_exec(wb, work);
2072 wrote += wb_writeback(wb, work);
2073 finish_writeback_work(wb, work);
2077 * Check for a flush-everything request
2079 wrote += wb_check_start_all(wb);
2082 * Check for periodic writeback, kupdated() style
2084 wrote += wb_check_old_data_flush(wb);
2085 wrote += wb_check_background_flush(wb);
2086 clear_bit(WB_writeback_running, &wb->state);
2092 * Handle writeback of dirty data for the device backed by this bdi. Also
2093 * reschedules periodically and does kupdated style flushing.
2095 void wb_workfn(struct work_struct *work)
2097 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2098 struct bdi_writeback, dwork);
2101 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2102 current->flags |= PF_SWAPWRITE;
2104 if (likely(!current_is_workqueue_rescuer() ||
2105 !test_bit(WB_registered, &wb->state))) {
2107 * The normal path. Keep writing back @wb until its
2108 * work_list is empty. Note that this path is also taken
2109 * if @wb is shutting down even when we're running off the
2110 * rescuer as work_list needs to be drained.
2113 pages_written = wb_do_writeback(wb);
2114 trace_writeback_pages_written(pages_written);
2115 } while (!list_empty(&wb->work_list));
2118 * bdi_wq can't get enough workers and we're running off
2119 * the emergency worker. Don't hog it. Hopefully, 1024 is
2120 * enough for efficient IO.
2122 pages_written = writeback_inodes_wb(wb, 1024,
2123 WB_REASON_FORKER_THREAD);
2124 trace_writeback_pages_written(pages_written);
2127 if (!list_empty(&wb->work_list))
2129 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2130 wb_wakeup_delayed(wb);
2132 current->flags &= ~PF_SWAPWRITE;
2136 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2137 * write back the whole world.
2139 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2140 enum wb_reason reason)
2142 struct bdi_writeback *wb;
2144 if (!bdi_has_dirty_io(bdi))
2147 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2148 wb_start_writeback(wb, reason);
2151 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2152 enum wb_reason reason)
2155 __wakeup_flusher_threads_bdi(bdi, reason);
2160 * Wakeup the flusher threads to start writeback of all currently dirty pages
2162 void wakeup_flusher_threads(enum wb_reason reason)
2164 struct backing_dev_info *bdi;
2167 * If we are expecting writeback progress we must submit plugged IO.
2169 if (blk_needs_flush_plug(current))
2170 blk_schedule_flush_plug(current);
2173 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2174 __wakeup_flusher_threads_bdi(bdi, reason);
2179 * Wake up bdi's periodically to make sure dirtytime inodes gets
2180 * written back periodically. We deliberately do *not* check the
2181 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2182 * kernel to be constantly waking up once there are any dirtytime
2183 * inodes on the system. So instead we define a separate delayed work
2184 * function which gets called much more rarely. (By default, only
2185 * once every 12 hours.)
2187 * If there is any other write activity going on in the file system,
2188 * this function won't be necessary. But if the only thing that has
2189 * happened on the file system is a dirtytime inode caused by an atime
2190 * update, we need this infrastructure below to make sure that inode
2191 * eventually gets pushed out to disk.
2193 static void wakeup_dirtytime_writeback(struct work_struct *w);
2194 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2196 static void wakeup_dirtytime_writeback(struct work_struct *w)
2198 struct backing_dev_info *bdi;
2201 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2202 struct bdi_writeback *wb;
2204 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2205 if (!list_empty(&wb->b_dirty_time))
2209 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2212 static int __init start_dirtytime_writeback(void)
2214 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2217 __initcall(start_dirtytime_writeback);
2219 int dirtytime_interval_handler(struct ctl_table *table, int write,
2220 void *buffer, size_t *lenp, loff_t *ppos)
2224 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2225 if (ret == 0 && write)
2226 mod_delayed_work(system_wq, &dirtytime_work, 0);
2230 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2232 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2233 struct dentry *dentry;
2234 const char *name = "?";
2236 dentry = d_find_alias(inode);
2238 spin_lock(&dentry->d_lock);
2239 name = (const char *) dentry->d_name.name;
2242 "%s(%d): dirtied inode %lu (%s) on %s\n",
2243 current->comm, task_pid_nr(current), inode->i_ino,
2244 name, inode->i_sb->s_id);
2246 spin_unlock(&dentry->d_lock);
2253 * __mark_inode_dirty - internal function to mark an inode dirty
2255 * @inode: inode to mark
2256 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2257 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2258 * with I_DIRTY_PAGES.
2260 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2261 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2263 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2264 * instead of calling this directly.
2266 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2267 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2268 * even if they are later hashed, as they will have been marked dirty already.
2270 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2272 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2273 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2274 * the kernel-internal blockdev inode represents the dirtying time of the
2275 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2276 * page->mapping->host, so the page-dirtying time is recorded in the internal
2279 void __mark_inode_dirty(struct inode *inode, int flags)
2281 struct super_block *sb = inode->i_sb;
2284 trace_writeback_mark_inode_dirty(inode, flags);
2286 if (flags & I_DIRTY_INODE) {
2288 * Notify the filesystem about the inode being dirtied, so that
2289 * (if needed) it can update on-disk fields and journal the
2290 * inode. This is only needed when the inode itself is being
2291 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2292 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2294 trace_writeback_dirty_inode_start(inode, flags);
2295 if (sb->s_op->dirty_inode)
2296 sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE);
2297 trace_writeback_dirty_inode(inode, flags);
2299 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2300 flags &= ~I_DIRTY_TIME;
2303 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2304 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2305 * in one call to __mark_inode_dirty().)
2307 dirtytime = flags & I_DIRTY_TIME;
2308 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2312 * Paired with smp_mb() in __writeback_single_inode() for the
2313 * following lockless i_state test. See there for details.
2317 if (((inode->i_state & flags) == flags) ||
2318 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2321 if (unlikely(block_dump))
2322 block_dump___mark_inode_dirty(inode);
2324 spin_lock(&inode->i_lock);
2325 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2326 goto out_unlock_inode;
2327 if ((inode->i_state & flags) != flags) {
2328 const int was_dirty = inode->i_state & I_DIRTY;
2330 inode_attach_wb(inode, NULL);
2332 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2333 if (flags & I_DIRTY_INODE)
2334 inode->i_state &= ~I_DIRTY_TIME;
2335 inode->i_state |= flags;
2338 * If the inode is queued for writeback by flush worker, just
2339 * update its dirty state. Once the flush worker is done with
2340 * the inode it will place it on the appropriate superblock
2341 * list, based upon its state.
2343 if (inode->i_state & I_SYNC_QUEUED)
2344 goto out_unlock_inode;
2347 * Only add valid (hashed) inodes to the superblock's
2348 * dirty list. Add blockdev inodes as well.
2350 if (!S_ISBLK(inode->i_mode)) {
2351 if (inode_unhashed(inode))
2352 goto out_unlock_inode;
2354 if (inode->i_state & I_FREEING)
2355 goto out_unlock_inode;
2358 * If the inode was already on b_dirty/b_io/b_more_io, don't
2359 * reposition it (that would break b_dirty time-ordering).
2362 struct bdi_writeback *wb;
2363 struct list_head *dirty_list;
2364 bool wakeup_bdi = false;
2366 wb = locked_inode_to_wb_and_lock_list(inode);
2368 inode->dirtied_when = jiffies;
2370 inode->dirtied_time_when = jiffies;
2372 if (inode->i_state & I_DIRTY)
2373 dirty_list = &wb->b_dirty;
2375 dirty_list = &wb->b_dirty_time;
2377 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2380 spin_unlock(&wb->list_lock);
2381 trace_writeback_dirty_inode_enqueue(inode);
2384 * If this is the first dirty inode for this bdi,
2385 * we have to wake-up the corresponding bdi thread
2386 * to make sure background write-back happens
2390 (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2391 wb_wakeup_delayed(wb);
2396 spin_unlock(&inode->i_lock);
2398 EXPORT_SYMBOL(__mark_inode_dirty);
2401 * The @s_sync_lock is used to serialise concurrent sync operations
2402 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2403 * Concurrent callers will block on the s_sync_lock rather than doing contending
2404 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2405 * has been issued up to the time this function is enter is guaranteed to be
2406 * completed by the time we have gained the lock and waited for all IO that is
2407 * in progress regardless of the order callers are granted the lock.
2409 static void wait_sb_inodes(struct super_block *sb)
2411 LIST_HEAD(sync_list);
2414 * We need to be protected against the filesystem going from
2415 * r/o to r/w or vice versa.
2417 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2419 mutex_lock(&sb->s_sync_lock);
2422 * Splice the writeback list onto a temporary list to avoid waiting on
2423 * inodes that have started writeback after this point.
2425 * Use rcu_read_lock() to keep the inodes around until we have a
2426 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2427 * the local list because inodes can be dropped from either by writeback
2431 spin_lock_irq(&sb->s_inode_wblist_lock);
2432 list_splice_init(&sb->s_inodes_wb, &sync_list);
2435 * Data integrity sync. Must wait for all pages under writeback, because
2436 * there may have been pages dirtied before our sync call, but which had
2437 * writeout started before we write it out. In which case, the inode
2438 * may not be on the dirty list, but we still have to wait for that
2441 while (!list_empty(&sync_list)) {
2442 struct inode *inode = list_first_entry(&sync_list, struct inode,
2444 struct address_space *mapping = inode->i_mapping;
2447 * Move each inode back to the wb list before we drop the lock
2448 * to preserve consistency between i_wb_list and the mapping
2449 * writeback tag. Writeback completion is responsible to remove
2450 * the inode from either list once the writeback tag is cleared.
2452 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2455 * The mapping can appear untagged while still on-list since we
2456 * do not have the mapping lock. Skip it here, wb completion
2459 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2462 spin_unlock_irq(&sb->s_inode_wblist_lock);
2464 spin_lock(&inode->i_lock);
2465 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2466 spin_unlock(&inode->i_lock);
2468 spin_lock_irq(&sb->s_inode_wblist_lock);
2472 spin_unlock(&inode->i_lock);
2476 * We keep the error status of individual mapping so that
2477 * applications can catch the writeback error using fsync(2).
2478 * See filemap_fdatawait_keep_errors() for details.
2480 filemap_fdatawait_keep_errors(mapping);
2487 spin_lock_irq(&sb->s_inode_wblist_lock);
2489 spin_unlock_irq(&sb->s_inode_wblist_lock);
2491 mutex_unlock(&sb->s_sync_lock);
2494 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2495 enum wb_reason reason, bool skip_if_busy)
2497 struct backing_dev_info *bdi = sb->s_bdi;
2498 DEFINE_WB_COMPLETION(done, bdi);
2499 struct wb_writeback_work work = {
2501 .sync_mode = WB_SYNC_NONE,
2502 .tagged_writepages = 1,
2508 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2510 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2512 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2513 wb_wait_for_completion(&done);
2517 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2518 * @sb: the superblock
2519 * @nr: the number of pages to write
2520 * @reason: reason why some writeback work initiated
2522 * Start writeback on some inodes on this super_block. No guarantees are made
2523 * on how many (if any) will be written, and this function does not wait
2524 * for IO completion of submitted IO.
2526 void writeback_inodes_sb_nr(struct super_block *sb,
2528 enum wb_reason reason)
2530 __writeback_inodes_sb_nr(sb, nr, reason, false);
2532 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2535 * writeback_inodes_sb - writeback dirty inodes from given super_block
2536 * @sb: the superblock
2537 * @reason: reason why some writeback work was initiated
2539 * Start writeback on some inodes on this super_block. No guarantees are made
2540 * on how many (if any) will be written, and this function does not wait
2541 * for IO completion of submitted IO.
2543 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2545 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2547 EXPORT_SYMBOL(writeback_inodes_sb);
2550 * try_to_writeback_inodes_sb - try to start writeback if none underway
2551 * @sb: the superblock
2552 * @reason: reason why some writeback work was initiated
2554 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2556 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2558 if (!down_read_trylock(&sb->s_umount))
2561 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2562 up_read(&sb->s_umount);
2564 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2567 * sync_inodes_sb - sync sb inode pages
2568 * @sb: the superblock
2570 * This function writes and waits on any dirty inode belonging to this
2573 void sync_inodes_sb(struct super_block *sb)
2575 struct backing_dev_info *bdi = sb->s_bdi;
2576 DEFINE_WB_COMPLETION(done, bdi);
2577 struct wb_writeback_work work = {
2579 .sync_mode = WB_SYNC_ALL,
2580 .nr_pages = LONG_MAX,
2583 .reason = WB_REASON_SYNC,
2588 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2589 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2590 * bdi_has_dirty() need to be written out too.
2592 if (bdi == &noop_backing_dev_info)
2594 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2596 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2597 bdi_down_write_wb_switch_rwsem(bdi);
2598 bdi_split_work_to_wbs(bdi, &work, false);
2599 wb_wait_for_completion(&done);
2600 bdi_up_write_wb_switch_rwsem(bdi);
2604 EXPORT_SYMBOL(sync_inodes_sb);
2607 * write_inode_now - write an inode to disk
2608 * @inode: inode to write to disk
2609 * @sync: whether the write should be synchronous or not
2611 * This function commits an inode to disk immediately if it is dirty. This is
2612 * primarily needed by knfsd.
2614 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2616 int write_inode_now(struct inode *inode, int sync)
2618 struct writeback_control wbc = {
2619 .nr_to_write = LONG_MAX,
2620 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2622 .range_end = LLONG_MAX,
2625 if (!mapping_can_writeback(inode->i_mapping))
2626 wbc.nr_to_write = 0;
2629 return writeback_single_inode(inode, &wbc);
2631 EXPORT_SYMBOL(write_inode_now);
2634 * sync_inode - write an inode and its pages to disk.
2635 * @inode: the inode to sync
2636 * @wbc: controls the writeback mode
2638 * sync_inode() will write an inode and its pages to disk. It will also
2639 * correctly update the inode on its superblock's dirty inode lists and will
2640 * update inode->i_state.
2642 * The caller must have a ref on the inode.
2644 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2646 return writeback_single_inode(inode, wbc);
2648 EXPORT_SYMBOL(sync_inode);
2651 * sync_inode_metadata - write an inode to disk
2652 * @inode: the inode to sync
2653 * @wait: wait for I/O to complete.
2655 * Write an inode to disk and adjust its dirty state after completion.
2657 * Note: only writes the actual inode, no associated data or other metadata.
2659 int sync_inode_metadata(struct inode *inode, int wait)
2661 struct writeback_control wbc = {
2662 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2663 .nr_to_write = 0, /* metadata-only */
2666 return sync_inode(inode, &wbc);
2668 EXPORT_SYMBOL(sync_inode_metadata);