4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include <linux/device.h>
30 #include <linux/memcontrol.h>
34 * 4MB minimal write chunk size
36 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
38 struct wb_completion {
43 * Passed into wb_writeback(), essentially a subset of writeback_control
45 struct wb_writeback_work {
47 struct super_block *sb;
48 enum writeback_sync_modes sync_mode;
49 unsigned int tagged_writepages:1;
50 unsigned int for_kupdate:1;
51 unsigned int range_cyclic:1;
52 unsigned int for_background:1;
53 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
54 unsigned int auto_free:1; /* free on completion */
55 enum wb_reason reason; /* why was writeback initiated? */
57 struct list_head list; /* pending work list */
58 struct wb_completion *done; /* set if the caller waits */
62 * If one wants to wait for one or more wb_writeback_works, each work's
63 * ->done should be set to a wb_completion defined using the following
64 * macro. Once all work items are issued with wb_queue_work(), the caller
65 * can wait for the completion of all using wb_wait_for_completion(). Work
66 * items which are waited upon aren't freed automatically on completion.
68 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
69 struct wb_completion cmpl = { \
70 .cnt = ATOMIC_INIT(1), \
75 * If an inode is constantly having its pages dirtied, but then the
76 * updates stop dirtytime_expire_interval seconds in the past, it's
77 * possible for the worst case time between when an inode has its
78 * timestamps updated and when they finally get written out to be two
79 * dirtytime_expire_intervals. We set the default to 12 hours (in
80 * seconds), which means most of the time inodes will have their
81 * timestamps written to disk after 12 hours, but in the worst case a
82 * few inodes might not their timestamps updated for 24 hours.
84 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
86 static inline struct inode *wb_inode(struct list_head *head)
88 return list_entry(head, struct inode, i_io_list);
92 * Include the creation of the trace points after defining the
93 * wb_writeback_work structure and inline functions so that the definition
94 * remains local to this file.
96 #define CREATE_TRACE_POINTS
97 #include <trace/events/writeback.h>
99 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
101 static bool wb_io_lists_populated(struct bdi_writeback *wb)
103 if (wb_has_dirty_io(wb)) {
106 set_bit(WB_has_dirty_io, &wb->state);
107 WARN_ON_ONCE(!wb->avg_write_bandwidth);
108 atomic_long_add(wb->avg_write_bandwidth,
109 &wb->bdi->tot_write_bandwidth);
114 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
116 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
117 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
118 clear_bit(WB_has_dirty_io, &wb->state);
119 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
120 &wb->bdi->tot_write_bandwidth) < 0);
125 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
126 * @inode: inode to be moved
127 * @wb: target bdi_writeback
128 * @head: one of @wb->b_{dirty|io|more_io}
130 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
131 * Returns %true if @inode is the first occupant of the !dirty_time IO
132 * lists; otherwise, %false.
134 static bool inode_io_list_move_locked(struct inode *inode,
135 struct bdi_writeback *wb,
136 struct list_head *head)
138 assert_spin_locked(&wb->list_lock);
140 list_move(&inode->i_io_list, head);
142 /* dirty_time doesn't count as dirty_io until expiration */
143 if (head != &wb->b_dirty_time)
144 return wb_io_lists_populated(wb);
146 wb_io_lists_depopulated(wb);
151 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
152 * @inode: inode to be removed
153 * @wb: bdi_writeback @inode is being removed from
155 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
156 * clear %WB_has_dirty_io if all are empty afterwards.
158 static void inode_io_list_del_locked(struct inode *inode,
159 struct bdi_writeback *wb)
161 assert_spin_locked(&wb->list_lock);
162 assert_spin_locked(&inode->i_lock);
164 inode->i_state &= ~I_SYNC_QUEUED;
165 list_del_init(&inode->i_io_list);
166 wb_io_lists_depopulated(wb);
169 static void wb_wakeup(struct bdi_writeback *wb)
171 spin_lock_bh(&wb->work_lock);
172 if (test_bit(WB_registered, &wb->state))
173 mod_delayed_work(bdi_wq, &wb->dwork, 0);
174 spin_unlock_bh(&wb->work_lock);
177 static void finish_writeback_work(struct bdi_writeback *wb,
178 struct wb_writeback_work *work)
180 struct wb_completion *done = work->done;
184 if (done && atomic_dec_and_test(&done->cnt))
185 wake_up_all(&wb->bdi->wb_waitq);
188 static void wb_queue_work(struct bdi_writeback *wb,
189 struct wb_writeback_work *work)
191 trace_writeback_queue(wb, work);
194 atomic_inc(&work->done->cnt);
196 spin_lock_bh(&wb->work_lock);
198 if (test_bit(WB_registered, &wb->state)) {
199 list_add_tail(&work->list, &wb->work_list);
200 mod_delayed_work(bdi_wq, &wb->dwork, 0);
202 finish_writeback_work(wb, work);
204 spin_unlock_bh(&wb->work_lock);
208 * wb_wait_for_completion - wait for completion of bdi_writeback_works
209 * @bdi: bdi work items were issued to
210 * @done: target wb_completion
212 * Wait for one or more work items issued to @bdi with their ->done field
213 * set to @done, which should have been defined with
214 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
215 * work items are completed. Work items which are waited upon aren't freed
216 * automatically on completion.
218 static void wb_wait_for_completion(struct backing_dev_info *bdi,
219 struct wb_completion *done)
221 atomic_dec(&done->cnt); /* put down the initial count */
222 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
225 #ifdef CONFIG_CGROUP_WRITEBACK
227 /* parameters for foreign inode detection, see wb_detach_inode() */
228 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
229 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
230 #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
231 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
233 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
234 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
235 /* each slot's duration is 2s / 16 */
236 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
237 /* if foreign slots >= 8, switch */
238 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
239 /* one round can affect upto 5 slots */
241 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
242 static struct workqueue_struct *isw_wq;
244 void __inode_attach_wb(struct inode *inode, struct page *page)
246 struct backing_dev_info *bdi = inode_to_bdi(inode);
247 struct bdi_writeback *wb = NULL;
249 if (inode_cgwb_enabled(inode)) {
250 struct cgroup_subsys_state *memcg_css;
253 memcg_css = mem_cgroup_css_from_page(page);
254 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
256 /* must pin memcg_css, see wb_get_create() */
257 memcg_css = task_get_css(current, memory_cgrp_id);
258 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
267 * There may be multiple instances of this function racing to
268 * update the same inode. Use cmpxchg() to tell the winner.
270 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
273 EXPORT_SYMBOL_GPL(__inode_attach_wb);
276 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
277 * @inode: inode of interest with i_lock held
279 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
280 * held on entry and is released on return. The returned wb is guaranteed
281 * to stay @inode's associated wb until its list_lock is released.
283 static struct bdi_writeback *
284 locked_inode_to_wb_and_lock_list(struct inode *inode)
285 __releases(&inode->i_lock)
286 __acquires(&wb->list_lock)
289 struct bdi_writeback *wb = inode_to_wb(inode);
292 * inode_to_wb() association is protected by both
293 * @inode->i_lock and @wb->list_lock but list_lock nests
294 * outside i_lock. Drop i_lock and verify that the
295 * association hasn't changed after acquiring list_lock.
298 spin_unlock(&inode->i_lock);
299 spin_lock(&wb->list_lock);
301 /* i_wb may have changed inbetween, can't use inode_to_wb() */
302 if (likely(wb == inode->i_wb)) {
303 wb_put(wb); /* @inode already has ref */
307 spin_unlock(&wb->list_lock);
310 spin_lock(&inode->i_lock);
315 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
316 * @inode: inode of interest
318 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
321 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
322 __acquires(&wb->list_lock)
324 spin_lock(&inode->i_lock);
325 return locked_inode_to_wb_and_lock_list(inode);
328 struct inode_switch_wbs_context {
330 struct bdi_writeback *new_wb;
332 struct rcu_head rcu_head;
333 struct work_struct work;
336 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
338 down_write(&bdi->wb_switch_rwsem);
341 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
343 up_write(&bdi->wb_switch_rwsem);
346 static void inode_switch_wbs_work_fn(struct work_struct *work)
348 struct inode_switch_wbs_context *isw =
349 container_of(work, struct inode_switch_wbs_context, work);
350 struct inode *inode = isw->inode;
351 struct backing_dev_info *bdi = inode_to_bdi(inode);
352 struct address_space *mapping = inode->i_mapping;
353 struct bdi_writeback *old_wb = inode->i_wb;
354 struct bdi_writeback *new_wb = isw->new_wb;
355 struct radix_tree_iter iter;
356 bool switched = false;
360 * If @inode switches cgwb membership while sync_inodes_sb() is
361 * being issued, sync_inodes_sb() might miss it. Synchronize.
363 down_read(&bdi->wb_switch_rwsem);
366 * By the time control reaches here, RCU grace period has passed
367 * since I_WB_SWITCH assertion and all wb stat update transactions
368 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
369 * synchronizing against mapping->tree_lock.
371 * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
372 * gives us exclusion against all wb related operations on @inode
373 * including IO list manipulations and stat updates.
375 if (old_wb < new_wb) {
376 spin_lock(&old_wb->list_lock);
377 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
379 spin_lock(&new_wb->list_lock);
380 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
382 spin_lock(&inode->i_lock);
383 spin_lock_irq(&mapping->tree_lock);
386 * Once I_FREEING is visible under i_lock, the eviction path owns
387 * the inode and we shouldn't modify ->i_io_list.
389 if (unlikely(inode->i_state & I_FREEING))
393 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
394 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
395 * pages actually under underwriteback.
397 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
398 PAGECACHE_TAG_DIRTY) {
399 struct page *page = radix_tree_deref_slot_protected(slot,
400 &mapping->tree_lock);
401 if (likely(page) && PageDirty(page)) {
402 __dec_wb_stat(old_wb, WB_RECLAIMABLE);
403 __inc_wb_stat(new_wb, WB_RECLAIMABLE);
407 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
408 PAGECACHE_TAG_WRITEBACK) {
409 struct page *page = radix_tree_deref_slot_protected(slot,
410 &mapping->tree_lock);
412 WARN_ON_ONCE(!PageWriteback(page));
413 __dec_wb_stat(old_wb, WB_WRITEBACK);
414 __inc_wb_stat(new_wb, WB_WRITEBACK);
421 * Transfer to @new_wb's IO list if necessary. The specific list
422 * @inode was on is ignored and the inode is put on ->b_dirty which
423 * is always correct including from ->b_dirty_time. The transfer
424 * preserves @inode->dirtied_when ordering.
426 if (!list_empty(&inode->i_io_list)) {
429 inode_io_list_del_locked(inode, old_wb);
430 inode->i_wb = new_wb;
431 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
432 if (time_after_eq(inode->dirtied_when,
435 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
437 inode->i_wb = new_wb;
440 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
441 inode->i_wb_frn_winner = 0;
442 inode->i_wb_frn_avg_time = 0;
443 inode->i_wb_frn_history = 0;
447 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
448 * ensures that the new wb is visible if they see !I_WB_SWITCH.
450 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
452 spin_unlock_irq(&mapping->tree_lock);
453 spin_unlock(&inode->i_lock);
454 spin_unlock(&new_wb->list_lock);
455 spin_unlock(&old_wb->list_lock);
457 up_read(&bdi->wb_switch_rwsem);
468 atomic_dec(&isw_nr_in_flight);
471 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
473 struct inode_switch_wbs_context *isw = container_of(rcu_head,
474 struct inode_switch_wbs_context, rcu_head);
476 /* needs to grab bh-unsafe locks, bounce to work item */
477 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
478 queue_work(isw_wq, &isw->work);
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)
500 * Avoid starting new switches while sync_inodes_sb() is in
501 * progress. Otherwise, if the down_write protected issue path
502 * blocks heavily, we might end up starting a large number of
503 * switches which will block on the rwsem.
505 if (!down_read_trylock(&bdi->wb_switch_rwsem))
508 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
512 /* find and pin the new wb */
514 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
515 if (memcg_css && !css_tryget(memcg_css))
521 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
526 /* while holding I_WB_SWITCH, no one else can update the association */
527 spin_lock(&inode->i_lock);
528 if (!(inode->i_sb->s_flags & MS_ACTIVE) ||
529 inode->i_state & (I_WB_SWITCH | I_FREEING) ||
530 inode_to_wb(inode) == isw->new_wb) {
531 spin_unlock(&inode->i_lock);
534 inode->i_state |= I_WB_SWITCH;
536 spin_unlock(&inode->i_lock);
541 * In addition to synchronizing among switchers, I_WB_SWITCH tells
542 * the RCU protected stat update paths to grab the mapping's
543 * tree_lock so that stat transfer can synchronize against them.
544 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
546 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
548 atomic_inc(&isw_nr_in_flight);
557 up_read(&bdi->wb_switch_rwsem);
561 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
562 * @wbc: writeback_control of interest
563 * @inode: target inode
565 * @inode is locked and about to be written back under the control of @wbc.
566 * Record @inode's writeback context into @wbc and unlock the i_lock. On
567 * writeback completion, wbc_detach_inode() should be called. This is used
568 * to track the cgroup writeback context.
570 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
573 if (!inode_cgwb_enabled(inode)) {
574 spin_unlock(&inode->i_lock);
578 wbc->wb = inode_to_wb(inode);
581 wbc->wb_id = wbc->wb->memcg_css->id;
582 wbc->wb_lcand_id = inode->i_wb_frn_winner;
583 wbc->wb_tcand_id = 0;
585 wbc->wb_lcand_bytes = 0;
586 wbc->wb_tcand_bytes = 0;
589 spin_unlock(&inode->i_lock);
592 * A dying wb indicates that either the blkcg associated with the
593 * memcg changed or the associated memcg is dying. In the first
594 * case, a replacement wb should already be available and we should
595 * refresh the wb immediately. In the second case, trying to
596 * refresh will keep failing.
598 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
599 inode_switch_wbs(inode, wbc->wb_id);
603 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
604 * @wbc: writeback_control of the just finished writeback
606 * To be called after a writeback attempt of an inode finishes and undoes
607 * wbc_attach_and_unlock_inode(). Can be called under any context.
609 * As concurrent write sharing of an inode is expected to be very rare and
610 * memcg only tracks page ownership on first-use basis severely confining
611 * the usefulness of such sharing, cgroup writeback tracks ownership
612 * per-inode. While the support for concurrent write sharing of an inode
613 * is deemed unnecessary, an inode being written to by different cgroups at
614 * different points in time is a lot more common, and, more importantly,
615 * charging only by first-use can too readily lead to grossly incorrect
616 * behaviors (single foreign page can lead to gigabytes of writeback to be
617 * incorrectly attributed).
619 * To resolve this issue, cgroup writeback detects the majority dirtier of
620 * an inode and transfers the ownership to it. To avoid unnnecessary
621 * oscillation, the detection mechanism keeps track of history and gives
622 * out the switch verdict only if the foreign usage pattern is stable over
623 * a certain amount of time and/or writeback attempts.
625 * On each writeback attempt, @wbc tries to detect the majority writer
626 * using Boyer-Moore majority vote algorithm. In addition to the byte
627 * count from the majority voting, it also counts the bytes written for the
628 * current wb and the last round's winner wb (max of last round's current
629 * wb, the winner from two rounds ago, and the last round's majority
630 * candidate). Keeping track of the historical winner helps the algorithm
631 * to semi-reliably detect the most active writer even when it's not the
634 * Once the winner of the round is determined, whether the winner is
635 * foreign or not and how much IO time the round consumed is recorded in
636 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
637 * over a certain threshold, the switch verdict is given.
639 void wbc_detach_inode(struct writeback_control *wbc)
641 struct bdi_writeback *wb = wbc->wb;
642 struct inode *inode = wbc->inode;
643 unsigned long avg_time, max_bytes, max_time;
650 history = inode->i_wb_frn_history;
651 avg_time = inode->i_wb_frn_avg_time;
653 /* pick the winner of this round */
654 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
655 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
657 max_bytes = wbc->wb_bytes;
658 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
659 max_id = wbc->wb_lcand_id;
660 max_bytes = wbc->wb_lcand_bytes;
662 max_id = wbc->wb_tcand_id;
663 max_bytes = wbc->wb_tcand_bytes;
667 * Calculate the amount of IO time the winner consumed and fold it
668 * into the running average kept per inode. If the consumed IO
669 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
670 * deciding whether to switch or not. This is to prevent one-off
671 * small dirtiers from skewing the verdict.
673 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
674 wb->avg_write_bandwidth);
676 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
677 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
679 avg_time = max_time; /* immediate catch up on first run */
681 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
685 * The switch verdict is reached if foreign wb's consume
686 * more than a certain proportion of IO time in a
687 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
688 * history mask where each bit represents one sixteenth of
689 * the period. Determine the number of slots to shift into
690 * history from @max_time.
692 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
693 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
695 if (wbc->wb_id != max_id)
696 history |= (1U << slots) - 1;
699 * Switch if the current wb isn't the consistent winner.
700 * If there are multiple closely competing dirtiers, the
701 * inode may switch across them repeatedly over time, which
702 * is okay. The main goal is avoiding keeping an inode on
703 * the wrong wb for an extended period of time.
705 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
706 inode_switch_wbs(inode, max_id);
710 * Multiple instances of this function may race to update the
711 * following fields but we don't mind occassional inaccuracies.
713 inode->i_wb_frn_winner = max_id;
714 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
715 inode->i_wb_frn_history = history;
722 * wbc_account_io - account IO issued during writeback
723 * @wbc: writeback_control of the writeback in progress
724 * @page: page being written out
725 * @bytes: number of bytes being written out
727 * @bytes from @page are about to written out during the writeback
728 * controlled by @wbc. Keep the book for foreign inode detection. See
729 * wbc_detach_inode().
731 void wbc_account_io(struct writeback_control *wbc, struct page *page,
734 struct cgroup_subsys_state *css;
738 * pageout() path doesn't attach @wbc to the inode being written
739 * out. This is intentional as we don't want the function to block
740 * behind a slow cgroup. Ultimately, we want pageout() to kick off
741 * regular writeback instead of writing things out itself.
746 css = mem_cgroup_css_from_page(page);
747 /* dead cgroups shouldn't contribute to inode ownership arbitration */
748 if (!(css->flags & CSS_ONLINE))
753 if (id == wbc->wb_id) {
754 wbc->wb_bytes += bytes;
758 if (id == wbc->wb_lcand_id)
759 wbc->wb_lcand_bytes += bytes;
761 /* Boyer-Moore majority vote algorithm */
762 if (!wbc->wb_tcand_bytes)
763 wbc->wb_tcand_id = id;
764 if (id == wbc->wb_tcand_id)
765 wbc->wb_tcand_bytes += bytes;
767 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
769 EXPORT_SYMBOL_GPL(wbc_account_io);
772 * inode_congested - test whether an inode is congested
773 * @inode: inode to test for congestion (may be NULL)
774 * @cong_bits: mask of WB_[a]sync_congested bits to test
776 * Tests whether @inode is congested. @cong_bits is the mask of congestion
777 * bits to test and the return value is the mask of set bits.
779 * If cgroup writeback is enabled for @inode, the congestion state is
780 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
781 * associated with @inode is congested; otherwise, the root wb's congestion
784 * @inode is allowed to be NULL as this function is often called on
785 * mapping->host which is NULL for the swapper space.
787 int inode_congested(struct inode *inode, int cong_bits)
790 * Once set, ->i_wb never becomes NULL while the inode is alive.
791 * Start transaction iff ->i_wb is visible.
793 if (inode && inode_to_wb_is_valid(inode)) {
794 struct bdi_writeback *wb;
795 struct wb_lock_cookie lock_cookie = {};
798 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
799 congested = wb_congested(wb, cong_bits);
800 unlocked_inode_to_wb_end(inode, &lock_cookie);
804 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
806 EXPORT_SYMBOL_GPL(inode_congested);
809 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
810 * @wb: target bdi_writeback to split @nr_pages to
811 * @nr_pages: number of pages to write for the whole bdi
813 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
814 * relation to the total write bandwidth of all wb's w/ dirty inodes on
817 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
819 unsigned long this_bw = wb->avg_write_bandwidth;
820 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
822 if (nr_pages == LONG_MAX)
826 * This may be called on clean wb's and proportional distribution
827 * may not make sense, just use the original @nr_pages in those
828 * cases. In general, we wanna err on the side of writing more.
830 if (!tot_bw || this_bw >= tot_bw)
833 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
837 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
838 * @bdi: target backing_dev_info
839 * @base_work: wb_writeback_work to issue
840 * @skip_if_busy: skip wb's which already have writeback in progress
842 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
843 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
844 * distributed to the busy wbs according to each wb's proportion in the
845 * total active write bandwidth of @bdi.
847 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
848 struct wb_writeback_work *base_work,
851 struct bdi_writeback *last_wb = NULL;
852 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
853 struct bdi_writeback, bdi_node);
858 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
859 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
860 struct wb_writeback_work fallback_work;
861 struct wb_writeback_work *work;
869 /* SYNC_ALL writes out I_DIRTY_TIME too */
870 if (!wb_has_dirty_io(wb) &&
871 (base_work->sync_mode == WB_SYNC_NONE ||
872 list_empty(&wb->b_dirty_time)))
874 if (skip_if_busy && writeback_in_progress(wb))
877 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
879 work = kmalloc(sizeof(*work), GFP_ATOMIC);
882 work->nr_pages = nr_pages;
884 wb_queue_work(wb, work);
888 /* alloc failed, execute synchronously using on-stack fallback */
889 work = &fallback_work;
891 work->nr_pages = nr_pages;
893 work->done = &fallback_work_done;
895 wb_queue_work(wb, work);
898 * Pin @wb so that it stays on @bdi->wb_list. This allows
899 * continuing iteration from @wb after dropping and
900 * regrabbing rcu read lock.
906 wb_wait_for_completion(bdi, &fallback_work_done);
916 * cgroup_writeback_umount - flush inode wb switches for umount
918 * This function is called when a super_block is about to be destroyed and
919 * flushes in-flight inode wb switches. An inode wb switch goes through
920 * RCU and then workqueue, so the two need to be flushed in order to ensure
921 * that all previously scheduled switches are finished. As wb switches are
922 * rare occurrences and synchronize_rcu() can take a while, perform
923 * flushing iff wb switches are in flight.
925 void cgroup_writeback_umount(void)
927 if (atomic_read(&isw_nr_in_flight)) {
929 * Use rcu_barrier() to wait for all pending callbacks to
930 * ensure that all in-flight wb switches are in the workqueue.
933 flush_workqueue(isw_wq);
937 static int __init cgroup_writeback_init(void)
939 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
944 fs_initcall(cgroup_writeback_init);
946 #else /* CONFIG_CGROUP_WRITEBACK */
948 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
949 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
951 static struct bdi_writeback *
952 locked_inode_to_wb_and_lock_list(struct inode *inode)
953 __releases(&inode->i_lock)
954 __acquires(&wb->list_lock)
956 struct bdi_writeback *wb = inode_to_wb(inode);
958 spin_unlock(&inode->i_lock);
959 spin_lock(&wb->list_lock);
963 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
964 __acquires(&wb->list_lock)
966 struct bdi_writeback *wb = inode_to_wb(inode);
968 spin_lock(&wb->list_lock);
972 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
977 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
978 struct wb_writeback_work *base_work,
983 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
984 base_work->auto_free = 0;
985 wb_queue_work(&bdi->wb, base_work);
989 #endif /* CONFIG_CGROUP_WRITEBACK */
991 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
992 bool range_cyclic, enum wb_reason reason)
994 struct wb_writeback_work *work;
996 if (!wb_has_dirty_io(wb))
1000 * This is WB_SYNC_NONE writeback, so if allocation fails just
1001 * wakeup the thread for old dirty data writeback
1003 work = kzalloc(sizeof(*work),
1004 GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
1006 trace_writeback_nowork(wb);
1011 work->sync_mode = WB_SYNC_NONE;
1012 work->nr_pages = nr_pages;
1013 work->range_cyclic = range_cyclic;
1014 work->reason = reason;
1015 work->auto_free = 1;
1017 wb_queue_work(wb, work);
1021 * wb_start_background_writeback - start background writeback
1022 * @wb: bdi_writback to write from
1025 * This makes sure WB_SYNC_NONE background writeback happens. When
1026 * this function returns, it is only guaranteed that for given wb
1027 * some IO is happening if we are over background dirty threshold.
1028 * Caller need not hold sb s_umount semaphore.
1030 void wb_start_background_writeback(struct bdi_writeback *wb)
1033 * We just wake up the flusher thread. It will perform background
1034 * writeback as soon as there is no other work to do.
1036 trace_writeback_wake_background(wb);
1041 * Remove the inode from the writeback list it is on.
1043 void inode_io_list_del(struct inode *inode)
1045 struct bdi_writeback *wb;
1047 wb = inode_to_wb_and_lock_list(inode);
1048 spin_lock(&inode->i_lock);
1049 inode_io_list_del_locked(inode, wb);
1050 spin_unlock(&inode->i_lock);
1051 spin_unlock(&wb->list_lock);
1055 * mark an inode as under writeback on the sb
1057 void sb_mark_inode_writeback(struct inode *inode)
1059 struct super_block *sb = inode->i_sb;
1060 unsigned long flags;
1062 if (list_empty(&inode->i_wb_list)) {
1063 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1064 if (list_empty(&inode->i_wb_list)) {
1065 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1066 trace_sb_mark_inode_writeback(inode);
1068 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1073 * clear an inode as under writeback on the sb
1075 void sb_clear_inode_writeback(struct inode *inode)
1077 struct super_block *sb = inode->i_sb;
1078 unsigned long flags;
1080 if (!list_empty(&inode->i_wb_list)) {
1081 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1082 if (!list_empty(&inode->i_wb_list)) {
1083 list_del_init(&inode->i_wb_list);
1084 trace_sb_clear_inode_writeback(inode);
1086 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1091 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1092 * furthest end of its superblock's dirty-inode list.
1094 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1095 * already the most-recently-dirtied inode on the b_dirty list. If that is
1096 * the case then the inode must have been redirtied while it was being written
1097 * out and we don't reset its dirtied_when.
1099 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1101 assert_spin_locked(&inode->i_lock);
1103 if (!list_empty(&wb->b_dirty)) {
1106 tail = wb_inode(wb->b_dirty.next);
1107 if (time_before(inode->dirtied_when, tail->dirtied_when))
1108 inode->dirtied_when = jiffies;
1110 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1111 inode->i_state &= ~I_SYNC_QUEUED;
1114 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1116 spin_lock(&inode->i_lock);
1117 redirty_tail_locked(inode, wb);
1118 spin_unlock(&inode->i_lock);
1122 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1124 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1126 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1129 static void inode_sync_complete(struct inode *inode)
1131 inode->i_state &= ~I_SYNC;
1132 /* If inode is clean an unused, put it into LRU now... */
1133 inode_add_lru(inode);
1134 /* Waiters must see I_SYNC cleared before being woken up */
1136 wake_up_bit(&inode->i_state, __I_SYNC);
1139 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1141 bool ret = time_after(inode->dirtied_when, t);
1142 #ifndef CONFIG_64BIT
1144 * For inodes being constantly redirtied, dirtied_when can get stuck.
1145 * It _appears_ to be in the future, but is actually in distant past.
1146 * This test is necessary to prevent such wrapped-around relative times
1147 * from permanently stopping the whole bdi writeback.
1149 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1154 #define EXPIRE_DIRTY_ATIME 0x0001
1157 * Move expired (dirtied before dirtied_before) dirty inodes from
1158 * @delaying_queue to @dispatch_queue.
1160 static int move_expired_inodes(struct list_head *delaying_queue,
1161 struct list_head *dispatch_queue,
1162 int flags, unsigned long dirtied_before)
1165 struct list_head *pos, *node;
1166 struct super_block *sb = NULL;
1167 struct inode *inode;
1171 while (!list_empty(delaying_queue)) {
1172 inode = wb_inode(delaying_queue->prev);
1173 if (inode_dirtied_after(inode, dirtied_before))
1175 list_move(&inode->i_io_list, &tmp);
1177 spin_lock(&inode->i_lock);
1178 if (flags & EXPIRE_DIRTY_ATIME)
1179 inode->i_state |= I_DIRTY_TIME_EXPIRED;
1180 inode->i_state |= I_SYNC_QUEUED;
1181 spin_unlock(&inode->i_lock);
1182 if (sb_is_blkdev_sb(inode->i_sb))
1184 if (sb && sb != inode->i_sb)
1189 /* just one sb in list, splice to dispatch_queue and we're done */
1191 list_splice(&tmp, dispatch_queue);
1195 /* Move inodes from one superblock together */
1196 while (!list_empty(&tmp)) {
1197 sb = wb_inode(tmp.prev)->i_sb;
1198 list_for_each_prev_safe(pos, node, &tmp) {
1199 inode = wb_inode(pos);
1200 if (inode->i_sb == sb)
1201 list_move(&inode->i_io_list, dispatch_queue);
1209 * Queue all expired dirty inodes for io, eldest first.
1211 * newly dirtied b_dirty b_io b_more_io
1212 * =============> gf edc BA
1214 * newly dirtied b_dirty b_io b_more_io
1215 * =============> g fBAedc
1217 * +--> dequeue for IO
1219 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1220 unsigned long dirtied_before)
1223 unsigned long time_expire_jif = dirtied_before;
1225 assert_spin_locked(&wb->list_lock);
1226 list_splice_init(&wb->b_more_io, &wb->b_io);
1227 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, dirtied_before);
1228 if (!work->for_sync)
1229 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1230 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1231 EXPIRE_DIRTY_ATIME, time_expire_jif);
1233 wb_io_lists_populated(wb);
1234 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1237 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1241 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1242 trace_writeback_write_inode_start(inode, wbc);
1243 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1244 trace_writeback_write_inode(inode, wbc);
1251 * Wait for writeback on an inode to complete. Called with i_lock held.
1252 * Caller must make sure inode cannot go away when we drop i_lock.
1254 static void __inode_wait_for_writeback(struct inode *inode)
1255 __releases(inode->i_lock)
1256 __acquires(inode->i_lock)
1258 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1259 wait_queue_head_t *wqh;
1261 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1262 while (inode->i_state & I_SYNC) {
1263 spin_unlock(&inode->i_lock);
1264 __wait_on_bit(wqh, &wq, bit_wait,
1265 TASK_UNINTERRUPTIBLE);
1266 spin_lock(&inode->i_lock);
1271 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1273 void inode_wait_for_writeback(struct inode *inode)
1275 spin_lock(&inode->i_lock);
1276 __inode_wait_for_writeback(inode);
1277 spin_unlock(&inode->i_lock);
1281 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1282 * held and drops it. It is aimed for callers not holding any inode reference
1283 * so once i_lock is dropped, inode can go away.
1285 static void inode_sleep_on_writeback(struct inode *inode)
1286 __releases(inode->i_lock)
1289 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1292 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1293 sleep = inode->i_state & I_SYNC;
1294 spin_unlock(&inode->i_lock);
1297 finish_wait(wqh, &wait);
1301 * Find proper writeback list for the inode depending on its current state and
1302 * possibly also change of its state while we were doing writeback. Here we
1303 * handle things such as livelock prevention or fairness of writeback among
1304 * inodes. This function can be called only by flusher thread - noone else
1305 * processes all inodes in writeback lists and requeueing inodes behind flusher
1306 * thread's back can have unexpected consequences.
1308 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1309 struct writeback_control *wbc)
1311 if (inode->i_state & I_FREEING)
1315 * Sync livelock prevention. Each inode is tagged and synced in one
1316 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1317 * the dirty time to prevent enqueue and sync it again.
1319 if ((inode->i_state & I_DIRTY) &&
1320 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1321 inode->dirtied_when = jiffies;
1323 if (wbc->pages_skipped) {
1325 * writeback is not making progress due to locked
1326 * buffers. Skip this inode for now.
1328 redirty_tail_locked(inode, wb);
1332 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1334 * We didn't write back all the pages. nfs_writepages()
1335 * sometimes bales out without doing anything.
1337 if (wbc->nr_to_write <= 0) {
1338 /* Slice used up. Queue for next turn. */
1339 requeue_io(inode, wb);
1342 * Writeback blocked by something other than
1343 * congestion. Delay the inode for some time to
1344 * avoid spinning on the CPU (100% iowait)
1345 * retrying writeback of the dirty page/inode
1346 * that cannot be performed immediately.
1348 redirty_tail_locked(inode, wb);
1350 } else if (inode->i_state & I_DIRTY) {
1352 * Filesystems can dirty the inode during writeback operations,
1353 * such as delayed allocation during submission or metadata
1354 * updates after data IO completion.
1356 redirty_tail_locked(inode, wb);
1357 } else if (inode->i_state & I_DIRTY_TIME) {
1358 inode->dirtied_when = jiffies;
1359 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1360 inode->i_state &= ~I_SYNC_QUEUED;
1362 /* The inode is clean. Remove from writeback lists. */
1363 inode_io_list_del_locked(inode, wb);
1368 * Write out an inode and its dirty pages. Do not update the writeback list
1369 * linkage. That is left to the caller. The caller is also responsible for
1370 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1373 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1375 struct address_space *mapping = inode->i_mapping;
1376 long nr_to_write = wbc->nr_to_write;
1380 WARN_ON(!(inode->i_state & I_SYNC));
1382 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1384 ret = do_writepages(mapping, wbc);
1387 * Make sure to wait on the data before writing out the metadata.
1388 * This is important for filesystems that modify metadata on data
1389 * I/O completion. We don't do it for sync(2) writeback because it has a
1390 * separate, external IO completion path and ->sync_fs for guaranteeing
1391 * inode metadata is written back correctly.
1393 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1394 int err = filemap_fdatawait(mapping);
1400 * Some filesystems may redirty the inode during the writeback
1401 * due to delalloc, clear dirty metadata flags right before
1404 spin_lock(&inode->i_lock);
1406 dirty = inode->i_state & I_DIRTY;
1407 if (inode->i_state & I_DIRTY_TIME) {
1408 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1409 wbc->sync_mode == WB_SYNC_ALL ||
1410 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1411 unlikely(time_after(jiffies,
1412 (inode->dirtied_time_when +
1413 dirtytime_expire_interval * HZ)))) {
1414 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1415 trace_writeback_lazytime(inode);
1418 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1419 inode->i_state &= ~dirty;
1422 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1423 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1424 * either they see the I_DIRTY bits cleared or we see the dirtied
1427 * I_DIRTY_PAGES is always cleared together above even if @mapping
1428 * still has dirty pages. The flag is reinstated after smp_mb() if
1429 * necessary. This guarantees that either __mark_inode_dirty()
1430 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1434 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1435 inode->i_state |= I_DIRTY_PAGES;
1437 spin_unlock(&inode->i_lock);
1439 if (dirty & I_DIRTY_TIME)
1440 mark_inode_dirty_sync(inode);
1441 /* Don't write the inode if only I_DIRTY_PAGES was set */
1442 if (dirty & ~I_DIRTY_PAGES) {
1443 int err = write_inode(inode, wbc);
1447 trace_writeback_single_inode(inode, wbc, nr_to_write);
1452 * Write out an inode's dirty pages. Either the caller has an active reference
1453 * on the inode or the inode has I_WILL_FREE set.
1455 * This function is designed to be called for writing back one inode which
1456 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1457 * and does more profound writeback list handling in writeback_sb_inodes().
1459 static int writeback_single_inode(struct inode *inode,
1460 struct writeback_control *wbc)
1462 struct bdi_writeback *wb;
1465 spin_lock(&inode->i_lock);
1466 if (!atomic_read(&inode->i_count))
1467 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1469 WARN_ON(inode->i_state & I_WILL_FREE);
1471 if (inode->i_state & I_SYNC) {
1472 if (wbc->sync_mode != WB_SYNC_ALL)
1475 * It's a data-integrity sync. We must wait. Since callers hold
1476 * inode reference or inode has I_WILL_FREE set, it cannot go
1479 __inode_wait_for_writeback(inode);
1481 WARN_ON(inode->i_state & I_SYNC);
1483 * Skip inode if it is clean and we have no outstanding writeback in
1484 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1485 * function since flusher thread may be doing for example sync in
1486 * parallel and if we move the inode, it could get skipped. So here we
1487 * make sure inode is on some writeback list and leave it there unless
1488 * we have completely cleaned the inode.
1490 if (!(inode->i_state & I_DIRTY_ALL) &&
1491 (wbc->sync_mode != WB_SYNC_ALL ||
1492 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1494 inode->i_state |= I_SYNC;
1495 wbc_attach_and_unlock_inode(wbc, inode);
1497 ret = __writeback_single_inode(inode, wbc);
1499 wbc_detach_inode(wbc);
1501 wb = inode_to_wb_and_lock_list(inode);
1502 spin_lock(&inode->i_lock);
1504 * If inode is clean, remove it from writeback lists. Otherwise don't
1505 * touch it. See comment above for explanation.
1507 if (!(inode->i_state & I_DIRTY_ALL))
1508 inode_io_list_del_locked(inode, wb);
1509 spin_unlock(&wb->list_lock);
1510 inode_sync_complete(inode);
1512 spin_unlock(&inode->i_lock);
1516 static long writeback_chunk_size(struct bdi_writeback *wb,
1517 struct wb_writeback_work *work)
1522 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1523 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1524 * here avoids calling into writeback_inodes_wb() more than once.
1526 * The intended call sequence for WB_SYNC_ALL writeback is:
1529 * writeback_sb_inodes() <== called only once
1530 * write_cache_pages() <== called once for each inode
1531 * (quickly) tag currently dirty pages
1532 * (maybe slowly) sync all tagged pages
1534 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1537 pages = min(wb->avg_write_bandwidth / 2,
1538 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1539 pages = min(pages, work->nr_pages);
1540 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1541 MIN_WRITEBACK_PAGES);
1548 * Write a portion of b_io inodes which belong to @sb.
1550 * Return the number of pages and/or inodes written.
1552 * NOTE! This is called with wb->list_lock held, and will
1553 * unlock and relock that for each inode it ends up doing
1556 static long writeback_sb_inodes(struct super_block *sb,
1557 struct bdi_writeback *wb,
1558 struct wb_writeback_work *work)
1560 struct writeback_control wbc = {
1561 .sync_mode = work->sync_mode,
1562 .tagged_writepages = work->tagged_writepages,
1563 .for_kupdate = work->for_kupdate,
1564 .for_background = work->for_background,
1565 .for_sync = work->for_sync,
1566 .range_cyclic = work->range_cyclic,
1568 .range_end = LLONG_MAX,
1570 unsigned long start_time = jiffies;
1572 long total_wrote = 0; /* count both pages and inodes */
1574 while (!list_empty(&wb->b_io)) {
1575 struct inode *inode = wb_inode(wb->b_io.prev);
1576 struct bdi_writeback *tmp_wb;
1579 if (inode->i_sb != sb) {
1582 * We only want to write back data for this
1583 * superblock, move all inodes not belonging
1584 * to it back onto the dirty list.
1586 redirty_tail(inode, wb);
1591 * The inode belongs to a different superblock.
1592 * Bounce back to the caller to unpin this and
1593 * pin the next superblock.
1599 * Don't bother with new inodes or inodes being freed, first
1600 * kind does not need periodic writeout yet, and for the latter
1601 * kind writeout is handled by the freer.
1603 spin_lock(&inode->i_lock);
1604 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1605 redirty_tail_locked(inode, wb);
1606 spin_unlock(&inode->i_lock);
1609 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1611 * If this inode is locked for writeback and we are not
1612 * doing writeback-for-data-integrity, move it to
1613 * b_more_io so that writeback can proceed with the
1614 * other inodes on s_io.
1616 * We'll have another go at writing back this inode
1617 * when we completed a full scan of b_io.
1619 spin_unlock(&inode->i_lock);
1620 requeue_io(inode, wb);
1621 trace_writeback_sb_inodes_requeue(inode);
1624 spin_unlock(&wb->list_lock);
1627 * We already requeued the inode if it had I_SYNC set and we
1628 * are doing WB_SYNC_NONE writeback. So this catches only the
1631 if (inode->i_state & I_SYNC) {
1632 /* Wait for I_SYNC. This function drops i_lock... */
1633 inode_sleep_on_writeback(inode);
1634 /* Inode may be gone, start again */
1635 spin_lock(&wb->list_lock);
1638 inode->i_state |= I_SYNC;
1639 wbc_attach_and_unlock_inode(&wbc, inode);
1641 write_chunk = writeback_chunk_size(wb, work);
1642 wbc.nr_to_write = write_chunk;
1643 wbc.pages_skipped = 0;
1646 * We use I_SYNC to pin the inode in memory. While it is set
1647 * evict_inode() will wait so the inode cannot be freed.
1649 __writeback_single_inode(inode, &wbc);
1651 wbc_detach_inode(&wbc);
1652 work->nr_pages -= write_chunk - wbc.nr_to_write;
1653 wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1654 wrote = wrote < 0 ? 0 : wrote;
1655 total_wrote += wrote;
1657 if (need_resched()) {
1659 * We're trying to balance between building up a nice
1660 * long list of IOs to improve our merge rate, and
1661 * getting those IOs out quickly for anyone throttling
1662 * in balance_dirty_pages(). cond_resched() doesn't
1663 * unplug, so get our IOs out the door before we
1666 blk_flush_plug(current);
1671 * Requeue @inode if still dirty. Be careful as @inode may
1672 * have been switched to another wb in the meantime.
1674 tmp_wb = inode_to_wb_and_lock_list(inode);
1675 spin_lock(&inode->i_lock);
1676 if (!(inode->i_state & I_DIRTY_ALL))
1678 requeue_inode(inode, tmp_wb, &wbc);
1679 inode_sync_complete(inode);
1680 spin_unlock(&inode->i_lock);
1682 if (unlikely(tmp_wb != wb)) {
1683 spin_unlock(&tmp_wb->list_lock);
1684 spin_lock(&wb->list_lock);
1688 * bail out to wb_writeback() often enough to check
1689 * background threshold and other termination conditions.
1692 if (time_is_before_jiffies(start_time + HZ / 10UL))
1694 if (work->nr_pages <= 0)
1701 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1702 struct wb_writeback_work *work)
1704 unsigned long start_time = jiffies;
1707 while (!list_empty(&wb->b_io)) {
1708 struct inode *inode = wb_inode(wb->b_io.prev);
1709 struct super_block *sb = inode->i_sb;
1711 if (!trylock_super(sb)) {
1713 * trylock_super() may fail consistently due to
1714 * s_umount being grabbed by someone else. Don't use
1715 * requeue_io() to avoid busy retrying the inode/sb.
1717 redirty_tail(inode, wb);
1720 wrote += writeback_sb_inodes(sb, wb, work);
1721 up_read(&sb->s_umount);
1723 /* refer to the same tests at the end of writeback_sb_inodes */
1725 if (time_is_before_jiffies(start_time + HZ / 10UL))
1727 if (work->nr_pages <= 0)
1731 /* Leave any unwritten inodes on b_io */
1735 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1736 enum wb_reason reason)
1738 struct wb_writeback_work work = {
1739 .nr_pages = nr_pages,
1740 .sync_mode = WB_SYNC_NONE,
1744 struct blk_plug plug;
1746 blk_start_plug(&plug);
1747 spin_lock(&wb->list_lock);
1748 if (list_empty(&wb->b_io))
1749 queue_io(wb, &work, jiffies);
1750 __writeback_inodes_wb(wb, &work);
1751 spin_unlock(&wb->list_lock);
1752 blk_finish_plug(&plug);
1754 return nr_pages - work.nr_pages;
1758 * Explicit flushing or periodic writeback of "old" data.
1760 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1761 * dirtying-time in the inode's address_space. So this periodic writeback code
1762 * just walks the superblock inode list, writing back any inodes which are
1763 * older than a specific point in time.
1765 * Try to run once per dirty_writeback_interval. But if a writeback event
1766 * takes longer than a dirty_writeback_interval interval, then leave a
1769 * dirtied_before takes precedence over nr_to_write. So we'll only write back
1770 * all dirty pages if they are all attached to "old" mappings.
1772 static long wb_writeback(struct bdi_writeback *wb,
1773 struct wb_writeback_work *work)
1775 unsigned long wb_start = jiffies;
1776 long nr_pages = work->nr_pages;
1777 unsigned long dirtied_before = jiffies;
1778 struct inode *inode;
1780 struct blk_plug plug;
1782 blk_start_plug(&plug);
1783 spin_lock(&wb->list_lock);
1786 * Stop writeback when nr_pages has been consumed
1788 if (work->nr_pages <= 0)
1792 * Background writeout and kupdate-style writeback may
1793 * run forever. Stop them if there is other work to do
1794 * so that e.g. sync can proceed. They'll be restarted
1795 * after the other works are all done.
1797 if ((work->for_background || work->for_kupdate) &&
1798 !list_empty(&wb->work_list))
1802 * For background writeout, stop when we are below the
1803 * background dirty threshold
1805 if (work->for_background && !wb_over_bg_thresh(wb))
1809 * Kupdate and background works are special and we want to
1810 * include all inodes that need writing. Livelock avoidance is
1811 * handled by these works yielding to any other work so we are
1814 if (work->for_kupdate) {
1815 dirtied_before = jiffies -
1816 msecs_to_jiffies(dirty_expire_interval * 10);
1817 } else if (work->for_background)
1818 dirtied_before = jiffies;
1820 trace_writeback_start(wb, work);
1821 if (list_empty(&wb->b_io))
1822 queue_io(wb, work, dirtied_before);
1824 progress = writeback_sb_inodes(work->sb, wb, work);
1826 progress = __writeback_inodes_wb(wb, work);
1827 trace_writeback_written(wb, work);
1829 wb_update_bandwidth(wb, wb_start);
1832 * Did we write something? Try for more
1834 * Dirty inodes are moved to b_io for writeback in batches.
1835 * The completion of the current batch does not necessarily
1836 * mean the overall work is done. So we keep looping as long
1837 * as made some progress on cleaning pages or inodes.
1842 * No more inodes for IO, bail
1844 if (list_empty(&wb->b_more_io))
1847 * Nothing written. Wait for some inode to
1848 * become available for writeback. Otherwise
1849 * we'll just busyloop.
1851 if (!list_empty(&wb->b_more_io)) {
1852 trace_writeback_wait(wb, work);
1853 inode = wb_inode(wb->b_more_io.prev);
1854 spin_lock(&inode->i_lock);
1855 spin_unlock(&wb->list_lock);
1856 /* This function drops i_lock... */
1857 inode_sleep_on_writeback(inode);
1858 spin_lock(&wb->list_lock);
1861 spin_unlock(&wb->list_lock);
1862 blk_finish_plug(&plug);
1864 return nr_pages - work->nr_pages;
1868 * Return the next wb_writeback_work struct that hasn't been processed yet.
1870 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1872 struct wb_writeback_work *work = NULL;
1874 spin_lock_bh(&wb->work_lock);
1875 if (!list_empty(&wb->work_list)) {
1876 work = list_entry(wb->work_list.next,
1877 struct wb_writeback_work, list);
1878 list_del_init(&work->list);
1880 spin_unlock_bh(&wb->work_lock);
1885 * Add in the number of potentially dirty inodes, because each inode
1886 * write can dirty pagecache in the underlying blockdev.
1888 static unsigned long get_nr_dirty_pages(void)
1890 return global_node_page_state(NR_FILE_DIRTY) +
1891 global_node_page_state(NR_UNSTABLE_NFS) +
1892 get_nr_dirty_inodes();
1895 static long wb_check_background_flush(struct bdi_writeback *wb)
1897 if (wb_over_bg_thresh(wb)) {
1899 struct wb_writeback_work work = {
1900 .nr_pages = LONG_MAX,
1901 .sync_mode = WB_SYNC_NONE,
1902 .for_background = 1,
1904 .reason = WB_REASON_BACKGROUND,
1907 return wb_writeback(wb, &work);
1913 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1915 unsigned long expired;
1919 * When set to zero, disable periodic writeback
1921 if (!dirty_writeback_interval)
1924 expired = wb->last_old_flush +
1925 msecs_to_jiffies(dirty_writeback_interval * 10);
1926 if (time_before(jiffies, expired))
1929 wb->last_old_flush = jiffies;
1930 nr_pages = get_nr_dirty_pages();
1933 struct wb_writeback_work work = {
1934 .nr_pages = nr_pages,
1935 .sync_mode = WB_SYNC_NONE,
1938 .reason = WB_REASON_PERIODIC,
1941 return wb_writeback(wb, &work);
1948 * Retrieve work items and do the writeback they describe
1950 static long wb_do_writeback(struct bdi_writeback *wb)
1952 struct wb_writeback_work *work;
1955 set_bit(WB_writeback_running, &wb->state);
1956 while ((work = get_next_work_item(wb)) != NULL) {
1957 trace_writeback_exec(wb, work);
1958 wrote += wb_writeback(wb, work);
1959 finish_writeback_work(wb, work);
1963 * Check for periodic writeback, kupdated() style
1965 wrote += wb_check_old_data_flush(wb);
1966 wrote += wb_check_background_flush(wb);
1967 clear_bit(WB_writeback_running, &wb->state);
1973 * Handle writeback of dirty data for the device backed by this bdi. Also
1974 * reschedules periodically and does kupdated style flushing.
1976 void wb_workfn(struct work_struct *work)
1978 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1979 struct bdi_writeback, dwork);
1982 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
1983 current->flags |= PF_SWAPWRITE;
1985 if (likely(!current_is_workqueue_rescuer() ||
1986 !test_bit(WB_registered, &wb->state))) {
1988 * The normal path. Keep writing back @wb until its
1989 * work_list is empty. Note that this path is also taken
1990 * if @wb is shutting down even when we're running off the
1991 * rescuer as work_list needs to be drained.
1994 pages_written = wb_do_writeback(wb);
1995 trace_writeback_pages_written(pages_written);
1996 } while (!list_empty(&wb->work_list));
1999 * bdi_wq can't get enough workers and we're running off
2000 * the emergency worker. Don't hog it. Hopefully, 1024 is
2001 * enough for efficient IO.
2003 pages_written = writeback_inodes_wb(wb, 1024,
2004 WB_REASON_FORKER_THREAD);
2005 trace_writeback_pages_written(pages_written);
2008 if (!list_empty(&wb->work_list))
2010 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2011 wb_wakeup_delayed(wb);
2013 current->flags &= ~PF_SWAPWRITE;
2017 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
2020 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
2022 struct backing_dev_info *bdi;
2025 * If we are expecting writeback progress we must submit plugged IO.
2027 if (blk_needs_flush_plug(current))
2028 blk_schedule_flush_plug(current);
2031 nr_pages = get_nr_dirty_pages();
2034 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2035 struct bdi_writeback *wb;
2037 if (!bdi_has_dirty_io(bdi))
2040 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2041 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
2048 * Wake up bdi's periodically to make sure dirtytime inodes gets
2049 * written back periodically. We deliberately do *not* check the
2050 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2051 * kernel to be constantly waking up once there are any dirtytime
2052 * inodes on the system. So instead we define a separate delayed work
2053 * function which gets called much more rarely. (By default, only
2054 * once every 12 hours.)
2056 * If there is any other write activity going on in the file system,
2057 * this function won't be necessary. But if the only thing that has
2058 * happened on the file system is a dirtytime inode caused by an atime
2059 * update, we need this infrastructure below to make sure that inode
2060 * eventually gets pushed out to disk.
2062 static void wakeup_dirtytime_writeback(struct work_struct *w);
2063 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2065 static void wakeup_dirtytime_writeback(struct work_struct *w)
2067 struct backing_dev_info *bdi;
2070 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2071 struct bdi_writeback *wb;
2073 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2074 if (!list_empty(&wb->b_dirty_time))
2078 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2081 static int __init start_dirtytime_writeback(void)
2083 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2086 __initcall(start_dirtytime_writeback);
2088 int dirtytime_interval_handler(struct ctl_table *table, int write,
2089 void __user *buffer, size_t *lenp, loff_t *ppos)
2093 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2094 if (ret == 0 && write)
2095 mod_delayed_work(system_wq, &dirtytime_work, 0);
2100 * __mark_inode_dirty - internal function
2101 * @inode: inode to mark
2102 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2103 * Mark an inode as dirty. Callers should use mark_inode_dirty or
2104 * mark_inode_dirty_sync.
2106 * Put the inode on the super block's dirty list.
2108 * CAREFUL! We mark it dirty unconditionally, but move it onto the
2109 * dirty list only if it is hashed or if it refers to a blockdev.
2110 * If it was not hashed, it will never be added to the dirty list
2111 * even if it is later hashed, as it will have been marked dirty already.
2113 * In short, make sure you hash any inodes _before_ you start marking
2116 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2117 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2118 * the kernel-internal blockdev inode represents the dirtying time of the
2119 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2120 * page->mapping->host, so the page-dirtying time is recorded in the internal
2123 void __mark_inode_dirty(struct inode *inode, int flags)
2125 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
2126 struct super_block *sb = inode->i_sb;
2129 trace_writeback_mark_inode_dirty(inode, flags);
2132 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2133 * dirty the inode itself
2135 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
2136 trace_writeback_dirty_inode_start(inode, flags);
2138 if (sb->s_op->dirty_inode)
2139 sb->s_op->dirty_inode(inode, flags);
2141 trace_writeback_dirty_inode(inode, flags);
2143 if (flags & I_DIRTY_INODE)
2144 flags &= ~I_DIRTY_TIME;
2145 dirtytime = flags & I_DIRTY_TIME;
2148 * Paired with smp_mb() in __writeback_single_inode() for the
2149 * following lockless i_state test. See there for details.
2153 if (((inode->i_state & flags) == flags) ||
2154 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2157 spin_lock(&inode->i_lock);
2158 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2159 goto out_unlock_inode;
2160 if ((inode->i_state & flags) != flags) {
2161 const int was_dirty = inode->i_state & I_DIRTY;
2163 inode_attach_wb(inode, NULL);
2165 if (flags & I_DIRTY_INODE)
2166 inode->i_state &= ~I_DIRTY_TIME;
2167 inode->i_state |= flags;
2170 * If the inode is queued for writeback by flush worker, just
2171 * update its dirty state. Once the flush worker is done with
2172 * the inode it will place it on the appropriate superblock
2173 * list, based upon its state.
2175 if (inode->i_state & I_SYNC_QUEUED)
2176 goto out_unlock_inode;
2179 * Only add valid (hashed) inodes to the superblock's
2180 * dirty list. Add blockdev inodes as well.
2182 if (!S_ISBLK(inode->i_mode)) {
2183 if (inode_unhashed(inode))
2184 goto out_unlock_inode;
2186 if (inode->i_state & I_FREEING)
2187 goto out_unlock_inode;
2190 * If the inode was already on b_dirty/b_io/b_more_io, don't
2191 * reposition it (that would break b_dirty time-ordering).
2194 struct bdi_writeback *wb;
2195 struct list_head *dirty_list;
2196 bool wakeup_bdi = false;
2198 wb = locked_inode_to_wb_and_lock_list(inode);
2200 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2201 !test_bit(WB_registered, &wb->state),
2202 "bdi-%s not registered\n", wb->bdi->name);
2204 inode->dirtied_when = jiffies;
2206 inode->dirtied_time_when = jiffies;
2208 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
2209 dirty_list = &wb->b_dirty;
2211 dirty_list = &wb->b_dirty_time;
2213 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2216 spin_unlock(&wb->list_lock);
2217 trace_writeback_dirty_inode_enqueue(inode);
2220 * If this is the first dirty inode for this bdi,
2221 * we have to wake-up the corresponding bdi thread
2222 * to make sure background write-back happens
2225 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2226 wb_wakeup_delayed(wb);
2231 spin_unlock(&inode->i_lock);
2233 #undef I_DIRTY_INODE
2235 EXPORT_SYMBOL(__mark_inode_dirty);
2238 * The @s_sync_lock is used to serialise concurrent sync operations
2239 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2240 * Concurrent callers will block on the s_sync_lock rather than doing contending
2241 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2242 * has been issued up to the time this function is enter is guaranteed to be
2243 * completed by the time we have gained the lock and waited for all IO that is
2244 * in progress regardless of the order callers are granted the lock.
2246 static void wait_sb_inodes(struct super_block *sb)
2248 LIST_HEAD(sync_list);
2251 * We need to be protected against the filesystem going from
2252 * r/o to r/w or vice versa.
2254 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2256 mutex_lock(&sb->s_sync_lock);
2259 * Splice the writeback list onto a temporary list to avoid waiting on
2260 * inodes that have started writeback after this point.
2262 * Use rcu_read_lock() to keep the inodes around until we have a
2263 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2264 * the local list because inodes can be dropped from either by writeback
2268 spin_lock_irq(&sb->s_inode_wblist_lock);
2269 list_splice_init(&sb->s_inodes_wb, &sync_list);
2272 * Data integrity sync. Must wait for all pages under writeback, because
2273 * there may have been pages dirtied before our sync call, but which had
2274 * writeout started before we write it out. In which case, the inode
2275 * may not be on the dirty list, but we still have to wait for that
2278 while (!list_empty(&sync_list)) {
2279 struct inode *inode = list_first_entry(&sync_list, struct inode,
2281 struct address_space *mapping = inode->i_mapping;
2284 * Move each inode back to the wb list before we drop the lock
2285 * to preserve consistency between i_wb_list and the mapping
2286 * writeback tag. Writeback completion is responsible to remove
2287 * the inode from either list once the writeback tag is cleared.
2289 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2292 * The mapping can appear untagged while still on-list since we
2293 * do not have the mapping lock. Skip it here, wb completion
2296 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2299 spin_unlock_irq(&sb->s_inode_wblist_lock);
2301 spin_lock(&inode->i_lock);
2302 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2303 spin_unlock(&inode->i_lock);
2305 spin_lock_irq(&sb->s_inode_wblist_lock);
2309 spin_unlock(&inode->i_lock);
2313 * We keep the error status of individual mapping so that
2314 * applications can catch the writeback error using fsync(2).
2315 * See filemap_fdatawait_keep_errors() for details.
2317 filemap_fdatawait_keep_errors(mapping);
2324 spin_lock_irq(&sb->s_inode_wblist_lock);
2326 spin_unlock_irq(&sb->s_inode_wblist_lock);
2328 mutex_unlock(&sb->s_sync_lock);
2331 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2332 enum wb_reason reason, bool skip_if_busy)
2334 DEFINE_WB_COMPLETION_ONSTACK(done);
2335 struct wb_writeback_work work = {
2337 .sync_mode = WB_SYNC_NONE,
2338 .tagged_writepages = 1,
2343 struct backing_dev_info *bdi = sb->s_bdi;
2345 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2347 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2349 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2350 wb_wait_for_completion(bdi, &done);
2354 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2355 * @sb: the superblock
2356 * @nr: the number of pages to write
2357 * @reason: reason why some writeback work initiated
2359 * Start writeback on some inodes on this super_block. No guarantees are made
2360 * on how many (if any) will be written, and this function does not wait
2361 * for IO completion of submitted IO.
2363 void writeback_inodes_sb_nr(struct super_block *sb,
2365 enum wb_reason reason)
2367 __writeback_inodes_sb_nr(sb, nr, reason, false);
2369 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2372 * writeback_inodes_sb - writeback dirty inodes from given super_block
2373 * @sb: the superblock
2374 * @reason: reason why some writeback work was initiated
2376 * Start writeback on some inodes on this super_block. No guarantees are made
2377 * on how many (if any) will be written, and this function does not wait
2378 * for IO completion of submitted IO.
2380 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2382 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2384 EXPORT_SYMBOL(writeback_inodes_sb);
2387 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2388 * @sb: the superblock
2389 * @nr: the number of pages to write
2390 * @reason: the reason of writeback
2392 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2393 * Returns 1 if writeback was started, 0 if not.
2395 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2396 enum wb_reason reason)
2398 if (!down_read_trylock(&sb->s_umount))
2401 __writeback_inodes_sb_nr(sb, nr, reason, true);
2402 up_read(&sb->s_umount);
2405 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2408 * try_to_writeback_inodes_sb - try to start writeback if none underway
2409 * @sb: the superblock
2410 * @reason: reason why some writeback work was initiated
2412 * Implement by try_to_writeback_inodes_sb_nr()
2413 * Returns 1 if writeback was started, 0 if not.
2415 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2417 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2419 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2422 * sync_inodes_sb - sync sb inode pages
2423 * @sb: the superblock
2425 * This function writes and waits on any dirty inode belonging to this
2428 void sync_inodes_sb(struct super_block *sb)
2430 DEFINE_WB_COMPLETION_ONSTACK(done);
2431 struct wb_writeback_work work = {
2433 .sync_mode = WB_SYNC_ALL,
2434 .nr_pages = LONG_MAX,
2437 .reason = WB_REASON_SYNC,
2440 struct backing_dev_info *bdi = sb->s_bdi;
2443 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2444 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2445 * bdi_has_dirty() need to be written out too.
2447 if (bdi == &noop_backing_dev_info)
2449 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2451 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2452 bdi_down_write_wb_switch_rwsem(bdi);
2453 bdi_split_work_to_wbs(bdi, &work, false);
2454 wb_wait_for_completion(bdi, &done);
2455 bdi_up_write_wb_switch_rwsem(bdi);
2459 EXPORT_SYMBOL(sync_inodes_sb);
2462 * write_inode_now - write an inode to disk
2463 * @inode: inode to write to disk
2464 * @sync: whether the write should be synchronous or not
2466 * This function commits an inode to disk immediately if it is dirty. This is
2467 * primarily needed by knfsd.
2469 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2471 int write_inode_now(struct inode *inode, int sync)
2473 struct writeback_control wbc = {
2474 .nr_to_write = LONG_MAX,
2475 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2477 .range_end = LLONG_MAX,
2480 if (!mapping_cap_writeback_dirty(inode->i_mapping))
2481 wbc.nr_to_write = 0;
2484 return writeback_single_inode(inode, &wbc);
2486 EXPORT_SYMBOL(write_inode_now);
2489 * sync_inode - write an inode and its pages to disk.
2490 * @inode: the inode to sync
2491 * @wbc: controls the writeback mode
2493 * sync_inode() will write an inode and its pages to disk. It will also
2494 * correctly update the inode on its superblock's dirty inode lists and will
2495 * update inode->i_state.
2497 * The caller must have a ref on the inode.
2499 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2501 return writeback_single_inode(inode, wbc);
2503 EXPORT_SYMBOL(sync_inode);
2506 * sync_inode_metadata - write an inode to disk
2507 * @inode: the inode to sync
2508 * @wait: wait for I/O to complete.
2510 * Write an inode to disk and adjust its dirty state after completion.
2512 * Note: only writes the actual inode, no associated data or other metadata.
2514 int sync_inode_metadata(struct inode *inode, int wait)
2516 struct writeback_control wbc = {
2517 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2518 .nr_to_write = 0, /* metadata-only */
2521 return sync_inode(inode, &wbc);
2523 EXPORT_SYMBOL(sync_inode_metadata);